From 5fb8f3fe6be23fbbacc4f0ad5f8e57cd7b28ccaf Mon Sep 17 00:00:00 2001 From: eloquentarduino Date: Fri, 29 Oct 2021 09:47:05 +0200 Subject: [PATCH] support Cortex-M + ESP, support Tf@2.4.0 on Cortex-M, support Tf@2.1.1 on ESP32 --- .DS_Store | Bin 6148 -> 6148 bytes README.md | 23 +- examples/SineExample/SineExample.ino | 28 +- library.json | 2 +- library.properties | 2 +- src/.DS_Store | Bin 8196 -> 6148 bytes src/EloquentTinyML.h | 229 +- src/TfLiteARM.h | 28 + src/TfLiteAbstract.h | 294 + src/TfLiteESP32.h | 28 + src/tensorflow/.DS_Store | Bin 6148 -> 0 bytes src/tensorflow/lite/.DS_Store | Bin 8196 -> 0 bytes src/tensorflow/lite/kernels/.DS_Store | Bin 6148 -> 0 bytes .../internal/reference/integer_ops/softmax.h | 102 - .../lite/micro/kernels/concatenation.cpp | 225 - src/tensorflow/lite/micro/kernels/mul.cpp | 174 - src/tensorflow/lite/micro/kernels/pad.cpp | 227 - src/tensorflow/lite/micro/micro_allocator.cpp | 531 - src/tensorflow/lite/micro/micro_allocator.h | 85 - .../lite/micro/micro_mutable_op_resolver.h | 113 - src/tensorflow/lite/type_to_tflitetype.h | 78 - .../tensorflow/core/public/version.h | 142 + .../tensorflow/lite/c/builtin_op_data.h | 487 + .../tensorflow/lite/c/c_api_types.h | 95 + .../tensorflow/lite/c/common.c | 27 +- src/tensorflow_arm/tensorflow/lite/c/common.h | 916 + .../lite/core/api/error_reporter.cpp | 5 +- .../tensorflow/lite/core/api/error_reporter.h | 62 + .../lite/core/api/flatbuffer_conversions.cpp | 1948 ++ .../lite/core/api/flatbuffer_conversions.h | 304 + .../tensorflow/lite/core/api/op_resolver.cpp | 72 + .../tensorflow/lite/core/api/op_resolver.h | 63 + .../tensorflow/lite/core/api/profiler.h | 197 + .../tensorflow/lite/core/api/tensor_utils.cpp | 53 + .../tensorflow/lite/core/api/tensor_utils.h | 5 +- .../experimental/microfrontend/lib/bits.h | 105 + .../experimental/microfrontend/lib/fft.cpp | 56 + .../lite/experimental/microfrontend/lib/fft.h | 53 + .../microfrontend/lib/fft_util.cpp | 75 + .../experimental/microfrontend/lib/fft_util.h | 37 + .../microfrontend/lib/filterbank.c | 137 + .../microfrontend/lib/filterbank.h | 66 + .../microfrontend/lib/filterbank_util.c | 223 + .../microfrontend/lib/filterbank_util.h | 53 + .../experimental/microfrontend/lib/frontend.c | 75 + .../experimental/microfrontend/lib/frontend.h | 67 + .../microfrontend/lib/frontend_util.c | 88 + .../microfrontend/lib/frontend_util.h | 55 + .../experimental/microfrontend/lib/log_lut.c | 33 + .../experimental/microfrontend/lib/log_lut.h | 43 + .../microfrontend/lib/log_scale.c | 86 + .../microfrontend/lib/log_scale.h | 42 + .../microfrontend/lib/log_scale_util.c | 30 + .../microfrontend/lib/log_scale_util.h | 48 + .../microfrontend/lib/noise_reduction.c | 54 + .../microfrontend/lib/noise_reduction.h | 49 + .../microfrontend/lib/noise_reduction_util.c | 48 + .../microfrontend/lib/noise_reduction_util.h | 53 + .../microfrontend/lib/pcan_gain_control.c | 59 + .../microfrontend/lib/pcan_gain_control.h | 50 + .../lib/pcan_gain_control_util.c | 95 + .../lib/pcan_gain_control_util.h | 60 + .../experimental/microfrontend/lib/window.c | 73 + .../experimental/microfrontend/lib/window.h | 52 + .../microfrontend/lib/window_util.c | 76 + .../microfrontend/lib/window_util.h | 48 + .../tensorflow/lite/kernels/internal/common.h | 1040 + .../lite/kernels/internal/compatibility.h | 115 + .../lite/kernels/internal/cppmath.h | 43 + .../tensorflow/lite/kernels/internal/max.h | 38 + .../tensorflow/lite/kernels/internal/min.h | 38 + .../kernels/internal/optimized/neon_check.h | 43 + .../lite/kernels/internal/portable_tensor.h | 125 + .../kernels/internal/quantization_util.cpp | 398 + .../lite/kernels/internal/quantization_util.h | 9 +- .../lite/kernels/internal/reference/add.h | 449 + .../kernels/internal/reference/arg_min_max.h | 5 +- .../internal/reference/binary_function.h | 83 + .../lite/kernels/internal/reference/ceil.h | 5 +- .../kernels/internal/reference/comparisons.h | 283 + .../internal/reference/concatenation.h | 31 +- .../lite/kernels/internal/reference/conv.h | 267 + .../internal/reference/depthwiseconv_float.h | 9 +- .../internal/reference/depthwiseconv_uint8.h | 300 + .../kernels/internal/reference/dequantize.h | 81 + .../lite/kernels/internal/reference/floor.h | 5 +- .../internal/reference/fully_connected.h | 323 + .../kernels/internal/reference/hard_swish.h | 169 + .../internal/reference/integer_ops/add.h | 147 + .../internal/reference/integer_ops/conv.h | 224 + .../reference/integer_ops/depthwise_conv.h | 292 + .../reference/integer_ops/fully_connected.h | 111 + .../reference/integer_ops/l2normalization.h | 68 + .../internal/reference/integer_ops/logistic.h | 109 + .../internal/reference/integer_ops/mean.h | 80 + .../internal/reference/integer_ops/mul.h | 76 +- .../internal/reference/integer_ops/pooling.h | 261 + .../internal/reference/integer_ops/tanh.h | 113 + .../reference/integer_ops/transpose_conv.h | 224 + .../internal/reference/l2normalization.h | 93 + .../kernels/internal/reference/logistic.h | 135 + .../internal/reference/maximum_minimum.h | 67 + .../lite/kernels/internal/reference/mul.h | 41 +- .../lite/kernels/internal/reference/neg.h | 5 +- .../lite/kernels/internal/reference/pad.h | 41 +- .../lite/kernels/internal/reference/pooling.h | 300 + .../lite/kernels/internal/reference/prelu.h | 112 + .../reference/process_broadcast_shapes.h | 141 + .../kernels/internal/reference/quantize.h | 58 + .../lite/kernels/internal/reference/reduce.h | 415 + .../kernels/internal/reference/requantize.h | 71 + .../reference/resize_nearest_neighbor.h | 104 + .../lite/kernels/internal/reference/round.h | 5 +- .../lite/kernels/internal/reference/softmax.h | 235 + .../internal/reference/strided_slice.h | 124 + .../lite/kernels/internal/reference/sub.h | 515 + .../lite/kernels/internal/reference/tanh.h | 132 + .../internal/reference/transpose_conv.h | 220 + .../kernels/internal/strided_slice_logic.h | 214 + .../lite/kernels/internal/tensor_ctypes.h | 7 +- .../tensorflow/lite/kernels/internal/types.h | 1172 ++ .../tensorflow/lite/kernels/kernel_util.cpp | 483 + .../tensorflow/lite/kernels/kernel_util.h | 296 + .../tensorflow/lite/kernels/op_macros.h | 86 + .../tensorflow/lite/kernels/padding.h | 5 +- .../lite/micro/all_ops_resolver.cpp | 87 + .../tensorflow/lite/micro/all_ops_resolver.h | 41 + .../lite/micro/arduino/debug_log.cpp | 41 + .../keyword_scrambled_model_data.cpp | 2848 +++ .../benchmarks/keyword_scrambled_model_data.h | 25 + .../tensorflow/lite/micro/compatibility.h | 3 + .../tensorflow/lite/micro/debug_log.h | 13 +- .../lite/micro/kernels/activation_utils.h | 60 + .../lite/micro/kernels/activations.cpp | 291 + .../lite/micro/kernels/arg_min_max.cpp | 136 + .../tensorflow/lite/micro/kernels/ceil.cpp | 79 + .../lite/micro/kernels/circular_buffer.cpp | 180 + .../lite/micro/kernels/cmsis-nn/add.cpp | 249 + .../lite/micro/kernels/cmsis-nn/conv.cpp | 463 + .../micro/kernels/cmsis-nn/depthwise_conv.cpp | 478 + .../kernels/cmsis-nn/fully_connected.cpp | 395 + .../lite/micro/kernels/cmsis-nn/mul.cpp | 229 + .../lite/micro/kernels/cmsis-nn/pooling.cpp | 400 + .../lite/micro/kernels/cmsis-nn/softmax.cpp | 235 + .../lite/micro/kernels/cmsis-nn/svdf.cpp | 479 + .../lite/micro/kernels/comparisons.cpp | 727 + .../lite/micro/kernels/concatenation.cpp | 279 + .../tensorflow/lite/micro/kernels/conv_test.h | 97 + .../lite/micro/kernels/conv_test_common.cpp | 255 + .../lite/micro/kernels/dequantize.cpp | 169 + .../micro/kernels/detection_postprocess.cpp | 808 + .../lite/micro/kernels/elementwise.cpp | 217 + .../tensorflow/lite/micro/kernels/ethosu.cpp | 30 + .../tensorflow/lite/micro/kernels/ethosu.h | 31 + .../kernels/flexbuffers_generated_data.cpp | 71 + .../kernels/flexbuffers_generated_data.h | 28 + .../tensorflow/lite/micro/kernels/floor.cpp | 60 + .../lite/micro/kernels/fully_connected.h | 53 + .../lite/micro/kernels/hard_swish.cpp | 145 + .../lite/micro/kernels/kernel_runner.cpp | 169 + .../lite/micro/kernels/kernel_runner.h | 87 + .../lite/micro/kernels/kernel_util.h | 78 + .../lite/micro/kernels/kernel_util_full.cpp | 44 + .../tensorflow/lite/micro/kernels/l2norm.cpp | 160 + .../tensorflow/lite/micro/kernels/logical.cpp | 108 + .../lite/micro/kernels/logistic.cpp | 153 + .../lite/micro/kernels/maximum_minimum.cpp | 151 + .../tensorflow/lite/micro/kernels/micro_ops.h | 104 + .../lite/micro/kernels/micro_utils.h | 6 + .../tensorflow/lite/micro/kernels/neg.cpp | 69 + .../tensorflow/lite/micro/kernels/pack.cpp | 130 + .../tensorflow/lite/micro/kernels/pad.cpp | 257 + .../tensorflow/lite/micro/kernels/prelu.cpp | 172 + .../lite/micro/kernels/quantize.cpp | 98 + .../tensorflow/lite/micro/kernels/quantize.h | 40 + .../lite/micro/kernels/quantize_common.cpp | 125 + .../tensorflow/lite/micro/kernels/reduce.cpp | 345 + .../tensorflow/lite/micro/kernels/reshape.cpp | 121 + .../micro/kernels/resize_nearest_neighbor.cpp | 124 + .../tensorflow/lite/micro/kernels/round.cpp | 79 + .../tensorflow/lite/micro/kernels/shape.cpp | 76 + .../tensorflow/lite/micro/kernels/split.cpp | 138 + .../tensorflow/lite/micro/kernels/split_v.cpp | 138 + .../lite/micro/kernels/strided_slice.cpp | 195 + .../tensorflow/lite/micro/kernels/sub.cpp | 259 + .../tensorflow/lite/micro/kernels/svdf.h | 54 + .../lite/micro/kernels/svdf_common.cpp | 180 + .../tensorflow/lite/micro/kernels/tanh.cpp | 161 + .../lite/micro/kernels/transpose_conv.cpp | 268 + .../tensorflow/lite/micro/kernels/unpack.cpp | 124 + .../tensorflow/lite/micro/memory_helpers.cpp | 161 + .../tensorflow/lite/micro/memory_helpers.h | 62 + .../memory_planner/greedy_memory_planner.cpp | 440 + .../memory_planner/greedy_memory_planner.h | 166 + .../memory_planner/linear_memory_planner.cpp | 57 + .../memory_planner/linear_memory_planner.h | 7 +- .../micro/memory_planner/memory_planner.h | 9 +- .../tensorflow/lite/micro/micro_allocator.cpp | 1161 ++ .../tensorflow/lite/micro/micro_allocator.h | 282 + .../lite/micro/micro_error_reporter.cpp | 44 + .../lite/micro/micro_error_reporter.h | 13 +- .../lite/micro/micro_interpreter.cpp | 456 + .../tensorflow/lite/micro/micro_interpreter.h | 214 + .../lite/micro/micro_mutable_op_resolver.h | 488 + .../tensorflow/lite/micro/micro_op_resolver.h | 76 + .../tensorflow/lite/micro/micro_profiler.cpp | 45 + .../tensorflow/lite/micro/micro_profiler.h | 74 + .../tensorflow/lite/micro/micro_string.cpp | 312 + .../tensorflow/lite/micro/micro_string.h | 36 + .../tensorflow/lite/micro/micro_time.cpp | 62 + .../tensorflow/lite/micro/micro_time.h | 34 + .../tensorflow/lite/micro/micro_utils.cpp | 83 + .../tensorflow/lite/micro/micro_utils.h | 137 + .../lite/micro/recording_micro_allocator.cpp | 247 + .../lite/micro/recording_micro_allocator.h | 128 + .../lite/micro/recording_micro_interpreter.h | 68 + .../recording_simple_memory_allocator.cpp | 87 + .../micro/recording_simple_memory_allocator.h | 67 + .../lite/micro/simple_memory_allocator.cpp | 152 + .../lite/micro/simple_memory_allocator.h | 115 + .../tensorflow/lite/micro/test_helpers.cpp | 1082 + .../tensorflow/lite/micro/test_helpers.h | 244 + .../lite/micro/testing/micro_test.h | 241 + .../lite/micro/testing/test_conv_model.cpp | 1802 ++ .../lite/micro/testing/test_conv_model.h | 26 + .../cmsis/CMSIS/Core/Include/cmsis_compiler.h | 286 + .../CMSIS/DSP/Include/arm_common_tables.h | 532 + .../CMSIS/DSP/Include/arm_helium_utils.h | 751 + .../cmsis/CMSIS/DSP/Include/arm_math.h | 249 + .../cmsis/CMSIS/DSP/Include/arm_math_memory.h | 243 + .../cmsis/CMSIS/DSP/Include/arm_math_types.h | 601 + .../DSP/Include/dsp/basic_math_functions.h | 702 + .../CMSIS/DSP/Include/dsp/bayes_functions.h | 89 + .../DSP/Include/dsp/complex_math_functions.h | 297 + .../DSP/Include/dsp/controller_functions.h | 793 + .../DSP/Include/dsp/distance_functions.h | 299 + .../DSP/Include/dsp/fast_math_functions.h | 290 + .../DSP/Include/dsp/filtering_functions.h | 2442 +++ .../DSP/Include/dsp/interpolation_functions.h | 321 + .../CMSIS/DSP/Include/dsp/matrix_functions.h | 600 + .../cmsis/CMSIS/DSP/Include/dsp/none.h | 579 + .../DSP/Include/dsp/statistics_functions.h | 486 + .../CMSIS/DSP/Include/dsp/support_functions.h | 429 + .../cmsis/CMSIS/DSP/Include/dsp/svm_defines.h | 45 + .../CMSIS/DSP/Include/dsp/svm_functions.h | 301 + .../DSP/Include/dsp/transform_functions.h | 594 + .../cmsis/CMSIS/DSP/Include/dsp/utils.h | 242 + .../cmsis/CMSIS/NN/Include/arm_nn_tables.h | 59 + .../cmsis/CMSIS/NN/Include/arm_nn_types.h | 133 + .../cmsis/CMSIS/NN/Include/arm_nnfunctions.h | 2127 ++ .../CMSIS/NN/Include/arm_nnsupportfunctions.h | 957 + .../arm_nn_activations_q15.c | 100 + .../arm_nn_activations_q7.c | 93 + .../Source/ActivationFunctions/arm_relu6_s8.c | 68 + .../Source/ActivationFunctions/arm_relu_q15.c | 107 + .../Source/ActivationFunctions/arm_relu_q7.c | 112 + .../arm_elementwise_add_s8.c | 261 + .../arm_elementwise_mul_s8.c | 205 + .../arm_concatenation_s8_w.c | 68 + .../arm_concatenation_s8_x.c | 77 + .../arm_concatenation_s8_y.c | 78 + .../arm_concatenation_s8_z.c | 77 + .../arm_convolve_1_x_n_s8.c | 204 + .../arm_convolve_1x1_HWC_q7_fast_nonsquare.c | 239 + .../arm_convolve_1x1_s8_fast.c | 189 + .../arm_convolve_HWC_q15_basic.c | 210 + .../arm_convolve_HWC_q15_fast.c | 258 + .../arm_convolve_HWC_q15_fast_nonsquare.c | 268 + .../arm_convolve_HWC_q7_RGB.c | 282 + .../arm_convolve_HWC_q7_basic.c | 234 + .../arm_convolve_HWC_q7_basic_nonsquare.c | 232 + .../arm_convolve_HWC_q7_fast.c | 411 + .../arm_convolve_HWC_q7_fast_nonsquare.c | 382 + .../ConvolutionFunctions/arm_convolve_s8.c | 385 + .../arm_convolve_wrapper_s8.c | 151 + .../arm_depthwise_conv_3x3_s8.c | 216 + .../arm_depthwise_conv_s8.c | 252 + .../arm_depthwise_conv_s8_opt.c | 428 + .../arm_depthwise_conv_u8_basic_ver1.c | 297 + .../arm_depthwise_conv_wrapper_s8.c | 136 + .../arm_depthwise_separable_conv_HWC_q7.c | 421 + ...epthwise_separable_conv_HWC_q7_nonsquare.c | 416 + .../arm_nn_depthwise_conv_s8_core.c | 222 + .../arm_nn_mat_mult_kernel_q7_q15.c | 182 + .../arm_nn_mat_mult_kernel_q7_q15_reordered.c | 132 + .../arm_nn_mat_mult_kernel_s8_s16.c | 394 + .../arm_nn_mat_mult_kernel_s8_s16_reordered.c | 204 + .../ConvolutionFunctions/arm_nn_mat_mult_s8.c | 184 + .../arm_fully_connected_mat_q7_vec_q15.c | 202 + .../arm_fully_connected_mat_q7_vec_q15_opt.c | 407 + .../arm_fully_connected_q15.c | 196 + .../arm_fully_connected_q15_opt.c | 335 + .../arm_fully_connected_q7.c | 201 + .../arm_fully_connected_q7_opt.c | 487 + .../arm_fully_connected_s8.c | 100 + .../arm_nn_accumulate_q7_to_q15.c | 83 + .../Source/NNSupportFunctions/arm_nn_add_q7.c | 84 + .../arm_nn_depthwise_conv_nt_t_padded_s8.c | 172 + .../arm_nn_depthwise_conv_nt_t_s8.c | 174 + .../arm_nn_mat_mul_core_1x_s8.c | 94 + .../arm_nn_mat_mul_core_4x_s8.c | 121 + .../arm_nn_mat_mult_nt_t_s8.c | 587 + .../NNSupportFunctions/arm_nn_mult_q15.c | 148 + .../NNSupportFunctions/arm_nn_mult_q7.c | 121 + .../arm_nn_vec_mat_mult_t_s8.c | 465 + .../Source/NNSupportFunctions/arm_nntables.c | 300 + .../arm_q7_to_q15_no_shift.c | 125 + .../arm_q7_to_q15_reordered_no_shift.c | 147 + .../arm_q7_to_q15_reordered_with_offset.c | 103 + .../arm_q7_to_q15_with_offset.c | 120 + .../Source/PoolingFunctions/arm_avgpool_s8.c | 366 + .../Source/PoolingFunctions/arm_max_pool_s8.c | 238 + .../Source/PoolingFunctions/arm_pool_q7_HWC.c | 457 + .../Source/ReshapeFunctions/arm_reshape_s8.c | 60 + .../NN/Source/SVDFunctions/arm_svdf_s8.c | 228 + .../Source/SoftmaxFunctions/arm_softmax_q15.c | 122 + .../Source/SoftmaxFunctions/arm_softmax_q7.c | 111 + .../Source/SoftmaxFunctions/arm_softmax_s8.c | 260 + .../Source/SoftmaxFunctions/arm_softmax_u8.c | 103 + .../arm_softmax_with_batch_q7.c | 78 + .../tools/make/downloads}/kissfft/COPYING | 0 .../make/downloads/kissfft/_kiss_fft_guts.h | 167 + .../tools/make/downloads/kissfft/kiss_fft.c | 411 + .../tools/make/downloads/kissfft/kiss_fft.h | 133 + .../make/downloads/kissfft/tools/kiss_fftr.c | 162 + .../make/downloads}/kissfft/tools/kiss_fftr.h | 5 +- .../lite/portable_type_to_tflitetype.h | 74 + .../tensorflow/lite/schema/schema_generated.h | 17150 ++++++++++++++++ .../tensorflow/lite/schema/schema_utils.cpp | 65 + .../tensorflow/lite/schema/schema_utils.h | 36 + .../tensorflow/lite/version.h | 5 +- .../third_party/flatbuffers/LICENSE.txt | 202 + .../flatbuffers/include/flatbuffers/base.h | 430 + .../include/flatbuffers/flatbuffers.h | 2830 +++ .../include/flatbuffers/flexbuffers.h | 1634 ++ .../include/flatbuffers/stl_emulation.h | 452 + .../flatbuffers/include/flatbuffers/util.h | 695 + .../third_party/gemmlowp/LICENSE | 202 + .../gemmlowp/fixedpoint/fixedpoint.h | 903 + .../gemmlowp/fixedpoint/fixedpoint_neon.h | 334 + .../gemmlowp/fixedpoint/fixedpoint_sse.h | 387 + .../gemmlowp/internal/detect_platform.h | 169 + .../third_party/kissfft/COPYING | 11 + .../third_party/kissfft/_kiss_fft_guts.h | 167 + .../third_party/kissfft/kiss_fft.h | 133 + .../third_party/kissfft/tools/kiss_fftr.h | 49 + .../ruy}/ruy/profiler/instrumentation.h | 9 +- .../tensorflow/core/public/version.h | 5 +- .../tensorflow/lite/c/builtin_op_data.h | 21 +- .../tensorflow/lite/c/c_api_internal.c | 188 + .../tensorflow/lite/c/c_api_internal.h} | 192 +- .../lite/core/api/error_reporter.cpp | 41 + .../tensorflow/lite/core/api/error_reporter.h | 3 + .../lite/core/api/flatbuffer_conversions.cpp | 32 +- .../lite/core/api/flatbuffer_conversions.h | 11 +- .../tensorflow/lite/core/api/op_resolver.cpp | 5 +- .../tensorflow/lite/core/api/op_resolver.h | 9 +- .../tensorflow/lite/core/api/tensor_utils.cpp | 7 +- .../tensorflow/lite/core/api/tensor_utils.h | 31 + .../lite/experimental/micro/compatibility.h | 35 + .../lite/experimental}/micro/debug_log.cpp | 9 +- .../lite/experimental/micro/debug_log.h | 26 + .../experimental}/micro/debug_log_numbers.cpp | 7 +- .../experimental}/micro/debug_log_numbers.h | 9 +- .../micro/kernels/activation_utils.h | 35 +- .../micro/kernels/activations.cpp | 31 +- .../lite/experimental}/micro/kernels/add.cpp | 48 +- .../micro/kernels/all_ops_resolver.cpp | 54 +- .../micro/kernels/all_ops_resolver.h | 13 +- .../micro/kernels/arg_min_max.cpp | 25 +- .../lite/experimental}/micro/kernels/ceil.cpp | 16 +- .../micro/kernels/comparisons.cpp | 47 +- .../lite/experimental}/micro/kernels/conv.cpp | 31 +- .../micro/kernels/depthwise_conv.cpp | 41 +- .../micro/kernels/dequantize.cpp | 20 +- .../micro/kernels/elementwise.cpp | 65 +- .../experimental}/micro/kernels/floor.cpp | 16 +- .../micro/kernels/fully_connected.cpp | 27 +- .../experimental}/micro/kernels/logical.cpp | 22 +- .../experimental}/micro/kernels/logistic.cpp | 25 +- .../micro/kernels/maximum_minimum.cpp | 37 +- .../experimental}/micro/kernels/micro_ops.h | 15 +- .../experimental/micro/kernels/micro_utils.h | 40 + .../lite/experimental}/micro/kernels/neg.cpp | 15 +- .../lite/experimental}/micro/kernels/pack.cpp | 15 +- .../experimental}/micro/kernels/pooling.cpp | 110 +- .../experimental}/micro/kernels/prelu.cpp | 25 +- .../experimental}/micro/kernels/quantize.cpp | 22 +- .../experimental}/micro/kernels/reshape.cpp | 21 +- .../experimental}/micro/kernels/round.cpp | 16 +- .../experimental}/micro/kernels/softmax.cpp | 82 +- .../experimental}/micro/kernels/split.cpp | 15 +- .../micro/kernels/strided_slice.cpp | 21 +- .../lite/experimental}/micro/kernels/svdf.cpp | 447 +- .../experimental}/micro/kernels/unpack.cpp | 16 +- .../experimental}/micro/memory_helpers.cpp | 7 +- .../lite/experimental}/micro/memory_helpers.h | 15 +- .../memory_planner/greedy_memory_planner.cpp | 13 +- .../memory_planner/greedy_memory_planner.h | 16 +- .../memory_planner/linear_memory_planner.cpp | 7 +- .../memory_planner/linear_memory_planner.h | 50 + .../micro/memory_planner/memory_planner.h | 74 + .../experimental/micro/micro_allocator.cpp | 504 + .../lite/experimental/micro/micro_allocator.h | 96 + .../micro/micro_error_reporter.cpp | 5 +- .../experimental/micro/micro_error_reporter.h | 39 + .../experimental}/micro/micro_interpreter.cpp | 121 +- .../experimental}/micro/micro_interpreter.h | 87 +- .../micro/micro_mutable_op_resolver.cpp | 83 + .../micro/micro_mutable_op_resolver.h | 49 + .../micro/micro_optional_debug_tools.cpp | 56 +- .../micro/micro_optional_debug_tools.h | 24 +- .../lite/experimental}/micro/micro_utils.cpp | 86 +- .../lite/experimental}/micro/micro_utils.h | 17 +- .../micro/simple_memory_allocator.cpp | 24 +- .../micro/simple_memory_allocator.h | 26 +- .../lite/experimental}/micro/test_helpers.cpp | 222 +- .../lite/experimental}/micro/test_helpers.h | 34 +- .../experimental}/micro/testing/micro_test.h | 17 +- .../experimental}/micro/testing/test_utils.h | 37 +- .../experimental/microfrontend/lib/bits.h | 105 + .../experimental/microfrontend/lib/fft.cpp | 56 + .../lite/experimental/microfrontend/lib/fft.h | 53 + .../microfrontend/lib/fft_util.cpp | 75 + .../experimental/microfrontend/lib/fft_util.h | 37 + .../microfrontend/lib/filterbank.c | 137 + .../microfrontend/lib/filterbank.h | 66 + .../microfrontend/lib/filterbank_util.c | 223 + .../microfrontend/lib/filterbank_util.h | 53 + .../experimental/microfrontend/lib/frontend.c | 75 + .../experimental/microfrontend/lib/frontend.h | 67 + .../microfrontend/lib/frontend_util.c | 88 + .../microfrontend/lib/frontend_util.h | 55 + .../experimental/microfrontend/lib/log_lut.c | 33 + .../experimental/microfrontend/lib/log_lut.h | 43 + .../microfrontend/lib/log_scale.c | 86 + .../microfrontend/lib/log_scale.h | 42 + .../microfrontend/lib/log_scale_util.c | 30 + .../microfrontend/lib/log_scale_util.h | 48 + .../microfrontend/lib/noise_reduction.c | 54 + .../microfrontend/lib/noise_reduction.h | 49 + .../microfrontend/lib/noise_reduction_util.c | 48 + .../microfrontend/lib/noise_reduction_util.h | 53 + .../microfrontend/lib/pcan_gain_control.c | 59 + .../microfrontend/lib/pcan_gain_control.h | 49 + .../lib/pcan_gain_control_util.c | 95 + .../lib/pcan_gain_control_util.h | 60 + .../experimental/microfrontend/lib/window.c | 73 + .../experimental/microfrontend/lib/window.h | 52 + .../microfrontend/lib/window_util.c | 76 + .../microfrontend/lib/window_util.h | 48 + .../tensorflow/lite/kernels/internal/common.h | 97 +- .../lite/kernels/internal/compatibility.h | 6 +- .../kernels/internal/optimized/neon_check.h | 3 + .../kernels/internal/quantization_util.cpp | 28 +- .../lite/kernels/internal/quantization_util.h | 295 + .../lite/kernels/internal/reference/add.h | 7 +- .../kernels/internal/reference/arg_min_max.h | 71 + .../internal/reference/binary_function.h | 9 +- .../lite/kernels/internal/reference/ceil.h | 40 + .../kernels/internal/reference/comparisons.h | 7 +- .../lite/kernels/internal/reference/conv.h | 7 +- .../internal/reference/depthwiseconv_float.h | 103 + .../internal/reference/depthwiseconv_uint8.h | 11 +- .../kernels/internal/reference/dequantize.h | 7 +- .../lite/kernels/internal/reference/floor.h | 42 + .../internal/reference/fully_connected.h | 11 +- .../internal/reference/integer_ops/add.h | 7 +- .../internal/reference/integer_ops/conv.h | 9 +- .../reference/integer_ops/depthwise_conv.h | 5 +- .../reference/integer_ops/fully_connected.h | 5 +- .../internal/reference/integer_ops/pooling.h | 5 +- .../kernels/internal/reference/logistic.h | 37 +- .../internal/reference/maximum_minimum.h | 7 +- .../lite/kernels/internal/reference/neg.h | 40 + .../lite/kernels/internal/reference/pooling.h | 11 +- .../lite/kernels/internal/reference/prelu.h | 9 +- .../reference/process_broadcast_shapes.h | 5 +- .../kernels/internal/reference/quantize.h | 13 +- .../lite/kernels/internal/reference/round.h | 54 + .../lite/kernels/internal/reference/softmax.h | 23 +- .../internal/reference/strided_slice.h | 9 +- .../tensorflow/lite/kernels/internal/round.h | 3 + .../kernels/internal/strided_slice_logic.h | 7 +- .../tensorflow/lite/kernels/internal/tensor.h | 11 +- .../lite/kernels/internal/tensor_ctypes.h | 50 + .../tensorflow/lite/kernels/internal/types.h | 18 +- .../tensorflow/lite/kernels/kernel_util.cpp | 112 +- .../tensorflow/lite/kernels/kernel_util.h | 39 +- .../tensorflow/lite/kernels/op_macros.h | 5 +- .../tensorflow/lite/kernels/padding.h | 83 + .../tensorflow/lite/schema/schema_generated.h | 714 +- .../tensorflow/lite/string_type.h | 3 + .../tensorflow/lite/string_util.h | 9 +- .../tensorflow/lite/version.h | 32 + .../third_party/flatbuffers/LICENSE.txt | 202 + .../flatbuffers/include}/flatbuffers/base.h | 8 +- .../include}/flatbuffers/flatbuffers.h | 5 +- .../include}/flatbuffers/stl_emulation.h | 3 + .../third_party/gemmlowp/LICENSE | 202 + .../gemmlowp}/fixedpoint/fixedpoint.h | 3 + .../gemmlowp}/fixedpoint/fixedpoint_sse.h | 3 + .../gemmlowp}/internal/detect_platform.h | 3 + .../third_party/kissfft/COPYING | 11 + .../third_party}/kissfft/_kiss_fft_guts.h | 3 + .../third_party/kissfft/kiss_fft.c | 411 + .../third_party}/kissfft/kiss_fft.h | 4 + .../third_party/kissfft/tools/kiss_fftr.c | 162 + .../third_party/kissfft/tools/kiss_fftr.h | 49 + 509 files changed, 102741 insertions(+), 4338 deletions(-) create mode 100644 src/TfLiteARM.h create mode 100644 src/TfLiteAbstract.h create mode 100644 src/TfLiteESP32.h delete mode 100644 src/tensorflow/.DS_Store delete mode 100644 src/tensorflow/lite/.DS_Store delete mode 100644 src/tensorflow/lite/kernels/.DS_Store delete mode 100644 src/tensorflow/lite/kernels/internal/reference/integer_ops/softmax.h delete mode 100644 src/tensorflow/lite/micro/kernels/concatenation.cpp delete mode 100644 src/tensorflow/lite/micro/kernels/mul.cpp delete mode 100644 src/tensorflow/lite/micro/kernels/pad.cpp delete mode 100644 src/tensorflow/lite/micro/micro_allocator.cpp delete mode 100644 src/tensorflow/lite/micro/micro_allocator.h delete mode 100644 src/tensorflow/lite/micro/micro_mutable_op_resolver.h delete mode 100644 src/tensorflow/lite/type_to_tflitetype.h create mode 100644 src/tensorflow_arm/tensorflow/core/public/version.h create mode 100644 src/tensorflow_arm/tensorflow/lite/c/builtin_op_data.h create mode 100644 src/tensorflow_arm/tensorflow/lite/c/c_api_types.h rename src/{ => tensorflow_arm}/tensorflow/lite/c/common.c (89%) create mode 100644 src/tensorflow_arm/tensorflow/lite/c/common.h rename src/{ => tensorflow_arm}/tensorflow/lite/core/api/error_reporter.cpp (90%) create mode 100644 src/tensorflow_arm/tensorflow/lite/core/api/error_reporter.h create mode 100644 src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.h create mode 100644 src/tensorflow_arm/tensorflow/lite/core/api/profiler.h create mode 100644 src/tensorflow_arm/tensorflow/lite/core/api/tensor_utils.cpp rename src/{ => tensorflow_arm}/tensorflow/lite/core/api/tensor_utils.h (89%) create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.h create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.c create mode 100644 src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/common.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/max.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/min.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/optimized/neon_check.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.cpp rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/quantization_util.h (98%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/add.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/arg_min_max.h (95%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/binary_function.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/ceil.h (91%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/comparisons.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/concatenation.h (82%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/conv.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h (94%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/dequantize.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/floor.h (91%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/fully_connected.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/hard_swish.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/add.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mean.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h (65%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/l2normalization.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/logistic.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/maximum_minimum.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/mul.h (84%) rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/neg.h (91%) rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/pad.h (80%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/pooling.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/prelu.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/quantize.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/reduce.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/requantize.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/reference/round.h (93%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/softmax.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/strided_slice.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/sub.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/tanh.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/transpose_conv.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/strided_slice_logic.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/internal/tensor_ctypes.h (89%) create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/internal/types.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/kernels/op_macros.h rename src/{ => tensorflow_arm}/tensorflow/lite/kernels/padding.h (96%) create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/arduino/debug_log.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.h rename src/{ => tensorflow_arm}/tensorflow/lite/micro/compatibility.h (96%) rename src/{ => tensorflow_arm}/tensorflow/lite/micro/debug_log.h (83%) create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/activation_utils.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/activations.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/arg_min_max.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/ceil.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/circular_buffer.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/add.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/conv.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/mul.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/svdf.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/comparisons.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/concatenation.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test_common.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/dequantize.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/detection_postprocess.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/elementwise.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/floor.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/fully_connected.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/hard_swish.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util_full.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/l2norm.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/logical.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/logistic.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/maximum_minimum.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/micro_ops.h rename src/{ => tensorflow_arm}/tensorflow/lite/micro/kernels/micro_utils.h (95%) create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/neg.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/pack.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/pad.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/prelu.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize_common.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/reduce.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/reshape.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/resize_nearest_neighbor.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/round.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/shape.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/split.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/split_v.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/strided_slice.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/sub.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf_common.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/tanh.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/transpose_conv.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/kernels/unpack.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.cpp rename src/{ => tensorflow_arm}/tensorflow/lite/micro/memory_planner/linear_memory_planner.h (89%) rename src/{ => tensorflow_arm}/tensorflow/lite/micro/memory_planner/memory_planner.h (91%) create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.cpp rename src/{ => tensorflow_arm}/tensorflow/lite/micro/micro_error_reporter.h (80%) create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_mutable_op_resolver.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_string.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_string.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_time.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_time.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_utils.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/micro_utils.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/recording_micro_interpreter.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/test_helpers.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/test_helpers.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/testing/micro_test.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/Core/Include/cmsis_compiler.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_helium_utils.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/bayes_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/complex_math_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/controller_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/distance_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/filtering_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/interpolation_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/matrix_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/statistics_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/support_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_defines.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/transform_functions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_tables.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q7.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu6_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q7.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_add_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_mul_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_w.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_x.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_y.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_z.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1_x_n_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_wrapper_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_3x3_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8_opt.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_u8_basic_ver1.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_wrapper_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_depthwise_conv_s8_core.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16_reordered.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_accumulate_q7_to_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_add_q7.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_padded_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_1x_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_4x_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_vec_mat_mult_t_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nntables.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_with_offset.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_with_offset.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_avgpool_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_max_pool_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ReshapeFunctions/arm_reshape_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SVDFunctions/arm_svdf_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c rename src/{ => tensorflow_arm/tensorflow/lite/micro/tools/make/downloads}/kissfft/COPYING (100%) create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/_kiss_fft_guts.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.c create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.h create mode 100644 src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/tools/kiss_fftr.c rename src/{ => tensorflow_arm/tensorflow/lite/micro/tools/make/downloads}/kissfft/tools/kiss_fftr.h (88%) create mode 100644 src/tensorflow_arm/tensorflow/lite/portable_type_to_tflitetype.h create mode 100644 src/tensorflow_arm/tensorflow/lite/schema/schema_generated.h create mode 100644 src/tensorflow_arm/tensorflow/lite/schema/schema_utils.cpp create mode 100644 src/tensorflow_arm/tensorflow/lite/schema/schema_utils.h rename src/{ => tensorflow_arm}/tensorflow/lite/version.h (91%) create mode 100644 src/tensorflow_arm/third_party/flatbuffers/LICENSE.txt create mode 100644 src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h create mode 100644 src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h create mode 100644 src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flexbuffers.h create mode 100644 src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/stl_emulation.h create mode 100644 src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/util.h create mode 100644 src/tensorflow_arm/third_party/gemmlowp/LICENSE create mode 100644 src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h create mode 100644 src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_neon.h create mode 100644 src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_sse.h create mode 100644 src/tensorflow_arm/third_party/gemmlowp/internal/detect_platform.h create mode 100644 src/tensorflow_arm/third_party/kissfft/COPYING create mode 100644 src/tensorflow_arm/third_party/kissfft/_kiss_fft_guts.h create mode 100644 src/tensorflow_arm/third_party/kissfft/kiss_fft.h create mode 100644 src/tensorflow_arm/third_party/kissfft/tools/kiss_fftr.h rename src/{tensorflow/lite/experimental => tensorflow_arm/third_party/ruy}/ruy/profiler/instrumentation.h (96%) rename src/{ => tensorflow_esp32}/tensorflow/core/public/version.h (98%) rename src/{ => tensorflow_esp32}/tensorflow/lite/c/builtin_op_data.h (94%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/c/c_api_internal.c rename src/{tensorflow/lite/c/common.h => tensorflow_esp32/tensorflow/lite/c/c_api_internal.h} (80%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/core/api/error_reporter.cpp rename src/{ => tensorflow_esp32}/tensorflow/lite/core/api/error_reporter.h (96%) rename src/{ => tensorflow_esp32}/tensorflow/lite/core/api/flatbuffer_conversions.cpp (97%) rename src/{ => tensorflow_esp32}/tensorflow/lite/core/api/flatbuffer_conversions.h (89%) rename src/{ => tensorflow_esp32}/tensorflow/lite/core/api/op_resolver.cpp (94%) rename src/{ => tensorflow_esp32}/tensorflow/lite/core/api/op_resolver.h (88%) rename src/{ => tensorflow_esp32}/tensorflow/lite/core/api/tensor_utils.cpp (90%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/debug_log.cpp (88%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.h rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/debug_log_numbers.cpp (96%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/debug_log_numbers.h (77%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/activation_utils.h (63%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/activations.cpp (85%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/add.cpp (82%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/all_ops_resolver.cpp (65%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/all_ops_resolver.h (67%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/arg_min_max.cpp (82%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/ceil.cpp (81%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/comparisons.cpp (90%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/conv.cpp (92%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/depthwise_conv.cpp (90%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/dequantize.cpp (80%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/elementwise.cpp (70%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/floor.cpp (73%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/fully_connected.cpp (90%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/logical.cpp (77%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/logistic.cpp (69%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/maximum_minimum.cpp (76%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/micro_ops.h (89%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_utils.h rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/neg.cpp (77%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/pack.cpp (89%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/pooling.cpp (69%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/prelu.cpp (84%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/quantize.cpp (84%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/reshape.cpp (84%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/round.cpp (81%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/softmax.cpp (72%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/split.cpp (90%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/strided_slice.cpp (91%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/svdf.cpp (52%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/kernels/unpack.cpp (88%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/memory_helpers.cpp (93%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/memory_helpers.h (77%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/memory_planner/greedy_memory_planner.cpp (97%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/memory_planner/greedy_memory_planner.h (92%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/memory_planner/linear_memory_planner.cpp (88%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/memory_planner.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.cpp create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.h rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/micro_error_reporter.cpp (93%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/micro_interpreter.cpp (70%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/micro_interpreter.h (56%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.cpp create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.h rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/micro_optional_debug_tools.cpp (76%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/micro_optional_debug_tools.h (63%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/micro_utils.cpp (71%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/micro_utils.h (87%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/simple_memory_allocator.cpp (71%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/simple_memory_allocator.h (74%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/test_helpers.cpp (65%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/test_helpers.h (81%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/testing/micro_test.h (95%) rename src/{tensorflow/lite => tensorflow_esp32/tensorflow/lite/experimental}/micro/testing/test_utils.h (90%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.cpp create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.h create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.c create mode 100644 src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/common.h (90%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/compatibility.h (96%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/optimized/neon_check.h (96%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/quantization_util.cpp (94%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/add.h (98%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/arg_min_max.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/binary_function.h (92%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/ceil.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/comparisons.h (98%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/conv.h (98%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h (96%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/dequantize.h (88%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/floor.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/fully_connected.h (97%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/integer_ops/add.h (97%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h (95%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h (98%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h (95%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h (98%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/logistic.h (69%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/maximum_minimum.h (92%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/neg.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/pooling.h (97%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/prelu.h (92%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h (97%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/quantize.h (83%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/round.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/softmax.h (90%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/reference/strided_slice.h (92%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/round.h (96%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/strided_slice_logic.h (97%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/tensor.h (93%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/internal/types.h (98%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/kernel_util.cpp (68%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/kernel_util.h (83%) rename src/{ => tensorflow_esp32}/tensorflow/lite/kernels/op_macros.h (93%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/kernels/padding.h rename src/{ => tensorflow_esp32}/tensorflow/lite/schema/schema_generated.h (95%) mode change 100755 => 100644 rename src/{ => tensorflow_esp32}/tensorflow/lite/string_type.h (94%) rename src/{ => tensorflow_esp32}/tensorflow/lite/string_util.h (94%) create mode 100644 src/tensorflow_esp32/tensorflow/lite/version.h create mode 100644 src/tensorflow_esp32/third_party/flatbuffers/LICENSE.txt rename src/{ => tensorflow_esp32/third_party/flatbuffers/include}/flatbuffers/base.h (98%) rename src/{ => tensorflow_esp32/third_party/flatbuffers/include}/flatbuffers/flatbuffers.h (99%) rename src/{ => tensorflow_esp32/third_party/flatbuffers/include}/flatbuffers/stl_emulation.h (99%) create mode 100644 src/tensorflow_esp32/third_party/gemmlowp/LICENSE rename src/{ => tensorflow_esp32/third_party/gemmlowp}/fixedpoint/fixedpoint.h (99%) rename src/{ => tensorflow_esp32/third_party/gemmlowp}/fixedpoint/fixedpoint_sse.h (99%) rename src/{ => tensorflow_esp32/third_party/gemmlowp}/internal/detect_platform.h (98%) create mode 100644 src/tensorflow_esp32/third_party/kissfft/COPYING rename src/{ => tensorflow_esp32/third_party}/kissfft/_kiss_fft_guts.h (99%) create mode 100644 src/tensorflow_esp32/third_party/kissfft/kiss_fft.c rename src/{ => tensorflow_esp32/third_party}/kissfft/kiss_fft.h (97%) create mode 100644 src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.c create mode 100644 src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.h diff --git a/.DS_Store b/.DS_Store index 578602b5a7937c550300ae5f47324030aa8357c4..bf6aca59f7513140ac9cd89b630e7d62df0d238c 100644 GIT binary patch delta 70 zcmZoMXffE}$i#SJvJ+E-hInM8XF_ge bRdr2m-OS0anB*B}ZRTSB!?Ky3<1aq|@_iN- diff --git a/README.md b/README.md index c2a4d89..976ed83 100644 --- a/README.md +++ b/README.md @@ -14,6 +14,14 @@ Clone this repo in you Arduino libraries folder. git clone https://github.com/eloquentarduino/EloquentTinyML.git ``` +## Export TensorFlow Lite model + +To run a model on your microcontroller, you should first have a model. + +I suggest you use [`tinymlgen`](https://github.com/eloquentarduino/tinymlgen) to complete this step: +it will export your TensorFlow Lite model to a C array ready to be loaded +by this library. + ## Use @@ -25,7 +33,7 @@ git clone https://github.com/eloquentarduino/EloquentTinyML.git #define NUMBER_OF_OUTPUTS 1 #define TENSOR_ARENA_SIZE 2*1024 -Eloquent::TinyML::TinyML< +Eloquent::TinyML::TfLite< NUMBER_OF_INPUTS, NUMBER_OF_OUTPUTS, TENSOR_ARENA_SIZE> ml; @@ -33,7 +41,7 @@ Eloquent::TinyML::TinyML< void setup() { Serial.begin(115200); - ml.begin(sine_model_quantized_tflite); + ml.begin(sine_model); } void loop() { @@ -50,4 +58,13 @@ void loop() { Serial.println(predicted); delay(1000); } -``` \ No newline at end of file +``` + +## Compatibility + +Latest version of this library (2.4.0) is compatible with Cortex-M and ESP32 chips and is built starting from: + + - [Arduino_TensorFlowLite library version 2.4.0-ALPHA](https://www.tensorflow.org/lite/microcontrollers/overview) + - [TensorFlowLite_ESP32 version 0.9.0](https://github.com/tanakamasayuki/Arduino_TensorFlowLite_ESP32) + +ESP32 support is stuck at TensorFlow 2.1.1 at the moment. \ No newline at end of file diff --git a/examples/SineExample/SineExample.ino b/examples/SineExample/SineExample.ino index 00fd343..d308278 100644 --- a/examples/SineExample/SineExample.ino +++ b/examples/SineExample/SineExample.ino @@ -4,7 +4,7 @@ #define NUMBER_OF_INPUTS 1 #define NUMBER_OF_OUTPUTS 1 -// in future projects you may need to tweek this value: it's a trial and error process +// in future projects you may need to tweak this value: it's a trial and error process #define TENSOR_ARENA_SIZE 2*1024 Eloquent::TinyML::TfLite ml; @@ -16,17 +16,19 @@ void setup() { } void loop() { - // pick up a random x and predict its sine - float x = 3.14 * random(100) / 100; - float y = sin(x); - float input[1] = { x }; - float predicted = ml.predict(input); + for (float i = 0; i < 10; i++) { + // pick x from 0 to PI + float x = 3.14 * i / 10; + float y = sin(x); + float input[1] = { x }; + float predicted = ml.predict(input); - Serial.print("sin("); - Serial.print(x); - Serial.print(") = "); - Serial.print(y); - Serial.print("\t predicted: "); - Serial.println(predicted); - delay(1000); + Serial.print("sin("); + Serial.print(x); + Serial.print(") = "); + Serial.print(y); + Serial.print("\t predicted: "); + Serial.println(predicted); + delay(1000); + } } diff --git a/library.json b/library.json index 46a022e..c5246c6 100644 --- a/library.json +++ b/library.json @@ -6,7 +6,7 @@ "type": "git", "url": "https://github.com/eloquentarduino/EloquentTinyML" }, - "version": "0.0.10", + "version": "2.4.0", "authors": { "name": "Simone Salerno", "url": "https://github.com/eloquentarduino" diff --git a/library.properties b/library.properties index 8686c4a..5c67e8e 100644 --- a/library.properties +++ b/library.properties @@ -1,5 +1,5 @@ name=EloquentTinyML -version=0.0.10 +version=2.4.0 author=Simone Salerno,eloquentarduino@gmail.com maintainer=Simone Salerno,eloquentarduino@gmail.com sentence=An eloquent interface to Tensorflow Lite for Microcontrollers diff --git a/src/.DS_Store b/src/.DS_Store index 4add1b3b4a20d22a03eb997fce872bdaf3850b2f..1ee74573d20d6aa7233e87615c615210b0ebb782 100644 GIT binary patch delta 364 zcmZp1XfcprU|?W$DortDU=RQ@Ie-{Mvv5r;6q~50$jG)aU^g=>8%TJvfPjbuLp(z& zLoq`EgE50qQh9MfQcivnQ1ZlN1Azt&@#<=0LmdSpGxJ&?P7ZNZ zLtD>;+{&uzn%cSseo)|PP$=ma(-?BP!a~f7ONmD%FTCifm+P5X6bI2?||?&G9^Qm;vU$OV$7Y literal 8196 zcmeHMO>7la6h2=mFf;sgK#)%P=}Q~_Ld8Cy1wr8P+J--ohPJfj$MRlhpd-_n`et4q zP^(oFG={iQKwYT|iDBW-!iBPNgK?oTihv6jE?nrwrO{~Ox%bX%pzt;>RKnck-tXRf z&Y649nQ!Li%n%W>MLSJYOGFx1f#xi3?oxzY)H_O5JX6gGwkHaxK*z~rEZw12gIg`Qw%6`mZyX|VR5jo^jcx(QiFNGs{FUYg}i~NY^hEa3#4OC65o<3uyrgNF2AMF|n zN?lPdDu&ry;jkEW=Axl|-0uzoZ?d1uA9S74fVrS8@S~jTI|XU%+77E6>2-ZOC}o49 zZ-??aT}mJq8Zc|e#@26aX-chM*E-&m8rzh9q$$<1Aw52>Ym1hrx9&bReEQ7T)8o&) zq5MPqj-Z+;{=QksSuY%IQ_fbsQg-&{q_f5>Yv(RwPoFy&_+9Q9XTYop+1K4X>?}B= z>{st#;1!4caKM~5nDYjUUM_NWc%D!W`y%I5BzlL!pb)i5#R1bShRy+p_1saXoweC? zzPQrKkUH?Sys2jR@&*}Da!_GW3>H4(?7uC+O>Z~PY(Nq)7Y}ERSdH1QJ8=aPG zEUUMSBR21DUto(>Z1a+)OG#tBWgQlhwt~a0t!Z+4vV%|(vS2r{Yt;nAM_{vg&C;95-dYKQdo@^v|=MN=)g|w!fqTuFZyu^hhgJM zco>F{5}w6#cpfj{9M0ncF5)e`j}P!6KEbEBf~)u(U*kJmS9`1~Gj^{Ums{NA-d!J0 zW^DX;R`uMCv32boM)p5t?Dw-3O&?gWa9QKZrgZC;?UT22%00Z5Ns>N_ACfm^UdY>c z8yN+mt-8g=(n^HxVcI#TG{JszrqWK;HE2v?Q!5WwXqs6i2gd5UH9GUwoTg-Jjb@D* zYEDMU_}DHM<3B=^d)`6#QTAM zqF?AY5H!q1tt4Fo8nGN}@i5X@hYi?@ZP<<-*n_>;hsV(=>30N2F@PbsILRaw6g-79 z7{fTu;zhiKm+>lIV=`XE>v$XQP9fgTGUXKFQYGSzb`^Xt@K3P3OxoK`lhM5rfoW=^ zwIcuTsQmr^UU4VcXA*%V0{ +#include #ifdef max #define REDEFINE_MAX @@ -8,230 +9,14 @@ #undef min #endif -#include -#include "tensorflow/lite/version.h" -#include "tensorflow/lite/micro/kernels/all_ops_resolver.h" -#include "tensorflow/lite/micro/micro_error_reporter.h" -#include "tensorflow/lite/micro/micro_interpreter.h" + +#if defined(ESP32) +#include "TfLiteESP32.h" +#else +#include "TfLiteARM.h" +#endif #ifdef REDEFINE_MAX #define max(a,b) ((a)>(b)?(a):(b)) #define min(a,b) ((a)<(b)?(a):(b)) #endif - - -namespace Eloquent { - namespace TinyML { - - enum TfLiteError { - OK, - VERSION_MISMATCH, - CANNOT_ALLOCATE_TENSORS, - NOT_INITIALIZED, - INVOKE_ERROR - }; - - /** - * Eloquent interface to Tensorflow Lite for Microcontrollers - * - * @tparam inputSize - * @tparam outputSize - * @tparam tensorArenaSize how much memory to allocate to the tensors - */ - template - class TfLite { - public: - /** - * Contructor - * @param modelData a model as exported by tinymlgen - */ - TfLite() : - failed(false) { - } - - ~TfLite() { - delete reporter; - delete interpreter; - } - - /** - * Inizialize NN - * - * @param modelData - * @return - */ - bool begin(const unsigned char *modelData) { - tflite::ops::micro::AllOpsResolver resolver; - reporter = new tflite::MicroErrorReporter(); - - model = tflite::GetModel(modelData); - - // assert model version and runtime version match - if (model->version() != TFLITE_SCHEMA_VERSION) { - failed = true; - error = VERSION_MISMATCH; - - reporter->Report( - "Model provided is schema version %d not equal " - "to supported version %d.", - model->version(), TFLITE_SCHEMA_VERSION); - - return false; - } - - interpreter = new tflite::MicroInterpreter(model, resolver, tensorArena, tensorArenaSize, reporter); - - if (interpreter->AllocateTensors() != kTfLiteOk) { - failed = true; - error = CANNOT_ALLOCATE_TENSORS; - - return false; - } - - input = interpreter->input(0); - output = interpreter->output(0); - error = OK; - - return true; - } - - /** - * Test if the initialization completed fine - */ - bool initialized() { - return !failed; - } - - /** - * - * @param input - * @param output - * @return - */ - uint8_t predict(uint8_t *input, uint8_t *output = NULL) { - // abort if initialization failed - if (!initialized()) - return sqrt(-1); - - memcpy(this->input->data.uint8, input, sizeof(uint8_t) * inputSize); - - if (interpreter->Invoke() != kTfLiteOk) { - reporter->Report("Inference failed"); - - return sqrt(-1); - } - - // copy output - if (output != NULL) { - for (uint16_t i = 0; i < outputSize; i++) - output[i] = this->output->data.uint8[i]; - } - - return this->output->data.uint8[0]; - } - - /** - * Run inference - * @return output[0], so you can use it directly if it's the only output - */ - float predict(float *input, float *output = NULL) { - // abort if initialization failed - if (!initialized()) { - error = NOT_INITIALIZED; - - return sqrt(-1); - } - - // copy input - for (size_t i = 0; i < inputSize; i++) - this->input->data.f[i] = input[i]; - - if (interpreter->Invoke() != kTfLiteOk) { - error = INVOKE_ERROR; - reporter->Report("Inference failed"); - - return sqrt(-1); - } - - // copy output - if (output != NULL) { - for (uint16_t i = 0; i < outputSize; i++) - output[i] = this->output->data.f[i]; - } - - return this->output->data.f[0]; - } - - /** - * Predict class - * @param input - * @return - */ - uint8_t predictClass(float *input) { - float output[outputSize]; - - predict(input, output); - - return probaToClass(output); - } - - /** - * Get class with highest probability - * @param output - * @return - */ - uint8_t probaToClass(float *output) { - uint8_t classIdx = 0; - float maxProba = output[0]; - - for (uint8_t i = 1; i < outputSize; i++) { - if (output[i] > maxProba) { - classIdx = i; - maxProba = output[i]; - } - } - - return classIdx; - } - - /** - * Get error - * @return - */ - TfLiteError getError() { - return error; - } - - /** - * Get error message - * @return - */ - const char* errorMessage() { - switch (error) { - case OK: - return "No error"; - case VERSION_MISMATCH: - return "Version mismatch"; - case CANNOT_ALLOCATE_TENSORS: - return "Cannot allocate tensors"; - case NOT_INITIALIZED: - return "Interpreter has not been initialized"; - case INVOKE_ERROR: - return "Interpreter invoke() returned an error"; - default: - return "Unknown error"; - } - } - - protected: - bool failed; - TfLiteError error; - uint8_t tensorArena[tensorArenaSize]; - tflite::ErrorReporter *reporter; - tflite::MicroInterpreter *interpreter; - TfLiteTensor *input; - TfLiteTensor *output; - const tflite::Model *model; - }; - } -} diff --git a/src/TfLiteARM.h b/src/TfLiteARM.h new file mode 100644 index 0000000..28510e8 --- /dev/null +++ b/src/TfLiteARM.h @@ -0,0 +1,28 @@ +// +// Created by Simone on 28/10/2021. +// + +#ifndef ELOQUENTTINYML_TFLITEARM_H +#define ELOQUENTTINYML_TFLITEARM_H + +#include "tensorflow_arm/tensorflow/lite/version.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_interpreter.h" +#include "TfLiteAbstract.h" + + +namespace Eloquent { + namespace TinyML { + + /** + * Run TensorFlow Lite models on ARM + */ + template + class TfLite : public TfLiteAbstract { + }; + } +} + +#endif //ELOQUENTTINYML_TFLITEESP32_H diff --git a/src/TfLiteAbstract.h b/src/TfLiteAbstract.h new file mode 100644 index 0000000..f6b5793 --- /dev/null +++ b/src/TfLiteAbstract.h @@ -0,0 +1,294 @@ +// +// Created by Simone on 28/10/2021. +// + +#ifndef ELOQUENTTINYML_TFLITEABSTRACT_H +#define ELOQUENTTINYML_TFLITEABSTRACT_H + + +namespace Eloquent { + namespace TinyML { + + /** + * + */ + enum TfLiteError { + OK, + VERSION_MISMATCH, + CANNOT_ALLOCATE_TENSORS, + NOT_INITIALIZED, + INVOKE_ERROR + }; + + /** + * Eloquent interface to Tensorflow Lite for Microcontrollers + * + * @tparam inputSize + * @tparam outputSize + * @tparam tensorArenaSize how much memory to allocate to the tensors + */ + template + class TfLiteAbstract { + public: + /** + * Contructor + */ + TfLiteAbstract() : + failed(false), + shouldRescaleInput(false), + shouldRescaleOutput(false) { + } + + /** + * Destructor + */ + ~TfLiteAbstract() { + delete interpreter; + delete model; + } + + /** + * Inizialize NN + * + * @param modelData + * @return + */ + bool begin(const unsigned char *modelData) { + model = tflite::GetModel(modelData); + + if (model->version() != TFLITE_SCHEMA_VERSION) + return this->abort(VERSION_MISMATCH, false); + + interpreter = new tflite::MicroInterpreter(model, opsResolver, tensorArena, tensorArenaSize, &errorReporter); + + if (interpreter->AllocateTensors() != kTfLiteOk) + return this->abort(CANNOT_ALLOCATE_TENSORS, false); + + input = interpreter->input(0); + output = interpreter->output(0); + error = OK; + failed = false; + + return true; + } + + /** + * Test if the initialization completed fine + */ + bool isInitialized() { + return !failed; + } + + /** + * + * @param on + */ + void turnInputScalingOn(bool on = true) { + shouldRescaleInput = on; + } + + /** + * + * @param on + */ + void turnOutputScalingOn(bool on = true) { + shouldRescaleOutput = on; + } + + /** + * + * @param input + * @param output + * @return + */ + uint8_t predict(uint8_t *input, uint8_t *output = NULL) { + if (!isInitialized()) + return this->abort(NOT_INITIALIZED, 255); + + memcpy(this->input->data.uint8, input, sizeof(uint8_t) * inputSize); + + if (interpreter->Invoke() != kTfLiteOk) + return this->abort(INVOKE_ERROR, 255); + + // copy output + if (output != NULL) { + for (uint16_t i = 0; i < outputSize; i++) { + uint8_t y = this->output->data.uint8[i]; + + output[i] = shouldRescaleOutput ? scaleOutput(y) : y; + } + } + + return this->output->data.uint8[0]; + } + + /** + * + * @param input + * @param output + * @return + */ + int8_t predict(int8_t *input, int8_t *output = NULL) { + if (!isInitialized()) + return this->abort(NOT_INITIALIZED, -127); + + memcpy(this->input->data.int8, input, sizeof(int8_t) * inputSize); + + if (interpreter->Invoke() != kTfLiteOk) + return this->abort(INVOKE_ERROR, -127); + + // copy output + if (output != NULL) { + for (uint16_t i = 0; i < outputSize; i++) { + int8_t y = this->output->data.int8[i]; + + output[i] = shouldRescaleOutput ? scaleOutput(y) : y; + } + } + + return this->output->data.int8[0]; + } + + /** + * Run inference + * @return output[0], so you can use it directly if it's the only output + */ + float predict(float *input, float *output = NULL) { + if (!isInitialized()) + return this->abort(NOT_INITIALIZED, sqrt(-1)); + + // copy input + for (size_t i = 0; i < inputSize; i++) + this->input->data.f[i] = input[i]; + + if (interpreter->Invoke() != kTfLiteOk) + return this->abort(INVOKE_ERROR, sqrt(-1)); + + // copy output + if (output != NULL) { + for (uint16_t i = 0; i < outputSize; i++) { + float y = this->output->data.f[i]; + + output[i] = shouldRescaleOutput ? scaleOutput(y) : y; + } + } + + return this->output->data.f[0]; + } + + /** + * Predict class + * @param input + * @return + */ + uint8_t predictClass(float *input) { + float output[outputSize]; + + predict(input, output); + + return probaToClass(output); + } + + /** + * Get class with highest probability + * @param output + * @return + */ + uint8_t probaToClass(float *output) { + uint8_t classIdx = 0; + float maxProba = output[0]; + + for (uint8_t i = 1; i < outputSize; i++) { + if (output[i] > maxProba) { + classIdx = i; + maxProba = output[i]; + } + } + + return classIdx; + } + + /** + * Apply model scaling to input + * @tparam T + * @param x + * @return + */ + template + T scaleInput(T x) { + return x / this->input->params.zero_point + this->input->params.scale; + } + + /** + * Apply model scaling to output + * @tparam T + * @param y + * @return + */ + template + T scaleOutput(T y) { + return (y - this->output->params.zero_point) * this->output->params.scale; + } + + /** + * Get error + * @return + */ + TfLiteError getError() { + return error; + } + + /** + * Get error message + * @return + */ + const char* errorMessage() { + switch (error) { + case OK: + return "No error"; + case VERSION_MISMATCH: + return "Version mismatch"; + case CANNOT_ALLOCATE_TENSORS: + return "Cannot allocate tensors"; + case NOT_INITIALIZED: + return "Interpreter has not been initialized"; + case INVOKE_ERROR: + return "Interpreter invoke() returned an error"; + default: + return "Unknown error"; + } + } + + protected: + bool failed; + bool shouldRescaleInput; + bool shouldRescaleOutput; + TfLiteError error; + uint8_t tensorArena[tensorArenaSize]; + tflite::MicroErrorReporter errorReporter; + tflite::MicroInterpreter *interpreter; + TfLiteTensor *input; + TfLiteTensor *output; + const tflite::Model *model; + OpsResolver opsResolver; + + /** + * Abort execution with given error code + * + * @tparam T + * @param errorCode + * @param rvalue + * @return + */ + template + T abort(TfLiteError errorCode, T rvalue) { + error = errorCode; + failed = true; + + return rvalue; + } + }; + } +} + +#endif //ELOQUENTTINYML_TFLITE_H diff --git a/src/TfLiteESP32.h b/src/TfLiteESP32.h new file mode 100644 index 0000000..161a817 --- /dev/null +++ b/src/TfLiteESP32.h @@ -0,0 +1,28 @@ +// +// Created by Simone on 28/10/2021. +// + +#ifndef ELOQUENTTINYML_TFLITEESP32_H +#define ELOQUENTTINYML_TFLITEESP32_H + +#include "tensorflow_esp32/tensorflow/lite/version.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.h" +#include "TfLiteAbstract.h" + + +namespace Eloquent { + namespace TinyML { + + /** + * Run TensorFlow Lite models on ESP32 + */ + template + class TfLite : public TfLiteAbstract { + }; + } +} + +#endif //ELOQUENTTINYML_TFLITEESP32_H diff --git a/src/tensorflow/.DS_Store b/src/tensorflow/.DS_Store deleted file mode 100644 index 3e9593f407680487fa549b786233c6449e0a6896..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 6148 zcmeHK%}T>S5Z-O0rij>sV2^w8)-V?fcq3OOj}6d;WRY%R1GKqa;qfFzyXxav1a=+qy|`|vn^%Aa2ps~mnnDeD?1@CwFsgZJ(r zj*|EWe3SS|93e443=jhg#DG2m&D90kGi{X^AO?QM0Gv{qWw&2Ft+f{S#aEtD!OZ9y(2RJPC?T6T8^I(9l!c4oJQ zT2e!ZnqV}>czwW2)CXRSf^T?x0`ahdStq>$dLgDNit zAX!Rkuh2c^0iI9PmkD1^3hB90Op!gncZHu~K)91V%DarIogpD)PzMME2+T!*=k5#H5M}OJW?h}X+i9{~N{^U^{k+wgatG7?eV1*UGxw?V0Ye`hP^;Q(%S{=Up7op!O{XpgyA4aT zM_cWjr8!=lq%e2`4X8EayH`RtPF$la^ftlMy3V%t=? zT2ZM~DNDQahGV2negCMTxr4naC!<@MVIAosa#^l%*wC|b1!8iSl^WKkDr=VAe#f2R z@a*1&C_TBfaWrSR$J9DW5@Xbc>tyx3bWo!w++%AzqUKq`wc(YL)X(*8S)FR3R#t~& zvYL{*970EgSgXWk)see~Xa@+_2Bk?-HF>|8O6#UjH7hNWx?Aqfi)jmGm9j;WdouR$ zFipl0y4G;Jq+XOB<^y-69Cz=CZqe9$YTIkW;a;xas%g3=Nm5-xKjfiTmsllGtP_GU z4-i@9r*W9AU>oW5?P7;XK~Q zhq#DK_za)p3Vy)P_yxb>H%y00LQ1G2R4JY}RM_sFh3EJ6RV0hqPG2NFw!d<1wnrY= zv2&NwHt*T~^%73;TdQtcS{+@!dTm4V=GGY^%v$@aIq1<&@XF&`fluCY7n4}*1tqed zXsM4~Z%e2P`Sy{fc<@xe2{m!O|v}IU~N<^@V2-|?o z*iLlajSlR=0ra2`{Wu66M~JwO5OrPTF@|GA-;+3n$MFQ7#Ityg$om3b#A`&}H}MwU z#yfZy?-7kJ5Q#s+$M}S3JcV!Y-7L~}6q6R;7n63@`BdEbhO|mVlp3D@Z@d2Y{~Iz^ z5F|h#Kwy3%fYSD4dn-L#;W(ZBm?2U`={iIgFTCHRke&-wybITR={U)mKMd(U>e~{& foD`BQRQ~r50e(hP{A)Kj{|i52gY*C9IRF0!+SSN; diff --git a/src/tensorflow/lite/kernels/.DS_Store b/src/tensorflow/lite/kernels/.DS_Store deleted file mode 100644 index 6c23b3d9cdce8d7973c726784bdcf2e865a2fd1e..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 6148 zcmeHK%}T>S5Z-O8rihq>V2^w8)m_gQQ%MFM&pB*%9WMXHKPi$QN4ABQR2jI zJe{=N@iq3&-FWC}`yz^hx$SzF{vhd^TPIN%yM8zr%VNJjfRvjnKkP?IJDP_5i5$mK z0Xb3Etfgu5pj8*m-TkaC(nIU0E?Rq5mQ{`I9qah4cmMd5K4&klid_yLqLfXIGk61| zu+Rr*90pN%2eDb*ET51VAO?tmC1yY$dB*w@uaTBU3=jjqV*t+w0g7mA%oWO`0~-8& z#CQb}1#EmvAPSAP##|vpK)5OeRHa-$F}NxRztDNM#$2H)XI#z<koX=H&|K kC1A*_7;^C{u7OGczd!@f)|e{<4+#ATNE)ai2L6VgLXD diff --git a/src/tensorflow/lite/kernels/internal/reference/integer_ops/softmax.h b/src/tensorflow/lite/kernels/internal/reference/integer_ops/softmax.h deleted file mode 100644 index 3f6bf1c..0000000 --- a/src/tensorflow/lite/kernels/internal/reference/integer_ops/softmax.h +++ /dev/null @@ -1,102 +0,0 @@ -/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ -#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_SOFTMAX_H_ -#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_SOFTMAX_H_ - -#include "tensorflow/lite/kernels/internal/common.h" - -namespace tflite { -namespace reference_integer_ops { - -// Quantized softmax with int8 input and output. -inline void Softmax(const SoftmaxParams& params, - const RuntimeShape& input_shape, const int8* input_data, - const RuntimeShape& output_shape, int8* output_data) { - const int32 input_beta_multiplier = params.input_multiplier; - const int32 input_beta_left_shift = params.input_left_shift; - const int diff_min = params.diff_min; - // The representation chosen for the input to the exp() function is Q5.26. - // We need to leave extra space since values that we skip might be as large as - // -32 before multiplying by input_beta_multiplier, and therefore as large as - // -16 afterwards. Note that exp(-8) is definitely not insignificant to - // accumulation, but exp(-16) definitely is. - static const int kScaledDiffIntegerBits = 5; - static const int kAccumulationIntegerBits = 12; - using FixedPointScaledDiff = - gemmlowp::FixedPoint; - using FixedPointAccum = gemmlowp::FixedPoint; - using FixedPoint0 = gemmlowp::FixedPoint; - - const int trailing_dim = input_shape.DimensionsCount() - 1; - const int outer_size = - MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); - const int depth = - MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); - - for (int i = 0; i < outer_size; ++i) { - int8 max_in_row = -128; - for (int c = 0; c < depth; ++c) { - max_in_row = std::max(max_in_row, input_data[i * depth + c]); - } - - FixedPointAccum sum_of_exps = FixedPointAccum::Zero(); - for (int c = 0; c < depth; ++c) { - int32 input_diff = - static_cast(input_data[i * depth + c]) - max_in_row; - if (input_diff >= diff_min) { - const int32 input_diff_rescaled = - MultiplyByQuantizedMultiplierGreaterThanOne( - input_diff, input_beta_multiplier, input_beta_left_shift); - const FixedPointScaledDiff scaled_diff_f8 = - FixedPointScaledDiff::FromRaw(input_diff_rescaled); - sum_of_exps = sum_of_exps + gemmlowp::Rescale( - exp_on_negative_values(scaled_diff_f8)); - } - } - - int num_bits_over_unit; - FixedPoint0 shifted_scale = FixedPoint0::FromRaw(GetReciprocal( - sum_of_exps.raw(), kAccumulationIntegerBits, &num_bits_over_unit)); - - for (int c = 0; c < depth; ++c) { - int32 input_diff = - static_cast(input_data[i * depth + c]) - max_in_row; - if (input_diff >= diff_min) { - const int32 input_diff_rescaled = - MultiplyByQuantizedMultiplierGreaterThanOne( - input_diff, input_beta_multiplier, input_beta_left_shift); - const FixedPointScaledDiff scaled_diff_f8 = - FixedPointScaledDiff::FromRaw(input_diff_rescaled); - - FixedPoint0 exp_in_0 = exp_on_negative_values(scaled_diff_f8); - const int32 unsat_output = gemmlowp::RoundingDivideByPOT( - (shifted_scale * exp_in_0).raw(), num_bits_over_unit + 31 - 8); - const int32 shifted_output = unsat_output - 128; - - output_data[i * depth + c] = static_cast( - std::max(std::min(shifted_output, static_cast(127)), - static_cast(-128))); - - } else { - output_data[i * depth + c] = -128; - } - } - } -} - -} // namespace reference_integer_ops -} // namespace tflite - -#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_SOFTMAX_H_ diff --git a/src/tensorflow/lite/micro/kernels/concatenation.cpp b/src/tensorflow/lite/micro/kernels/concatenation.cpp deleted file mode 100644 index cd810f7..0000000 --- a/src/tensorflow/lite/micro/kernels/concatenation.cpp +++ /dev/null @@ -1,225 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/concatenation.h" -#include - -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/internal/types.h" -#include "tensorflow/lite/kernels/kernel_util.h" - -namespace tflite { -namespace ops { -namespace micro { -namespace concatenation { - -constexpr int kMaxInputNum = 10; // Maximum number of input tensors -constexpr int kOutputTensor = 0; - -TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { - // This function only checks the types. Additional shape validations are - // performed in the reference implementation called during Eval(). - const TfLiteConcatenationParams* params = - reinterpret_cast(node->builtin_data); - - TfLiteType input_type = GetInput(context, node, 0)->type; - TfLiteType output_type = GetOutput(context, node, kOutputTensor)->type; - - // Check activation and input type - TF_LITE_ENSURE_EQ(context, params->activation, kTfLiteActNone); - TF_LITE_ENSURE(context, - input_type == kTfLiteFloat32 || input_type == kTfLiteUInt8 || - input_type == kTfLiteInt8 || input_type == kTfLiteInt32 || - input_type == kTfLiteInt64); - - // Output type must match input type - TF_LITE_ENSURE_EQ(context, output_type, input_type); - - // This implementation does not support large number of input tensors - const int num_inputs = NumInputs(node); - TF_LITE_ENSURE(context, num_inputs <= kMaxInputNum); - - // Shapes with dimensions >4 are not yet supported with static allocation. - for (int i = 0; i < num_inputs; ++i) { - const TfLiteTensor* input = GetInput(context, node, i); - int num_dimensions = NumDimensions(input); - - if (num_dimensions > 4) { - context->ReportError( - context, - "Op Concatenation does not currently support num dimensions >4 " - "Tensor '%s' has %d dimensions.", - input->name, num_dimensions); - return kTfLiteError; - } - } - - return kTfLiteOk; -} - -// Handles negative axis index, coerces to positive index value. -inline int CalculatePositiveAxis(int axis, const TfLiteTensor* output_tensor) { - if (axis >= 0) { - return axis; - } else { - return NumDimensions(output_tensor) + axis; - } -} - -// The following functions are helpers to get tensor data in the format that the -// reference op implementation expects. They provide the same functionality as -// class VectorOfTensors and class VectorOfQuantizedTensors in TFLite. - -// Gets shapes from a list of tensors. -inline void GetAllTensorShapes(const TfLiteContext& context, - const TfLiteIntArray& tensor_list, - RuntimeShape all_shapes[kMaxInputNum]) { - for (int i = 0; i < tensor_list.size; ++i) { - const TfLiteTensor* t = &context.tensors[tensor_list.data[i]]; - RuntimeShape shape = GetTensorShape(t); - all_shapes[i].ReplaceWith(shape.DimensionsCount(), shape.DimsData()); - } -} - -// Get shape pointers from a list of shapes. -inline void GetShapesPointers(const RuntimeShape* shapes, size_t num, - const RuntimeShape* pointers[]) { - for (size_t i = 0; i < num; ++i) { - pointers[i] = &shapes[i]; - } -} - -// Gets data pointers from a list of tensors. -template -inline void GetAllTensorData(const TfLiteContext& context, - const TfLiteIntArray& tensor_list, - T* all_data[kMaxInputNum]) { - for (int i = 0; i < tensor_list.size; ++i) { - const TfLiteTensor* t = &context.tensors[tensor_list.data[i]]; - all_data[i] = GetTensorData(t); - } -} - -// Gets scale and zero point from a list of tensors -inline void GetAllQuantizationParam(const TfLiteContext& context, - const TfLiteIntArray& tensor_list, - float scales[kMaxInputNum], - int32 zero_points[kMaxInputNum]) { - for (int i = 0; i < tensor_list.size; ++i) { - const TfLiteTensor* t = &context.tensors[tensor_list.data[i]]; - scales[i] = t->params.scale; - zero_points[i] = t->params.zero_point; - } -} - -template -void EvalUnquantized(TfLiteContext* context, TfLiteNode* node) { - // Collect the shapes and data pointer of input tensors - RuntimeShape inputs_shape[kMaxInputNum]; - const RuntimeShape* inputs_shape_ptr[kMaxInputNum]; - const data_type* inputs_data[kMaxInputNum]; - GetAllTensorShapes(*context, *node->inputs, inputs_shape); - GetShapesPointers(inputs_shape, node->inputs->size, inputs_shape_ptr); - GetAllTensorData(*context, *node->inputs, inputs_data); - - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - - const TfLiteConcatenationParams* params = - reinterpret_cast(node->builtin_data); - - ConcatenationParams op_params; - op_params.axis = CalculatePositiveAxis(params->axis, output); - op_params.inputs_count = NumInputs(node); - - reference_ops::Concatenation(op_params, inputs_shape_ptr, inputs_data, - GetTensorShape(output), - GetTensorData(output)); -} - -void EvalQuantizedUInt8(TfLiteContext* context, TfLiteNode* node) { - // Collect the shapes and data pointer of input tensors - RuntimeShape inputs_shape[kMaxInputNum]; - const RuntimeShape* inputs_shape_ptr[kMaxInputNum]; - const uint8_t* inputs_data[kMaxInputNum]; - float inputs_scale[kMaxInputNum]; - int32 inputs_zero_point[kMaxInputNum]; - GetAllTensorShapes(*context, *node->inputs, inputs_shape); - GetShapesPointers(inputs_shape, node->inputs->size, inputs_shape_ptr); - GetAllTensorData(*context, *node->inputs, inputs_data); - GetAllQuantizationParam(*context, *node->inputs, inputs_scale, - inputs_zero_point); - - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - - const TfLiteConcatenationParams* params = - reinterpret_cast(node->builtin_data); - - ConcatenationParams op_params; - op_params.axis = CalculatePositiveAxis(params->axis, output); - op_params.inputs_count = NumInputs(node); - op_params.input_zeropoint = inputs_zero_point; - op_params.input_scale = inputs_scale; - op_params.output_zeropoint = output->params.zero_point; - op_params.output_scale = output->params.scale; - - reference_ops::ConcatenationWithScaling( - op_params, inputs_shape_ptr, inputs_data, GetTensorShape(output), - GetTensorData(output)); -} - -TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { - TfLiteType output_type = GetOutput(context, node, kOutputTensor)->type; - - switch (output_type) { // Already know in/outtypes are same. - case kTfLiteFloat32: - EvalUnquantized(context, node); - break; - case kTfLiteInt32: - EvalUnquantized(context, node); - break; - case kTfLiteUInt8: - EvalQuantizedUInt8(context, node); - break; - case kTfLiteInt8: - EvalUnquantized(context, node); - break; - case kTfLiteInt64: - EvalUnquantized(context, node); - break; - - default: - context->ReportError( - context, "Op Concatenation does not currently support Type '%s'.", - TfLiteTypeGetName(output_type)); - return kTfLiteError; - } - - return kTfLiteOk; -} - -} // namespace concatenation - -TfLiteRegistration* Register_CONCATENATION() { - static TfLiteRegistration r = {}; - r.prepare = concatenation::Prepare; - r.invoke = concatenation::Eval; - return &r; -} - -} // namespace micro -} // namespace ops -} // namespace tflite diff --git a/src/tensorflow/lite/micro/kernels/mul.cpp b/src/tensorflow/lite/micro/kernels/mul.cpp deleted file mode 100644 index 2322b59..0000000 --- a/src/tensorflow/lite/micro/kernels/mul.cpp +++ /dev/null @@ -1,174 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/kernels/internal/reference/mul.h" - -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/reference/integer_ops/mul.h" -#include "tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" - -namespace tflite { -namespace ops { -namespace micro { -namespace mul { - -constexpr int kInput1Tensor = 0; -constexpr int kInput2Tensor = 1; -constexpr int kOutputTensor = 0; - -struct OpData { - int32_t output_activation_min; - int32_t output_activation_max; - - int32_t output_multiplier; - int output_shift; -}; - -TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node, - TfLiteMulParams* params, OpData* data) { - const TfLiteTensor* input1 = GetInput(context, node, kInput1Tensor); - const TfLiteTensor* input2 = GetInput(context, node, kInput2Tensor); - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - - TF_LITE_ENSURE_EQ(context, NumInputs(node), 2); - TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); - - TF_LITE_ENSURE_EQ(context, input1->type, input2->type); - - TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( - context, params->activation, output, &data->output_activation_min, - &data->output_activation_max)); - - if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { - double real_multiplier = static_cast(input1->params.scale) * - static_cast(input2->params.scale) / - static_cast(output->params.scale); - QuantizeMultiplier(real_multiplier, &data->output_multiplier, - &data->output_shift); - } - - return kTfLiteOk; -} - -TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { - return kTfLiteOk; -} - -void EvalQuantized(TfLiteContext* context, TfLiteNode* node, - TfLiteMulParams* params, OpData* data, - const TfLiteTensor* input1, const TfLiteTensor* input2, - TfLiteTensor* output) { - if (output->type == kTfLiteInt8 || output->type == kTfLiteUInt8) { - tflite::ArithmeticParams op_params; - SetActivationParams(data->output_activation_min, - data->output_activation_max, &op_params); - op_params.input1_offset = -input1->params.zero_point; - op_params.input2_offset = -input2->params.zero_point; - op_params.output_offset = output->params.zero_point; - op_params.output_multiplier = data->output_multiplier; - op_params.output_shift = data->output_shift; - bool need_broadcast = reference_ops::ProcessBroadcastShapes( - GetTensorShape(input1), GetTensorShape(input2), &op_params); - -#define TF_LITE_MUL(type, opname, dtype) \ - type::opname(op_params, GetTensorShape(input1), \ - GetTensorData(input1), GetTensorShape(input2), \ - GetTensorData(input2), GetTensorShape(output), \ - GetTensorData(output)); - - if (output->type == kTfLiteInt8) { - if (need_broadcast) { - TF_LITE_MUL(reference_integer_ops, BroadcastMul4DSlow, int8_t); - } else { - TF_LITE_MUL(reference_integer_ops, Mul, int8_t); - } - } else if (output->type == kTfLiteUInt8) { - if (need_broadcast) { - TF_LITE_MUL(reference_ops, BroadcastMul4DSlow, uint8_t); - } else { - TF_LITE_MUL(reference_ops, Mul, uint8_t); - } - } -#undef TF_LITE_MUL - } -} - -void EvalFloat(TfLiteContext* context, TfLiteNode* node, - TfLiteMulParams* params, OpData* data, - const TfLiteTensor* input1, const TfLiteTensor* input2, - TfLiteTensor* output) { - float output_activation_min, output_activation_max; - CalculateActivationRange(params->activation, &output_activation_min, - &output_activation_max); - tflite::ArithmeticParams op_params; - SetActivationParams(output_activation_min, output_activation_max, &op_params); - - bool need_broadcast = reference_ops::ProcessBroadcastShapes( - GetTensorShape(input1), GetTensorShape(input2), &op_params); -#define TF_LITE_MUL(opname) \ - reference_ops::opname(op_params, GetTensorShape(input1), \ - GetTensorData(input1), GetTensorShape(input2), \ - GetTensorData(input2), GetTensorShape(output), \ - GetTensorData(output)); - - if (need_broadcast) { - TF_LITE_MUL(BroadcastMul4DSlow); - } else { - TF_LITE_MUL(Mul); - } -#undef TF_LITE_MUL -} - -TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { - auto* params = reinterpret_cast(node->builtin_data); - OpData data; - - const TfLiteTensor* input1 = GetInput(context, node, kInput1Tensor); - const TfLiteTensor* input2 = GetInput(context, node, kInput2Tensor); - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - - CalculateOpData(context, node, params, &data); - - switch (input1->type) { - case kTfLiteUInt8: - case kTfLiteInt8: - EvalQuantized(context, node, params, &data, input1, input2, output); - break; - case kTfLiteFloat32: - EvalFloat(context, node, params, &data, input1, input2, output); - break; - default: - context->ReportError(context, "Type %d not currently supported.", - input1->type); - return kTfLiteError; - } - - return kTfLiteOk; -} -} // namespace mul - -TfLiteRegistration* Register_MUL() { - static TfLiteRegistration r = {}; - r.prepare = mul::Prepare; - r.invoke = mul::Eval; - return &r; -} - -} // namespace micro -} // namespace ops -} // namespace tflite diff --git a/src/tensorflow/lite/micro/kernels/pad.cpp b/src/tensorflow/lite/micro/kernels/pad.cpp deleted file mode 100644 index 5ed2bac..0000000 --- a/src/tensorflow/lite/micro/kernels/pad.cpp +++ /dev/null @@ -1,227 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/pad.h" - -#include - -#include "tensorflow/lite/kernels/internal/types.h" - -#ifdef MEMORY_SANITIZER -#include -#else -#define __msan_check_mem_is_initialized(ptr, size) -#endif - -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" - -namespace tflite { -namespace ops { -namespace micro { -namespace pad { - -struct PadContext { - PadContext(TfLiteContext* context, TfLiteNode* node) { - input = GetInput(context, node, 0); - paddings = GetInput(context, node, 1); - constant_values = nullptr; - if (NumInputs(node) == 3) { - constant_values = GetOptionalInputTensor(context, node, 2); - } else { - constant_values = nullptr; - } - output = GetOutput(context, node, 0); - dims = NumDimensions(input); - - resizing_category = ResizingCategory::kGenericResize; - const int paddings_total = GetTensorShape(paddings).FlatSize(); - const int32* paddings_data = GetTensorData(paddings); - // Paddings will be a n,2 array, and we need to detect 4D arrays with the - // pattern { {0,0}, {a, b}, {c, d}, {0,0} }. - if (IsConstantTensor(paddings) && paddings_total == 8 && - (paddings_data[0] == 0 && paddings_data[1] == 0) && - (paddings_data[6] == 0 && paddings_data[7] == 0)) { - resizing_category = ResizingCategory::kImageStyle; - } - } - const TfLiteTensor* constant_values; - const TfLiteTensor* input; - const TfLiteTensor* paddings; - TfLiteTensor* output; - int dims; - ResizingCategory resizing_category; -}; - -TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { - TF_LITE_ENSURE(context, NumInputs(node) == 2 || NumInputs(node) == 3); - TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); - - PadContext op_context(context, node); - TF_LITE_ENSURE_EQ(context, op_context.input->type, op_context.output->type); - if (op_context.constant_values != nullptr) { - TF_LITE_ENSURE_EQ(context, op_context.input->type, - op_context.constant_values->type); - } - - // There must be a pair of paddings for each output dimension. - TF_LITE_ENSURE_EQ(context, GetTensorShape(op_context.paddings).FlatSize(), - op_context.output->dims->size * 2); - - // On Micro, outputs must be properly sized by the converter. - const int32* paddings_data = GetTensorData(op_context.paddings); - for (int i = 0; i < op_context.output->dims->size; i++) { - int output_dim = op_context.output->dims->data[i]; - int expected_dim = op_context.input->dims->data[i] + paddings_data[i * 2] + - paddings_data[i * 2 + 1]; - TF_LITE_ENSURE_EQ(context, output_dim, expected_dim); - } - - // Current implementations rely on the inputs being <= 4D. - TF_LITE_ENSURE( - context, op_context.dims <= reference_ops::PadKernelMaxDimensionCount()); - TF_LITE_ENSURE(context, IsConstantTensor(op_context.paddings)); - return kTfLiteOk; -} - -TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { - PadContext op_context(context, node); - - if (op_context.constant_values != nullptr) { - // Ensure that constant_values is a scalar. - TF_LITE_ENSURE_EQ(context, NumElements(op_context.constant_values), 1); - } - - // Create before and after padding arrays that are accepted by the kernel. - const int32* paddings_data = GetTensorData(op_context.paddings); - - tflite::PadParams op_params; - memset(&op_params, 0, sizeof(PadParams)); - op_params.left_padding_count = op_context.dims; - op_params.right_padding_count = op_context.dims; - - for (int idx = op_context.dims - 1; idx >= 0; --idx) { - op_params.left_padding[idx] = paddings_data[idx * 2]; - op_params.right_padding[idx] = paddings_data[idx * 2 + 1]; - } - -#define TF_LITE_PAD(type, op_name, scalar, pad_value) \ - const scalar pad_value_copy = pad_value; \ - \ - type::op_name(op_params, GetTensorShape(op_context.input), \ - GetTensorData(op_context.input), &pad_value_copy, \ - GetTensorShape(op_context.output), \ - GetTensorData(op_context.output)) - switch (op_context.input->type) { - case kTfLiteFloat32: { - float pad_value = op_context.constant_values == nullptr - ? 0.f - : *GetTensorData(op_context.constant_values); - if (op_context.resizing_category == ResizingCategory::kImageStyle) { - TF_LITE_PAD(reference_ops, PadImageStyle, float, pad_value); - } else { - TF_LITE_PAD(reference_ops, Pad, float, pad_value); - } - } break; - case kTfLiteUInt8: { - uint8_t pad_value; - if (op_context.constant_values == nullptr) { - // Quantized Pad requires that 0 is represented in the quantized - // range. - TF_LITE_ENSURE(context, op_context.output->params.zero_point >= - std::numeric_limits::min()); - TF_LITE_ENSURE(context, op_context.output->params.zero_point <= - std::numeric_limits::max()); - pad_value = static_cast(op_context.output->params.zero_point); - } else { - // Quantized Pad requires that 'constant_values' is represented in the - // same quantized range as the input and output tensors. - TF_LITE_ENSURE_EQ(context, op_context.output->params.zero_point, - op_context.constant_values->params.zero_point); - TF_LITE_ENSURE_EQ( - context, static_cast(op_context.output->params.scale), - static_cast(op_context.constant_values->params.scale)); - pad_value = *GetTensorData(op_context.constant_values); - } - if (op_context.resizing_category == ResizingCategory::kImageStyle) { - TF_LITE_PAD(reference_ops, PadImageStyle, uint8_t, pad_value); - } else { - TF_LITE_PAD(reference_ops, Pad, uint8_t, pad_value); - } - } break; - case kTfLiteInt8: { - int8_t pad_value; - if (op_context.constant_values == nullptr) { - // Quantized Pad requires that 0 is represented in the quantized - // range. - TF_LITE_ENSURE(context, op_context.output->params.zero_point >= - std::numeric_limits::min()); - TF_LITE_ENSURE(context, op_context.output->params.zero_point <= - std::numeric_limits::max()); - pad_value = static_cast(op_context.output->params.zero_point); - } else { - // Quantized Pad requires that 'constant_values' is represented in the - // same quantized range as the input and output tensors. - TF_LITE_ENSURE_EQ(context, op_context.output->params.zero_point, - op_context.constant_values->params.zero_point); - TF_LITE_ENSURE(context, op_context.output->params.scale == - op_context.constant_values->params.scale); - pad_value = *GetTensorData(op_context.constant_values); - } - if (op_context.resizing_category == ResizingCategory::kImageStyle) { - TF_LITE_PAD(reference_ops, PadImageStyle, int8_t, pad_value); - } else { - TF_LITE_PAD(reference_ops, Pad, int8_t, pad_value); - } - } break; - case kTfLiteInt32: { - int32_t pad_value = - op_context.constant_values == nullptr - ? 0 - : *GetTensorData(op_context.constant_values); - TF_LITE_PAD(reference_ops, Pad, int32_t, pad_value); - } break; - default: - - context->ReportError(context, "Type %s not currently supported by Pad.", - TfLiteTypeGetName(op_context.input->type)); - return kTfLiteError; - } -#undef TF_LITE_PAD - return kTfLiteOk; -} - -} // namespace pad - -TfLiteRegistration* Register_PAD() { - static TfLiteRegistration r = {}; - r.prepare = pad::Prepare; - r.invoke = pad::Eval; - return &r; -} - -// Also register Pad as PadV2. -TfLiteRegistration* Register_PADV2() { - static TfLiteRegistration r = {}; - r.prepare = pad::Prepare; - r.invoke = pad::Eval; - return &r; -} - -} // namespace micro -} // namespace ops -} // namespace tflite diff --git a/src/tensorflow/lite/micro/micro_allocator.cpp b/src/tensorflow/lite/micro/micro_allocator.cpp deleted file mode 100644 index 895739e..0000000 --- a/src/tensorflow/lite/micro/micro_allocator.cpp +++ /dev/null @@ -1,531 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/lite/micro/micro_allocator.h" - -#include - -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/core/api/flatbuffer_conversions.h" -#include "tensorflow/lite/core/api/op_resolver.h" -#include "tensorflow/lite/core/api/tensor_utils.h" -#include "tensorflow/lite/micro/compatibility.h" -#include "tensorflow/lite/micro/memory_helpers.h" -#include "tensorflow/lite/micro/memory_planner/greedy_memory_planner.h" -#include "tensorflow/lite/micro/simple_memory_allocator.h" - -namespace tflite { - -namespace { -// Used to hold information used during allocation calculations. -struct AllocationInfo { - size_t bytes; - int first_created; - int last_used; - bool needs_allocating; - void** output_ptr; -}; - -// We align tensor buffers to 16-byte boundaries, since this is a common -// requirement for SIMD extensions. -constexpr int kBufferAlignment = 16; - -// If building with GNU clib from GCC 4.8.x or lower, `max_align_t` is not a -// member of `std`. If using a newer version of clib, we import `max_align_t` -// into the local anonymous namespace to be able to use it like the global -// `max_align_t` from the older clib. -#if defined(__GNUC__) && defined(__GNUC_PREREQ) -#if __GNUC_PREREQ(4, 9) -using std::max_align_t; -#endif -#else -// We assume other compiler/clib configurations don't have this issue. -using std::max_align_t; -#endif - -class MicroBuiltinDataAllocator : public BuiltinDataAllocator { - public: - explicit MicroBuiltinDataAllocator(SimpleMemoryAllocator* memory_allocator) - : memory_allocator_(memory_allocator) {} - - void* Allocate(size_t size) override { - // Align to an address that is proper for all primitive types, but no more - // than the size. - return memory_allocator_->AllocateFromTail( - size, std::min(size, alignof(max_align_t))); - } - void Deallocate(void* data) override { - // Do not deallocate, builtin data needs to be available for the life time - // of the model. - } - - private: - SimpleMemoryAllocator* memory_allocator_; - - TF_LITE_REMOVE_VIRTUAL_DELETE -}; - -TfLiteStatus AllocateVariables( - const flatbuffers::Vector>* flatbuffer_tensors, - TfLiteTensor* runtime_tensors, SimpleMemoryAllocator* allocator) { - for (size_t i = 0; i < flatbuffer_tensors->size(); ++i) { - if (flatbuffer_tensors->Get(i)->is_variable()) { - runtime_tensors[i].data.uint8 = allocator->AllocateFromTail( - runtime_tensors[i].bytes, kBufferAlignment); - // Allocation failure. - if (runtime_tensors[i].data.uint8 == nullptr) { - return kTfLiteError; - } - } - tflite::ResetVariableTensor(&(runtime_tensors[i])); - } - return kTfLiteOk; -} - -AllocationInfo* AllocateAndCalculateAllocationInfo( - ErrorReporter* error_reporter, size_t allocation_info_size, - const SubGraph* subgraph, TfLiteTensor* runtime_tensors, - SimpleMemoryAllocator* allocator) { - AllocationInfo* allocation_info = reinterpret_cast( - allocator->AllocateFromTail(sizeof(AllocationInfo) * allocation_info_size, - alignof(AllocationInfo))); - if (allocation_info == nullptr) { - error_reporter->Report( - "Failed to allocate memory for allocation_info, %d bytes required", - sizeof(TfLiteTensor) * allocation_info_size); - return nullptr; - } - - // Set up the runtime data structures for all tensors. - for (size_t i = 0; i < allocation_info_size; ++i) { - AllocationInfo* current = &allocation_info[i]; - // TfLiteTensor.uint8 field is deprecated so use .data field instead. - current->output_ptr = &(runtime_tensors[i].data.data); - current->bytes = runtime_tensors[i].bytes; - current->first_created = -1; - current->last_used = -1; - current->needs_allocating = (runtime_tensors[i].data.raw == nullptr) && - (!subgraph->tensors()->Get(i)->is_variable()); - } - - for (size_t i = 0; i < subgraph->inputs()->size(); ++i) { - const int tensor_index = subgraph->inputs()->Get(i); - AllocationInfo* current = &allocation_info[tensor_index]; - current->first_created = 0; - } - - // Mark all outputs as persistent to the end of the invocation. - for (size_t i = 0; i < subgraph->outputs()->size(); ++i) { - const int tensor_index = subgraph->outputs()->Get(i); - AllocationInfo* current = &allocation_info[tensor_index]; - current->last_used = subgraph->operators()->size() - 1; - } - - // Figure out when the first and last use of each tensor is. - for (int i = (subgraph->operators()->size() - 1); i >= 0; --i) { - const auto* op = subgraph->operators()->Get(i); - for (size_t n = 0; n < op->inputs()->size(); ++n) { - const int tensor_index = op->inputs()->Get(n); - AllocationInfo* current = &allocation_info[tensor_index]; - if (((current->last_used == -1) || (current->last_used > i))) { - current->last_used = i; - } - } - for (size_t n = 0; n < op->outputs()->size(); ++n) { - const int tensor_index = op->outputs()->Get(n); - AllocationInfo* current = &allocation_info[tensor_index]; - if ((current->first_created == -1) || (current->first_created < i)) { - current->first_created = i; - } - } - } - - // Work out which tensors need to be allocated. - for (size_t i = 0; i < allocation_info_size; ++i) { - AllocationInfo* current = &allocation_info[i]; - const bool is_read_only = - (current->first_created == -1) && (current->last_used != -1); - if (is_read_only) { - current->needs_allocating = false; - } - const bool has_partial_lifetime = - !is_read_only && - ((current->first_created == -1) || (current->last_used == -1)); - if (has_partial_lifetime && current->needs_allocating) { - error_reporter->Report( - "Logic error in memory planner, tensor %d has an invalid lifetime: " - "first_created: %d, last_used: %d", - i, current->first_created, current->last_used); - return nullptr; - } - } // namespace - - return allocation_info; -} // namespace tflite - -TfLiteStatus CreatePlan(ErrorReporter* error_reporter, MemoryPlanner* planner, - const AllocationInfo* allocation_info, - size_t allocation_info_size) { - // Add the tensors to our allocation plan. - for (size_t i = 0; i < allocation_info_size; ++i) { - const AllocationInfo* current = &allocation_info[i]; - if (current->needs_allocating) { - size_t aligned_bytes_required = - AlignSizeUp(current->bytes, kBufferAlignment); - TF_LITE_ENSURE_STATUS( - planner->AddBuffer(error_reporter, aligned_bytes_required, - current->first_created, current->last_used)); - } - } - return kTfLiteOk; -} - -TfLiteStatus CommitPlan(ErrorReporter* error_reporter, MemoryPlanner* planner, - uint8_t* starting_point, - AllocationInfo* allocation_info, - size_t allocation_info_size) { - // Figure out the actual memory addresses for each buffer, based on the plan. - int planner_index = 0; - for (size_t i = 0; i < allocation_info_size; ++i) { - AllocationInfo* current = &allocation_info[i]; - if (current->needs_allocating) { - int offset = -1; - TF_LITE_ENSURE_STATUS( - planner->GetOffsetForBuffer(error_reporter, planner_index, &offset)); - *current->output_ptr = reinterpret_cast(starting_point + offset); - ++planner_index; - } - } - return kTfLiteOk; -} -} // namespace - -namespace internal { - -TfLiteStatus InitializeRuntimeTensor( - SimpleMemoryAllocator* allocator, const tflite::Tensor& flatbuffer_tensor, - const flatbuffers::Vector>* buffers, - ErrorReporter* error_reporter, TfLiteTensor* result) { - *result = {}; - // Make sure the serialized type is one we know how to deal with, and convert - // it from a flatbuffer enum into a constant used by the kernel C API. - TF_LITE_ENSURE_STATUS(ConvertTensorType(flatbuffer_tensor.type(), - &result->type, error_reporter)); - // Make sure we remember if the serialized tensor is designated as a variable. - result->is_variable = flatbuffer_tensor.is_variable(); - - // We need to figure out where the actual contents of this tensor are stored - // in memory. We'll check to see if there's a serialized buffer (pretty much - // the same as a constant op in TensorFlow) associated with this tensor first, - // and if there is update the runtime structure to point to its location in - // memory. - // First see if there's any buffer information in the serialized tensor. - if (auto* buffer = (*buffers)[flatbuffer_tensor.buffer()]) { - // If we've found a buffer, does it have any data? - if (auto* array = buffer->data()) { - // If it has any data, is the data size larger than zero? - if (array->size()) { - // We've found a buffer with valid data, so update the runtime tensor - // data structure to point to it. - result->data.raw = - const_cast(reinterpret_cast(array->data())); - // We set the data from a serialized buffer, so record tha. - result->allocation_type = kTfLiteMmapRo; - } - } - // TODO(petewarden): It's not clear in what circumstances we could have a - // buffer in the serialized tensor, but it doesn't have any data in it. Is - // that a validly-generated file, and if so what does it mean, or is it an - // error condition? It would be good to tighten up the specification to make - // it less ambiguous. - } - - // TODO(petewarden): Some of these paths aren't getting enough testing - // coverage, so we should figure out some tests that exercise them. - if (!result->data.raw) { - // The tensor contents haven't been set from a serialized buffer, so - // make a note that they will be allocated from memory. The actual - // allocation won't happen until later. - result->allocation_type = kTfLiteArenaRw; - } - - // Figure out what the size in bytes of the buffer is and store it. - size_t type_size; - TF_LITE_ENSURE_STATUS(BytesRequiredForTensor( - flatbuffer_tensor, &result->bytes, &type_size, error_reporter)); - - // TFLM doesn't allow reshaping the tensor which requires dynamic memory - // allocation so it is safe to drop the const qualifier. In the future, if we - // really want to update the tensor shape, we can always pass in a new - // TfLiteIntArray - especially we have to do so if the dimension is changed. - result->dims = const_cast( - reinterpret_cast(flatbuffer_tensor.shape())); - - // Copy the quantization information from the serialized data. - const auto* src_quantization = flatbuffer_tensor.quantization(); - if (src_quantization && src_quantization->scale() && - (src_quantization->scale()->size() > 0) && - src_quantization->zero_point() && - (src_quantization->zero_point()->size() > 0)) { - // Always populate the TfLiteTensor.params field, even if there are - // per-channel quantization parameters. - result->params.scale = src_quantization->scale()->Get(0); - // Note that the zero_point field in the FlatBuffers schema is a 64-bit - // integer, but the zero_point field in the TfLiteQuantizationParams struct - // is a 32-bit integer. - result->params.zero_point = - static_cast(src_quantization->zero_point()->Get(0)); - - // Populate per-channel quantization params. - int channels = src_quantization->scale()->size(); - TfLiteAffineQuantization* quantization = - reinterpret_cast( - allocator->AllocateFromTail(sizeof(TfLiteAffineQuantization), - alignof(TfLiteAffineQuantization))); - quantization->zero_point = - reinterpret_cast(allocator->AllocateFromTail( - TfLiteIntArrayGetSizeInBytes(channels), alignof(TfLiteIntArray))); - quantization->scale = reinterpret_cast( - allocator->AllocateFromTail(TfLiteFloatArrayGetSizeInBytes(channels), - alignof(TfLiteFloatArray))); - quantization->zero_point->size = channels; - quantization->scale->size = channels; - int* zero_point_data = quantization->zero_point->data; - float* scale_data = quantization->scale->data; - for (int i = 0; i < channels; i++) { - zero_point_data[i] = src_quantization->zero_point()->Get(i); - scale_data[i] = src_quantization->scale()->Get(i); - } - // TODO(rocky): Need to add a micro_allocator test case that fails when - // this is not copied: - quantization->quantized_dimension = src_quantization->quantized_dimension(); - - result->quantization = {kTfLiteAffineQuantization, quantization}; - } - // Copy the name, if there is one. - if (flatbuffer_tensor.name()->c_str() != nullptr) { - result->name = flatbuffer_tensor.name()->c_str(); - } else { - result->name = ""; - } - return kTfLiteOk; -} -} // namespace internal - -TfLiteStatus MicroAllocator::Init() { - auto* subgraphs = model_->subgraphs(); - if (subgraphs->size() != 1) { - error_reporter_->Report("Only 1 subgraph is currently supported.\n"); - return kTfLiteError; - } - subgraph_ = (*subgraphs)[0]; - tensors_ = subgraph_->tensors(); - operators_ = subgraph_->operators(); - - context_->tensors_size = tensors_->size(); - context_->tensors = - reinterpret_cast(memory_allocator_->AllocateFromTail( - sizeof(TfLiteTensor) * context_->tensors_size, - alignof(TfLiteTensor))); - if (context_->tensors == nullptr) { - error_reporter_->Report( - "Failed to allocate memory for context->tensors, %d bytes required", - sizeof(TfLiteTensor) * context_->tensors_size); - } - - // Initialize runtime tensors in context_ using the flatbuffer. - for (size_t i = 0; i < tensors_->size(); ++i) { - TfLiteStatus status = internal::InitializeRuntimeTensor( - memory_allocator_, *tensors_->Get(i), model_->buffers(), - error_reporter_, &context_->tensors[i]); - if (status == kTfLiteError) { - error_reporter_->Report("Failed to initialize tensor %d", i); - return kTfLiteError; - } - } - - return kTfLiteOk; -} - -MicroAllocator::MicroAllocator(TfLiteContext* context, const Model* model, - uint8_t* tensor_arena, size_t arena_size, - ErrorReporter* error_reporter) - : model_(model), error_reporter_(error_reporter), context_(context) { - uint8_t* aligned_arena = AlignPointerUp(tensor_arena, kBufferAlignment); - size_t aligned_arena_size = tensor_arena + arena_size - aligned_arena; - // Creates a root memory allocator managing the arena. The allocator itself - // also locates in the arena buffer. This allocator doesn't need to be - // destructed as it's the root allocator. - SimpleMemoryAllocator* aligned_allocator = - CreateInPlaceSimpleMemoryAllocator(aligned_arena, aligned_arena_size); - memory_allocator_ = aligned_allocator; - TfLiteStatus status = Init(); - // TODO(b/147871299): Consider improving this code. A better way of handling - // failures in the constructor is to have a static function that returns a - // pointer to the class. If allocation failed, a nullptr will be returned. - if (status != kTfLiteOk) { - error_reporter_->Report("MicroAllocator: Failed to initialize."); - active_ = false; - } else { - active_ = true; - } -} - -TfLiteStatus MicroAllocator::AllocateNodeAndRegistrations( - const OpResolver& op_resolver, - NodeAndRegistration** node_and_registrations) { - if (!active_) { - return kTfLiteError; - } - - auto* output = reinterpret_cast( - memory_allocator_->AllocateFromTail( - sizeof(NodeAndRegistration) * operators_->size(), - alignof(NodeAndRegistration))); - if (output == nullptr) { - error_reporter_->Report( - "Failed to allocate memory for node_and_registrations."); - return kTfLiteError; - } - TfLiteStatus status = kTfLiteOk; - auto* opcodes = model_->operator_codes(); - MicroBuiltinDataAllocator builtin_data_allocator(memory_allocator_); - for (size_t i = 0; i < operators_->size(); ++i) { - const auto* op = operators_->Get(i); - size_t index = op->opcode_index(); - if (index >= opcodes->size()) { - error_reporter_->Report("Missing registration for opcode_index %d\n", - index); - return kTfLiteError; - } - auto* opcode = (*opcodes)[index]; - status = GetRegistrationFromOpCode(opcode, op_resolver, error_reporter_, - &(output[i].registration)); - if (status != kTfLiteOk) { - error_reporter_->Report("Failed to get registration from op code %d\n", - opcode); - return status; - } - const auto* registration = output[i].registration; - if (registration == nullptr) { - error_reporter_->Report("Skipping op for opcode_index %d\n", index); - return kTfLiteError; - } - BuiltinOperator op_type = - static_cast(registration->builtin_code); - - if (op_type != BuiltinOperator_CUSTOM && op->custom_options()) { - error_reporter_->Report( - "Unsupported behavior: found builtin operator %s with custom " - "options.\n", - EnumNameBuiltinOperator(op_type)); - return kTfLiteError; - } - - const char* custom_data = nullptr; - size_t custom_data_size = 0; - unsigned char* builtin_data = nullptr; - if (op->custom_options()) { - custom_data = reinterpret_cast(op->custom_options()->data()); - custom_data_size = op->custom_options()->size(); - } else { - TF_LITE_ENSURE_STATUS(ParseOpData(op, op_type, error_reporter_, - &builtin_data_allocator, - (void**)(&builtin_data))); - } - - // Disregard const qualifier to workaround with existing API. - TfLiteIntArray* inputs_array = const_cast( - reinterpret_cast(op->inputs())); - TfLiteIntArray* outputs_array = const_cast( - reinterpret_cast(op->outputs())); - - TfLiteNode* node = &(output[i].node); - *node = {}; - node->inputs = inputs_array; - node->outputs = outputs_array; - // This is OK for now as temporary array is not in used. - node->temporaries = nullptr; - node->user_data = nullptr; // Will be filled in after `init` - node->builtin_data = reinterpret_cast(builtin_data); - node->custom_initial_data = custom_data; - node->custom_initial_data_size = custom_data_size; - node->delegate = nullptr; - } - *node_and_registrations = output; - return kTfLiteOk; -} - -TfLiteStatus MicroAllocator::FinishTensorAllocation() { - if (!active_) { - return kTfLiteError; - } - - // Create static memory plan. AllocationInfo is needed for creating the plan - // but is thrown away afterwards. - { - SimpleMemoryAllocator tmp_allocator = - memory_allocator_->CreateChildAllocator(); - size_t allocation_info_size = tensors_->size(); - AllocationInfo* allocation_info = AllocateAndCalculateAllocationInfo( - error_reporter_, allocation_info_size, subgraph_, context_->tensors, - &tmp_allocator); - if (allocation_info == nullptr) { - return kTfLiteError; - } - - uint8_t* aligned_arena = memory_allocator_->GetBuffer(); - size_t arena_size = memory_allocator_->GetMaxBufferSize(); - - // Remaining arena size that memory planner can use for calculating offsets. - // The remaining size should always be a positive number since the parent - // allocator is always bigger than the child allocator. - size_t remaining_arena_size = arena_size - tmp_allocator.GetDataSize(); - GreedyMemoryPlanner planner(aligned_arena, remaining_arena_size); - TF_LITE_ENSURE_STATUS(CreatePlan(error_reporter_, &planner, allocation_info, - allocation_info_size)); - - // Actual size available for placing tensors. This includes memory held by - // the tensor info array, which will be released. - size_t actual_available_arena_size = - arena_size - memory_allocator_->GetDataSize(); - // Make sure we have enough arena size. - if (planner.GetMaximumMemorySize() > actual_available_arena_size) { - error_reporter_->Report( - "Arena size is too small for activation buffers. Needed %d but only " - "%d was available.", - planner.GetMaximumMemorySize(), remaining_arena_size); - return kTfLiteError; - } - - TF_LITE_ENSURE_STATUS(CommitPlan(error_reporter_, &planner, aligned_arena, - allocation_info, allocation_info_size)); - } - - // Data in variables need to be kept for the next invocation so allocating - // them from the tail (persistent area). - if (AllocateVariables(tensors_, context_->tensors, memory_allocator_) != - kTfLiteOk) { - error_reporter_->Report( - "Failed to allocate variables. Please increase arena size."); - return kTfLiteError; - } - - active_ = false; - return kTfLiteOk; -} - -} // namespace tflite diff --git a/src/tensorflow/lite/micro/micro_allocator.h b/src/tensorflow/lite/micro/micro_allocator.h deleted file mode 100644 index 428f4b2..0000000 --- a/src/tensorflow/lite/micro/micro_allocator.h +++ /dev/null @@ -1,85 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_MICRO_ALLOCATOR_H_ -#define TENSORFLOW_LITE_MICRO_MICRO_ALLOCATOR_H_ - -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/core/api/flatbuffer_conversions.h" -#include "tensorflow/lite/micro/simple_memory_allocator.h" -#include "tensorflow/lite/schema/schema_generated.h" - -namespace tflite { - -// Namespace used for unittests. -namespace internal { -// Sets up all of the data structure members for a runtime tensor -// based on the contents of a serialized tensor. -TfLiteStatus InitializeRuntimeTensor( - SimpleMemoryAllocator* allocator, const tflite::Tensor& flatbuffer_tensor, - const flatbuffers::Vector>* buffers, - ErrorReporter* error_reporter, TfLiteTensor* result); -} // namespace internal - -typedef struct { - TfLiteNode node; - const TfLiteRegistration* registration; -} NodeAndRegistration; - -// Allocator responsible for allocating memory for all intermediate tensors -// necessary to invoke a model. -class MicroAllocator { - public: - // The lifetime of the model, tensor allocator and error reporter must be at - // least as long as that of the allocator object, since the allocator needs - // them to be accessible during its entire lifetime. - - // Note: Please use __declspec(align(16)) to make sure tensor_arena is 16 - // bytes aligned, otherwise some head room will be wasted. - MicroAllocator(TfLiteContext* context, const Model* model, - uint8_t* tensor_arena, size_t arena_size, - ErrorReporter* error_reporter); - - // Runs through the model and allocates all necessary input, output and - // intermediate tensors. - // WARNING: doing any allocation after calling this method has the risk of - // corrupting tensor data so this method should be the last method to be - // called in this class. - TfLiteStatus FinishTensorAllocation(); - - // Run through the model to allocate nodes and registrations. We need to keep - // them for the entire life time of the model to allow persistent tensors. - // This method needs to be called before FinishTensorAllocation method. - TfLiteStatus AllocateNodeAndRegistrations( - const OpResolver& op_resolver, - NodeAndRegistration** node_and_registrations); - - private: - TfLiteStatus Init(); - - const Model* model_; - SimpleMemoryAllocator* memory_allocator_; - ErrorReporter* error_reporter_; - TfLiteContext* context_; - // Indicating if the allocator is ready for allocation. - bool active_ = false; - - const SubGraph* subgraph_; - const flatbuffers::Vector>* operators_; - const flatbuffers::Vector>* tensors_; -}; - -} // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_MICRO_ALLOCATOR_H_ diff --git a/src/tensorflow/lite/micro/micro_mutable_op_resolver.h b/src/tensorflow/lite/micro/micro_mutable_op_resolver.h deleted file mode 100644 index 21066cf..0000000 --- a/src/tensorflow/lite/micro/micro_mutable_op_resolver.h +++ /dev/null @@ -1,113 +0,0 @@ -/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ -#define TENSORFLOW_LITE_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ - -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/op_resolver.h" -#include "tensorflow/lite/micro/compatibility.h" -#include "tensorflow/lite/schema/schema_generated.h" - -#ifndef TFLITE_REGISTRATIONS_MAX -#define TFLITE_REGISTRATIONS_MAX (128) -#endif - -namespace tflite { - -// Op versions discussed in this file are enumerated here: -// tensorflow/lite/tools/versioning/op_version.cc - -template -class MicroOpResolver : public OpResolver { - public: - const TfLiteRegistration* FindOp(tflite::BuiltinOperator op, - int version) const override { - for (unsigned int i = 0; i < registrations_len_; ++i) { - const TfLiteRegistration& registration = registrations_[i]; - if ((registration.builtin_code == op) && - (registration.version == version)) { - return ®istration; - } - } - return nullptr; - } - - const TfLiteRegistration* FindOp(const char* op, int version) const override { - for (unsigned int i = 0; i < registrations_len_; ++i) { - const TfLiteRegistration& registration = registrations_[i]; - if ((registration.builtin_code == BuiltinOperator_CUSTOM) && - (strcmp(registration.custom_name, op) == 0) && - (registration.version == version)) { - return ®istration; - } - } - return nullptr; - } - - void AddBuiltin(tflite::BuiltinOperator op, TfLiteRegistration* registration, - int min_version = 1, int max_version = 1) { - for (int version = min_version; version <= max_version; ++version) { - if (registrations_len_ >= tOpCount) { - // TODO(b/147748244) - Add error reporting hooks so we can report this! - return; - } - TfLiteRegistration* new_registration = - ®istrations_[registrations_len_]; - registrations_len_ += 1; - - *new_registration = *registration; - new_registration->builtin_code = op; - new_registration->version = version; - } - } - - void AddCustom(const char* name, TfLiteRegistration* registration, - int min_version = 1, int max_version = 1) { - for (int version = min_version; version <= max_version; ++version) { - if (registrations_len_ >= tOpCount) { - // TODO(b/147748244) - Add error reporting hooks so we can report this! - return; - } - TfLiteRegistration* new_registration = - ®istrations_[registrations_len_]; - registrations_len_ += 1; - - *new_registration = *registration; - new_registration->builtin_code = BuiltinOperator_CUSTOM; - new_registration->custom_name = name; - new_registration->version = version; - } - } - - unsigned int GetRegistrationLength() { return registrations_len_; } - - private: - TfLiteRegistration registrations_[tOpCount]; - unsigned int registrations_len_ = 0; - - TF_LITE_REMOVE_VIRTUAL_DELETE -}; - -// TODO(b/147854028): Consider switching all uses of MicroMutableOpResolver to -// MicroOpResolver. -class MicroMutableOpResolver - : public MicroOpResolver { - private: - TF_LITE_REMOVE_VIRTUAL_DELETE -}; - -}; // namespace tflite - -#endif // TENSORFLOW_LITE_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ diff --git a/src/tensorflow/lite/type_to_tflitetype.h b/src/tensorflow/lite/type_to_tflitetype.h deleted file mode 100644 index 28efb96..0000000 --- a/src/tensorflow/lite/type_to_tflitetype.h +++ /dev/null @@ -1,78 +0,0 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ -#ifndef TENSORFLOW_LITE_TYPE_TO_TFLITETYPE_H_ -#define TENSORFLOW_LITE_TYPE_TO_TFLITETYPE_H_ - -// Arduino build defines abs as a macro here. That is invalid C++, and breaks -// libc++'s header, undefine it. -#ifdef abs -#undef abs -#endif - -#include -#include - -#include "tensorflow/lite/c/common.h" - -namespace tflite { - -// Map statically from a c++ type to a TfLiteType. Used in interpreter for safe -// casts. -template -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteNoType; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteInt32; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteInt16; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteInt64; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteFloat32; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteUInt8; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteInt8; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteBool; -} -template <> -constexpr TfLiteType typeToTfLiteType>() { - return kTfLiteComplex64; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteString; -} -template <> -constexpr TfLiteType typeToTfLiteType() { - return kTfLiteFloat16; -} -} // namespace tflite -#endif // TENSORFLOW_LITE_TYPE_TO_TFLITETYPE_H_ diff --git a/src/tensorflow_arm/tensorflow/core/public/version.h b/src/tensorflow_arm/tensorflow/core/public/version.h new file mode 100644 index 0000000..8c036c1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/core/public/version.h @@ -0,0 +1,142 @@ +#if !defined(ESP32) +/* Copyright 2015 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CORE_PUBLIC_VERSION_H_ +#define TENSORFLOW_CORE_PUBLIC_VERSION_H_ + +// TensorFlow uses semantic versioning, see http://semver.org/. + +// Also update tensorflow/tensorflow.bzl and +// tensorflow/tools/pip_package/setup.py +#define TF_MAJOR_VERSION 2 +#define TF_MINOR_VERSION 5 +#define TF_PATCH_VERSION 0 + +// TF_VERSION_SUFFIX is non-empty for pre-releases (e.g. "-alpha", "-alpha.1", +// "-beta", "-rc", "-rc.1") +#define TF_VERSION_SUFFIX "" + +#define TF_STR_HELPER(x) #x +#define TF_STR(x) TF_STR_HELPER(x) + +// e.g. "0.5.0" or "0.6.0-alpha". +#define TF_VERSION_STRING \ + (TF_STR(TF_MAJOR_VERSION) "." TF_STR(TF_MINOR_VERSION) "." TF_STR( \ + TF_PATCH_VERSION) TF_VERSION_SUFFIX) + +// GraphDef compatibility versions (the versions field in graph.proto). +// +// Each graph has producer and min_consumer versions, and each +// consumer has its own version and a min_producer. In addition, graphs can +// mark specific consumer versions as bad (to prevent bugs from executing). +// A consumer will execute a graph if the consumer's version is at least the +// graph's min_consumer, the graph's producer version is at least the consumer's +// min_producer, and the consumer version isn't specifically disallowed by the +// graph. +// +// By default, newly created graphs have producer version TF_GRAPH_DEF_VERSION +// min_consumer TF_GRAPH_DEF_MIN_CONSUMER, and no other bad consumer versions. +// +// Version history: +// +// 0. Graphs created before GraphDef versioning +// 1. First real version (2dec2015) +// 2. adjust_contrast only takes float, doesn't perform clamping (11dec2015) +// 3. Remove TileGrad, since it was equivalent to reduce_sum (30dec2015) +// 4. When support for this version is removed, we can safely make AttrValue +// parsing more strict with respect to empty list values (see +// 111635679, 7jan2016). +// 5. Graphs are wholly-validated during Session::Create() (7jan2016). +// 6. TensorFlow is scalar strict within Google (27jan2016). +// 7. Remove TopK in favor of TopKV2 (5feb2016). +// 8. Replace RandomCrop from C++ with pure Python (5feb2016). +// 9. Deprecate batch_norm_with_global_normalization (16feb2016). +// 10. Deprecate conv3d_backprop_{filter,input} (10jun2016). +// 11. Deprecate {batch}_self_adjoint_eig (3aug2016). +// 12. Graph consumers understand the node_def field of FunctionDef (22aug2016). +// 13. Deprecate multiple batch linear algebra ops (9sep2016). +// 14. Deprecate batch_matrix_* ops. (10sep2016). +// 15. Deprecate batch_fft_* ops. (14sep2016). +// 16. Deprecate tensor_array (v1) ops in favor of v2 (10nov2016). +// 17. Deprecate inv (11nov2016). +// 17. Expose reverse_v2 (10nov2016) +// 18. Add VariableV2 (30nov2016) +// 19. Deprecated ops created by models moved out of core SkipGram, NegTrain. +// (08dec2016) +// 20. Catch all version 1.0 changes to Python API generation. SplitV is now +// used for tf.split, ReverseV2 is now used by tf.reverse, ConcatV2 is +// now used by tf.concat. Graphs use flooring +// division and mod semantics. TensorArrayV3. (12dec2016) +// Also considered the version for when it is required for reduction +// ops' indices to be scalar or vector, and not higher rank. +// Some earlier graph def versions allowed this. +// 21. Dropped FunctionDef.Node support, switched to node_def introduced +// in version 12. (11jan2017) +// 22. Placeholder now can specify and enforce scalar and partial +// shapes, particularly when restoring a graph from GraphDef +// produced at version 22 or later. (04/10/2016) +// 23. Remove NonMaxSuppression in favor of NonMaxSuppressionV2. +// 24. Deprecate lookup ops (v1) ops in favor of v2 (30may2017) +// 25. Deprecate stack (v1) ops in favor of v2 (2017/6/15). +// 25. Deprecate RandomPoisson (v1) ops in favor of v2 (2017/10/25). +// 26. Add a bool 'stripped_default_attrs' to MetaInfoDef indicating +// whether default-valued attrs have been stripped from the nodes in the +// GraphDef. (7dec2017) +// 27. Deprecate TensorArray ops v2 in favor of v3 and deprecated io_ops +// deprecated in favor of V2 ops. (2018/01/23) +// 28. Deprecate MatrixExponential op in favor of Python implementation. +// (2018/08/21). +// (2019/02/15). Added `control_ret` field to FunctionDef proto, and +// `control_output` field to OpDef proto. +// 29. Deprecate StatefulStandardNormal op in favor of StatefulStandardNormalV2. +// (2019/03/25). +// (2019/04/17). Added `arg_attr` field to FunctionDefProto. +// 30. (2019/05/09) First date based GraphDef version. GraphDef +// versions advance by 1 each day after this point. + +#define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0 +#define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0 +#define TF_GRAPH_DEF_VERSION 642 // Updated: 2021/1/10 + +// Checkpoint compatibility versions (the versions field in SavedSliceMeta). +// +// The checkpoint versions have the same semantics as GraphDef versions, but the +// numbering scheme is separate. We have no plans to ever deprecate checkpoint +// versions, but it's good to have this in place in case we ever need to. +// +// Version history: +// +// 0. Checkpoints saved before checkpoint versioning. +// 1. First real version (10feb2015). +#define TF_CHECKPOINT_VERSION_MIN_PRODUCER 0 +#define TF_CHECKPOINT_VERSION_MIN_CONSUMER 0 +#define TF_CHECKPOINT_VERSION 1 + +/// Version query functions (defined in generated version_info.cc) + +// Host compiler version (declared elsewhere to be __VERSION__) +extern const char* tf_compiler_version(); +// The git commit designator when tensorflow was built +// If no git repository, this will be "internal". +extern const char* tf_git_version(); +// Value of the _GLIBCXX_USE_CXX11_ABI flag, or 0 if it's not set. +extern int tf_cxx11_abi_flag(); +// Returns 1 if build is monolithic, or 0 otherwise. +extern int tf_monolithic_build(); + +#endif // TENSORFLOW_CORE_PUBLIC_VERSION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/c/builtin_op_data.h b/src/tensorflow_arm/tensorflow/lite/c/builtin_op_data.h new file mode 100644 index 0000000..992aecb --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/c/builtin_op_data.h @@ -0,0 +1,487 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_C_BUILTIN_OP_DATA_H_ +#define TENSORFLOW_LITE_C_BUILTIN_OP_DATA_H_ + +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus + +// TfLiteReshapeParams can't have dynamic data so we fix the maximum possible +// number of dimensions. +#define TFLITE_RESHAPE_PARAMS_MAX_DIMENSION_COUNT 8 + +// TODO(aselle): Consider using "if this then that" for testing. + +// Useful placeholder to put in otherwise empty structs to avoid size warnings. +typedef struct { + char dummy; +} EmptyStructPlaceholder; + +// IMPORTANT: All new members of structs must be added at the end to ensure +// backwards compatibility. + +// Possible padding types (for convolutions) +typedef enum { + kTfLitePaddingUnknown = 0, + kTfLitePaddingSame, + kTfLitePaddingValid, +} TfLitePadding; + +typedef enum { + kTfLiteMirrorPaddingUnknown = 0, + kTfLiteMirrorPaddingReflect, + kTfLiteMirrorPaddingSymmetric, +} TfLiteMirrorPaddingMode; + +// TODO(b/130259536): We should move this out of builtin_op_data. +typedef struct { + int width; + int height; + int width_offset; + int height_offset; +} TfLitePaddingValues; + +typedef struct { + TfLiteMirrorPaddingMode mode; +} TfLiteMirrorPaddingParams; + +// Possible fused activation functions. +// TODO(aselle): rename to TfLiteActivation +typedef enum { + kTfLiteActNone = 0, + kTfLiteActRelu, + kTfLiteActReluN1To1, // min(max(-1, x), 1) + kTfLiteActRelu6, // min(max(0, x), 6) + kTfLiteActTanh, + kTfLiteActSignBit, + kTfLiteActSigmoid, +} TfLiteFusedActivation; + +typedef struct { + // Parameters for CONV_2D version 1. + TfLitePadding padding; + int stride_width; + int stride_height; + TfLiteFusedActivation activation; + + // Parameters for CONV_2D version 2. + // Note: Version 2 supports dilation values not equal to 1. + int dilation_width_factor; + int dilation_height_factor; +} TfLiteConvParams; + +typedef struct { + TfLitePadding padding; + int stride_width; + int stride_height; + int filter_width; + int filter_height; + TfLiteFusedActivation activation; + struct { + TfLitePaddingValues padding; + } computed; +} TfLitePoolParams; + +typedef struct { + // Parameters for DepthwiseConv version 1 or above. + TfLitePadding padding; + int stride_width; + int stride_height; + // `depth_multiplier` is redundant. It's used by CPU kernels in + // TensorFlow 2.0 or below, but ignored in versions above. + // + // The information can be deduced from the shape of input and the shape of + // weights. Since the TFLiteConverter toolchain doesn't support partially + // specified shapes, relying on `depth_multiplier` stops us from supporting + // graphs with dynamic shape tensors. + // + // Note: Some of the delegates (e.g. NNAPI, GPU) are still relying on this + // field. + int depth_multiplier; + TfLiteFusedActivation activation; + // Parameters for DepthwiseConv version 2 or above. + int dilation_width_factor; + int dilation_height_factor; +} TfLiteDepthwiseConvParams; + +typedef struct { + int rank; + TfLiteFusedActivation activation; + + // Parameter for SVDF version 4. + bool asymmetric_quantize_inputs; +} TfLiteSVDFParams; + +typedef struct { + TfLiteFusedActivation activation; + + // Parameter for RNN version 3. + bool asymmetric_quantize_inputs; +} TfLiteRNNParams; + +typedef struct { + bool time_major; + TfLiteFusedActivation activation; + + // Parameter for Sequence RNN version 3. + bool asymmetric_quantize_inputs; +} TfLiteSequenceRNNParams; + +typedef struct { + bool time_major; + TfLiteFusedActivation activation; + bool merge_outputs; + + // Parameter for Bidirectional RNN verison 3. + bool asymmetric_quantize_inputs; +} TfLiteBidirectionalSequenceRNNParams; + +typedef enum { + kTfLiteFullyConnectedWeightsFormatDefault = 0, + kTfLiteFullyConnectedWeightsFormatShuffled4x16Int8 = 1, +} TfLiteFullyConnectedWeightsFormat; + +typedef struct { + // Parameters for FullyConnected version 1 or above. + TfLiteFusedActivation activation; + + // Parameters for FullyConnected version 2 or above. + TfLiteFullyConnectedWeightsFormat weights_format; + + // Parameters for FullyConnected version 5 or above. + // If set to true, then the number of dimensions in the input and the output + // tensors are the same. Furthermore, all but the last dimension of the input + // and output shapes will be equal. + bool keep_num_dims; + + // Parameters for FullyConnected version 7 or above. + // If set to true and the weights are quantized, then non constant inputs + // are quantized at evaluation time with asymmetric quantization. + bool asymmetric_quantize_inputs; +} TfLiteFullyConnectedParams; + +typedef enum { + kTfLiteLshProjectionUnknown = 0, + kTfLiteLshProjectionSparse = 1, + kTfLiteLshProjectionDense = 2, +} TfLiteLSHProjectionType; + +typedef struct { + TfLiteLSHProjectionType type; +} TfLiteLSHProjectionParams; + +typedef struct { + float beta; +} TfLiteSoftmaxParams; + +typedef struct { + int axis; + TfLiteFusedActivation activation; +} TfLiteConcatenationParams; + +typedef struct { + TfLiteFusedActivation activation; + // Parameter added for the version 4. + bool pot_scale_int16; +} TfLiteAddParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLiteSpaceToBatchNDParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLiteBatchToSpaceNDParams; + +typedef struct { + bool adj_x; + bool adj_y; + // Parameters for BatchMatMul version 4 or above. + // If set to true and the weights are quantized, then non constant inputs + // are quantized at evaluation time with asymmetric quantization. + bool asymmetric_quantize_inputs; +} TfLiteBatchMatMulParams; + +typedef struct { + TfLiteFusedActivation activation; +} TfLiteMulParams; + +typedef struct { + TfLiteFusedActivation activation; + // Parameter added for the version 5. + bool pot_scale_int16; +} TfLiteSubParams; + +typedef struct { + TfLiteFusedActivation activation; +} TfLiteDivParams; + +typedef struct { + TfLiteFusedActivation activation; +} TfLiteL2NormParams; + +typedef struct { + int radius; + float bias; + float alpha; + float beta; +} TfLiteLocalResponseNormParams; + +typedef enum { + kTfLiteLSTMFullKernel = 0, + kTfLiteLSTMBasicKernel +} TfLiteLSTMKernelType; + +typedef struct { + // Parameters for LSTM version 1. + TfLiteFusedActivation activation; + float cell_clip; + float proj_clip; + + // Parameters for LSTM version 2. + // kTfLiteLSTMBasicKernel is only supported in version 2 or above. + TfLiteLSTMKernelType kernel_type; + + // Parameters for LSTM version 4. + bool asymmetric_quantize_inputs; +} TfLiteLSTMParams; + +typedef struct { + // Parameters needed for the underlying LSTM. + TfLiteFusedActivation activation; + float cell_clip; + float proj_clip; + + // If set to true then the first dimension is time, otherwise batch. + bool time_major; + + // Parameter for unidirectional sequence RNN version 3. + bool asymmetric_quantize_inputs; +} TfLiteUnidirectionalSequenceLSTMParams; + +typedef struct { + // Parameters supported by version 1: + // Parameters inherited for the LSTM kernel. + TfLiteFusedActivation activation; + float cell_clip; + float proj_clip; + + // If true, store the outputs of both directions in the first output. + bool merge_outputs; + + // Parameters supported by version 2: + // If set to true then the first dimension is time, otherwise batch. + bool time_major; + + // Parameters supported by version 4: + // If set to true, then hybrid ops use asymmetric quantization for inputs. + bool asymmetric_quantize_inputs; +} TfLiteBidirectionalSequenceLSTMParams; + +typedef struct { + bool align_corners; + // half_pixel_centers assumes pixels are of half the actual dimensions, and + // yields more accurate resizes. Corresponds to the same argument for the + // original TensorFlow op in TF2.0. + bool half_pixel_centers; +} TfLiteResizeBilinearParams; + +typedef struct { + bool align_corners; + bool half_pixel_centers; +} TfLiteResizeNearestNeighborParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLitePadParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLitePadV2Params; + +typedef struct { + // TODO(ahentz): We can't have dynamic data in this struct, at least not yet. + // For now we will fix the maximum possible number of dimensions. + int shape[TFLITE_RESHAPE_PARAMS_MAX_DIMENSION_COUNT]; + int num_dimensions; +} TfLiteReshapeParams; + +typedef struct { + int ngram_size; + int max_skip_size; + bool include_all_ngrams; +} TfLiteSkipGramParams; + +typedef struct { + int block_size; +} TfLiteSpaceToDepthParams; + +typedef struct { + int block_size; +} TfLiteDepthToSpaceParams; + +typedef struct { + TfLiteType in_data_type; + TfLiteType out_data_type; +} TfLiteCastParams; + +typedef enum { + kTfLiteCombinerTypeSum = 0, + kTfLiteCombinerTypeMean = 1, + kTfLiteCombinerTypeSqrtn = 2, +} TfLiteCombinerType; + +typedef struct { + TfLiteCombinerType combiner; +} TfLiteEmbeddingLookupSparseParams; + +typedef struct { + int axis; +} TfLiteGatherParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLiteTransposeParams; + +typedef struct { + bool keep_dims; +} TfLiteReducerParams; + +typedef struct { + int num_splits; +} TfLiteSplitParams; + +typedef struct { + int num_splits; +} TfLiteSplitVParams; + +typedef struct { + // TODO(ahentz): We can't have dynamic data in this struct, at least not yet. + // For now we will fix the maximum possible number of dimensions. + int squeeze_dims[8]; + int num_squeeze_dims; +} TfLiteSqueezeParams; + +typedef struct { + int begin_mask; + int end_mask; + int ellipsis_mask; + int new_axis_mask; + int shrink_axis_mask; +} TfLiteStridedSliceParams; + +typedef struct { + TfLiteType output_type; +} TfLiteArgMaxParams; + +typedef struct { + TfLiteType output_type; +} TfLiteArgMinParams; + +typedef struct { + TfLitePadding padding; + int stride_width; + int stride_height; +} TfLiteTransposeConvParams; + +typedef struct { + bool validate_indices; +} TfLiteSparseToDenseParams; + +typedef struct { + TfLiteType out_type; +} TfLiteShapeParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLiteRankParams; + +typedef struct { + // Parameters supported by version 1: + float min; + float max; + int num_bits; + + // Parameters supported by version 2: + bool narrow_range; +} TfLiteFakeQuantParams; + +typedef struct { + int values_count; + int axis; +} TfLitePackParams; + +typedef struct { + int axis; +} TfLiteOneHotParams; + +typedef struct { + int num; + int axis; +} TfLiteUnpackParams; + +typedef struct { + float alpha; +} TfLiteLeakyReluParams; + +typedef struct { + TfLiteType index_out_type; +} TfLiteUniqueParams; + +typedef struct { + int seq_dim; + int batch_dim; +} TfLiteReverseSequenceParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLiteMatrixDiagParams; + +typedef struct { + EmptyStructPlaceholder placeholder; +} TfLiteMatrixSetDiagParams; + +typedef struct { + int then_subgraph_index; + int else_subgraph_index; +} TfLiteIfParams; + +typedef struct { + int cond_subgraph_index; + int body_subgraph_index; +} TfLiteWhileParams; + +typedef struct { + bool exclusive; + bool reverse; +} TfLiteCumsumParams; + +typedef struct { + int init_subgraph_index; +} TfLiteCallOnceParams; + +#ifdef __cplusplus +} // extern "C" +#endif // __cplusplus + +#endif // TENSORFLOW_LITE_C_BUILTIN_OP_DATA_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/c/c_api_types.h b/src/tensorflow_arm/tensorflow/lite/c/c_api_types.h new file mode 100644 index 0000000..70bdc67 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/c/c_api_types.h @@ -0,0 +1,95 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// This file declares types used by the pure C inference API defined in c_api.h, +// some of which are also used in the C++ and C kernel and interpreter APIs. + +#ifndef TENSORFLOW_LITE_C_C_API_TYPES_H_ +#define TENSORFLOW_LITE_C_C_API_TYPES_H_ + +#include + +#ifdef __cplusplus +extern "C" { +#endif + +// Define TFL_CAPI_EXPORT macro to export a function properly with a shared +// library. +#ifdef SWIG +#define TFL_CAPI_EXPORT +#else +#if defined(_WIN32) +#ifdef TFL_COMPILE_LIBRARY +#define TFL_CAPI_EXPORT __declspec(dllexport) +#else +#define TFL_CAPI_EXPORT __declspec(dllimport) +#endif // TFL_COMPILE_LIBRARY +#else +#define TFL_CAPI_EXPORT __attribute__((visibility("default"))) +#endif // _WIN32 +#endif // SWIG + +typedef enum TfLiteStatus { + kTfLiteOk = 0, + + // Generally referring to an error in the runtime (i.e. interpreter) + kTfLiteError = 1, + + // Generally referring to an error from a TfLiteDelegate itself. + kTfLiteDelegateError = 2, + + // Generally referring to an error in applying a delegate due to + // incompatibility between runtime and delegate, e.g., this error is returned + // when trying to apply a TfLite delegate onto a model graph that's already + // immutable. + kTfLiteApplicationError = 3 +} TfLiteStatus; + +// Types supported by tensor +typedef enum { + kTfLiteNoType = 0, + kTfLiteFloat32 = 1, + kTfLiteInt32 = 2, + kTfLiteUInt8 = 3, + kTfLiteInt64 = 4, + kTfLiteString = 5, + kTfLiteBool = 6, + kTfLiteInt16 = 7, + kTfLiteComplex64 = 8, + kTfLiteInt8 = 9, + kTfLiteFloat16 = 10, + kTfLiteFloat64 = 11, + kTfLiteComplex128 = 12, + kTfLiteUInt64 = 13, +} TfLiteType; + +// Legacy. Will be deprecated in favor of TfLiteAffineQuantization. +// If per-layer quantization is specified this field will still be populated in +// addition to TfLiteAffineQuantization. +// Parameters for asymmetric quantization. Quantized values can be converted +// back to float using: +// real_value = scale * (quantized_value - zero_point) +typedef struct TfLiteQuantizationParams { + float scale; + int32_t zero_point; +} TfLiteQuantizationParams; + +#ifdef __cplusplus +} // extern C +#endif +#endif // TENSORFLOW_LITE_C_C_API_TYPES_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/c/common.c b/src/tensorflow_arm/tensorflow/lite/c/common.c similarity index 89% rename from src/tensorflow/lite/c/common.c rename to src/tensorflow_arm/tensorflow/lite/c/common.c index 1721e75..c455be8 100644 --- a/src/tensorflow/lite/c/common.c +++ b/src/tensorflow_arm/tensorflow/lite/c/common.c @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,7 +14,9 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/c/c_api_types.h" + #ifndef TF_LITE_STATIC_MEMORY #include #include @@ -79,7 +82,8 @@ TfLiteFloatArray* TfLiteFloatArrayCreate(int size) { void TfLiteFloatArrayFree(TfLiteFloatArray* a) { free(a); } void TfLiteTensorDataFree(TfLiteTensor* t) { - if (t->allocation_type == kTfLiteDynamic) { + if (t->allocation_type == kTfLiteDynamic || + t->allocation_type == kTfLitePersistentRo) { free(t->data.raw); } t->data.raw = NULL; @@ -119,7 +123,8 @@ void TfLiteSparsityFree(TfLiteSparsity* sparsity) { } if (sparsity->dim_metadata) { - for (int i = 0; i < sparsity->dim_metadata_size; i++) { + int i = 0; + for (; i < sparsity->dim_metadata_size; i++) { TfLiteDimensionMetadata metadata = sparsity->dim_metadata[i]; if (metadata.format == kTfLiteDimSparseCSR) { TfLiteIntArrayFree(metadata.array_segments); @@ -140,6 +145,11 @@ void TfLiteTensorFree(TfLiteTensor* t) { if (t->dims) TfLiteIntArrayFree(t->dims); t->dims = NULL; + if (t->dims_signature) { + TfLiteIntArrayFree((TfLiteIntArray *) t->dims_signature); + } + t->dims_signature = NULL; + TfLiteQuantizationFree(&t->quantization); TfLiteSparsityFree(t->sparsity); t->sparsity = NULL; @@ -166,7 +176,8 @@ void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims, } void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor) { - if (tensor->allocation_type != kTfLiteDynamic) { + if (tensor->allocation_type != kTfLiteDynamic && + tensor->allocation_type != kTfLitePersistentRo) { return; } // TODO(b/145340303): Tensor data should be aligned. @@ -195,14 +206,20 @@ const char* TfLiteTypeGetName(TfLiteType type) { return "INT8"; case kTfLiteInt64: return "INT64"; + case kTfLiteUInt64: + return "UINT64"; case kTfLiteBool: return "BOOL"; case kTfLiteComplex64: return "COMPLEX64"; + case kTfLiteComplex128: + return "COMPLEX128"; case kTfLiteString: return "STRING"; case kTfLiteFloat16: return "FLOAT16"; + case kTfLiteFloat64: + return "FLOAT64"; } return "Unknown type"; } @@ -218,3 +235,5 @@ TfLiteDelegate TfLiteDelegateCreate() { }; return d; } + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/c/common.h b/src/tensorflow_arm/tensorflow/lite/c/common.h new file mode 100644 index 0000000..45cf837 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/c/common.h @@ -0,0 +1,916 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// This file defines common C types and APIs for implementing operations, +// delegates and other constructs in TensorFlow Lite. The actual operations and +// delegates can be defined using C++, but the interface between the interpreter +// and the operations are C. +// +// Summary of abstractions +// TF_LITE_ENSURE - Self-sufficient error checking +// TfLiteStatus - Status reporting +// TfLiteIntArray - stores tensor shapes (dims), +// TfLiteContext - allows an op to access the tensors +// TfLiteTensor - tensor (a multidimensional array) +// TfLiteNode - a single node or operation +// TfLiteRegistration - the implementation of a conceptual operation. +// TfLiteDelegate - allows delegation of nodes to alternative backends. +// +// Some abstractions in this file are created and managed by Interpreter. +// +// NOTE: The order of values in these structs are "semi-ABI stable". New values +// should be added only to the end of structs and never reordered. + +#ifndef TENSORFLOW_LITE_C_COMMON_H_ +#define TENSORFLOW_LITE_C_COMMON_H_ + +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/c_api_types.h" + +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus + +// The list of external context types known to TF Lite. This list exists solely +// to avoid conflicts and to ensure ops can share the external contexts they +// need. Access to the external contexts is controlled by one of the +// corresponding support files. +typedef enum TfLiteExternalContextType { + kTfLiteEigenContext = 0, // include eigen_support.h to use. + kTfLiteGemmLowpContext = 1, // include gemm_support.h to use. + kTfLiteEdgeTpuContext = 2, // Placeholder for Edge TPU support. + kTfLiteCpuBackendContext = 3, // include cpu_backend_context.h to use. + kTfLiteMaxExternalContexts = 4 +} TfLiteExternalContextType; + +// Forward declare so dependent structs and methods can reference these types +// prior to the struct definitions. +struct TfLiteContext; +struct TfLiteDelegate; +struct TfLiteRegistration; + +// An external context is a collection of information unrelated to the TF Lite +// framework, but useful to a subset of the ops. TF Lite knows very little +// about the actual contexts, but it keeps a list of them, and is able to +// refresh them if configurations like the number of recommended threads +// change. +typedef struct TfLiteExternalContext { + TfLiteExternalContextType type; + TfLiteStatus (*Refresh)(struct TfLiteContext* context); +} TfLiteExternalContext; + +#define kTfLiteOptionalTensor (-1) + +// Fixed size list of integers. Used for dimensions and inputs/outputs tensor +// indices +typedef struct TfLiteIntArray { + int size; +// gcc 6.1+ have a bug where flexible members aren't properly handled +// https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c +#if (!defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \ + __GNUC_MINOR__ >= 1) || \ + defined(HEXAGON) || (__clang_major__ == 7 && __clang_minor__ == 1) + int data[0]; +#else + int data[]; +#endif +} TfLiteIntArray; + +// Given the size (number of elements) in a TfLiteIntArray, calculate its size +// in bytes. +int TfLiteIntArrayGetSizeInBytes(int size); + +#ifndef TF_LITE_STATIC_MEMORY +// Create a array of a given `size` (uninitialized entries). +// This returns a pointer, that you must free using TfLiteIntArrayFree(). +TfLiteIntArray* TfLiteIntArrayCreate(int size); +#endif + +// Check if two intarrays are equal. Returns 1 if they are equal, 0 otherwise. +int TfLiteIntArrayEqual(const TfLiteIntArray* a, const TfLiteIntArray* b); + +// Check if an intarray equals an array. Returns 1 if equals, 0 otherwise. +int TfLiteIntArrayEqualsArray(const TfLiteIntArray* a, int b_size, + const int b_data[]); + +#ifndef TF_LITE_STATIC_MEMORY +// Create a copy of an array passed as `src`. +// You are expected to free memory with TfLiteIntArrayFree +TfLiteIntArray* TfLiteIntArrayCopy(const TfLiteIntArray* src); + +// Free memory of array `a`. +void TfLiteIntArrayFree(TfLiteIntArray* a); +#endif // TF_LITE_STATIC_MEMORY + +// Fixed size list of floats. Used for per-channel quantization. +typedef struct TfLiteFloatArray { + int size; +// gcc 6.1+ have a bug where flexible members aren't properly handled +// https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c +// This also applies to the toolchain used for Qualcomm Hexagon DSPs. +#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \ + __GNUC_MINOR__ >= 1 + float data[0]; +#else + float data[]; +#endif +} TfLiteFloatArray; + +// Given the size (number of elements) in a TfLiteFloatArray, calculate its size +// in bytes. +int TfLiteFloatArrayGetSizeInBytes(int size); + +#ifndef TF_LITE_STATIC_MEMORY +// Create a array of a given `size` (uninitialized entries). +// This returns a pointer, that you must free using TfLiteFloatArrayFree(). +TfLiteFloatArray* TfLiteFloatArrayCreate(int size); + +// Free memory of array `a`. +void TfLiteFloatArrayFree(TfLiteFloatArray* a); +#endif // TF_LITE_STATIC_MEMORY + +// Since we must not depend on any libraries, define a minimal subset of +// error macros while avoiding names that have pre-conceived meanings like +// assert and check. + +// Try to make all reporting calls through TF_LITE_KERNEL_LOG rather than +// calling the context->ReportError function directly, so that message strings +// can be stripped out if the binary size needs to be severely optimized. +#ifndef TF_LITE_STRIP_ERROR_STRINGS +#define TF_LITE_KERNEL_LOG(context, ...) \ + do { \ + (context)->ReportError((context), __VA_ARGS__); \ + } while (false) + +#define TF_LITE_MAYBE_KERNEL_LOG(context, ...) \ + do { \ + if ((context) != nullptr) { \ + (context)->ReportError((context), __VA_ARGS__); \ + } \ + } while (false) +#else // TF_LITE_STRIP_ERROR_STRINGS +#define TF_LITE_KERNEL_LOG(context, ...) +#define TF_LITE_MAYBE_KERNEL_LOG(context, ...) +#endif // TF_LITE_STRIP_ERROR_STRINGS + +// Check whether value is true, and if not return kTfLiteError from +// the current function (and report the error string msg). +#define TF_LITE_ENSURE_MSG(context, value, msg) \ + do { \ + if (!(value)) { \ + TF_LITE_KERNEL_LOG((context), __FILE__ " " msg); \ + return kTfLiteError; \ + } \ + } while (0) + +// Check whether the value `a` is true, and if not return kTfLiteError from +// the current function, while also reporting the location of the error. +#define TF_LITE_ENSURE(context, a) \ + do { \ + if (!(a)) { \ + TF_LITE_KERNEL_LOG((context), "%s:%d %s was not true.", __FILE__, \ + __LINE__, #a); \ + return kTfLiteError; \ + } \ + } while (0) + +#define TF_LITE_ENSURE_STATUS(a) \ + do { \ + const TfLiteStatus s = (a); \ + if (s != kTfLiteOk) { \ + return s; \ + } \ + } while (0) + +// Check whether the value `a == b` is true, and if not return kTfLiteError from +// the current function, while also reporting the location of the error. +// `a` and `b` may be evaluated more than once, so no side effects or +// extremely expensive computations should be done. +// NOTE: Use TF_LITE_ENSURE_TYPES_EQ if comparing TfLiteTypes. +#define TF_LITE_ENSURE_EQ(context, a, b) \ + do { \ + if ((a) != (b)) { \ + TF_LITE_KERNEL_LOG((context), "%s:%d %s != %s (%d != %d)", __FILE__, \ + __LINE__, #a, #b, (a), (b)); \ + return kTfLiteError; \ + } \ + } while (0) + +#define TF_LITE_ENSURE_TYPES_EQ(context, a, b) \ + do { \ + if ((a) != (b)) { \ + TF_LITE_KERNEL_LOG((context), "%s:%d %s != %s (%s != %s)", __FILE__, \ + __LINE__, #a, #b, TfLiteTypeGetName(a), \ + TfLiteTypeGetName(b)); \ + return kTfLiteError; \ + } \ + } while (0) + +#define TF_LITE_ENSURE_NEAR(context, a, b, epsilon) \ + do { \ + auto delta = ((a) > (b)) ? ((a) - (b)) : ((b) - (a)); \ + if (delta > epsilon) { \ + TF_LITE_KERNEL_LOG((context), "%s:%d %s not near %s (%f != %f)", \ + __FILE__, __LINE__, #a, #b, static_cast(a), \ + static_cast(b)); \ + return kTfLiteError; \ + } \ + } while (0) + +#define TF_LITE_ENSURE_OK(context, status) \ + do { \ + const TfLiteStatus s = (status); \ + if ((s) != kTfLiteOk) { \ + return s; \ + } \ + } while (0) + +// Single-precision complex data type compatible with the C99 definition. +typedef struct TfLiteComplex64 { + float re, im; // real and imaginary parts, respectively. +} TfLiteComplex64; + +// Double-precision complex data type compatible with the C99 definition. +typedef struct TfLiteComplex128 { + double re, im; // real and imaginary parts, respectively. +} TfLiteComplex128; + +// Half precision data type compatible with the C99 definition. +typedef struct TfLiteFloat16 { + uint16_t data; +} TfLiteFloat16; + +// Return the name of a given type, for error reporting purposes. +const char* TfLiteTypeGetName(TfLiteType type); + +// SupportedQuantizationTypes. +typedef enum TfLiteQuantizationType { + // No quantization. + kTfLiteNoQuantization = 0, + // Affine quantization (with support for per-channel quantization). + // Corresponds to TfLiteAffineQuantization. + kTfLiteAffineQuantization = 1, +} TfLiteQuantizationType; + +// Structure specifying the quantization used by the tensor, if-any. +typedef struct TfLiteQuantization { + // The type of quantization held by params. + TfLiteQuantizationType type; + // Holds an optional reference to a quantization param structure. The actual + // type depends on the value of the `type` field (see the comment there for + // the values and corresponding types). + void* params; +} TfLiteQuantization; + +// Parameters for asymmetric quantization across a dimension (i.e per output +// channel quantization). +// quantized_dimension specifies which dimension the scales and zero_points +// correspond to. +// For a particular value in quantized_dimension, quantized values can be +// converted back to float using: +// real_value = scale * (quantized_value - zero_point) +typedef struct TfLiteAffineQuantization { + TfLiteFloatArray* scale; + TfLiteIntArray* zero_point; + int32_t quantized_dimension; +} TfLiteAffineQuantization; + +/* A union of pointers that points to memory for a given tensor. */ +typedef union TfLitePtrUnion { + /* Do not access these members directly, if possible, use + * GetTensorData(tensor) instead, otherwise only access .data, as other + * members are deprecated. */ + int32_t* i32; + int64_t* i64; + uint64_t* u64; + float* f; + TfLiteFloat16* f16; + double* f64; + char* raw; + const char* raw_const; + uint8_t* uint8; + bool* b; + int16_t* i16; + TfLiteComplex64* c64; + TfLiteComplex128* c128; + int8_t* int8; + /* Only use this member. */ + void* data; +} TfLitePtrUnion; + +// Memory allocation strategies. +// * kTfLiteMmapRo: Read-only memory-mapped data, or data externally allocated. +// * kTfLiteArenaRw: Arena allocated with no guarantees about persistence, +// and available during eval. +// * kTfLiteArenaRwPersistent: Arena allocated but persistent across eval, and +// only available during eval. +// * kTfLiteDynamic: Allocated during eval, or for string tensors. +// * kTfLitePersistentRo: Allocated and populated during prepare. This is +// useful for tensors that can be computed during prepare and treated +// as constant inputs for downstream ops (also in prepare). +// * kTfLiteCustom: Custom memory allocation provided by the user. See +// TfLiteCustomAllocation below. +typedef enum TfLiteAllocationType { + kTfLiteMemNone = 0, + kTfLiteMmapRo, + kTfLiteArenaRw, + kTfLiteArenaRwPersistent, + kTfLiteDynamic, + kTfLitePersistentRo, + kTfLiteCustom, +} TfLiteAllocationType; + +// The delegates should use zero or positive integers to represent handles. +// -1 is reserved from unallocated status. +typedef int TfLiteBufferHandle; +enum { + kTfLiteNullBufferHandle = -1, +}; + +// Storage format of each dimension in a sparse tensor. +typedef enum TfLiteDimensionType { + kTfLiteDimDense = 0, + kTfLiteDimSparseCSR, +} TfLiteDimensionType; + +// Metadata to encode each dimension in a sparse tensor. +typedef struct TfLiteDimensionMetadata { + TfLiteDimensionType format; + int dense_size; + TfLiteIntArray* array_segments; + TfLiteIntArray* array_indices; +} TfLiteDimensionMetadata; + +// Parameters used to encode a sparse tensor. For detailed explanation of each +// field please refer to lite/schema/schema.fbs. +typedef struct TfLiteSparsity { + TfLiteIntArray* traversal_order; + TfLiteIntArray* block_map; + TfLiteDimensionMetadata* dim_metadata; + int dim_metadata_size; +} TfLiteSparsity; + +// Defines a custom memory allocation not owned by the runtime. +// `data` should be aligned to kDefaultTensorAlignment defined in +// lite/util.h. (Currently 64 bytes) +// NOTE: See Interpreter.SetCustomAllocationForTensor for details on usage. +typedef struct TfLiteCustomAllocation { + void* data; + size_t bytes; +} TfLiteCustomAllocation; + +// A tensor in the interpreter system which is a wrapper around a buffer of +// data including a dimensionality (or NULL if not currently defined). +#ifndef TF_LITE_STATIC_MEMORY +typedef struct TfLiteTensor { + // The data type specification for data stored in `data`. This affects + // what member of `data` union should be used. + TfLiteType type; + // A union of data pointers. The appropriate type should be used for a typed + // tensor based on `type`. + TfLitePtrUnion data; + // A pointer to a structure representing the dimensionality interpretation + // that the buffer should have. NOTE: the product of elements of `dims` + // and the element datatype size should be equal to `bytes` below. + TfLiteIntArray* dims; + // Quantization information. + TfLiteQuantizationParams params; + // How memory is mapped + // kTfLiteMmapRo: Memory mapped read only. + // i.e. weights + // kTfLiteArenaRw: Arena allocated read write memory + // (i.e. temporaries, outputs). + TfLiteAllocationType allocation_type; + // The number of bytes required to store the data of this Tensor. I.e. + // (bytes of each element) * dims[0] * ... * dims[n-1]. For example, if + // type is kTfLiteFloat32 and dims = {3, 2} then + // bytes = sizeof(float) * 3 * 2 = 4 * 3 * 2 = 24. + size_t bytes; + + // An opaque pointer to a tflite::MMapAllocation + const void* allocation; + + // Null-terminated name of this tensor. + const char* name; + + // The delegate which knows how to handle `buffer_handle`. + // WARNING: This is an experimental interface that is subject to change. + struct TfLiteDelegate* delegate; + + // An integer buffer handle that can be handled by `delegate`. + // The value is valid only when delegate is not null. + // WARNING: This is an experimental interface that is subject to change. + TfLiteBufferHandle buffer_handle; + + // If the delegate uses its own buffer (e.g. GPU memory), the delegate is + // responsible to set data_is_stale to true. + // `delegate->CopyFromBufferHandle` can be called to copy the data from + // delegate buffer. + // WARNING: This is an // experimental interface that is subject to change. + bool data_is_stale; + + // True if the tensor is a variable. + bool is_variable; + + // Quantization information. Replaces params field above. + TfLiteQuantization quantization; + + // Parameters used to encode a sparse tensor. + // This is optional. The field is NULL if a tensor is dense. + // WARNING: This is an experimental interface that is subject to change. + TfLiteSparsity* sparsity; + + // Optional. Encodes shapes with unknown dimensions with -1. This field is + // only populated when unknown dimensions exist in a read-write tensor (i.e. + // an input or output tensor). (e.g. `dims` contains [1, 1, 1, 3] and + // `dims_signature` contains [1, -1, -1, 3]). + const TfLiteIntArray* dims_signature; +} TfLiteTensor; + +// A structure representing an instance of a node. +// This structure only exhibits the inputs, outputs and user defined data, not +// other features like the type. +typedef struct TfLiteNode { + // Inputs to this node expressed as indices into the simulator's tensors. + TfLiteIntArray* inputs; + + // Outputs to this node expressed as indices into the simulator's tensors. + TfLiteIntArray* outputs; + + // intermediate tensors to this node expressed as indices into the simulator's + // tensors. + TfLiteIntArray* intermediates; + + // Temporary tensors uses during the computations. This usually contains no + // tensors, but ops are allowed to change that if they need scratch space of + // any sort. + TfLiteIntArray* temporaries; + + // Opaque data provided by the node implementer through `Registration.init`. + void* user_data; + + // Opaque data provided to the node if the node is a builtin. This is usually + // a structure defined in builtin_op_data.h + void* builtin_data; + + // Custom initial data. This is the opaque data provided in the flatbuffer. + // WARNING: This is an experimental interface that is subject to change. + const void* custom_initial_data; + int custom_initial_data_size; + + // The pointer to the delegate. This is non-null only when the node is + // created by calling `interpreter.ModifyGraphWithDelegate`. + // WARNING: This is an experimental interface that is subject to change. + struct TfLiteDelegate* delegate; +} TfLiteNode; +#else // defined(TF_LITE_STATIC_MEMORY)? +// NOTE: This flag is opt-in only at compile time. +// +// Specific reduced TfLiteTensor struct for TF Micro runtime. This struct +// contains only the minimum fields required to initialize and prepare a micro +// inference graph. The fields in this struct have been ordered from +// largest-to-smallest for optimal struct sizeof. +// +// This struct does not use: +// - allocation +// - buffer_handle +// - data_is_stale +// - delegate +// - dims_signature +// - name +// - sparsity +typedef struct TfLiteTensor { + // TODO(b/155784997): Consider consolidating these quantization fields: + // Quantization information. Replaces params field above. + TfLiteQuantization quantization; + + // Quantization information. + TfLiteQuantizationParams params; + + // A union of data pointers. The appropriate type should be used for a typed + // tensor based on `type`. + TfLitePtrUnion data; + + // A pointer to a structure representing the dimensionality interpretation + // that the buffer should have. NOTE: the product of elements of `dims` + // and the element datatype size should be equal to `bytes` below. + TfLiteIntArray* dims; + + // The number of bytes required to store the data of this Tensor. I.e. + // (bytes of each element) * dims[0] * ... * dims[n-1]. For example, if + // type is kTfLiteFloat32 and dims = {3, 2} then + // bytes = sizeof(float) * 3 * 2 = 4 * 3 * 2 = 24. + size_t bytes; + + // The data type specification for data stored in `data`. This affects + // what member of `data` union should be used. + TfLiteType type; + + // How memory is mapped + // kTfLiteMmapRo: Memory mapped read only. + // i.e. weights + // kTfLiteArenaRw: Arena allocated read write memory + // (i.e. temporaries, outputs). + TfLiteAllocationType allocation_type; + + // True if the tensor is a variable. + bool is_variable; +} TfLiteTensor; + +// Specific reduced TfLiteNode struct for TF Micro runtime. This struct contains +// only the minimum fields required to represent a node. +// +// This struct does not use: +// - delegate +// - intermediates +// - temporaries +typedef struct TfLiteNode { + // Inputs to this node expressed as indices into the simulator's tensors. + TfLiteIntArray* inputs; + + // Outputs to this node expressed as indices into the simulator's tensors. + TfLiteIntArray* outputs; + + // Opaque data provided by the node implementer through `Registration.init`. + void* user_data; + + // Opaque data provided to the node if the node is a builtin. This is usually + // a structure defined in builtin_op_data.h + void* builtin_data; + + // Custom initial data. This is the opaque data provided in the flatbuffer. + // WARNING: This is an experimental interface that is subject to change. + const void* custom_initial_data; + int custom_initial_data_size; +} TfLiteNode; +#endif // TF_LITE_STATIC_MEMORY + +// Light-weight tensor struct for TF Micro runtime. Provides the minimal amount +// of information required for a kernel to run during TfLiteRegistration::Eval. +// TODO(b/160955687): Move this field into TF_LITE_STATIC_MEMORY when TFLM +// builds with this flag by default internally. +typedef struct TfLiteEvalTensor { + // A union of data pointers. The appropriate type should be used for a typed + // tensor based on `type`. + TfLitePtrUnion data; + + // A pointer to a structure representing the dimensionality interpretation + // that the buffer should have. + TfLiteIntArray* dims; + + // The data type specification for data stored in `data`. This affects + // what member of `data` union should be used. + TfLiteType type; +} TfLiteEvalTensor; + +#ifndef TF_LITE_STATIC_MEMORY +// Free data memory of tensor `t`. +void TfLiteTensorDataFree(TfLiteTensor* t); + +// Free quantization data. +void TfLiteQuantizationFree(TfLiteQuantization* quantization); + +// Free sparsity parameters. +void TfLiteSparsityFree(TfLiteSparsity* sparsity); + +// Free memory of tensor `t`. +void TfLiteTensorFree(TfLiteTensor* t); + +// Set all of a tensor's fields (and free any previously allocated data). +void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims, + TfLiteQuantizationParams quantization, char* buffer, + size_t size, TfLiteAllocationType allocation_type, + const void* allocation, bool is_variable, + TfLiteTensor* tensor); + +// Resize the allocated data of a (dynamic) tensor. Tensors with allocation +// types other than kTfLiteDynamic will be ignored. +void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor); +#endif // TF_LITE_STATIC_MEMORY + +// WARNING: This is an experimental interface that is subject to change. +// +// Currently, TfLiteDelegateParams has to be allocated in a way that it's +// trivially destructable. It will be stored as `builtin_data` field in +// `TfLiteNode` of the delegate node. +// +// See also the `CreateDelegateParams` function in `interpreter.cc` details. +typedef struct TfLiteDelegateParams { + struct TfLiteDelegate* delegate; + TfLiteIntArray* nodes_to_replace; + TfLiteIntArray* input_tensors; + TfLiteIntArray* output_tensors; +} TfLiteDelegateParams; + +typedef struct TfLiteContext { + // Number of tensors in the context. + size_t tensors_size; + + // The execution plan contains a list of the node indices in execution + // order. execution_plan->size is the current number of nodes. And, + // execution_plan->data[0] is the first node that needs to be run. + // TfLiteDelegates can traverse the current execution plan by iterating + // through each member of this array and using GetNodeAndRegistration() to + // access details about a node. i.e. + // TfLiteIntArray* execution_plan; + // TF_LITE_ENSURE_STATUS(context->GetExecutionPlan(context, &execution_plan)); + // for (int exec_index = 0; exec_index < execution_plan->size; exec_index++) { + // int node_index = execution_plan->data[exec_index]; + // TfLiteNode* node; + // TfLiteRegistration* reg; + // context->GetNodeAndRegistration(context, node_index, &node, ®); + // } + // WARNING: This is an experimental interface that is subject to change. + TfLiteStatus (*GetExecutionPlan)(struct TfLiteContext* context, + TfLiteIntArray** execution_plan); + + // An array of tensors in the interpreter context (of length `tensors_size`) + TfLiteTensor* tensors; + + // opaque full context ptr (an opaque c++ data structure) + void* impl_; + + // Request memory pointer be resized. Updates dimensions on the tensor. + // NOTE: ResizeTensor takes ownership of newSize. + TfLiteStatus (*ResizeTensor)(struct TfLiteContext*, TfLiteTensor* tensor, + TfLiteIntArray* new_size); + // Request that an error be reported with format string msg. + void (*ReportError)(struct TfLiteContext*, const char* msg, ...); + + // Add `tensors_to_add` tensors, preserving pre-existing Tensor entries. If + // non-null, the value pointed to by `first_new_tensor_index` will be set to + // the index of the first new tensor. + TfLiteStatus (*AddTensors)(struct TfLiteContext*, int tensors_to_add, + int* first_new_tensor_index); + + // Get a Tensor node by node_index. + // WARNING: This is an experimental interface that is subject to change. + TfLiteStatus (*GetNodeAndRegistration)( + struct TfLiteContext*, int node_index, TfLiteNode** node, + struct TfLiteRegistration** registration); + + // Replace ops with one or more stub delegate operations. This function + // does not take ownership of `nodes_to_replace`. + TfLiteStatus (*ReplaceNodeSubsetsWithDelegateKernels)( + struct TfLiteContext*, struct TfLiteRegistration registration, + const TfLiteIntArray* nodes_to_replace, struct TfLiteDelegate* delegate); + + // Number of threads that are recommended to subsystems like gemmlowp and + // eigen. + int recommended_num_threads; + + // Access external contexts by type. + // WARNING: This is an experimental interface that is subject to change. + TfLiteExternalContext* (*GetExternalContext)(struct TfLiteContext*, + TfLiteExternalContextType); + // Set the value of a external context. Does not take ownership of the + // pointer. + // WARNING: This is an experimental interface that is subject to change. + void (*SetExternalContext)(struct TfLiteContext*, TfLiteExternalContextType, + TfLiteExternalContext*); + + // Flag for allowing float16 precision for FP32 calculation. + // default: false. + // WARNING: This is an experimental API and subject to change. + bool allow_fp32_relax_to_fp16; + + // Pointer to the op-level profiler, if set; nullptr otherwise. + void* profiler; + + // Allocate persistent buffer which has the same life time as the interpreter. + // Returns nullptr on failure. + // The memory is allocated from heap for TFL, and from tail in TFLM. + // This method is only available in Init or Prepare stage. + // WARNING: This is an experimental interface that is subject to change. + void* (*AllocatePersistentBuffer)(struct TfLiteContext* ctx, size_t bytes); + + // Allocate a buffer which will be deallocated right after invoke phase. + // The memory is allocated from heap in TFL, and from volatile arena in TFLM. + // This method is only available in invoke stage. + // NOTE: If possible use RequestScratchBufferInArena method to avoid memory + // allocation during inference time. + // WARNING: This is an experimental interface that is subject to change. + TfLiteStatus (*AllocateBufferForEval)(struct TfLiteContext* ctx, size_t bytes, + void** ptr); + + // Request a scratch buffer in the arena through static memory planning. + // This method is only available in Prepare stage and the buffer is allocated + // by the interpreter between Prepare and Eval stage. In Eval stage, + // GetScratchBuffer API can be used to fetch the address. + // WARNING: This is an experimental interface that is subject to change. + TfLiteStatus (*RequestScratchBufferInArena)(struct TfLiteContext* ctx, + size_t bytes, int* buffer_idx); + + // Get the scratch buffer pointer. + // This method is only available in Eval stage. + // WARNING: This is an experimental interface that is subject to change. + void* (*GetScratchBuffer)(struct TfLiteContext* ctx, int buffer_idx); + + // Resize the memory pointer of the `tensor`. This method behaves the same as + // `ResizeTensor`, except that it makes a copy of the shape array internally + // so the shape array could be deallocated right afterwards. + // WARNING: This is an experimental interface that is subject to change. + TfLiteStatus (*ResizeTensorExplicit)(struct TfLiteContext* ctx, + TfLiteTensor* tensor, int dims, + const int* shape); + + // This method provides a preview of post-delegation partitioning. Each + // TfLiteDelegateParams in the referenced array corresponds to one instance of + // the delegate kernel. + // Example usage: + // + // TfLiteIntArray* nodes_to_replace = ...; + // TfLiteDelegateParams* params_array; + // int num_partitions = 0; + // TF_LITE_ENSURE_STATUS(context->PreviewDelegatePartitioning( + // context, delegate, nodes_to_replace, ¶ms_array, &num_partitions)); + // for (int idx = 0; idx < num_partitions; idx++) { + // const auto& partition_params = params_array[idx]; + // ... + // } + // + // NOTE: The context owns the memory referenced by partition_params_array. It + // will be cleared with another call to PreviewDelegateParitioning, or after + // TfLiteDelegateParams::Prepare returns. + // + // WARNING: This is an experimental interface that is subject to change. + TfLiteStatus (*PreviewDelegatePartitioning)( + struct TfLiteContext* context, const TfLiteIntArray* nodes_to_replace, + TfLiteDelegateParams** partition_params_array, int* num_partitions); + + // Returns a TfLiteTensor struct for a given index. + // WARNING: This is an experimental interface that is subject to change. + // WARNING: This method may not be available on all platforms. + TfLiteTensor* (*GetTensor)(const struct TfLiteContext* context, + int tensor_idx); + + // Returns a TfLiteEvalTensor struct for a given index. + // WARNING: This is an experimental interface that is subject to change. + // WARNING: This method may not be available on all platforms. + TfLiteEvalTensor* (*GetEvalTensor)(const struct TfLiteContext* context, + int tensor_idx); +} TfLiteContext; + +typedef struct TfLiteRegistration { + // Initializes the op from serialized data. + // If a built-in op: + // `buffer` is the op's params data (TfLiteLSTMParams*). + // `length` is zero. + // If custom op: + // `buffer` is the op's `custom_options`. + // `length` is the size of the buffer. + // + // Returns a type-punned (i.e. void*) opaque data (e.g. a primitive pointer + // or an instance of a struct). + // + // The returned pointer will be stored with the node in the `user_data` field, + // accessible within prepare and invoke functions below. + // NOTE: if the data is already in the desired format, simply implement this + // function to return `nullptr` and implement the free function to be a no-op. + void* (*init)(TfLiteContext* context, const char* buffer, size_t length); + + // The pointer `buffer` is the data previously returned by an init invocation. + void (*free)(TfLiteContext* context, void* buffer); + + // prepare is called when the inputs this node depends on have been resized. + // context->ResizeTensor() can be called to request output tensors to be + // resized. + // + // Returns kTfLiteOk on success. + TfLiteStatus (*prepare)(TfLiteContext* context, TfLiteNode* node); + + // Execute the node (should read node->inputs and output to node->outputs). + // Returns kTfLiteOk on success. + TfLiteStatus (*invoke)(TfLiteContext* context, TfLiteNode* node); + + // profiling_string is called during summarization of profiling information + // in order to group executions together. Providing a value here will cause a + // given op to appear multiple times is the profiling report. This is + // particularly useful for custom ops that can perform significantly + // different calculations depending on their `user-data`. + const char* (*profiling_string)(const TfLiteContext* context, + const TfLiteNode* node); + + // Builtin codes. If this kernel refers to a builtin this is the code + // of the builtin. This is so we can do marshaling to other frameworks like + // NN API. + // Note: It is the responsibility of the registration binder to set this + // properly. + int32_t builtin_code; + + // Custom op name. If the op is a builtin, this will be null. + // Note: It is the responsibility of the registration binder to set this + // properly. + // WARNING: This is an experimental interface that is subject to change. + const char* custom_name; + + // The version of the op. + // Note: It is the responsibility of the registration binder to set this + // properly. + int version; +} TfLiteRegistration; + +// The flags used in `TfLiteDelegate`. Note that this is a bitmask, so the +// values should be 1, 2, 4, 8, ...etc. +typedef enum TfLiteDelegateFlags { + kTfLiteDelegateFlagsNone = 0, + // The flag is set if the delegate can handle dynamic sized tensors. + // For example, the output shape of a `Resize` op with non-constant shape + // can only be inferred when the op is invoked. + // In this case, the Delegate is responsible for calling + // `SetTensorToDynamic` to mark the tensor as a dynamic tensor, and calling + // `ResizeTensor` when invoking the op. + // + // If the delegate isn't capable to handle dynamic tensors, this flag need + // to be set to false. + kTfLiteDelegateFlagsAllowDynamicTensors = 1, + + // This flag can be used by delegates (that allow dynamic tensors) to ensure + // applicable tensor shapes are automatically propagated in the case of tensor + // resizing. + // This means that non-dynamic (allocation_type != kTfLiteDynamic) I/O tensors + // of a delegate kernel will have correct shapes before its Prepare() method + // is called. The runtime leverages TFLite builtin ops in the original + // execution plan to propagate shapes. + // + // A few points to note: + // 1. This requires kTfLiteDelegateFlagsAllowDynamicTensors. If that flag is + // false, this one is redundant since the delegate kernels are re-initialized + // every time tensors are resized. + // 2. Enabling this flag adds some overhead to AllocateTensors(), since extra + // work is required to prepare the original execution plan. + // 3. This flag requires that the original execution plan only have ops with + // valid registrations (and not 'dummy' custom ops like with Flex). + // WARNING: This feature is experimental and subject to change. + kTfLiteDelegateFlagsRequirePropagatedShapes = 2 +} TfLiteDelegateFlags; + +// WARNING: This is an experimental interface that is subject to change. +typedef struct TfLiteDelegate { + // Data that delegate needs to identify itself. This data is owned by the + // delegate. The delegate is owned in the user code, so the delegate is + // responsible for doing this when it is destroyed. + void* data_; + + // Invoked by ModifyGraphWithDelegate. This prepare is called, giving the + // delegate a view of the current graph through TfLiteContext*. It typically + // will look at the nodes and call ReplaceNodeSubsetsWithDelegateKernels() + // to ask the TensorFlow lite runtime to create macro-nodes to represent + // delegated subgraphs of the original graph. + TfLiteStatus (*Prepare)(TfLiteContext* context, + struct TfLiteDelegate* delegate); + + // Copy the data from delegate buffer handle into raw memory of the given + // 'tensor'. Note that the delegate is allowed to allocate the raw bytes as + // long as it follows the rules for kTfLiteDynamic tensors, in which case this + // cannot be null. + TfLiteStatus (*CopyFromBufferHandle)(TfLiteContext* context, + struct TfLiteDelegate* delegate, + TfLiteBufferHandle buffer_handle, + TfLiteTensor* tensor); + + // Copy the data from raw memory of the given 'tensor' to delegate buffer + // handle. This can be null if the delegate doesn't use its own buffer. + TfLiteStatus (*CopyToBufferHandle)(TfLiteContext* context, + struct TfLiteDelegate* delegate, + TfLiteBufferHandle buffer_handle, + TfLiteTensor* tensor); + + // Free the Delegate Buffer Handle. Note: This only frees the handle, but + // this doesn't release the underlying resource (e.g. textures). The + // resources are either owned by application layer or the delegate. + // This can be null if the delegate doesn't use its own buffer. + void (*FreeBufferHandle)(TfLiteContext* context, + struct TfLiteDelegate* delegate, + TfLiteBufferHandle* handle); + + // Bitmask flags. See the comments in `TfLiteDelegateFlags`. + int64_t flags; +} TfLiteDelegate; + +// Build a 'null' delegate, with all the fields properly set to their default +// values. +TfLiteDelegate TfLiteDelegateCreate(); + +#ifdef __cplusplus +} // extern "C" +#endif // __cplusplus +#endif // TENSORFLOW_LITE_C_COMMON_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/error_reporter.cpp b/src/tensorflow_arm/tensorflow/lite/core/api/error_reporter.cpp similarity index 90% rename from src/tensorflow/lite/core/api/error_reporter.cpp rename to src/tensorflow_arm/tensorflow/lite/core/api/error_reporter.cpp index 7070eaa..13e213b 100644 --- a/src/tensorflow/lite/core/api/error_reporter.cpp +++ b/src/tensorflow_arm/tensorflow/lite/core/api/error_reporter.cpp @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,7 +13,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" #include namespace tflite { @@ -36,3 +37,5 @@ int ErrorReporter::ReportError(void*, const char* format, ...) { } } // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/core/api/error_reporter.h b/src/tensorflow_arm/tensorflow/lite/core/api/error_reporter.h new file mode 100644 index 0000000..b0c230b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/core/api/error_reporter.h @@ -0,0 +1,62 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_CORE_API_ERROR_REPORTER_H_ +#define TENSORFLOW_LITE_CORE_API_ERROR_REPORTER_H_ + +#include + +namespace tflite { + +/// A functor that reports error to supporting system. Invoked similar to +/// printf. +/// +/// Usage: +/// ErrorReporter foo; +/// foo.Report("test %d", 5); +/// or +/// va_list args; +/// foo.Report("test %d", args); // where args is va_list +/// +/// Subclass ErrorReporter to provide another reporting destination. +/// For example, if you have a GUI program, you might redirect to a buffer +/// that drives a GUI error log box. +class ErrorReporter { + public: + virtual ~ErrorReporter() {} + virtual int Report(const char* format, va_list args) = 0; + int Report(const char* format, ...); + int ReportError(void*, const char* format, ...); +}; + +} // namespace tflite + +// You should not make bare calls to the error reporter, instead use the +// TF_LITE_REPORT_ERROR macro, since this allows message strings to be +// stripped when the binary size has to be optimized. If you are looking to +// reduce binary size, define TF_LITE_STRIP_ERROR_STRINGS when compiling and +// every call will be stubbed out, taking no memory. +#ifndef TF_LITE_STRIP_ERROR_STRINGS +#define TF_LITE_REPORT_ERROR(reporter, ...) \ + do { \ + static_cast(reporter)->Report(__VA_ARGS__); \ + } while (false) +#else // TF_LITE_STRIP_ERROR_STRINGS +#define TF_LITE_REPORT_ERROR(reporter, ...) +#endif // TF_LITE_STRIP_ERROR_STRINGS + +#endif // TENSORFLOW_LITE_CORE_API_ERROR_REPORTER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.cpp b/src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.cpp new file mode 100644 index 0000000..9ab51d3 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.cpp @@ -0,0 +1,1948 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h" + +#include +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +namespace { + +// Utility class for safely allocating POD data. This is useful for avoiding +// leaks in cases where op params are allocated but fail to propagate to the +// parsed op data (e.g., when model parameters are invalid). +class SafeBuiltinDataAllocator { + public: + class BuiltinDataDeleter { + public: + explicit BuiltinDataDeleter(BuiltinDataAllocator* allocator) + : allocator_(allocator) {} + + void operator()(void* data) { allocator_->Deallocate(data); } + + private: + BuiltinDataAllocator* allocator_; + }; + + template + using BuiltinDataPtr = std::unique_ptr; + + explicit SafeBuiltinDataAllocator(BuiltinDataAllocator* allocator) + : allocator_(allocator) {} + + template + BuiltinDataPtr Allocate() { + return BuiltinDataPtr(allocator_->AllocatePOD(), + BuiltinDataDeleter(allocator_)); + } + + private: + BuiltinDataAllocator* allocator_; +}; + +// All the Parse functions take some pointers as params and this function has +// the common DCHECKs to catch if any of those are nullptr. +void CheckParsePointerParams(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + TFLITE_DCHECK(op != nullptr); + TFLITE_DCHECK(error_reporter != nullptr); + TFLITE_DCHECK(allocator != nullptr); + TFLITE_DCHECK(builtin_data != nullptr); +} + +// Copies the contents from the flatbuffer int vector `flatbuffer` into the +// int array `buffer`. `flat_vector` and `buffer` represent the same +// configuration operation for a given operation. +TfLiteStatus FlatBufferIntVectorToArray( + int max_size_of_buffer, const flatbuffers::Vector* flat_vector, + int* buffer, ErrorReporter* error_reporter, const char* op_name) { + if (!flat_vector) { + TF_LITE_REPORT_ERROR(error_reporter, + "Input array not provided for operation '%s'.\n", + op_name); + return kTfLiteError; + } else { + size_t num_dimensions = flat_vector->size(); + if (num_dimensions > max_size_of_buffer / sizeof(int)) { + TF_LITE_REPORT_ERROR( + error_reporter, + "Found too many dimensions in the input array of operation '%s'.\n", + op_name); + return kTfLiteError; + } else { + for (size_t i = 0; i < num_dimensions; ++i) { + buffer[i] = flat_vector->Get(i); + } + } + } + return kTfLiteOk; +} + +// Converts the flatbuffer activation to what is used at runtime. +TfLiteFusedActivation ConvertActivation(ActivationFunctionType activation) { + switch (activation) { + case ActivationFunctionType_NONE: + return kTfLiteActNone; + case ActivationFunctionType_RELU: + return kTfLiteActRelu; + case ActivationFunctionType_RELU_N1_TO_1: + return kTfLiteActReluN1To1; + case ActivationFunctionType_RELU6: + return kTfLiteActRelu6; + case ActivationFunctionType_TANH: + return kTfLiteActTanh; + case ActivationFunctionType_SIGN_BIT: + return kTfLiteActSignBit; + } + return kTfLiteActNone; +} + +// Converts the flatbuffer padding enum to what is used at runtime. +TfLitePadding ConvertPadding(Padding padding) { + switch (padding) { + case Padding_SAME: + return kTfLitePaddingSame; + case Padding_VALID: + return kTfLitePaddingValid; + } + return kTfLitePaddingUnknown; +} + +#ifndef TF_LITE_STATIC_MEMORY +TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + auto parseLSHProjectionType = [](LSHProjectionType type) { + switch (type) { + case LSHProjectionType_SPARSE: + return kTfLiteLshProjectionSparse; + case LSHProjectionType_DENSE: + return kTfLiteLshProjectionDense; + default: + return kTfLiteLshProjectionUnknown; + } + }; + auto parseCombinerType = [](CombinerType type) { + switch (type) { + case CombinerType_MEAN: + return kTfLiteCombinerTypeMean; + case CombinerType_SQRTN: + return kTfLiteCombinerTypeSqrtn; + case CombinerType_SUM: + default: + return kTfLiteCombinerTypeSum; + } + }; + + SafeBuiltinDataAllocator safe_allocator(allocator); + *builtin_data = nullptr; + switch (op_type) { + case BuiltinOperator_ABS: { + return ParseAbs(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_ADD: { + return ParseAdd(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_ARG_MAX: { + return ParseArgMax(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_ARG_MIN: { + return ParseArgMin(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_AVERAGE_POOL_2D: { + return ParsePool(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_CEIL: { + return ParseCeil(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_CONCATENATION: { + return ParseConcatenation(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_CONV_2D: { + return ParseConv2D(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_DEPTHWISE_CONV_2D: { + return ParseDepthwiseConv2D(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_DEQUANTIZE: { + return ParseDequantize(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_DIV: { + return ParseDiv(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_EXP: { + return ParseExp(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_FILL: { + return ParseFill(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_FLOOR: { + return ParseFloor(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_FLOOR_DIV: { + return ParseFloorDiv(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_FLOOR_MOD: { + return ParseFloorMod(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_FULLY_CONNECTED: { + return ParseFullyConnected(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_GREATER: { + return ParseGreater(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_GREATER_EQUAL: { + return ParseGreaterEqual(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_HARD_SWISH: { + return ParseHardSwish(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_L2_NORMALIZATION: { + return ParseL2Normalization(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_L2_POOL_2D: { + return ParsePool(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_LESS: { + return ParseLess(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_LESS_EQUAL: { + return ParseLessEqual(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_LOG: { + return ParseLog(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_LOGICAL_AND: { + return ParseLogicalAnd(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_LOGICAL_NOT: { + return ParseLogicalNot(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_LOGICAL_OR: { + return ParseLogicalOr(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_LOGISTIC: { + return ParseLogistic(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_MAXIMUM: { + return ParseMaximum(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_MAX_POOL_2D: { + return ParsePool(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_MEAN: { + return ParseReducer(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_MINIMUM: { + return ParseMinimum(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_MUL: { + return ParseMul(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_NEG: { + return ParseNeg(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_NOT_EQUAL: { + return ParseNotEqual(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_PACK: { + return ParsePack(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_PAD: { + return ParsePad(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_PADV2: { + return ParsePadV2(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_POW: { + return ParsePow(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_PRELU: { + return ParsePrelu(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_QUANTIZE: { + return ParseQuantize(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_REDUCE_ANY: { + return ParseReducer(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_REDUCE_MAX: { + return ParseReducer(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_REDUCE_MIN: { + return ParseReducer(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_REDUCE_PROD: { + return ParseReducer(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_RELU: { + return ParseRelu(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_RELU6: { + return ParseRelu6(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_RESHAPE: { + return ParseReshape(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_RESIZE_BILINEAR: { + return ParseResizeBilinear(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_RESIZE_NEAREST_NEIGHBOR: { + return ParseResizeNearestNeighbor(op, error_reporter, allocator, + builtin_data); + } + + case BuiltinOperator_ROUND: { + return ParseRound(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_RSQRT: { + return ParseRsqrt(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SHAPE: { + return ParseShape(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SIN: { + return ParseSin(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SOFTMAX: { + return ParseSoftmax(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SPACE_TO_DEPTH: { + return ParseSpaceToDepth(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SPLIT: { + return ParseSplit(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SPLIT_V: { + return ParseSplitV(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SQRT: { + return ParseSqrt(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SQUARE: { + return ParseSquare(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_STRIDED_SLICE: { + return ParseStridedSlice(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SUB: { + return ParseSub(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SUM: { + return ParseReducer(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_SVDF: { + return ParseSvdf(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_TANH: { + return ParseTanh(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_TRANSPOSE_CONV: { + return ParseTransposeConv(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_UNPACK: { + return ParseUnpack(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_ZEROS_LIKE: { + return ParseZerosLike(op, error_reporter, allocator, builtin_data); + } + + case BuiltinOperator_CAST: { + return ParseCast(op, error_reporter, allocator, builtin_data); + } + case BuiltinOperator_LSH_PROJECTION: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* lshParams = + op->builtin_options_as_LSHProjectionOptions()) { + params->type = parseLSHProjectionType(lshParams->type()); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* sequence_rnn_params = + op->builtin_options_as_SequenceRNNOptions()) { + params->activation = + ConvertActivation(sequence_rnn_params->fused_activation_function()); + params->time_major = sequence_rnn_params->time_major(); + params->asymmetric_quantize_inputs = + sequence_rnn_params->asymmetric_quantize_inputs(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN: { + auto params = + safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* bidi_sequence_rnn_params = + op->builtin_options_as_BidirectionalSequenceRNNOptions()) { + params->activation = ConvertActivation( + bidi_sequence_rnn_params->fused_activation_function()); + params->time_major = bidi_sequence_rnn_params->time_major(); + params->merge_outputs = bidi_sequence_rnn_params->merge_outputs(); + params->asymmetric_quantize_inputs = + bidi_sequence_rnn_params->asymmetric_quantize_inputs(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_RNN: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* rnn_params = op->builtin_options_as_RNNOptions()) { + params->activation = + ConvertActivation(rnn_params->fused_activation_function()); + params->asymmetric_quantize_inputs = + rnn_params->asymmetric_quantize_inputs(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_EMBEDDING_LOOKUP_SPARSE: { + auto params = + safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* embedding_params = + op->builtin_options_as_EmbeddingLookupSparseOptions()) { + params->combiner = parseCombinerType(embedding_params->combiner()); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + + case BuiltinOperator_HASHTABLE_LOOKUP: + // no-op. + return kTfLiteOk; + + case BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* schema_params = + op->builtin_options_as_LocalResponseNormalizationOptions()) { + params->radius = schema_params->radius(); + params->bias = schema_params->bias(); + params->alpha = schema_params->alpha(); + params->beta = schema_params->beta(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_LSTM: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* lstm_params = op->builtin_options_as_LSTMOptions()) { + params->activation = + ConvertActivation(lstm_params->fused_activation_function()); + params->cell_clip = lstm_params->cell_clip(); + params->proj_clip = lstm_params->proj_clip(); + switch (lstm_params->kernel_type()) { + case LSTMKernelType_FULL: + params->kernel_type = kTfLiteLSTMFullKernel; + break; + case LSTMKernelType_BASIC: + params->kernel_type = kTfLiteLSTMBasicKernel; + break; + default: + TF_LITE_REPORT_ERROR(error_reporter, + "Unhandled LSTM kernel type: %d", + lstm_params->kernel_type()); + return kTfLiteError; + } + params->asymmetric_quantize_inputs = + lstm_params->asymmetric_quantize_inputs(); + } else { + TF_LITE_REPORT_ERROR(error_reporter, + "No valid LSTM builtin options exist"); + return kTfLiteError; + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM: { + auto params = + safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* seq_lstm_params = + op->builtin_options_as_UnidirectionalSequenceLSTMOptions()) { + params->activation = + ConvertActivation(seq_lstm_params->fused_activation_function()); + params->cell_clip = seq_lstm_params->cell_clip(); + params->proj_clip = seq_lstm_params->proj_clip(); + params->time_major = seq_lstm_params->time_major(); + params->asymmetric_quantize_inputs = + seq_lstm_params->asymmetric_quantize_inputs(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM: { + auto params = + safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* bidi_lstm_params = + op->builtin_options_as_BidirectionalSequenceLSTMOptions()) { + params->activation = + ConvertActivation(bidi_lstm_params->fused_activation_function()); + params->cell_clip = bidi_lstm_params->cell_clip(); + params->proj_clip = bidi_lstm_params->proj_clip(); + params->merge_outputs = bidi_lstm_params->merge_outputs(); + params->time_major = bidi_lstm_params->time_major(); + params->asymmetric_quantize_inputs = + bidi_lstm_params->asymmetric_quantize_inputs(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_SKIP_GRAM: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* skip_gram_params = + op->builtin_options_as_SkipGramOptions()) { + params->ngram_size = skip_gram_params->ngram_size(); + params->max_skip_size = skip_gram_params->max_skip_size(); + params->include_all_ngrams = skip_gram_params->include_all_ngrams(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + + case BuiltinOperator_DEPTH_TO_SPACE: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* schema_params = + op->builtin_options_as_DepthToSpaceOptions()) { + params->block_size = schema_params->block_size(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_GATHER: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + params->axis = 0; + if (const auto* gather_params = op->builtin_options_as_GatherOptions()) { + params->axis = gather_params->axis(); + } + + *builtin_data = params.release(); + return kTfLiteOk; + } + + case BuiltinOperator_SQUEEZE: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* schema_params = op->builtin_options_as_SqueezeOptions()) { + const auto* squeeze_dims = schema_params->squeeze_dims(); + if (squeeze_dims != nullptr) { + TF_LITE_ENSURE_STATUS(FlatBufferIntVectorToArray( + sizeof(params->squeeze_dims), squeeze_dims, params->squeeze_dims, + error_reporter, "squeeze")); + params->num_squeeze_dims = squeeze_dims->size(); + } else { + params->num_squeeze_dims = 0; + } + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_SPARSE_TO_DENSE: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* sparse_to_dense_params = + op->builtin_options_as_SparseToDenseOptions()) { + params->validate_indices = sparse_to_dense_params->validate_indices(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_DELEGATE: { + // TODO(ycling): Revisit when supporting saving delegated models. + TF_LITE_REPORT_ERROR(error_reporter, + "DELEGATE op shouldn't exist in model."); + return kTfLiteError; + } + case BuiltinOperator_FAKE_QUANT: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* schema_params = + op->builtin_options_as_FakeQuantOptions()) { + params->min = schema_params->min(); + params->max = schema_params->max(); + params->num_bits = schema_params->num_bits(); + params->narrow_range = schema_params->narrow_range(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_ONE_HOT: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* schema_params = op->builtin_options_as_OneHotOptions()) { + params->axis = schema_params->axis(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_LEAKY_RELU: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* leaky_relu_params = + op->builtin_options_as_LeakyReluOptions()) { + params->alpha = leaky_relu_params->alpha(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_MIRROR_PAD: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + const auto* mirror_pad_params = op->builtin_options_as_MirrorPadOptions(); + if (mirror_pad_params != nullptr) { + params->mode = + mirror_pad_params->mode() == tflite::MirrorPadMode_REFLECT + ? TfLiteMirrorPaddingMode::kTfLiteMirrorPaddingReflect + : TfLiteMirrorPaddingMode::kTfLiteMirrorPaddingSymmetric; + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_UNIQUE: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + const auto* unique_params = op->builtin_options_as_UniqueOptions(); + if (unique_params != nullptr) { + params->index_out_type = + unique_params->idx_out_type() == tflite::TensorType_INT64 + ? TfLiteType::kTfLiteInt64 + : TfLiteType::kTfLiteInt32; + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_REVERSE_SEQUENCE: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* reverse_seq_params = + op->builtin_options_as_ReverseSequenceOptions()) { + params->seq_dim = reverse_seq_params->seq_dim(); + params->batch_dim = reverse_seq_params->batch_dim(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_IF: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* if_params = op->builtin_options_as_IfOptions()) { + params->then_subgraph_index = if_params->then_subgraph_index(); + params->else_subgraph_index = if_params->else_subgraph_index(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_WHILE: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* while_params = op->builtin_options_as_WhileOptions()) { + params->cond_subgraph_index = while_params->cond_subgraph_index(); + params->body_subgraph_index = while_params->body_subgraph_index(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_BATCH_MATMUL: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* bmm_params = + op->builtin_options_as_BatchMatMulOptions()) { + params->adj_x = bmm_params->adj_x(); + params->adj_y = bmm_params->adj_y(); + params->asymmetric_quantize_inputs = + bmm_params->asymmetric_quantize_inputs(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_CALL_ONCE: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* call_once_params = + op->builtin_options_as_CallOnceOptions()) { + params->init_subgraph_index = call_once_params->init_subgraph_index(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + case BuiltinOperator_CUMSUM: { + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* cumsum_params = op->builtin_options_as_CumsumOptions()) { + params->exclusive = cumsum_params->exclusive(); + params->reverse = cumsum_params->reverse(); + } + *builtin_data = params.release(); + return kTfLiteOk; + } + // Below are the ops with no builtin_data structure. + case BuiltinOperator_BATCH_TO_SPACE_ND: + // TODO(aselle): Implement call in BuiltinOptions, but nullptrs are + // ok for now, since there is no call implementation either. + case BuiltinOperator_CALL: + case BuiltinOperator_CONCAT_EMBEDDINGS: + case BuiltinOperator_COS: + case BuiltinOperator_CUSTOM: + case BuiltinOperator_ELU: + case BuiltinOperator_EMBEDDING_LOOKUP: + case BuiltinOperator_EQUAL: + case BuiltinOperator_EXPAND_DIMS: + case BuiltinOperator_LOG_SOFTMAX: + case BuiltinOperator_MATRIX_DIAG: + case BuiltinOperator_MATRIX_SET_DIAG: + case BuiltinOperator_RELU_N1_TO_1: + case BuiltinOperator_SELECT: + case BuiltinOperator_SELECT_V2: + case BuiltinOperator_SLICE: + case BuiltinOperator_SPACE_TO_BATCH_ND: + case BuiltinOperator_TILE: + case BuiltinOperator_TOPK_V2: + case BuiltinOperator_TRANSPOSE: + case BuiltinOperator_RANGE: + case BuiltinOperator_SQUARED_DIFFERENCE: + case BuiltinOperator_REVERSE_V2: + case BuiltinOperator_ADD_N: + case BuiltinOperator_GATHER_ND: + case BuiltinOperator_WHERE: + case BuiltinOperator_RANK: + case BuiltinOperator_NON_MAX_SUPPRESSION_V4: + case BuiltinOperator_NON_MAX_SUPPRESSION_V5: + case BuiltinOperator_SCATTER_ND: + case BuiltinOperator_DENSIFY: + case BuiltinOperator_SEGMENT_SUM: + case BuiltinOperator_BROADCAST_TO: + case BuiltinOperator_RFFT2D: + return kTfLiteOk; + case BuiltinOperator_PLACEHOLDER_FOR_GREATER_OP_CODES: + return kTfLiteError; + } + return kTfLiteError; +} // NOLINT[readability/fn_size] +#endif // !defined(TF_LITE_STATIC_MEMORY) +} // namespace + +TfLiteStatus ConvertTensorType(TensorType tensor_type, TfLiteType* type, + ErrorReporter* error_reporter) { + switch (tensor_type) { + case TensorType_FLOAT16: + *type = kTfLiteFloat16; + return kTfLiteOk; + case TensorType_FLOAT32: + *type = kTfLiteFloat32; + return kTfLiteOk; + case TensorType_FLOAT64: + *type = kTfLiteFloat64; + return kTfLiteOk; + case TensorType_INT16: + *type = kTfLiteInt16; + return kTfLiteOk; + case TensorType_INT32: + *type = kTfLiteInt32; + return kTfLiteOk; + case TensorType_UINT8: + *type = kTfLiteUInt8; + return kTfLiteOk; + case TensorType_INT8: + *type = kTfLiteInt8; + return kTfLiteOk; + case TensorType_INT64: + *type = kTfLiteInt64; + return kTfLiteOk; + case TensorType_UINT64: + *type = kTfLiteUInt64; + return kTfLiteOk; + case TensorType_STRING: + *type = kTfLiteString; + return kTfLiteOk; + case TensorType_BOOL: + *type = kTfLiteBool; + return kTfLiteOk; + case TensorType_COMPLEX64: + *type = kTfLiteComplex64; + return kTfLiteOk; + case TensorType_COMPLEX128: + *type = kTfLiteComplex128; + return kTfLiteOk; + default: + *type = kTfLiteNoType; + TF_LITE_REPORT_ERROR(error_reporter, + "Unsupported data type %d in tensor\n", tensor_type); + return kTfLiteError; + } +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseAbs(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseAdd(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const AddOptions* schema_params = op->builtin_options_as_AddOptions(); + + if (schema_params != nullptr) { + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + params->pot_scale_int16 = schema_params->pot_scale_int16(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseArgMax(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ArgMaxOptions* schema_params = op->builtin_options_as_ArgMaxOptions(); + + if (schema_params != nullptr) { + TF_LITE_ENSURE_STATUS(ConvertTensorType( + schema_params->output_type(), ¶ms->output_type, error_reporter)); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseArgMin(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ArgMinOptions* schema_params = op->builtin_options_as_ArgMinOptions(); + + if (schema_params != nullptr) { + TF_LITE_ENSURE_STATUS(ConvertTensorType( + schema_params->output_type(), ¶ms->output_type, error_reporter)); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseCast(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* schema_params = op->builtin_options_as_CastOptions()) { + TF_LITE_ENSURE_STATUS(ConvertTensorType( + schema_params->in_data_type(), ¶ms->in_data_type, error_reporter)); + TF_LITE_ENSURE_STATUS(ConvertTensorType(schema_params->out_data_type(), + ¶ms->out_data_type, + error_reporter)); + } + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseCeil(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseConcatenation(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ConcatenationOptions* schema_params = + op->builtin_options_as_ConcatenationOptions(); + + if (schema_params != nullptr) { + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + params->axis = schema_params->axis(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseConv2D(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const Conv2DOptions* schema_params = op->builtin_options_as_Conv2DOptions(); + + if (schema_params != nullptr) { + params->padding = ConvertPadding(schema_params->padding()); + params->stride_width = schema_params->stride_w(); + params->stride_height = schema_params->stride_h(); + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + + params->dilation_width_factor = schema_params->dilation_w_factor(); + params->dilation_height_factor = schema_params->dilation_h_factor(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseCos(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseDepthwiseConv2D(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const DepthwiseConv2DOptions* schema_params = + op->builtin_options_as_DepthwiseConv2DOptions(); + + if (schema_params != nullptr) { + params->padding = ConvertPadding(schema_params->padding()); + params->stride_width = schema_params->stride_w(); + params->stride_height = schema_params->stride_h(); + params->depth_multiplier = schema_params->depth_multiplier(); + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + + params->dilation_width_factor = schema_params->dilation_w_factor(); + params->dilation_height_factor = schema_params->dilation_h_factor(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseDequantize(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseDiv(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + auto params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + if (const auto* schema_params = op->builtin_options_as_DivOptions()) { + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + } + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseEqual(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseExp(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseFill(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseFloor(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseFloorDiv(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseFloorMod(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseFullyConnected(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const FullyConnectedOptions* schema_params = + op->builtin_options_as_FullyConnectedOptions(); + + if (schema_params != nullptr) { + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + params->keep_num_dims = schema_params->keep_num_dims(); + params->asymmetric_quantize_inputs = + schema_params->asymmetric_quantize_inputs(); + + switch (schema_params->weights_format()) { + case FullyConnectedOptionsWeightsFormat_DEFAULT: + params->weights_format = kTfLiteFullyConnectedWeightsFormatDefault; + break; + case FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8: + params->weights_format = + kTfLiteFullyConnectedWeightsFormatShuffled4x16Int8; + break; + default: + TF_LITE_REPORT_ERROR(error_reporter, + "Unhandled fully-connected weights format."); + return kTfLiteError; + } + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseGreater(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseGreaterEqual(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseHardSwish(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseL2Normalization(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const L2NormOptions* schema_params = op->builtin_options_as_L2NormOptions(); + + if (schema_params != nullptr) { + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseLess(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseLessEqual(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseLog(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseLogicalAnd(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseLogicalNot(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseLogicalOr(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseLogistic(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseMaximum(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseMinimum(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseMul(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const MulOptions* schema_params = op->builtin_options_as_MulOptions(); + + if (schema_params != nullptr) { + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseNeg(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseNotEqual(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +TfLiteStatus ParsePack(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const PackOptions* schema_params = op->builtin_options_as_PackOptions(); + + if (schema_params != nullptr) { + params->values_count = schema_params->values_count(); + params->axis = schema_params->axis(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParsePad(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParsePadV2(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParsePool(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const Pool2DOptions* schema_params = op->builtin_options_as_Pool2DOptions(); + + if (schema_params != nullptr) { + params->padding = ConvertPadding(schema_params->padding()); + params->stride_width = schema_params->stride_w(); + params->stride_height = schema_params->stride_h(); + params->filter_width = schema_params->filter_width(); + params->filter_height = schema_params->filter_height(); + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParsePow(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParsePrelu(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseQuantize(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseReducer(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ReducerOptions* schema_params = op->builtin_options_as_ReducerOptions(); + + if (schema_params != nullptr) { + params->keep_dims = schema_params->keep_dims(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseRelu(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseRelu6(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseReshape(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ReshapeOptions* schema_params = op->builtin_options_as_ReshapeOptions(); + + if (schema_params != nullptr) { + const flatbuffers::Vector* new_shape = schema_params->new_shape(); + if (new_shape != nullptr) { + TF_LITE_ENSURE_STATUS( + FlatBufferIntVectorToArray(sizeof(params->shape), new_shape, + params->shape, error_reporter, "reshape")); + params->num_dimensions = new_shape->size(); + } else { + // TODO(b/157480169) TODO(b/147203660): We should either return + // kTfLiteError or fill in some reasonable defaults in the params struct. + // We are not doing so until we better undertand the ramifications of + // changing the legacy behavior. + } + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseResizeBilinear(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ResizeBilinearOptions* schema_params = + op->builtin_options_as_ResizeBilinearOptions(); + + if (schema_params != nullptr) { + params->align_corners = schema_params->align_corners(); + params->half_pixel_centers = schema_params->half_pixel_centers(); + } else { + params->align_corners = false; + params->half_pixel_centers = false; + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseResizeNearestNeighbor(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ResizeNearestNeighborOptions* schema_params = + op->builtin_options_as_ResizeNearestNeighborOptions(); + + if (schema_params != nullptr) { + params->align_corners = schema_params->align_corners(); + params->half_pixel_centers = schema_params->half_pixel_centers(); + } else { + params->align_corners = false; + params->half_pixel_centers = false; + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseRound(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseRsqrt(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseShape(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const ShapeOptions* schema_params = op->builtin_options_as_ShapeOptions(); + + if (schema_params != nullptr) { + TF_LITE_ENSURE_STATUS(ConvertTensorType(schema_params->out_type(), + ¶ms->out_type, error_reporter)); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseSin(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseSoftmax(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const SoftmaxOptions* schema_params = op->builtin_options_as_SoftmaxOptions(); + + if (schema_params != nullptr) { + params->beta = schema_params->beta(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseSpaceToDepth(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const auto* schema_params = op->builtin_options_as_SpaceToDepthOptions(); + if (schema_params != nullptr) { + params->block_size = schema_params->block_size(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseSplit(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const SplitOptions* schema_params = op->builtin_options_as_SplitOptions(); + + if (schema_params != nullptr) { + params->num_splits = schema_params->num_splits(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseSplitV(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + SafeBuiltinDataAllocator safe_allocator(allocator); + + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const SplitVOptions* schema_params = op->builtin_options_as_SplitVOptions(); + + if (schema_params != nullptr) { + params->num_splits = schema_params->num_splits(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseSqrt(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseSquare(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseStridedSlice(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const StridedSliceOptions* schema_params = + op->builtin_options_as_StridedSliceOptions(); + + if (schema_params != nullptr) { + params->begin_mask = schema_params->begin_mask(); + params->end_mask = schema_params->end_mask(); + params->ellipsis_mask = schema_params->ellipsis_mask(); + params->new_axis_mask = schema_params->new_axis_mask(); + params->shrink_axis_mask = schema_params->shrink_axis_mask(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseSub(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const SubOptions* schema_params = op->builtin_options_as_SubOptions(); + + if (schema_params != nullptr) { + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + params->pot_scale_int16 = schema_params->pot_scale_int16(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseSvdf(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const SVDFOptions* schema_params = op->builtin_options_as_SVDFOptions(); + if (schema_params != nullptr) { + params->rank = schema_params->rank(); + params->activation = + ConvertActivation(schema_params->fused_activation_function()); + params->asymmetric_quantize_inputs = + schema_params->asymmetric_quantize_inputs(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseTanh(const Operator*, ErrorReporter*, BuiltinDataAllocator*, + void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseTransposeConv(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + const TransposeConvOptions* transpose_conv_params = + op->builtin_options_as_TransposeConvOptions(); + if (transpose_conv_params != nullptr) { + params->padding = ConvertPadding(transpose_conv_params->padding()); + params->stride_width = transpose_conv_params->stride_w(); + params->stride_height = transpose_conv_params->stride_h(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + *builtin_data = params.release(); + return kTfLiteOk; +} + +TfLiteStatus ParseUnpack(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { + CheckParsePointerParams(op, error_reporter, allocator, builtin_data); + + SafeBuiltinDataAllocator safe_allocator(allocator); + std::unique_ptr + params = safe_allocator.Allocate(); + TF_LITE_ENSURE(error_reporter, params != nullptr); + + const UnpackOptions* schema_params = op->builtin_options_as_UnpackOptions(); + + if (schema_params != nullptr) { + params->num = schema_params->num(); + params->axis = schema_params->axis(); + } else { + // TODO(b/157480169): We should either return kTfLiteError or fill in some + // reasonable defaults in the params struct. We are not doing so until we + // better undertand the ramifications of changing the legacy behavior. + } + + *builtin_data = params.release(); + return kTfLiteOk; +} + +// We have this parse function instead of directly returning kTfLiteOk from the +// switch-case in ParseOpData because this function is used as part of the +// selective registration for the OpResolver implementation in micro. +TfLiteStatus ParseZerosLike(const Operator*, ErrorReporter*, + BuiltinDataAllocator*, void**) { + return kTfLiteOk; +} + +TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data) { +// TODO(b/145762662): It would be preferable to have the build graph for TF Lite +// Micro not have the ParseOpData function at all. This would require splitting +// the current file into two separate files, one of which defines the +// ParseOpData function and the other that defines the operator specific parse +// functions (e.g. ParseAdd). +// +// Such a split was attempted but was not worth the effort at the time because +// of the following reasons: +// * We could either duplicate the functions and the SafeBuiltinDataAllocator +// class in the anonymous namespace of this file, or attempt to make a common +// library with these helper functions and class. +// * Making a common library with a separate build target was not feasible as +// it introduced circular dependencies due to the ErrorReporter and a common +// .cc and .h within the same api build target the also cause circular +// dependencies due to the BuiltinDataAllocator class. +// * If all the builtin operators were to have their own parse functions, or we +// were ok with some amount of code duplication, then this split of the .cc +// files would be a lot more feasible. +#ifdef TF_LITE_STATIC_MEMORY + TF_LITE_REPORT_ERROR( + error_reporter, + "ParseOpData is unsupported on TfLiteMicro, please use the operator " + "specific parse functions (e.g. ParseAdd etc.).\n"); + return kTfLiteError; +#else + return ParseOpDataTfLite(op, op_type, error_reporter, allocator, + builtin_data); +#endif +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h b/src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h new file mode 100644 index 0000000..ea6c66b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h @@ -0,0 +1,304 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_CORE_API_FLATBUFFER_CONVERSIONS_H_ +#define TENSORFLOW_LITE_CORE_API_FLATBUFFER_CONVERSIONS_H_ + +// These functions transform codes and data structures that are defined in the +// flatbuffer serialization format into in-memory values that are used by the +// runtime API and interpreter. + +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +// Interface class for builtin data allocations. +class BuiltinDataAllocator { + public: + virtual void* Allocate(size_t size, size_t alignment_hint) = 0; + virtual void Deallocate(void* data) = 0; + + // Allocate a structure, but make sure it is a POD structure that doesn't + // require constructors to run. The reason we do this, is that Interpreter's C + // extension part will take ownership so destructors will not be run during + // deallocation. + template + T* AllocatePOD() { + // TODO(b/154346074): Change this to is_trivially_destructible when all + // platform targets support that properly. + static_assert(std::is_pod::value, "Builtin data structure must be POD."); + void* allocated_memory = this->Allocate(sizeof(T), alignof(T)); + return new (allocated_memory) T(); + } + + virtual ~BuiltinDataAllocator() {} +}; + +// Parse the appropriate data out of the op. +// +// This handles builtin data explicitly as there are flatbuffer schemas. +// If it returns kTfLiteOk, it passes the data out with `builtin_data`. The +// calling function has to pass in an allocator object, and this allocator +// will be called to reserve space for the output data. If the calling +// function's allocator reserves memory on the heap, then it's the calling +// function's responsibility to free it. +// If it returns kTfLiteError, `builtin_data` will be `nullptr`. +TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +// Converts the tensor data type used in the flat buffer to the representation +// used by the runtime. +TfLiteStatus ConvertTensorType(TensorType tensor_type, TfLiteType* type, + ErrorReporter* error_reporter); + +TfLiteStatus ParseAbs(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseAdd(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseArgMax(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseArgMin(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseCeil(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseCast(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseConcatenation(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseConv2D(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseCos(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseDepthwiseConv2D(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseDequantize(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseDiv(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseEqual(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseExp(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseFill(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseFloor(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseFloorDiv(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseFloorMod(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseFullyConnected(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseGreater(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseGreaterEqual(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseHardSwish(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseL2Normalization(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseLess(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseLessEqual(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseLog(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseLogicalAnd(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseLogicalNot(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseLogicalOr(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseLogistic(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseMaximum(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseMinimum(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseMul(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseNeg(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseNotEqual(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParsePack(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParsePad(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParsePadV2(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParsePool(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParsePow(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParsePrelu(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseQuantize(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseReducer(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseRelu(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseRelu6(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseReshape(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseResizeBilinear(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseResizeNearestNeighbor(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseRound(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseRsqrt(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseShape(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseSin(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseSoftmax(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseSpaceToDepth(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseSplit(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseSplitV(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseSqrt(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseSquare(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseStridedSlice(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseSub(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseSvdf(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseTanh(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseTransposeConv(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +TfLiteStatus ParseUnpack(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, void** builtin_data); + +TfLiteStatus ParseZerosLike(const Operator* op, ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_CORE_API_FLATBUFFER_CONVERSIONS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.cpp b/src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.cpp new file mode 100644 index 0000000..d9e756f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.cpp @@ -0,0 +1,72 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/core/api/op_resolver.h" + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_utils.h" + +namespace tflite { + +TfLiteStatus GetRegistrationFromOpCode( + const OperatorCode* opcode, const OpResolver& op_resolver, + ErrorReporter* error_reporter, const TfLiteRegistration** registration) { + TfLiteStatus status = kTfLiteOk; + *registration = nullptr; + auto builtin_code = GetBuiltinCode(opcode); + int version = opcode->version(); + + if (builtin_code > BuiltinOperator_MAX || + builtin_code < BuiltinOperator_MIN) { + TF_LITE_REPORT_ERROR( + error_reporter, + "Op builtin_code out of range: %d. Are you using old TFLite binary " + "with newer model?", + builtin_code); + status = kTfLiteError; + } else if (builtin_code != BuiltinOperator_CUSTOM) { + *registration = op_resolver.FindOp(builtin_code, version); + if (*registration == nullptr) { + TF_LITE_REPORT_ERROR( + error_reporter, + "Didn't find op for builtin opcode '%s' version '%d'. " + "An older version of this builtin might be supported. " + "Are you using an old TFLite binary with a newer model?\n", + EnumNameBuiltinOperator(builtin_code), version); + status = kTfLiteError; + } + } else if (!opcode->custom_code()) { + TF_LITE_REPORT_ERROR( + error_reporter, + "Operator with CUSTOM builtin_code has no custom_code.\n"); + status = kTfLiteError; + } else { + const char* name = opcode->custom_code()->c_str(); + *registration = op_resolver.FindOp(name, version); + if (*registration == nullptr) { + // Do not report error for unresolved custom op, we do the final check + // while preparing ops. + status = kTfLiteError; + } + } + return status; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.h b/src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.h new file mode 100644 index 0000000..286a788 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/core/api/op_resolver.h @@ -0,0 +1,63 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_ +#define TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_ + +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +/// Abstract interface that returns TfLiteRegistrations given op codes or custom +/// op names. This is the mechanism that ops being referenced in the flatbuffer +/// model are mapped to executable function pointers (TfLiteRegistrations). +class OpResolver { + public: + /// Finds the op registration for a builtin operator by enum code. + virtual const TfLiteRegistration* FindOp(tflite::BuiltinOperator op, + int version) const = 0; + /// Finds the op registration of a custom operator by op name. + virtual const TfLiteRegistration* FindOp(const char* op, + int version) const = 0; + + // Returns optional delegates for resolving and handling ops in the flatbuffer + // model. This may be used in addition to the standard TfLiteRegistration + // lookup for graph resolution. + using TfLiteDelegatePtrVector = + std::vector>; + virtual TfLiteDelegatePtrVector GetDelegates(int num_threads) const { + return TfLiteDelegatePtrVector(); + } + + virtual ~OpResolver() {} +}; + +// Handles the logic for converting between an OperatorCode structure extracted +// from a flatbuffer and information about a registered operator +// implementation. +TfLiteStatus GetRegistrationFromOpCode(const OperatorCode* opcode, + const OpResolver& op_resolver, + ErrorReporter* error_reporter, + const TfLiteRegistration** registration); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/core/api/profiler.h b/src/tensorflow_arm/tensorflow/lite/core/api/profiler.h new file mode 100644 index 0000000..5e33ede --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/core/api/profiler.h @@ -0,0 +1,197 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_CORE_API_PROFILER_H_ +#define TENSORFLOW_LITE_CORE_API_PROFILER_H_ + +#include + +namespace tflite { + +// A simple utility for enabling profiled event tracing in TensorFlow Lite. +class Profiler { + public: + // As certain Profiler instance might be only interested in certain event + // types, we define each event type value to allow a Profiler to use + // bitmasking bitwise operations to determine whether an event should be + // recorded or not. + enum class EventType { + // Default event type, the metadata field has no special significance. + DEFAULT = 1, + + // The event is an operator invocation and the event_metadata field is the + // index of operator node. + OPERATOR_INVOKE_EVENT = 2, + + // The event is an invocation for an internal operator of a TFLite delegate. + // The event_metadata field is the index of operator node that's specific to + // the delegate. + DELEGATE_OPERATOR_INVOKE_EVENT = 4, + + // The event is a recording of runtime instrumentation such as the overall + // TFLite runtime status, the TFLite delegate status (if a delegate + // is applied), and the overall model inference latency etc. + // Note, the delegate status and overall status are stored as separate + // event_metadata fields. In particular, the delegate status is encoded + // as DelegateStatus::full_status(). + GENERAL_RUNTIME_INSTRUMENTATION_EVENT = 8, + }; + + virtual ~Profiler() {} + + // Signals the beginning of an event and returns a handle to the profile + // event. The `event_metadata1` and `event_metadata2` have different + // interpretations based on the actual Profiler instance and the `event_type`. + // For example, as for the 'SubgraphAwareProfiler' defined in + // lite/core/subgraph.h, when the event_type is OPERATOR_INVOKE_EVENT, + // `event_metadata1` represents the index of a TFLite node, and + // `event_metadata2` represents the index of the subgraph that this event + // comes from. + virtual uint32_t BeginEvent(const char* tag, EventType event_type, + int64_t event_metadata1, + int64_t event_metadata2) = 0; + // Similar w/ the above, but `event_metadata2` defaults to 0. + uint32_t BeginEvent(const char* tag, EventType event_type, + int64_t event_metadata) { + return BeginEvent(tag, event_type, event_metadata, /*event_metadata2*/ 0); + } + + // Signals an end to the specified profile event with 'event_metadata's, This + // is useful when 'event_metadata's are not available when the event begins + // or when one wants to overwrite the 'event_metadata's set at the beginning. + virtual void EndEvent(uint32_t event_handle, int64_t event_metadata1, + int64_t event_metadata2) {} + // Signals an end to the specified profile event. + virtual void EndEvent(uint32_t event_handle) = 0; + + // Appends an event of type 'event_type' with 'tag' and 'event_metadata' + // which started at 'start' and ended at 'end' + // Note: + // In cases were ProfileSimmarizer and tensorflow::StatsCalculator are used + // they assume the value is in "usec", if in any case subclasses + // didn't put usec, then the values are not meaningful. + // TODO karimnosseir: Revisit and make the function more clear. + void AddEvent(const char* tag, EventType event_type, uint64_t start, + uint64_t end, int64_t event_metadata) { + AddEvent(tag, event_type, start, end, event_metadata, + /*event_metadata2*/ 0); + } + + virtual void AddEvent(const char* tag, EventType event_type, uint64_t start, + uint64_t end, int64_t event_metadata1, + int64_t event_metadata2) {} + + protected: + friend class ScopedProfile; +}; + +// Adds a profile event to `profiler` that begins with the construction +// of the object and ends when the object goes out of scope. +// The lifetime of tag should be at least the lifetime of `profiler`. +// `profiler` may be null, in which case nothing is profiled. +class ScopedProfile { + public: + ScopedProfile(Profiler* profiler, const char* tag, + Profiler::EventType event_type = Profiler::EventType::DEFAULT, + int64_t event_metadata = 0) + : profiler_(profiler), event_handle_(0) { + if (profiler) { + event_handle_ = profiler_->BeginEvent(tag, event_type, event_metadata); + } + } + + ~ScopedProfile() { + if (profiler_) { + profiler_->EndEvent(event_handle_); + } + } + + protected: + Profiler* profiler_; + uint32_t event_handle_; +}; + +class ScopedOperatorProfile : public ScopedProfile { + public: + ScopedOperatorProfile(Profiler* profiler, const char* tag, int node_index) + : ScopedProfile(profiler, tag, Profiler::EventType::OPERATOR_INVOKE_EVENT, + static_cast(node_index)) {} +}; + +class ScopedDelegateOperatorProfile : public ScopedProfile { + public: + ScopedDelegateOperatorProfile(Profiler* profiler, const char* tag, + int node_index) + : ScopedProfile(profiler, tag, + Profiler::EventType::DELEGATE_OPERATOR_INVOKE_EVENT, + static_cast(node_index)) {} +}; + +class ScopedRuntimeInstrumentationProfile : public ScopedProfile { + public: + ScopedRuntimeInstrumentationProfile(Profiler* profiler, const char* tag) + : ScopedProfile( + profiler, tag, + Profiler::EventType::GENERAL_RUNTIME_INSTRUMENTATION_EVENT, -1) {} + + void set_runtime_status(int64_t delegate_status, int64_t interpreter_status) { + if (profiler_) { + delegate_status_ = delegate_status; + interpreter_status_ = interpreter_status; + } + } + + ~ScopedRuntimeInstrumentationProfile() { + if (profiler_) { + profiler_->EndEvent(event_handle_, delegate_status_, interpreter_status_); + } + } + + private: + int64_t delegate_status_; + int64_t interpreter_status_; +}; + +} // namespace tflite + +#define TFLITE_VARNAME_UNIQ_IMPL(name, ctr) name##ctr +#define TFLITE_VARNAME_UNIQ(name, ctr) TFLITE_VARNAME_UNIQ_IMPL(name, ctr) + +#define TFLITE_SCOPED_TAGGED_DEFAULT_PROFILE(profiler, tag) \ + tflite::ScopedProfile TFLITE_VARNAME_UNIQ(_profile_, __COUNTER__)( \ + (profiler), (tag)) + +#define TFLITE_SCOPED_TAGGED_OPERATOR_PROFILE(profiler, tag, node_index) \ + tflite::ScopedOperatorProfile TFLITE_VARNAME_UNIQ(_profile_, __COUNTER__)( \ + (profiler), (tag), (node_index)) + +#define TFLITE_SCOPED_DELEGATE_OPERATOR_PROFILE(profiler, tag, node_index) \ + tflite::ScopedDelegateOperatorProfile TFLITE_VARNAME_UNIQ( \ + _profile_, __COUNTER__)((profiler), (tag), (node_index)) + +#define TFLITE_ADD_RUNTIME_INSTRUMENTATION_EVENT( \ + profiler, tag, event_metadata1, event_metadata2) \ + do { \ + if (profiler) { \ + const auto handle = profiler->BeginEvent( \ + tag, Profiler::EventType::GENERAL_RUNTIME_INSTRUMENTATION_EVENT, \ + event_metadata1, event_metadata2); \ + profiler->EndEvent(handle); \ + } \ + } while (false); + +#endif // TENSORFLOW_LITE_CORE_API_PROFILER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/core/api/tensor_utils.cpp b/src/tensorflow_arm/tensorflow/lite/core/api/tensor_utils.cpp new file mode 100644 index 0000000..3e6447b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/core/api/tensor_utils.cpp @@ -0,0 +1,53 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/core/api/tensor_utils.h" + +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +TfLiteStatus ResetVariableTensor(TfLiteTensor* tensor) { + if (!tensor->is_variable) { + return kTfLiteOk; + } + // TODO(b/115961645): Implement - If a variable tensor has a buffer, reset it + // to the value of the buffer. + int value = 0; + if (tensor->type == kTfLiteInt8) { + value = tensor->params.zero_point; + } + // TODO(b/139446230): Provide a platform header to better handle these + // specific scenarios. +#if __ANDROID__ || defined(__x86_64__) || defined(__i386__) || \ + defined(__i386) || defined(__x86__) || defined(__X86__) || \ + defined(_X86_) || defined(_M_IX86) || defined(_M_X64) + memset(tensor->data.raw, value, tensor->bytes); +#else + char* raw_ptr = tensor->data.raw; + for (size_t i = 0; i < tensor->bytes; ++i) { + *raw_ptr = value; + raw_ptr++; + } +#endif + return kTfLiteOk; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/tensor_utils.h b/src/tensorflow_arm/tensorflow/lite/core/api/tensor_utils.h similarity index 89% rename from src/tensorflow/lite/core/api/tensor_utils.h rename to src/tensorflow_arm/tensorflow/lite/core/api/tensor_utils.h index 9f1cf94..3d8b1d7 100644 --- a/src/tensorflow/lite/core/api/tensor_utils.h +++ b/src/tensorflow_arm/tensorflow/lite/core/api/tensor_utils.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -16,7 +17,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_ #define TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_ -#include "tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" namespace tflite { @@ -26,3 +27,5 @@ TfLiteStatus ResetVariableTensor(TfLiteTensor* tensor); } // namespace tflite #endif // TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h new file mode 100644 index 0000000..c36b093 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h @@ -0,0 +1,105 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_ + +#ifdef __cplusplus +#include + +extern "C" { +#endif + +static inline int CountLeadingZeros32Slow(uint64_t n) { + int zeroes = 28; + if (n >> 16) zeroes -= 16, n >>= 16; + if (n >> 8) zeroes -= 8, n >>= 8; + if (n >> 4) zeroes -= 4, n >>= 4; + return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[n] + zeroes; +} + +static inline int CountLeadingZeros32(uint32_t n) { +#if defined(_MSC_VER) + unsigned long result = 0; // NOLINT(runtime/int) + if (_BitScanReverse(&result, n)) { + return 31 - result; + } + return 32; +#elif defined(__GNUC__) + + // Handle 0 as a special case because __builtin_clz(0) is undefined. + if (n == 0) { + return 32; + } + return __builtin_clz(n); +#else + return CountLeadingZeros32Slow(n); +#endif +} + +static inline int MostSignificantBit32(uint32_t n) { + return 32 - CountLeadingZeros32(n); +} + +static inline int CountLeadingZeros64Slow(uint64_t n) { + int zeroes = 60; + if (n >> 32) zeroes -= 32, n >>= 32; + if (n >> 16) zeroes -= 16, n >>= 16; + if (n >> 8) zeroes -= 8, n >>= 8; + if (n >> 4) zeroes -= 4, n >>= 4; + return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[n] + zeroes; +} + +static inline int CountLeadingZeros64(uint64_t n) { +#if defined(_MSC_VER) && defined(_M_X64) + // MSVC does not have __builtin_clzll. Use _BitScanReverse64. + unsigned long result = 0; // NOLINT(runtime/int) + if (_BitScanReverse64(&result, n)) { + return 63 - result; + } + return 64; +#elif defined(_MSC_VER) + // MSVC does not have __builtin_clzll. Compose two calls to _BitScanReverse + unsigned long result = 0; // NOLINT(runtime/int) + if ((n >> 32) && _BitScanReverse(&result, n >> 32)) { + return 31 - result; + } + if (_BitScanReverse(&result, n)) { + return 63 - result; + } + return 64; +#elif defined(__GNUC__) + + // Handle 0 as a special case because __builtin_clzll(0) is undefined. + if (n == 0) { + return 64; + } + return __builtin_clzll(n); +#else + return CountLeadingZeros64Slow(n); +#endif +} + +static inline int MostSignificantBit64(uint64_t n) { + return 64 - CountLeadingZeros64(n); +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.cpp b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.cpp new file mode 100644 index 0000000..853290b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.cpp @@ -0,0 +1,56 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h" + +#include + +#define FIXED_POINT 16 +#include "tensorflow_arm/third_party/kissfft/kiss_fft.h" +#include "tensorflow_arm/third_party/kissfft/tools/kiss_fftr.h" + +void FftCompute(struct FftState* state, const int16_t* input, + int input_scale_shift) { + const size_t input_size = state->input_size; + const size_t fft_size = state->fft_size; + + int16_t* fft_input = state->input; + // First, scale the input by the given shift. + size_t i; + for (i = 0; i < input_size; ++i) { + fft_input[i] = static_cast(static_cast(input[i]) + << input_scale_shift); + } + // Zero out whatever else remains in the top part of the input. + for (; i < fft_size; ++i) { + fft_input[i] = 0; + } + + // Apply the FFT. + kiss_fftr(reinterpret_cast(state->scratch), + state->input, + reinterpret_cast(state->output)); +} + +void FftInit(struct FftState* state) { + // All the initialization is done in FftPopulateState() +} + +void FftReset(struct FftState* state) { + memset(state->input, 0, state->fft_size * sizeof(*state->input)); + memset(state->output, 0, (state->fft_size / 2 + 1) * sizeof(*state->output)); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h new file mode 100644 index 0000000..01c945f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h @@ -0,0 +1,53 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_ + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct complex_int16_t { + int16_t real; + int16_t imag; +}; + +struct FftState { + int16_t* input; + struct complex_int16_t* output; + size_t fft_size; + size_t input_size; + void* scratch; + size_t scratch_size; +}; + +void FftCompute(struct FftState* state, const int16_t* input, + int input_scale_shift); + +void FftInit(struct FftState* state); + +void FftReset(struct FftState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp new file mode 100644 index 0000000..b0b42c6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp @@ -0,0 +1,75 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.h" + +#include + +#define FIXED_POINT 16 +#include "tensorflow_arm/third_party/kissfft/kiss_fft.h" +#include "tensorflow_arm/third_party/kissfft/tools/kiss_fftr.h" + +int FftPopulateState(struct FftState* state, size_t input_size) { + state->input_size = input_size; + state->fft_size = 1; + while (state->fft_size < state->input_size) { + state->fft_size <<= 1; + } + + state->input = reinterpret_cast( + malloc(state->fft_size * sizeof(*state->input))); + if (state->input == nullptr) { + fprintf(stderr, "Failed to alloc fft input buffer\n"); + return 0; + } + + state->output = reinterpret_cast( + malloc((state->fft_size / 2 + 1) * sizeof(*state->output) * 2)); + if (state->output == nullptr) { + fprintf(stderr, "Failed to alloc fft output buffer\n"); + return 0; + } + + // Ask kissfft how much memory it wants. + size_t scratch_size = 0; + kiss_fftr_cfg kfft_cfg = kiss_fftr_alloc( + state->fft_size, 0, nullptr, &scratch_size); + if (kfft_cfg != nullptr) { + fprintf(stderr, "Kiss memory sizing failed.\n"); + return 0; + } + state->scratch = malloc(scratch_size); + if (state->scratch == nullptr) { + fprintf(stderr, "Failed to alloc fft scratch buffer\n"); + return 0; + } + state->scratch_size = scratch_size; + // Let kissfft configure the scratch space we just allocated + kfft_cfg = kiss_fftr_alloc(state->fft_size, 0, + state->scratch, &scratch_size); + if (kfft_cfg != state->scratch) { + fprintf(stderr, "Kiss memory preallocation strategy failed.\n"); + return 0; + } + return 1; +} + +void FftFreeStateContents(struct FftState* state) { + free(state->input); + free(state->output); + free(state->scratch); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.h new file mode 100644 index 0000000..18f6629 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.h @@ -0,0 +1,37 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_ + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h" + +#ifdef __cplusplus +extern "C" { +#endif + +// Prepares and FFT for the given input size. +int FftPopulateState(struct FftState* state, size_t input_size); + +// Frees any allocated buffers. +void FftFreeStateContents(struct FftState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.c new file mode 100644 index 0000000..abc822f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.c @@ -0,0 +1,137 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.h" + +#include + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +void FilterbankConvertFftComplexToEnergy(struct FilterbankState* state, + struct complex_int16_t* fft_output, + int32_t* energy) { + const int end_index = state->end_index; + int i; + energy += state->start_index; + fft_output += state->start_index; + for (i = state->start_index; i < end_index; ++i) { + const int32_t real = fft_output->real; + const int32_t imag = fft_output->imag; + fft_output++; + const uint32_t mag_squared = (real * real) + (imag * imag); + *energy++ = mag_squared; + } +} + +void FilterbankAccumulateChannels(struct FilterbankState* state, + const int32_t* energy) { + uint64_t* work = state->work; + uint64_t weight_accumulator = 0; + uint64_t unweight_accumulator = 0; + + const int16_t* channel_frequency_starts = state->channel_frequency_starts; + const int16_t* channel_weight_starts = state->channel_weight_starts; + const int16_t* channel_widths = state->channel_widths; + + int num_channels_plus_1 = state->num_channels + 1; + int i; + for (i = 0; i < num_channels_plus_1; ++i) { + const int32_t* magnitudes = energy + *channel_frequency_starts++; + const int16_t* weights = state->weights + *channel_weight_starts; + const int16_t* unweights = state->unweights + *channel_weight_starts++; + const int width = *channel_widths++; + int j; + for (j = 0; j < width; ++j) { + weight_accumulator += *weights++ * ((uint64_t)*magnitudes); + unweight_accumulator += *unweights++ * ((uint64_t)*magnitudes); + ++magnitudes; + } + *work++ = weight_accumulator; + weight_accumulator = unweight_accumulator; + unweight_accumulator = 0; + } +} + +static uint16_t Sqrt32(uint32_t num) { + if (num == 0) { + return 0; + } + uint32_t res = 0; + int max_bit_number = 32 - MostSignificantBit32(num); + max_bit_number |= 1; + uint32_t bit = 1U << (31 - max_bit_number); + int iterations = (31 - max_bit_number) / 2 + 1; + while (iterations--) { + if (num >= res + bit) { + num -= res + bit; + res = (res >> 1U) + bit; + } else { + res >>= 1U; + } + bit >>= 2U; + } + // Do rounding - if we have the bits. + if (num > res && res != 0xFFFF) { + ++res; + } + return res; +} + +static uint32_t Sqrt64(uint64_t num) { + // Take a shortcut and just use 32 bit operations if the upper word is all + // clear. This will cause a slight off by one issue for numbers close to 2^32, + // but it probably isn't going to matter (and gives us a big performance win). + if ((num >> 32) == 0) { + return Sqrt32((uint32_t)num); + } + uint64_t res = 0; + int max_bit_number = 64 - MostSignificantBit64(num); + max_bit_number |= 1; + uint64_t bit = 1ULL << (63 - max_bit_number); + int iterations = (63 - max_bit_number) / 2 + 1; + while (iterations--) { + if (num >= res + bit) { + num -= res + bit; + res = (res >> 1U) + bit; + } else { + res >>= 1U; + } + bit >>= 2U; + } + // Do rounding - if we have the bits. + if (num > res && res != 0xFFFFFFFFLL) { + ++res; + } + return res; +} + +uint32_t* FilterbankSqrt(struct FilterbankState* state, int scale_down_shift) { + const int num_channels = state->num_channels; + const uint64_t* work = state->work + 1; + // Reuse the work buffer since we're fine clobbering it at this point to hold + // the output. + uint32_t* output = (uint32_t*)state->work; + int i; + for (i = 0; i < num_channels; ++i) { + *output++ = Sqrt64(*work++) >> scale_down_shift; + } + return (uint32_t*)state->work; +} + +void FilterbankReset(struct FilterbankState* state) { + memset(state->work, 0, (state->num_channels + 1) * sizeof(*state->work)); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.h new file mode 100644 index 0000000..bdbbafc --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.h @@ -0,0 +1,66 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h" + +#define kFilterbankBits 12 + +#ifdef __cplusplus +extern "C" { +#endif + +struct FilterbankState { + int num_channels; + int start_index; + int end_index; + int16_t* channel_frequency_starts; + int16_t* channel_weight_starts; + int16_t* channel_widths; + int16_t* weights; + int16_t* unweights; + uint64_t* work; +}; + +// Converts the relevant complex values of an FFT output into energy (the +// square magnitude). +void FilterbankConvertFftComplexToEnergy(struct FilterbankState* state, + struct complex_int16_t* fft_output, + int32_t* energy); + +// Computes the mel-scale filterbank on the given energy array. Output is cached +// internally - to fetch it, you need to call FilterbankSqrt. +void FilterbankAccumulateChannels(struct FilterbankState* state, + const int32_t* energy); + +// Applies an integer square root to the 64 bit intermediate values of the +// filterbank, and returns a pointer to them. Memory will be invalidated the +// next time FilterbankAccumulateChannels is called. +uint32_t* FilterbankSqrt(struct FilterbankState* state, int scale_down_shift); + +void FilterbankReset(struct FilterbankState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c new file mode 100644 index 0000000..741e122 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c @@ -0,0 +1,223 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h" + +#include +#include +#include + +#define kFilterbankIndexAlignment 4 +#define kFilterbankChannelBlockSize 4 + +void FilterbankFillConfigWithDefaults(struct FilterbankConfig* config) { + config->num_channels = 32; + config->lower_band_limit = 125.0f; + config->upper_band_limit = 7500.0f; + config->output_scale_shift = 7; +} + +static float FreqToMel(float freq) { return 1127.0 * log1p(freq / 700.0); } + +static void CalculateCenterFrequencies(const int num_channels, + const float lower_frequency_limit, + const float upper_frequency_limit, + float* center_frequencies) { + assert(lower_frequency_limit >= 0.0f); + assert(upper_frequency_limit > lower_frequency_limit); + + const float mel_low = FreqToMel(lower_frequency_limit); + const float mel_hi = FreqToMel(upper_frequency_limit); + const float mel_span = mel_hi - mel_low; + const float mel_spacing = mel_span / ((float)num_channels); + int i; + for (i = 0; i < num_channels; ++i) { + center_frequencies[i] = mel_low + (mel_spacing * (i + 1)); + } +} + +static void QuantizeFilterbankWeights(const float float_weight, int16_t* weight, + int16_t* unweight) { + *weight = floor(float_weight * (1 << kFilterbankBits) + 0.5); + *unweight = floor((1.0 - float_weight) * (1 << kFilterbankBits) + 0.5); +} + +int FilterbankPopulateState(const struct FilterbankConfig* config, + struct FilterbankState* state, int sample_rate, + int spectrum_size) { + state->num_channels = config->num_channels; + const int num_channels_plus_1 = config->num_channels + 1; + + // How should we align things to index counts given the byte alignment? + const int index_alignment = + (kFilterbankIndexAlignment < sizeof(int16_t) + ? 1 + : kFilterbankIndexAlignment / sizeof(int16_t)); + + state->channel_frequency_starts = + malloc(num_channels_plus_1 * sizeof(*state->channel_frequency_starts)); + state->channel_weight_starts = + malloc(num_channels_plus_1 * sizeof(*state->channel_weight_starts)); + state->channel_widths = + malloc(num_channels_plus_1 * sizeof(*state->channel_widths)); + state->work = malloc(num_channels_plus_1 * sizeof(*state->work)); + + float* center_mel_freqs = + malloc(num_channels_plus_1 * sizeof(*center_mel_freqs)); + int16_t* actual_channel_starts = + malloc(num_channels_plus_1 * sizeof(*actual_channel_starts)); + int16_t* actual_channel_widths = + malloc(num_channels_plus_1 * sizeof(*actual_channel_widths)); + + if (state->channel_frequency_starts == NULL || + state->channel_weight_starts == NULL || state->channel_widths == NULL || + center_mel_freqs == NULL || actual_channel_starts == NULL || + actual_channel_widths == NULL) { + free(center_mel_freqs); + free(actual_channel_starts); + free(actual_channel_widths); + fprintf(stderr, "Failed to allocate channel buffers\n"); + return 0; + } + + CalculateCenterFrequencies(num_channels_plus_1, config->lower_band_limit, + config->upper_band_limit, center_mel_freqs); + + // Always exclude DC. + const float hz_per_sbin = 0.5 * sample_rate / ((float)spectrum_size - 1); + state->start_index = 1.5 + config->lower_band_limit / hz_per_sbin; + state->end_index = 0; // Initialized to zero here, but actually set below. + + // For each channel, we need to figure out what frequencies belong to it, and + // how much padding we need to add so that we can efficiently multiply the + // weights and unweights for accumulation. To simplify the multiplication + // logic, all channels will have some multiplication to do (even if there are + // no frequencies that accumulate to that channel) - they will be directed to + // a set of zero weights. + int chan_freq_index_start = state->start_index; + int weight_index_start = 0; + int needs_zeros = 0; + + int chan; + for (chan = 0; chan < num_channels_plus_1; ++chan) { + // Keep jumping frequencies until we overshoot the bound on this channel. + int freq_index = chan_freq_index_start; + while (FreqToMel((freq_index)*hz_per_sbin) <= center_mel_freqs[chan]) { + ++freq_index; + } + + const int width = freq_index - chan_freq_index_start; + actual_channel_starts[chan] = chan_freq_index_start; + actual_channel_widths[chan] = width; + + if (width == 0) { + // This channel doesn't actually get anything from the frequencies, it's + // always zero. We need then to insert some 'zero' weights into the + // output, and just redirect this channel to do a single multiplication at + // this point. For simplicity, the zeros are placed at the beginning of + // the weights arrays, so we have to go and update all the other + // weight_starts to reflect this shift (but only once). + state->channel_frequency_starts[chan] = 0; + state->channel_weight_starts[chan] = 0; + state->channel_widths[chan] = kFilterbankChannelBlockSize; + if (!needs_zeros) { + needs_zeros = 1; + int j; + for (j = 0; j < chan; ++j) { + state->channel_weight_starts[j] += kFilterbankChannelBlockSize; + } + weight_index_start += kFilterbankChannelBlockSize; + } + } else { + // How far back do we need to go to ensure that we have the proper + // alignment? + const int aligned_start = + (chan_freq_index_start / index_alignment) * index_alignment; + const int aligned_width = (chan_freq_index_start - aligned_start + width); + const int padded_width = + (((aligned_width - 1) / kFilterbankChannelBlockSize) + 1) * + kFilterbankChannelBlockSize; + + state->channel_frequency_starts[chan] = aligned_start; + state->channel_weight_starts[chan] = weight_index_start; + state->channel_widths[chan] = padded_width; + weight_index_start += padded_width; + } + chan_freq_index_start = freq_index; + } + + // Allocate the two arrays to store the weights - weight_index_start contains + // the index of what would be the next set of weights that we would need to + // add, so that's how many weights we need to allocate. + state->weights = calloc(weight_index_start, sizeof(*state->weights)); + state->unweights = calloc(weight_index_start, sizeof(*state->unweights)); + + // If the alloc failed, we also need to nuke the arrays. + if (state->weights == NULL || state->unweights == NULL) { + free(center_mel_freqs); + free(actual_channel_starts); + free(actual_channel_widths); + fprintf(stderr, "Failed to allocate weights or unweights\n"); + return 0; + } + + // Next pass, compute all the weights. Since everything has been memset to + // zero, we only need to fill in the weights that correspond to some frequency + // for a channel. + const float mel_low = FreqToMel(config->lower_band_limit); + for (chan = 0; chan < num_channels_plus_1; ++chan) { + int frequency = actual_channel_starts[chan]; + const int num_frequencies = actual_channel_widths[chan]; + const int frequency_offset = + frequency - state->channel_frequency_starts[chan]; + const int weight_start = state->channel_weight_starts[chan]; + const float denom_val = (chan == 0) ? mel_low : center_mel_freqs[chan - 1]; + + int j; + for (j = 0; j < num_frequencies; ++j, ++frequency) { + const float weight = + (center_mel_freqs[chan] - FreqToMel(frequency * hz_per_sbin)) / + (center_mel_freqs[chan] - denom_val); + + // Make the float into an integer for the weights (and unweights). + const int weight_index = weight_start + frequency_offset + j; + QuantizeFilterbankWeights(weight, state->weights + weight_index, + state->unweights + weight_index); + } + if (frequency > state->end_index) { + state->end_index = frequency; + } + } + + free(center_mel_freqs); + free(actual_channel_starts); + free(actual_channel_widths); + if (state->end_index >= spectrum_size) { + fprintf(stderr, "Filterbank end_index is above spectrum size.\n"); + return 0; + } + return 1; +} + +void FilterbankFreeStateContents(struct FilterbankState* state) { + free(state->channel_frequency_starts); + free(state->channel_weight_starts); + free(state->channel_widths); + free(state->weights); + free(state->unweights); + free(state->work); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h new file mode 100644 index 0000000..d1853a9 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h @@ -0,0 +1,53 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_ + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct FilterbankConfig { + // number of frequency channel buckets for filterbank + int num_channels; + // maximum frequency to include + float upper_band_limit; + // minimum frequency to include + float lower_band_limit; + // unused + int output_scale_shift; +}; + +// Fills the frontendConfig with "sane" defaults. +void FilterbankFillConfigWithDefaults(struct FilterbankConfig* config); + +// Allocates any buffers. +int FilterbankPopulateState(const struct FilterbankConfig* config, + struct FilterbankState* state, int sample_rate, + int spectrum_size); + +// Frees any allocated buffers. +void FilterbankFreeStateContents(struct FilterbankState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.c new file mode 100644 index 0000000..7199649 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.c @@ -0,0 +1,75 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.h" + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +struct FrontendOutput FrontendProcessSamples(struct FrontendState* state, + const int16_t* samples, + size_t num_samples, + size_t* num_samples_read) { + struct FrontendOutput output; + output.values = NULL; + output.size = 0; + + // Try to apply the window - if it fails, return and wait for more data. + if (!WindowProcessSamples(&state->window, samples, num_samples, + num_samples_read)) { + return output; + } + + // Apply the FFT to the window's output (and scale it so that the fixed point + // FFT can have as much resolution as possible). + int input_shift = + 15 - MostSignificantBit32(state->window.max_abs_output_value); + FftCompute(&state->fft, state->window.output, input_shift); + + // We can re-ruse the fft's output buffer to hold the energy. + int32_t* energy = (int32_t*)state->fft.output; + + FilterbankConvertFftComplexToEnergy(&state->filterbank, state->fft.output, + energy); + + FilterbankAccumulateChannels(&state->filterbank, energy); + uint32_t* scaled_filterbank = FilterbankSqrt(&state->filterbank, input_shift); + + // Apply noise reduction. + NoiseReductionApply(&state->noise_reduction, scaled_filterbank); + + if (state->pcan_gain_control.enable_pcan) { + PcanGainControlApply(&state->pcan_gain_control, scaled_filterbank); + } + + // Apply the log and scale. + int correction_bits = + MostSignificantBit32(state->fft.fft_size) - 1 - (kFilterbankBits / 2); + uint16_t* logged_filterbank = + LogScaleApply(&state->log_scale, scaled_filterbank, + state->filterbank.num_channels, correction_bits); + + output.size = state->filterbank.num_channels; + output.values = logged_filterbank; + return output; +} + +void FrontendReset(struct FrontendState* state) { + WindowReset(&state->window); + FftReset(&state->fft); + FilterbankReset(&state->filterbank); + NoiseReductionReset(&state->noise_reduction); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.h new file mode 100644 index 0000000..285378d --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.h @@ -0,0 +1,67 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct FrontendState { + struct WindowState window; + struct FftState fft; + struct FilterbankState filterbank; + struct NoiseReductionState noise_reduction; + struct PcanGainControlState pcan_gain_control; + struct LogScaleState log_scale; +}; + +struct FrontendOutput { + const uint16_t* values; + size_t size; +}; + +// Main entry point to processing frontend samples. Updates num_samples_read to +// contain the number of samples that have been consumed from the input array. +// Returns a struct containing the generated output. If not enough samples were +// added to generate a feature vector, the returned size will be 0 and the +// values pointer will be NULL. Note that the output pointer will be invalidated +// as soon as FrontendProcessSamples is called again, so copy the contents +// elsewhere if you need to use them later. +struct FrontendOutput FrontendProcessSamples(struct FrontendState* state, + const int16_t* samples, + size_t num_samples, + size_t* num_samples_read); + +void FrontendReset(struct FrontendState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c new file mode 100644 index 0000000..a17f388 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c @@ -0,0 +1,88 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h" + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +void FrontendFillConfigWithDefaults(struct FrontendConfig* config) { + WindowFillConfigWithDefaults(&config->window); + FilterbankFillConfigWithDefaults(&config->filterbank); + NoiseReductionFillConfigWithDefaults(&config->noise_reduction); + PcanGainControlFillConfigWithDefaults(&config->pcan_gain_control); + LogScaleFillConfigWithDefaults(&config->log_scale); +} + +int FrontendPopulateState(const struct FrontendConfig* config, + struct FrontendState* state, int sample_rate) { + memset(state, 0, sizeof(*state)); + + if (!WindowPopulateState(&config->window, &state->window, sample_rate)) { + fprintf(stderr, "Failed to populate window state\n"); + return 0; + } + + if (!FftPopulateState(&state->fft, state->window.size)) { + fprintf(stderr, "Failed to populate fft state\n"); + return 0; + } + FftInit(&state->fft); + + if (!FilterbankPopulateState(&config->filterbank, &state->filterbank, + sample_rate, state->fft.fft_size / 2 + 1)) { + fprintf(stderr, "Failed to populate filterbank state\n"); + return 0; + } + + if (!NoiseReductionPopulateState(&config->noise_reduction, + &state->noise_reduction, + state->filterbank.num_channels)) { + fprintf(stderr, "Failed to populate noise reduction state\n"); + return 0; + } + + int input_correction_bits = + MostSignificantBit32(state->fft.fft_size) - 1 - (kFilterbankBits / 2); + if (!PcanGainControlPopulateState( + &config->pcan_gain_control, &state->pcan_gain_control, + state->noise_reduction.estimate, state->filterbank.num_channels, + state->noise_reduction.smoothing_bits, input_correction_bits)) { + fprintf(stderr, "Failed to populate pcan gain control state\n"); + return 0; + } + + if (!LogScalePopulateState(&config->log_scale, &state->log_scale)) { + fprintf(stderr, "Failed to populate log scale state\n"); + return 0; + } + + FrontendReset(state); + + // All good, return a true value. + return 1; +} + +void FrontendFreeStateContents(struct FrontendState* state) { + WindowFreeStateContents(&state->window); + FftFreeStateContents(&state->fft); + FilterbankFreeStateContents(&state->filterbank); + NoiseReductionFreeStateContents(&state->noise_reduction); + PcanGainControlFreeStateContents(&state->pcan_gain_control); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h new file mode 100644 index 0000000..c632e40 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h @@ -0,0 +1,55 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_ + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/fft_util.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/frontend.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct FrontendConfig { + struct WindowConfig window; + struct FilterbankConfig filterbank; + struct NoiseReductionConfig noise_reduction; + struct PcanGainControlConfig pcan_gain_control; + struct LogScaleConfig log_scale; +}; + +// Fills the frontendConfig with "sane" defaults. +void FrontendFillConfigWithDefaults(struct FrontendConfig* config); + +// Allocates any buffers. +int FrontendPopulateState(const struct FrontendConfig* config, + struct FrontendState* state, int sample_rate); + +// Frees any allocated buffers. +void FrontendFreeStateContents(struct FrontendState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.c new file mode 100644 index 0000000..cffffd1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.c @@ -0,0 +1,33 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.h" +const uint16_t kLogLut[] +#ifndef _MSC_VER + __attribute__((aligned(4))) +#endif // _MSV_VER + = {0, 224, 442, 654, 861, 1063, 1259, 1450, 1636, 1817, 1992, 2163, + 2329, 2490, 2646, 2797, 2944, 3087, 3224, 3358, 3487, 3611, 3732, 3848, + 3960, 4068, 4172, 4272, 4368, 4460, 4549, 4633, 4714, 4791, 4864, 4934, + 5001, 5063, 5123, 5178, 5231, 5280, 5326, 5368, 5408, 5444, 5477, 5507, + 5533, 5557, 5578, 5595, 5610, 5622, 5631, 5637, 5640, 5641, 5638, 5633, + 5626, 5615, 5602, 5586, 5568, 5547, 5524, 5498, 5470, 5439, 5406, 5370, + 5332, 5291, 5249, 5203, 5156, 5106, 5054, 5000, 4944, 4885, 4825, 4762, + 4697, 4630, 4561, 4490, 4416, 4341, 4264, 4184, 4103, 4020, 3935, 3848, + 3759, 3668, 3575, 3481, 3384, 3286, 3186, 3084, 2981, 2875, 2768, 2659, + 2549, 2437, 2323, 2207, 2090, 1971, 1851, 1729, 1605, 1480, 1353, 1224, + 1094, 963, 830, 695, 559, 421, 282, 142, 0, 0}; + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.h new file mode 100644 index 0000000..3639778 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.h @@ -0,0 +1,43 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_ + +#include + +#ifdef __cplusplus +extern "C" { +#endif + +// Number of segments in the log lookup table. The table will be kLogSegments+1 +// in length (with some padding). +#define kLogSegments 128 +#define kLogSegmentsLog2 7 + +// Scale used by lookup table. +#define kLogScale 65536 +#define kLogScaleLog2 16 +#define kLogCoeff 45426 + +extern const uint16_t kLogLut[]; + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.c new file mode 100644 index 0000000..d5a6648 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.c @@ -0,0 +1,86 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.h" + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h" +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_lut.h" + +#define kuint16max 0x0000FFFF + +// The following functions implement integer logarithms of various sizes. The +// approximation is calculated according to method described in +// www.inti.gob.ar/electronicaeinformatica/instrumentacion/utic/ +// publicaciones/SPL2007/Log10-spl07.pdf +// It first calculates log2 of the input and then converts it to natural +// logarithm. + +static uint32_t Log2FractionPart(const uint32_t x, const uint32_t log2x) { + // Part 1 + int32_t frac = x - (1LL << log2x); + if (log2x < kLogScaleLog2) { + frac <<= kLogScaleLog2 - log2x; + } else { + frac >>= log2x - kLogScaleLog2; + } + // Part 2 + const uint32_t base_seg = frac >> (kLogScaleLog2 - kLogSegmentsLog2); + const uint32_t seg_unit = + (((uint32_t)1) << kLogScaleLog2) >> kLogSegmentsLog2; + + const int32_t c0 = kLogLut[base_seg]; + const int32_t c1 = kLogLut[base_seg + 1]; + const int32_t seg_base = seg_unit * base_seg; + const int32_t rel_pos = ((c1 - c0) * (frac - seg_base)) >> kLogScaleLog2; + return frac + c0 + rel_pos; +} + +static uint32_t Log(const uint32_t x, const uint32_t scale_shift) { + const uint32_t integer = MostSignificantBit32(x) - 1; + const uint32_t fraction = Log2FractionPart(x, integer); + const uint32_t log2 = (integer << kLogScaleLog2) + fraction; + const uint32_t round = kLogScale / 2; + const uint32_t loge = (((uint64_t)kLogCoeff) * log2 + round) >> kLogScaleLog2; + // Finally scale to our output scale + const uint32_t loge_scaled = ((loge << scale_shift) + round) >> kLogScaleLog2; + return loge_scaled; +} + +uint16_t* LogScaleApply(struct LogScaleState* state, uint32_t* signal, + int signal_size, int correction_bits) { + const int scale_shift = state->scale_shift; + uint16_t* output = (uint16_t*)signal; + uint16_t* ret = output; + int i; + for (i = 0; i < signal_size; ++i) { + uint32_t value = *signal++; + if (state->enable_log) { + if (correction_bits < 0) { + value >>= -correction_bits; + } else { + value <<= correction_bits; + } + if (value > 1) { + value = Log(value, scale_shift); + } else { + value = 0; + } + } + *output++ = (value < kuint16max) ? value : kuint16max; + } + return ret; +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.h new file mode 100644 index 0000000..d5a2e18 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.h @@ -0,0 +1,42 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_ + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct LogScaleState { + int enable_log; + int scale_shift; +}; + +// Applies a fixed point logarithm to the signal and converts it to 16 bit. Note +// that the signal array will be modified. +uint16_t* LogScaleApply(struct LogScaleState* state, uint32_t* signal, + int signal_size, int correction_bits); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c new file mode 100644 index 0000000..8449af0 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c @@ -0,0 +1,30 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h" + +void LogScaleFillConfigWithDefaults(struct LogScaleConfig* config) { + config->enable_log = 1; + config->scale_shift = 6; +} + +int LogScalePopulateState(const struct LogScaleConfig* config, + struct LogScaleState* state) { + state->enable_log = config->enable_log; + state->scale_shift = config->scale_shift; + return 1; +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h new file mode 100644 index 0000000..6255ec2 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h @@ -0,0 +1,48 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/log_scale.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct LogScaleConfig { + // set to false (0) to disable this module + int enable_log; + // scale results by 2^(scale_shift) + int scale_shift; +}; + +// Populates the LogScaleConfig with "sane" default values. +void LogScaleFillConfigWithDefaults(struct LogScaleConfig* config); + +// Allocates any buffers. +int LogScalePopulateState(const struct LogScaleConfig* config, + struct LogScaleState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c new file mode 100644 index 0000000..8e91dd1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c @@ -0,0 +1,54 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h" + +#include + +void NoiseReductionApply(struct NoiseReductionState* state, uint32_t* signal) { + int i; + for (i = 0; i < state->num_channels; ++i) { + const uint32_t smoothing = + ((i & 1) == 0) ? state->even_smoothing : state->odd_smoothing; + const uint32_t one_minus_smoothing = (1 << kNoiseReductionBits) - smoothing; + + // Update the estimate of the noise. + const uint32_t signal_scaled_up = signal[i] << state->smoothing_bits; + uint32_t estimate = + (((uint64_t)signal_scaled_up * smoothing) + + ((uint64_t)state->estimate[i] * one_minus_smoothing)) >> + kNoiseReductionBits; + state->estimate[i] = estimate; + + // Make sure that we can't get a negative value for the signal - estimate. + if (estimate > signal_scaled_up) { + estimate = signal_scaled_up; + } + + const uint32_t floor = + ((uint64_t)signal[i] * state->min_signal_remaining) >> + kNoiseReductionBits; + const uint32_t subtracted = + (signal_scaled_up - estimate) >> state->smoothing_bits; + const uint32_t output = subtracted > floor ? subtracted : floor; + signal[i] = output; + } +} + +void NoiseReductionReset(struct NoiseReductionState* state) { + memset(state->estimate, 0, sizeof(*state->estimate) * state->num_channels); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h new file mode 100644 index 0000000..201ca9a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h @@ -0,0 +1,49 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_ + +#define kNoiseReductionBits 14 + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct NoiseReductionState { + int smoothing_bits; + uint16_t even_smoothing; + uint16_t odd_smoothing; + uint16_t min_signal_remaining; + int num_channels; + uint32_t* estimate; +}; + +// Removes stationary noise from each channel of the signal using a low pass +// filter. +void NoiseReductionApply(struct NoiseReductionState* state, uint32_t* signal); + +void NoiseReductionReset(struct NoiseReductionState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c new file mode 100644 index 0000000..2b9f69f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c @@ -0,0 +1,48 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h" + +#include + +void NoiseReductionFillConfigWithDefaults(struct NoiseReductionConfig* config) { + config->smoothing_bits = 10; + config->even_smoothing = 0.025; + config->odd_smoothing = 0.06; + config->min_signal_remaining = 0.05; +} + +int NoiseReductionPopulateState(const struct NoiseReductionConfig* config, + struct NoiseReductionState* state, + int num_channels) { + state->smoothing_bits = config->smoothing_bits; + state->odd_smoothing = config->odd_smoothing * (1 << kNoiseReductionBits); + state->even_smoothing = config->even_smoothing * (1 << kNoiseReductionBits); + state->min_signal_remaining = + config->min_signal_remaining * (1 << kNoiseReductionBits); + state->num_channels = num_channels; + state->estimate = calloc(state->num_channels, sizeof(*state->estimate)); + if (state->estimate == NULL) { + fprintf(stderr, "Failed to alloc estimate buffer\n"); + return 0; + } + return 1; +} + +void NoiseReductionFreeStateContents(struct NoiseReductionState* state) { + free(state->estimate); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h new file mode 100644 index 0000000..092eba5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h @@ -0,0 +1,53 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_ + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct NoiseReductionConfig { + // scale the signal up by 2^(smoothing_bits) before reduction + int smoothing_bits; + // smoothing coefficient for even-numbered channels + float even_smoothing; + // smoothing coefficient for odd-numbered channels + float odd_smoothing; + // fraction of signal to preserve (1.0 disables this module) + float min_signal_remaining; +}; + +// Populates the NoiseReductionConfig with "sane" default values. +void NoiseReductionFillConfigWithDefaults(struct NoiseReductionConfig* config); + +// Allocates any buffers. +int NoiseReductionPopulateState(const struct NoiseReductionConfig* config, + struct NoiseReductionState* state, + int num_channels); + +// Frees any allocated buffers. +void NoiseReductionFreeStateContents(struct NoiseReductionState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c new file mode 100644 index 0000000..3ee19cf --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c @@ -0,0 +1,59 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h" + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +int16_t WideDynamicFunction(const uint32_t x, const int16_t* lut) { + if (x <= 2) { + return lut[x]; + } + + const int16_t interval = MostSignificantBit32(x); + lut += 4 * interval - 6; + + const int16_t frac = + ((interval < 11) ? (x << (11 - interval)) : (x >> (interval - 11))) & + 0x3FF; + + int32_t result = ((int32_t)lut[2] * frac) >> 5; + result += (int32_t)((uint32_t)lut[1] << 5); + result *= frac; + result = (result + (1 << 14)) >> 15; + result += lut[0]; + return (int16_t)result; +} + +uint32_t PcanShrink(const uint32_t x) { + if (x < (2 << kPcanSnrBits)) { + return (x * x) >> (2 + 2 * kPcanSnrBits - kPcanOutputBits); + } else { + return (x >> (kPcanSnrBits - kPcanOutputBits)) - (1 << kPcanOutputBits); + } +} + +void PcanGainControlApply(struct PcanGainControlState* state, + uint32_t* signal) { + int i; + for (i = 0; i < state->num_channels; ++i) { + const uint32_t gain = + WideDynamicFunction(state->noise_estimate[i], state->gain_lut); + const uint32_t snr = ((uint64_t)signal[i] * gain) >> state->snr_shift; + signal[i] = PcanShrink(snr); + } +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h new file mode 100644 index 0000000..f0cd58e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h @@ -0,0 +1,50 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_ + +#include +#include + +#define kPcanSnrBits 12 +#define kPcanOutputBits 6 + +#ifdef __cplusplus +extern "C" { +#endif + +// Details at https://research.google/pubs/pub45911.pdf +struct PcanGainControlState { + int enable_pcan; + uint32_t* noise_estimate; + int num_channels; + int16_t* gain_lut; + int32_t snr_shift; +}; + +int16_t WideDynamicFunction(const uint32_t x, const int16_t* lut); + +uint32_t PcanShrink(const uint32_t x); + +void PcanGainControlApply(struct PcanGainControlState* state, uint32_t* signal); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c new file mode 100644 index 0000000..74e3b17 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c @@ -0,0 +1,95 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h" + +#include +#include + +#define kint16max 0x00007FFF + +void PcanGainControlFillConfigWithDefaults( + struct PcanGainControlConfig* config) { + config->enable_pcan = 0; + config->strength = 0.95; + config->offset = 80.0; + config->gain_bits = 21; +} + +int16_t PcanGainLookupFunction(const struct PcanGainControlConfig* config, + int32_t input_bits, uint32_t x) { + const float x_as_float = ((float)x) / ((uint32_t)1 << input_bits); + const float gain_as_float = + ((uint32_t)1 << config->gain_bits) * + powf(x_as_float + config->offset, -config->strength); + + if (gain_as_float > kint16max) { + return kint16max; + } + return (int16_t)(gain_as_float + 0.5f); +} + +int PcanGainControlPopulateState(const struct PcanGainControlConfig* config, + struct PcanGainControlState* state, + uint32_t* noise_estimate, + const int num_channels, + const uint16_t smoothing_bits, + const int32_t input_correction_bits) { + state->enable_pcan = config->enable_pcan; + if (!state->enable_pcan) { + return 1; + } + state->noise_estimate = noise_estimate; + state->num_channels = num_channels; + state->gain_lut = malloc(kWideDynamicFunctionLUTSize * sizeof(int16_t)); + if (state->gain_lut == NULL) { + fprintf(stderr, "Failed to allocate gain LUT\n"); + return 0; + } + state->snr_shift = config->gain_bits - input_correction_bits - kPcanSnrBits; + + const int32_t input_bits = smoothing_bits - input_correction_bits; + state->gain_lut[0] = PcanGainLookupFunction(config, input_bits, 0); + state->gain_lut[1] = PcanGainLookupFunction(config, input_bits, 1); + state->gain_lut -= 6; + int interval; + for (interval = 2; interval <= kWideDynamicFunctionBits; ++interval) { + const uint32_t x0 = (uint32_t)1 << (interval - 1); + const uint32_t x1 = x0 + (x0 >> 1); + const uint32_t x2 = + (interval == kWideDynamicFunctionBits) ? x0 + (x0 - 1) : 2 * x0; + + const int16_t y0 = PcanGainLookupFunction(config, input_bits, x0); + const int16_t y1 = PcanGainLookupFunction(config, input_bits, x1); + const int16_t y2 = PcanGainLookupFunction(config, input_bits, x2); + + const int32_t diff1 = (int32_t)y1 - y0; + const int32_t diff2 = (int32_t)y2 - y0; + const int32_t a1 = 4 * diff1 - diff2; + const int32_t a2 = diff2 - a1; + + state->gain_lut[4 * interval] = y0; + state->gain_lut[4 * interval + 1] = (int16_t)a1; + state->gain_lut[4 * interval + 2] = (int16_t)a2; + } + state->gain_lut += 6; + return 1; +} + +void PcanGainControlFreeStateContents(struct PcanGainControlState* state) { + free(state->gain_lut); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h new file mode 100644 index 0000000..7a3df0e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h @@ -0,0 +1,60 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_ + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h" + +#define kWideDynamicFunctionBits 32 +#define kWideDynamicFunctionLUTSize (4 * kWideDynamicFunctionBits - 3) + +#ifdef __cplusplus +extern "C" { +#endif + +struct PcanGainControlConfig { + // set to false (0) to disable this module + int enable_pcan; + // gain normalization exponent (0.0 disables, 1.0 full strength) + float strength; + // positive value added in the normalization denominator + float offset; + // number of fractional bits in the gain + int gain_bits; +}; + +void PcanGainControlFillConfigWithDefaults( + struct PcanGainControlConfig* config); + +int16_t PcanGainLookupFunction(const struct PcanGainControlConfig* config, + int32_t input_bits, uint32_t x); + +int PcanGainControlPopulateState(const struct PcanGainControlConfig* config, + struct PcanGainControlState* state, + uint32_t* noise_estimate, + const int num_channels, + const uint16_t smoothing_bits, + const int32_t input_correction_bits); + +void PcanGainControlFreeStateContents(struct PcanGainControlState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.c new file mode 100644 index 0000000..5c780f6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.c @@ -0,0 +1,73 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.h" + +#include + +int WindowProcessSamples(struct WindowState* state, const int16_t* samples, + size_t num_samples, size_t* num_samples_read) { + const int size = state->size; + + // Copy samples from the samples buffer over to our local input. + size_t max_samples_to_copy = state->size - state->input_used; + if (max_samples_to_copy > num_samples) { + max_samples_to_copy = num_samples; + } + memcpy(state->input + state->input_used, samples, + max_samples_to_copy * sizeof(*samples)); + *num_samples_read = max_samples_to_copy; + state->input_used += max_samples_to_copy; + + if (state->input_used < state->size) { + // We don't have enough samples to compute a window. + return 0; + } + + // Apply the window to the input. + const int16_t* coefficients = state->coefficients; + const int16_t* input = state->input; + int16_t* output = state->output; + int i; + int16_t max_abs_output_value = 0; + for (i = 0; i < size; ++i) { + int16_t new_value = + (((int32_t)*input++) * *coefficients++) >> kFrontendWindowBits; + *output++ = new_value; + if (new_value < 0) { + new_value = -new_value; + } + if (new_value > max_abs_output_value) { + max_abs_output_value = new_value; + } + } + // Shuffle the input down by the step size, and update how much we have used. + memmove(state->input, state->input + state->step, + sizeof(*state->input) * (state->size - state->step)); + state->input_used -= state->step; + state->max_abs_output_value = max_abs_output_value; + + // Indicate that the output buffer is valid for the next stage. + return 1; +} + +void WindowReset(struct WindowState* state) { + memset(state->input, 0, state->size * sizeof(*state->input)); + memset(state->output, 0, state->size * sizeof(*state->output)); + state->input_used = 0; + state->max_abs_output_value = 0; +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.h new file mode 100644 index 0000000..d6e7fa6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.h @@ -0,0 +1,52 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_ + +#include +#include + +#define kFrontendWindowBits 12 + +#ifdef __cplusplus +extern "C" { +#endif + +struct WindowState { + size_t size; + int16_t* coefficients; + size_t step; + + int16_t* input; + size_t input_used; + int16_t* output; + int16_t max_abs_output_value; +}; + +// Applies a window to the samples coming in, stepping forward at the given +// rate. +int WindowProcessSamples(struct WindowState* state, const int16_t* samples, + size_t num_samples, size_t* num_samples_read); + +void WindowReset(struct WindowState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.c b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.c new file mode 100644 index 0000000..281675d --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.c @@ -0,0 +1,76 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.h" + +#include +#include +#include +#include + +// Some platforms don't have M_PI +#ifndef M_PI +#define M_PI 3.14159265358979323846 +#endif + +void WindowFillConfigWithDefaults(struct WindowConfig* config) { + config->size_ms = 25; + config->step_size_ms = 10; +} + +int WindowPopulateState(const struct WindowConfig* config, + struct WindowState* state, int sample_rate) { + state->size = config->size_ms * sample_rate / 1000; + state->step = config->step_size_ms * sample_rate / 1000; + + state->coefficients = malloc(state->size * sizeof(*state->coefficients)); + if (state->coefficients == NULL) { + fprintf(stderr, "Failed to allocate window coefficients\n"); + return 0; + } + + // Populate the window values. + const float arg = M_PI * 2.0 / ((float)state->size); + int i; + for (i = 0; i < state->size; ++i) { + float float_value = 0.5 - (0.5 * cos(arg * (i + 0.5))); + // Scale it to fixed point and round it. + state->coefficients[i] = + floor(float_value * (1 << kFrontendWindowBits) + 0.5); + } + + state->input_used = 0; + state->input = malloc(state->size * sizeof(*state->input)); + if (state->input == NULL) { + fprintf(stderr, "Failed to allocate window input\n"); + return 0; + } + + state->output = malloc(state->size * sizeof(*state->output)); + if (state->output == NULL) { + fprintf(stderr, "Failed to allocate window output\n"); + return 0; + } + + return 1; +} + +void WindowFreeStateContents(struct WindowState* state) { + free(state->coefficients); + free(state->input); + free(state->output); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.h b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.h new file mode 100644 index 0000000..c41003b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window_util.h @@ -0,0 +1,48 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_ + +#include "tensorflow_arm/tensorflow/lite/experimental/microfrontend/lib/window.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct WindowConfig { + // length of window frame in milliseconds + size_t size_ms; + // length of step for next frame in milliseconds + size_t step_size_ms; +}; + +// Populates the WindowConfig with "sane" default values. +void WindowFillConfigWithDefaults(struct WindowConfig* config); + +// Allocates any buffers. +int WindowPopulateState(const struct WindowConfig* config, + struct WindowState* state, int sample_rate); + +// Frees any allocated buffers. +void WindowFreeStateContents(struct WindowState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/common.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/common.h new file mode 100644 index 0000000..73eedb9 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/common.h @@ -0,0 +1,1040 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_COMMON_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_COMMON_H_ + +#ifndef ALLOW_SLOW_GENERIC_DEPTHWISECONV_FALLBACK +#ifdef GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK +#define ALLOW_SLOW_GENERIC_DEPTHWISECONV_FALLBACK +#endif +#endif + +#include + +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/optimized/neon_check.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +constexpr int kReverseShift = -1; + +inline void GetActivationMinMax(FusedActivationFunctionType ac, + float* output_activation_min, + float* output_activation_max) { + switch (ac) { + case FusedActivationFunctionType::kNone: + *output_activation_min = std::numeric_limits::lowest(); + *output_activation_max = std::numeric_limits::max(); + break; + case FusedActivationFunctionType::kRelu: + *output_activation_min = 0.f; + *output_activation_max = std::numeric_limits::max(); + break; + case FusedActivationFunctionType::kRelu1: + *output_activation_min = -1.f; + *output_activation_max = 1.f; + break; + case FusedActivationFunctionType::kRelu6: + *output_activation_min = 0.f; + *output_activation_max = 6.f; + break; + } +} + +template +inline T ActivationFunctionWithMinMax(T x, T output_activation_min, + T output_activation_max) { + using std::max; + using std::min; + return min(max(x, output_activation_min), output_activation_max); +} + +// Legacy function, left for compatibility only. +template +float ActivationFunction(float x) { + float output_activation_min, output_activation_max; + GetActivationMinMax(Ac, &output_activation_min, &output_activation_max); + return ActivationFunctionWithMinMax(x, output_activation_min, + output_activation_max); +} + +inline void BiasAndClamp(float clamp_min, float clamp_max, int bias_size, + const float* bias_data, int array_size, + float* array_data) { + // Note: see b/132215220: in May 2019 we thought it would be OK to replace + // this with the Eigen one-liner: + // return (array.colwise() + bias).cwiseMin(clamp_max).cwiseMin(clamp_max). + // This turned out to severely regress performance: +4ms (i.e. 8%) on + // MobileNet v2 / 1.0 / 224. So we keep custom NEON code for now. + TFLITE_DCHECK_EQ((array_size % bias_size), 0); +#ifdef USE_NEON + float* array_ptr = array_data; + float* array_end_ptr = array_ptr + array_size; + const auto clamp_min_vec = vdupq_n_f32(clamp_min); + const auto clamp_max_vec = vdupq_n_f32(clamp_max); + for (; array_ptr != array_end_ptr; array_ptr += bias_size) { + int i = 0; + for (; i <= bias_size - 16; i += 16) { + auto b0 = vld1q_f32(bias_data + i); + auto b1 = vld1q_f32(bias_data + i + 4); + auto b2 = vld1q_f32(bias_data + i + 8); + auto b3 = vld1q_f32(bias_data + i + 12); + auto a0 = vld1q_f32(array_ptr + i); + auto a1 = vld1q_f32(array_ptr + i + 4); + auto a2 = vld1q_f32(array_ptr + i + 8); + auto a3 = vld1q_f32(array_ptr + i + 12); + auto x0 = vaddq_f32(a0, b0); + auto x1 = vaddq_f32(a1, b1); + auto x2 = vaddq_f32(a2, b2); + auto x3 = vaddq_f32(a3, b3); + x0 = vmaxq_f32(clamp_min_vec, x0); + x1 = vmaxq_f32(clamp_min_vec, x1); + x2 = vmaxq_f32(clamp_min_vec, x2); + x3 = vmaxq_f32(clamp_min_vec, x3); + x0 = vminq_f32(clamp_max_vec, x0); + x1 = vminq_f32(clamp_max_vec, x1); + x2 = vminq_f32(clamp_max_vec, x2); + x3 = vminq_f32(clamp_max_vec, x3); + vst1q_f32(array_ptr + i, x0); + vst1q_f32(array_ptr + i + 4, x1); + vst1q_f32(array_ptr + i + 8, x2); + vst1q_f32(array_ptr + i + 12, x3); + } + for (; i <= bias_size - 4; i += 4) { + auto b = vld1q_f32(bias_data + i); + auto a = vld1q_f32(array_ptr + i); + auto x = vaddq_f32(a, b); + x = vmaxq_f32(clamp_min_vec, x); + x = vminq_f32(clamp_max_vec, x); + vst1q_f32(array_ptr + i, x); + } + for (; i < bias_size; i++) { + array_ptr[i] = ActivationFunctionWithMinMax(array_ptr[i] + bias_data[i], + clamp_min, clamp_max); + } + } +#else // not NEON + for (int array_offset = 0; array_offset < array_size; + array_offset += bias_size) { + for (int i = 0; i < bias_size; i++) { + array_data[array_offset + i] = ActivationFunctionWithMinMax( + array_data[array_offset + i] + bias_data[i], clamp_min, clamp_max); + } + } +#endif +} + +inline int32_t MultiplyByQuantizedMultiplierSmallerThanOneExp( + int32_t x, int32_t quantized_multiplier, int left_shift) { + using gemmlowp::RoundingDivideByPOT; + using gemmlowp::SaturatingRoundingDoublingHighMul; + return RoundingDivideByPOT( + SaturatingRoundingDoublingHighMul(x, quantized_multiplier), -left_shift); +} + +inline int32_t MultiplyByQuantizedMultiplierGreaterThanOne( + int32_t x, int32_t quantized_multiplier, int left_shift) { + using gemmlowp::SaturatingRoundingDoublingHighMul; + return SaturatingRoundingDoublingHighMul(x * (1 << left_shift), + quantized_multiplier); +} + +inline int32_t MultiplyByQuantizedMultiplier(int32_t x, + int32_t quantized_multiplier, + int shift) { + using gemmlowp::RoundingDivideByPOT; + using gemmlowp::SaturatingRoundingDoublingHighMul; + int left_shift = shift > 0 ? shift : 0; + int right_shift = shift > 0 ? 0 : -shift; + return RoundingDivideByPOT(SaturatingRoundingDoublingHighMul( + x * (1 << left_shift), quantized_multiplier), + right_shift); +} + +inline int32_t MultiplyByQuantizedMultiplier(int64_t x, + int32_t quantized_multiplier, + int shift) { + // Inputs: + // - quantized_multiplier has fixed point at bit 31 + // - shift is -31 to +7 (negative for right shift) + // + // Assumptions: The following input ranges are assumed + // - quantize_scale>=0 (the usual range is (1<<30) to (1>>31)-1) + // - scaling is chosen so final scaled result fits in int32_t + // - input x is in the range -(1<<47) <= x < (1<<47) + assert(quantized_multiplier >= 0); + assert(shift >= -31 && shift < 8); + assert(x >= -(static_cast(1) << 47) && + x < (static_cast(1) << 47)); + + int32_t reduced_multiplier = (quantized_multiplier < 0x7FFF0000) + ? ((quantized_multiplier + (1 << 15)) >> 16) + : 0x7FFF; + int total_shift = 15 - shift; + x = (x * (int64_t)reduced_multiplier) + ((int64_t)1 << (total_shift - 1)); + int32_t result = x >> total_shift; + return result; +} + +#ifdef USE_NEON +// Round uses ARM's rounding shift right. +inline int32x4x4_t MultiplyByQuantizedMultiplier4Rows( + int32x4x4_t input_val, int32_t quantized_multiplier, int shift) { + const int left_shift = std::max(shift, 0); + const int right_shift = std::min(shift, 0); + int32x4x4_t result; + + int32x4_t multiplier_dup = vdupq_n_s32(quantized_multiplier); + int32x4_t left_shift_dup = vdupq_n_s32(left_shift); + int32x4_t right_shift_dup = vdupq_n_s32(right_shift); + + result.val[0] = + vrshlq_s32(vqrdmulhq_s32(vshlq_s32(input_val.val[0], left_shift_dup), + multiplier_dup), + right_shift_dup); + + result.val[1] = + vrshlq_s32(vqrdmulhq_s32(vshlq_s32(input_val.val[1], left_shift_dup), + multiplier_dup), + right_shift_dup); + + result.val[2] = + vrshlq_s32(vqrdmulhq_s32(vshlq_s32(input_val.val[2], left_shift_dup), + multiplier_dup), + right_shift_dup); + + result.val[3] = + vrshlq_s32(vqrdmulhq_s32(vshlq_s32(input_val.val[3], left_shift_dup), + multiplier_dup), + right_shift_dup); + + return result; +} +#endif + +template +int CountLeadingZeros(T integer_input) { + static_assert(std::is_unsigned::value, + "Only unsigned integer types handled."); +#if defined(__GNUC__) + return integer_input ? __builtin_clz(integer_input) + : std::numeric_limits::digits; +#else + if (integer_input == 0) { + return std::numeric_limits::digits; + } + + const T one_in_leading_positive = static_cast(1) + << (std::numeric_limits::digits - 1); + int leading_zeros = 0; + while (integer_input < one_in_leading_positive) { + integer_input <<= 1; + ++leading_zeros; + } + return leading_zeros; +#endif +} + +template +inline int CountLeadingSignBits(T integer_input) { + static_assert(std::is_signed::value, "Only signed integer types handled."); +#if defined(__GNUC__) && !defined(__clang__) + return integer_input ? __builtin_clrsb(integer_input) + : std::numeric_limits::digits; +#else + using U = typename std::make_unsigned::type; + return integer_input >= 0 + ? CountLeadingZeros(static_cast(integer_input)) - 1 + : integer_input != std::numeric_limits::min() + ? CountLeadingZeros(2 * static_cast(-integer_input) - 1) + : 0; +#endif +} + +// Use "count leading zeros" helper functions to do a fast Floor(log_2(x)). +template +inline Integer FloorLog2(Integer n) { + static_assert(std::is_integral::value, ""); + static_assert(std::is_signed::value, ""); + static_assert(sizeof(Integer) == 4 || sizeof(Integer) == 8, ""); + TFLITE_CHECK_GT(n, 0); + if (sizeof(Integer) == 4) { + return 30 - CountLeadingSignBits(n); + } else { + return 62 - CountLeadingSignBits(n); + } +} + +// generate INT16 LUT for function(), e.g., table exp(x) and 1/(1+x) used in +// softmax +// func - the function to build the LUT for (e.g exp(x)) +// min,max - table limits +// table - pointer to buffer +// num - number of elements in the LUT +inline void gen_lut(double (*func)(double), double min, double max, + int16_t* table, const int num) { + // size of table should equal to num + 1 + // last element only for slope calculation + double step = (max - min) / (num - 1); + double half_step = step / 2.0; + for (int i = 0; i < num - 1; i++) { + double sample_val = TfLiteRound(func(min + i * step) * 32768.0); + double midpoint_interp_val = + TfLiteRound((func(min + (i + 1) * step) * 32768.0 + + TfLiteRound(func(min + i * step) * 32768.0)) / + 2.0); + double midpoint_val = + TfLiteRound(func(min + i * step + half_step) * 32768.0); + double midpoint_err = midpoint_interp_val - midpoint_val; + double bias = TfLiteRound(midpoint_err / 2.0); + table[i] = std::min(std::max(sample_val - bias, -32768.0), + 32767.0); + } + table[num - 1] = std::min( + std::max(TfLiteRound(func(max) * 32768.0), -32768.0), 32767.0); +} + +// generate INT16 LUT for function(), e.g., table exp(x) and 1/(1+x) used in +// softmax +// func - the function to build the LUT for (e.g exp(x)) +// min,max - table limits +// table - pointer to buffer +// num - number of elements in the LUT +inline void gen_lut(float (*func)(float), float min, float max, int16_t* table, + const int num) { + // size of table should equal to num + 1 + // last element only for slope calculation + float step = (max - min) / (num - 1); + float half_step = step / 2.0f; + for (int i = 0; i < num - 1; i++) { + float sample_val = TfLiteRound(func(min + i * step) * 32768.0f); + float midpoint_interp_val = + TfLiteRound((func(min + (i + 1) * step) * 32768.0f + + TfLiteRound(func(min + i * step) * 32768.0f)) / + 2.0f); + float midpoint_val = + TfLiteRound(func(min + i * step + half_step) * 32768.0f); + float midpoint_err = midpoint_interp_val - midpoint_val; + float bias = TfLiteRound(midpoint_err / 2.0f); + table[i] = std::min(std::max(sample_val - bias, -32768.0f), + 32767.0f); + } + table[num - 1] = std::min( + std::max(TfLiteRound(func(max) * 32768.0f), -32768.0f), 32767.0f); +} + +// int16_t func table lookup, e.g., lookup exp() and 1/(1+x) used in softmax +inline int16_t generic_int16_table_lookup(int16_t value, const int16_t* lut) { + // 512 base value, lut[513] only for calculate slope + uint16_t index = static_cast(256 + (value >> 7)); + assert(index < 512 && "LUT index out of range."); + int16_t offset = value & 0x7f; + + // base and slope are Q0.15 + int16_t base = lut[index]; + int16_t slope = lut[index + 1] - lut[index]; + + // Q0.15 * Q0.7 = Q0.22 + // Round and convert from Q0.22 to Q0.15 + int32_t delta = (static_cast(slope) * offset + 64) >> 7; + + // Q0.15 + Q0.15 + return base + delta; +} + +// Table of sigmoid(i/24) at 0.16 format - 256 elements. + +// We use combined sigmoid and tanh look-up table, since +// tanh(x) = 2*sigmoid(2*x) -1. +// Both functions are symmetric, so the LUT table is only needed +// for the absolute value of the input. +static const uint16_t sigmoid_table_uint16[256] = { + 32768, 33451, 34133, 34813, 35493, 36169, 36843, 37513, 38180, 38841, 39498, + 40149, 40794, 41432, 42064, 42688, 43304, 43912, 44511, 45102, 45683, 46255, + 46817, 47369, 47911, 48443, 48964, 49475, 49975, 50464, 50942, 51409, 51865, + 52311, 52745, 53169, 53581, 53983, 54374, 54755, 55125, 55485, 55834, 56174, + 56503, 56823, 57133, 57433, 57724, 58007, 58280, 58544, 58800, 59048, 59288, + 59519, 59743, 59959, 60168, 60370, 60565, 60753, 60935, 61110, 61279, 61441, + 61599, 61750, 61896, 62036, 62172, 62302, 62428, 62549, 62666, 62778, 62886, + 62990, 63090, 63186, 63279, 63368, 63454, 63536, 63615, 63691, 63765, 63835, + 63903, 63968, 64030, 64090, 64148, 64204, 64257, 64308, 64357, 64405, 64450, + 64494, 64536, 64576, 64614, 64652, 64687, 64721, 64754, 64786, 64816, 64845, + 64873, 64900, 64926, 64950, 64974, 64997, 65019, 65039, 65060, 65079, 65097, + 65115, 65132, 65149, 65164, 65179, 65194, 65208, 65221, 65234, 65246, 65258, + 65269, 65280, 65291, 65301, 65310, 65319, 65328, 65337, 65345, 65352, 65360, + 65367, 65374, 65381, 65387, 65393, 65399, 65404, 65410, 65415, 65420, 65425, + 65429, 65433, 65438, 65442, 65445, 65449, 65453, 65456, 65459, 65462, 65465, + 65468, 65471, 65474, 65476, 65479, 65481, 65483, 65485, 65488, 65489, 65491, + 65493, 65495, 65497, 65498, 65500, 65501, 65503, 65504, 65505, 65507, 65508, + 65509, 65510, 65511, 65512, 65513, 65514, 65515, 65516, 65517, 65517, 65518, + 65519, 65520, 65520, 65521, 65522, 65522, 65523, 65523, 65524, 65524, 65525, + 65525, 65526, 65526, 65526, 65527, 65527, 65528, 65528, 65528, 65529, 65529, + 65529, 65529, 65530, 65530, 65530, 65530, 65531, 65531, 65531, 65531, 65531, + 65532, 65532, 65532, 65532, 65532, 65532, 65533, 65533, 65533, 65533, 65533, + 65533, 65533, 65533, 65534, 65534, 65534, 65534, 65534, 65534, 65534, 65534, + 65534, 65534, 65535}; + +// TODO(b/77858996): Add these to gemmlowp. +template +IntegerType SaturatingAddNonGemmlowp(IntegerType a, IntegerType b) { + static_assert(std::is_same::value, "unimplemented"); + return a; +} + +template <> +inline std::int32_t SaturatingAddNonGemmlowp(std::int32_t a, std::int32_t b) { + std::int64_t a64 = a; + std::int64_t b64 = b; + std::int64_t sum = a64 + b64; + return static_cast(std::min( + static_cast(std::numeric_limits::max()), + std::max( + static_cast(std::numeric_limits::min()), + sum))); +} + +template +gemmlowp::FixedPoint SaturatingAddNonGemmlowp( + gemmlowp::FixedPoint a, + gemmlowp::FixedPoint b) { + return gemmlowp::FixedPoint::FromRaw( + SaturatingAddNonGemmlowp(a.raw(), b.raw())); +} + +template +IntegerType SaturatingSub(IntegerType a, IntegerType b) { + static_assert(std::is_same::value, "unimplemented"); + return a; +} + +template <> +inline std::int16_t SaturatingSub(std::int16_t a, std::int16_t b) { + std::int32_t a32 = a; + std::int32_t b32 = b; + std::int32_t diff = a32 - b32; + return static_cast( + std::min(static_cast(32767), + std::max(static_cast(-32768), diff))); +} + +template <> +inline std::int32_t SaturatingSub(std::int32_t a, std::int32_t b) { + std::int64_t a64 = a; + std::int64_t b64 = b; + std::int64_t diff = a64 - b64; + return static_cast(std::min( + static_cast(std::numeric_limits::max()), + std::max( + static_cast(std::numeric_limits::min()), + diff))); +} + +template +gemmlowp::FixedPoint SaturatingSub( + gemmlowp::FixedPoint a, + gemmlowp::FixedPoint b) { + return gemmlowp::FixedPoint::FromRaw( + SaturatingSub(a.raw(), b.raw())); +} +// End section to be moved to gemmlowp. + +template +IntegerType SaturatingRoundingMultiplyByPOTParam(IntegerType x, int exponent) { + if (exponent == 0) { + return x; + } + using ScalarIntegerType = + typename gemmlowp::FixedPointRawTypeTraits::ScalarRawType; + const IntegerType min = + gemmlowp::Dup(std::numeric_limits::min()); + const IntegerType max = + gemmlowp::Dup(std::numeric_limits::max()); + const int ScalarIntegerTypeBits = 8 * sizeof(ScalarIntegerType); + + const std::int32_t threshold = + ((1 << (ScalarIntegerTypeBits - 1 - exponent)) - 1); + const IntegerType positive_mask = + gemmlowp::MaskIfGreaterThan(x, gemmlowp::Dup(threshold)); + const IntegerType negative_mask = + gemmlowp::MaskIfLessThan(x, gemmlowp::Dup(-threshold)); + + IntegerType result = gemmlowp::ShiftLeft(x, exponent); + result = gemmlowp::SelectUsingMask(positive_mask, max, result); + result = gemmlowp::SelectUsingMask(negative_mask, min, result); + return result; +} + +// If we want to leave IntegerBits fixed, then multiplication +// by a power of two has to be saturating/rounding, not exact anymore. +template +gemmlowp::FixedPoint +SaturatingRoundingMultiplyByPOTParam( + gemmlowp::FixedPoint a, int exponent) { + return gemmlowp::FixedPoint::FromRaw( + SaturatingRoundingMultiplyByPOTParam(a.raw(), exponent)); +} + +// Convert int32_t multiplier to int16_t with rounding. +inline void DownScaleInt32ToInt16Multiplier(int32_t multiplier_int32_t, + int16_t* multiplier_int16_t) { + TFLITE_DCHECK_GE(multiplier_int32_t, 0); + static constexpr int32_t kRoundingOffset = 1 << 15; + if (multiplier_int32_t >= + std::numeric_limits::max() - kRoundingOffset) { + *multiplier_int16_t = std::numeric_limits::max(); + return; + } + const int32_t result = (multiplier_int32_t + kRoundingOffset) >> 16; + TFLITE_DCHECK_LE(result << 16, multiplier_int32_t + kRoundingOffset); + TFLITE_DCHECK_GT(result << 16, multiplier_int32_t - kRoundingOffset); + *multiplier_int16_t = result; + TFLITE_DCHECK_EQ(*multiplier_int16_t, result); +} + +// Minimum output bits to accommodate log of maximum input range. It actually +// does not matter if one considers, say, [-64,64] or [-64,64). +// +// For example, run this through Octave: +// [0:127; ... +// ceil(log(abs( log(2.^(0:127))+1 ))/log(2)); ... +// ceil(log(abs( log(2.^(0:127))+1 ))/log(2))] +constexpr int min_log_x_output_bits(int input_bits) { + return input_bits > 90 ? 7 + : input_bits > 44 ? 6 + : input_bits > 21 ? 5 + : input_bits > 10 ? 4 + : input_bits > 4 ? 3 + : input_bits > 1 ? 2 + : 1; +} + +// Although currently the name of this function says that it cannot handle +// values less than 1, in practice it can handle as low as 1/x_max, where +// x_max is the largest representable input. In other words, the output range +// is symmetric. +template +inline gemmlowp::FixedPoint +log_x_for_x_greater_than_or_equal_to_1_impl( + gemmlowp::FixedPoint input_val) { + // assert(__builtin_clz(0u) >= std::numeric_limits::digits - 1); + // assert(__builtin_clz(0u) <= std::numeric_limits::digits); + using FixedPoint0 = gemmlowp::FixedPoint; + // The reason for accumulating the result with an extra bit of headroom is + // that z_pow_2_adj * log_2 might be saturated, and adding num_scaled * + // recip_denom will otherwise introduce an error. + static constexpr int kAccumIntegerBits = OutputIntegerBits + 1; + using FixedPointAccum = gemmlowp::FixedPoint; + + const FixedPoint0 log_2 = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( + FixedPoint0, 1488522236, std::log(2.0)); + const FixedPoint0 sqrt_sqrt_half = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( + FixedPoint0, 1805811301, std::sqrt(std::sqrt(0.5))); + const FixedPoint0 sqrt_half = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( + FixedPoint0, 1518500250, std::sqrt(0.5)); + const FixedPoint0 one_quarter = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(FixedPoint0, 536870912, 1.0 / 4.0); + + const FixedPoint0 alpha_n = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( + FixedPoint0, 117049297, 11.0 / 240.0 * std::sqrt(std::sqrt(2.0))); + const FixedPoint0 alpha_d = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( + FixedPoint0, 127690142, 1.0 / 20.0 * std::sqrt(std::sqrt(2.0))); + const FixedPoint0 alpha_i = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( + FixedPoint0, 1057819769, + 2.0 / std::sqrt(std::sqrt(2.0)) - std::sqrt(std::sqrt(2.0))); + const FixedPoint0 alpha_f = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( + FixedPoint0, 638450708, 1.0 / 4.0 * std::sqrt(std::sqrt(2.0))); + + const FixedPointAccum shifted_quarter = + gemmlowp::Rescale(one_quarter); + + // Reinterpret the input value as Q0.31, because we will figure out the + // required shift "ourselves" instead of using, say, Rescale. + FixedPoint0 z_a = FixedPoint0::FromRaw(input_val.raw()); + // z_a_pow_2 = input_integer_bits - z_a_headroom; + int z_a_headroom_plus_1 = CountLeadingZeros(static_cast(z_a.raw())); + FixedPoint0 r_a_tmp = + SaturatingRoundingMultiplyByPOTParam(z_a, (z_a_headroom_plus_1 - 1)); + const int32_t r_a_raw = + SaturatingRoundingMultiplyByPOTParam((r_a_tmp * sqrt_half).raw(), 1); + // z_pow_2_adj = max(z_pow_2_a - 0.75, z_pow_2_b - 0.25); + // z_pow_2_adj = max(InputIntegerBits - z_a_headroom_plus_1 + 0.25, + // InputIntegerBits - z_b_headroom - 0.25); + const FixedPointAccum z_a_pow_2_adj = SaturatingAddNonGemmlowp( + FixedPointAccum::FromRaw(SaturatingRoundingMultiplyByPOTParam( + InputIntegerBits - z_a_headroom_plus_1, 31 - kAccumIntegerBits)), + shifted_quarter); + + // z_b is treated like z_a, but premultiplying by sqrt(0.5). + FixedPoint0 z_b = z_a * sqrt_half; + int z_b_headroom = CountLeadingZeros(static_cast(z_b.raw())) - 1; + const int32_t r_b_raw = + SaturatingRoundingMultiplyByPOTParam(z_a.raw(), z_b_headroom); + const FixedPointAccum z_b_pow_2_adj = SaturatingSub( + FixedPointAccum::FromRaw(SaturatingRoundingMultiplyByPOTParam( + InputIntegerBits - z_b_headroom, 31 - kAccumIntegerBits)), + shifted_quarter); + + const FixedPoint0 r = FixedPoint0::FromRaw(std::min(r_a_raw, r_b_raw)); + const FixedPointAccum z_pow_2_adj = FixedPointAccum::FromRaw( + std::max(z_a_pow_2_adj.raw(), z_b_pow_2_adj.raw())); + + const FixedPoint0 p = gemmlowp::RoundingHalfSum(r, sqrt_sqrt_half); + FixedPoint0 q = r - sqrt_sqrt_half; + q = q + q; + + const FixedPoint0 common_sq = q * q; + const FixedPoint0 num = q * r + q * common_sq * alpha_n; + const FixedPoint0 denom_minus_one_0 = + p * (alpha_i + q + alpha_d * common_sq) + alpha_f * q; + const FixedPoint0 recip_denom = + one_over_one_plus_x_for_x_in_0_1(denom_minus_one_0); + + const FixedPointAccum num_scaled = gemmlowp::Rescale(num); + return gemmlowp::Rescale(z_pow_2_adj * log_2 + + num_scaled * recip_denom); +} + +template +inline gemmlowp::FixedPoint +log_x_for_x_greater_than_or_equal_to_1( + gemmlowp::FixedPoint input_val) { + static_assert( + OutputIntegerBits >= min_log_x_output_bits(InputIntegerBits), + "Output integer bits must be sufficient to accommodate logs of inputs."); + return log_x_for_x_greater_than_or_equal_to_1_impl( + input_val); +} + +inline int32_t GetReciprocal(int32_t x, int x_integer_digits, + int* num_bits_over_unit) { + int headroom_plus_one = CountLeadingZeros(static_cast(x)); + // This is the number of bits to the left of the binary point above 1.0. + // Consider x=1.25. In that case shifted_scale=0.8 and + // no later adjustment will be needed. + *num_bits_over_unit = x_integer_digits - headroom_plus_one; + const int32_t shifted_sum_minus_one = + static_cast((static_cast(x) << headroom_plus_one) - + (static_cast(1) << 31)); + + gemmlowp::FixedPoint shifted_scale = + gemmlowp::one_over_one_plus_x_for_x_in_0_1( + gemmlowp::FixedPoint::FromRaw(shifted_sum_minus_one)); + return shifted_scale.raw(); +} + +inline void GetInvSqrtQuantizedMultiplierExp(int32_t input, int reverse_shift, + int32_t* output_inv_sqrt, + int* output_shift) { + TFLITE_DCHECK_GE(input, 0); + if (input <= 1) { + // Handle the input value 1 separately to avoid overflow in that case + // in the general computation below (b/143972021). Also handle 0 as if it + // were a 1. 0 is an invalid input here (divide by zero) and 1 is a valid + // but rare/unrealistic input value. We can expect both to occur in some + // incompletely trained models, but probably not in fully trained models. + *output_inv_sqrt = std::numeric_limits::max(); + *output_shift = 0; + return; + } + TFLITE_DCHECK_GT(input, 1); + *output_shift = 11; + while (input >= (1 << 29)) { + input /= 4; + ++*output_shift; + } + const unsigned max_left_shift_bits = + CountLeadingZeros(static_cast(input)) - 1; + const unsigned max_left_shift_bit_pairs = max_left_shift_bits / 2; + const unsigned left_shift_bit_pairs = max_left_shift_bit_pairs - 1; + *output_shift -= left_shift_bit_pairs; + input <<= 2 * left_shift_bit_pairs; + TFLITE_DCHECK_GE(input, (1 << 27)); + TFLITE_DCHECK_LT(input, (1 << 29)); + using gemmlowp::FixedPoint; + using gemmlowp::Rescale; + using gemmlowp::SaturatingRoundingMultiplyByPOT; + // Using 3 integer bits gives us enough room for the internal arithmetic in + // this Newton-Raphson iteration. + using F3 = FixedPoint; + using F0 = FixedPoint; + const F3 fixedpoint_input = F3::FromRaw(input >> 1); + const F3 fixedpoint_half_input = + SaturatingRoundingMultiplyByPOT<-1>(fixedpoint_input); + const F3 fixedpoint_half_three = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F3, (1 << 28) + (1 << 27), 1.5); + // Newton-Raphson iteration + // Naive unoptimized starting guess: x = 1 + F3 x = F3::One(); + // Naive unoptimized number of iterations: 5 + for (int i = 0; i < 5; i++) { + const F3 x3 = Rescale<3>(x * x * x); + x = Rescale<3>(fixedpoint_half_three * x - fixedpoint_half_input * x3); + } + const F0 fixedpoint_half_sqrt_2 = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F0, 1518500250, std::sqrt(2.) / 2.); + x = x * fixedpoint_half_sqrt_2; + *output_inv_sqrt = x.raw(); + if (*output_shift < 0) { + *output_inv_sqrt <<= -*output_shift; + *output_shift = 0; + } + // Convert right shift (right is positive) to left shift. + *output_shift *= reverse_shift; +} + +// DO NOT USE THIS STRUCT FOR NEW FUNCTIONALITY BEYOND IMPLEMENTING +// BROADCASTING. +// +// NdArrayDesc describes the shape and memory layout of an N-dimensional +// rectangular array of numbers. +// +// NdArrayDesc is basically identical to Dims defined in types.h. +// However, as Dims is to be deprecated, this class exists as an adaptor +// to enable simple unoptimized implementations of element-wise broadcasting +// operations. +template +struct NdArrayDesc { + // The "extent" of each dimension. Indices along dimension d must be in the + // half-open interval [0, extents[d]). + int extents[N]; + + // The number of *elements* (not bytes) between consecutive indices of each + // dimension. + int strides[N]; +}; + +// DO NOT USE THIS FUNCTION FOR NEW FUNCTIONALITY BEYOND IMPLEMENTING +// BROADCASTING. +// +// Same as Offset(), except takes as NdArrayDesc instead of Dims. +inline int SubscriptToIndex(const NdArrayDesc<4>& desc, int i0, int i1, int i2, + int i3) { + TFLITE_DCHECK(i0 >= 0 && i0 < desc.extents[0]); + TFLITE_DCHECK(i1 >= 0 && i1 < desc.extents[1]); + TFLITE_DCHECK(i2 >= 0 && i2 < desc.extents[2]); + TFLITE_DCHECK(i3 >= 0 && i3 < desc.extents[3]); + return i0 * desc.strides[0] + i1 * desc.strides[1] + i2 * desc.strides[2] + + i3 * desc.strides[3]; +} + +inline int SubscriptToIndex(const NdArrayDesc<5>& desc, int indexes[5]) { + return indexes[0] * desc.strides[0] + indexes[1] * desc.strides[1] + + indexes[2] * desc.strides[2] + indexes[3] * desc.strides[3] + + indexes[4] * desc.strides[4]; +} + +inline int SubscriptToIndex(const NdArrayDesc<8>& desc, int indexes[8]) { + return indexes[0] * desc.strides[0] + indexes[1] * desc.strides[1] + + indexes[2] * desc.strides[2] + indexes[3] * desc.strides[3] + + indexes[4] * desc.strides[4] + indexes[5] * desc.strides[5] + + indexes[6] * desc.strides[6] + indexes[7] * desc.strides[7]; +} + +// Given the dimensions of the operands for an element-wise binary broadcast, +// adjusts them so that they can be directly iterated over with simple loops. +// Returns the adjusted dims as instances of NdArrayDesc in 'desc0_out' and +// 'desc1_out'. 'desc0_out' and 'desc1_out' cannot be nullptr. +// +// This function assumes that the two input shapes are compatible up to +// broadcasting and the shorter one has already been prepended with 1s to be the +// same length. E.g., if shape0 is (1, 16, 16, 64) and shape1 is (1, 64), +// shape1 must already have been prepended to be (1, 1, 1, 64). Recall that +// Dims refer to shapes in reverse order. In this case, input0_dims will be +// (64, 16, 16, 1) and input1_dims will be (64, 1, 1, 1). +// +// When two shapes are compatible up to broadcasting, for each dimension d, +// the input extents are either equal, or one of them is 1. +// +// This function performs the following for each dimension d: +// - If the extents are equal, then do nothing since the loop that walks over +// both of the input arrays is correct. +// - Otherwise, one (and only one) of the extents must be 1. Say extent0 is 1 +// and extent1 is e1. Then set extent0 to e1 and stride0 *to 0*. This allows +// array0 to be referenced *at any index* in dimension d and still access the +// same slice. +template +inline void NdArrayDescsForElementwiseBroadcast(const Dims& input0_dims, + const Dims& input1_dims, + NdArrayDesc* desc0_out, + NdArrayDesc* desc1_out) { + TFLITE_DCHECK(desc0_out != nullptr); + TFLITE_DCHECK(desc1_out != nullptr); + + // Copy dims to desc. + for (int i = 0; i < N; ++i) { + desc0_out->extents[i] = input0_dims.sizes[i]; + desc0_out->strides[i] = input0_dims.strides[i]; + desc1_out->extents[i] = input1_dims.sizes[i]; + desc1_out->strides[i] = input1_dims.strides[i]; + } + + // Walk over each dimension. If the extents are equal do nothing. + // Otherwise, set the desc with extent 1 to have extent equal to the other and + // stride 0. + for (int i = 0; i < N; ++i) { + const int extent0 = ArraySize(input0_dims, i); + const int extent1 = ArraySize(input1_dims, i); + if (extent0 != extent1) { + if (extent0 == 1) { + desc0_out->strides[i] = 0; + desc0_out->extents[i] = extent1; + } else { + TFLITE_DCHECK_EQ(extent1, 1); + desc1_out->strides[i] = 0; + desc1_out->extents[i] = extent0; + } + } + } +} + +// Copies dims to desc, calculating strides. +template +inline void CopyDimsToDesc(const RuntimeShape& input_shape, + NdArrayDesc* desc_out) { + int desc_stride = 1; + for (int i = N - 1; i >= 0; --i) { + desc_out->extents[i] = input_shape.Dims(i); + desc_out->strides[i] = desc_stride; + desc_stride *= input_shape.Dims(i); + } +} + +template +inline void NdArrayDescsForElementwiseBroadcast( + const RuntimeShape& input0_shape, const RuntimeShape& input1_shape, + NdArrayDesc* desc0_out, NdArrayDesc* desc1_out) { + TFLITE_DCHECK(desc0_out != nullptr); + TFLITE_DCHECK(desc1_out != nullptr); + + auto extended_input0_shape = RuntimeShape::ExtendedShape(N, input0_shape); + auto extended_input1_shape = RuntimeShape::ExtendedShape(N, input1_shape); + + // Copy dims to desc, calculating strides. + CopyDimsToDesc(extended_input0_shape, desc0_out); + CopyDimsToDesc(extended_input1_shape, desc1_out); + + // Walk over each dimension. If the extents are equal do nothing. + // Otherwise, set the desc with extent 1 to have extent equal to the other and + // stride 0. + for (int i = 0; i < N; ++i) { + const int extent0 = extended_input0_shape.Dims(i); + const int extent1 = extended_input1_shape.Dims(i); + if (extent0 != extent1) { + if (extent0 == 1) { + desc0_out->strides[i] = 0; + desc0_out->extents[i] = extent1; + } else { + TFLITE_DCHECK_EQ(extent1, 1); + desc1_out->strides[i] = 0; + desc1_out->extents[i] = extent0; + } + } + } +} + +template +inline void NdArrayDescsForElementwiseBroadcast( + const RuntimeShape& input0_shape, const RuntimeShape& input1_shape, + const RuntimeShape& input2_shape, NdArrayDesc* desc0_out, + NdArrayDesc* desc1_out, NdArrayDesc* desc2_out) { + TFLITE_DCHECK(desc0_out != nullptr); + TFLITE_DCHECK(desc1_out != nullptr); + TFLITE_DCHECK(desc2_out != nullptr); + + auto extended_input0_shape = RuntimeShape::ExtendedShape(N, input0_shape); + auto extended_input1_shape = RuntimeShape::ExtendedShape(N, input1_shape); + auto extended_input2_shape = RuntimeShape::ExtendedShape(N, input2_shape); + + // Copy dims to desc, calculating strides. + CopyDimsToDesc(extended_input0_shape, desc0_out); + CopyDimsToDesc(extended_input1_shape, desc1_out); + CopyDimsToDesc(extended_input2_shape, desc2_out); + + // Walk over each dimension. If the extents are equal do nothing. + // Otherwise, set the desc with extent 1 to have extent equal to the other and + // stride 0. + for (int i = 0; i < N; ++i) { + const int extent0 = extended_input0_shape.Dims(i); + const int extent1 = extended_input1_shape.Dims(i); + const int extent2 = extended_input2_shape.Dims(i); + + int extent = extent0; + if (extent1 != 1) extent = extent1; + if (extent2 != 1) extent = extent2; + + TFLITE_DCHECK(extent0 == 1 || extent0 == extent); + TFLITE_DCHECK(extent1 == 1 || extent1 == extent); + TFLITE_DCHECK(extent2 == 1 || extent2 == extent); + + if (!(extent0 == extent1 && extent1 == extent2)) { + if (extent0 == 1) { + desc0_out->strides[i] = 0; + desc0_out->extents[i] = extent; + } + if (extent1 == 1) { + desc1_out->strides[i] = 0; + desc1_out->extents[i] = extent; + } + if (extent2 == 1) { + desc2_out->strides[i] = 0; + desc2_out->extents[i] = extent; + } + } + } +} + +// Detailed implementation of NDOpsHelper, the indexes must be a zero array. +// This implementation is equivalent to N nested loops. Ex, if N=4, it can be +// re-writen as: +// for (int b = 0; b < output.extents[0]; ++b) { +// for (int y = 0; y < output.extents[1]; ++y) { +// for (int x = 0; x < output.extents[2]; ++x) { +// for (int c = 0; c < output.extents[3]; ++c) { +// calc({b,y,x,c}); +// } +// } +// } +// } +template +typename std::enable_if::type NDOpsHelperImpl( + const NdArrayDesc& output, const Calc& calc, int indexes[N]) { + for (indexes[DIM] = 0; indexes[DIM] < output.extents[DIM]; ++indexes[DIM]) { + NDOpsHelperImpl(output, calc, indexes); + } +} + +template +typename std::enable_if::type NDOpsHelperImpl( + const NdArrayDesc& output, const Calc& calc, int indexes[N]) { + for (indexes[DIM] = 0; indexes[DIM] < output.extents[DIM]; ++indexes[DIM]) { + calc(indexes); + } +} + +// Execute the calc function in the innermost iteration based on the shape of +// the output. The calc function should take a single argument of type int[N]. +template +inline void NDOpsHelper(const NdArrayDesc& output, const Calc& calc) { + int indexes[N] = {0}; + NDOpsHelperImpl(output, calc, indexes); +} +// Copied from gemmlowp::RoundDown when we dropped direct dependency on +// gemmlowp. +// +// Returns the runtime argument rounded down to the nearest multiple of +// the fixed Modulus. +template +Integer RoundDown(Integer i) { + return i - (i % Modulus); +} + +// Copied from gemmlowp::RoundUp when we dropped direct dependency on +// gemmlowp. +// +// Returns the runtime argument rounded up to the nearest multiple of +// the fixed Modulus. +template +Integer RoundUp(Integer i) { + return RoundDown(i + Modulus - 1); +} + +// Copied from gemmlowp::CeilQuotient when we dropped direct dependency on +// gemmlowp. +// +// Returns the quotient a / b rounded up ('ceil') to the nearest integer. +template +Integer CeilQuotient(Integer a, Integer b) { + return (a + b - 1) / b; +} + +// This function is a copy of gemmlowp::HowManyThreads, copied when we dropped +// the direct dependency of internal/optimized/ on gemmlowp. +// +// It computes a reasonable number of threads to use for a GEMM of shape +// (rows, cols, depth). +// +// TODO(b/131910176): get rid of this function by switching each call site +// to its own more sensible logic for its own workload. +template +inline int LegacyHowManyThreads(int max_num_threads, int rows, int cols, + int depth) { + // Early-exit in the default case where multi-threading is disabled. + if (max_num_threads == 1) { + return 1; + } + + // Ensure that each thread has KernelRows rows to process, if at all possible. + int thread_count = std::min(max_num_threads, rows / KernelRows); + + // Limit the number of threads according to the overall size of the problem. + if (thread_count > 1) { + // Empirically determined value. + static constexpr std::uint64_t min_cubic_size_per_thread = 64 * 1024; + + // We can only multiply two out of three sizes without risking overflow + const std::uint64_t cubic_size = + std::uint64_t(rows) * std::uint64_t(cols) * std::uint64_t(depth); + + thread_count = std::min( + thread_count, static_cast(cubic_size / min_cubic_size_per_thread)); + } + + if (thread_count < 1) { + thread_count = 1; + } + + assert(thread_count > 0 && thread_count <= max_num_threads); + return thread_count; +} + +template +void optimized_ops_preload_l1_stream(const T* ptr) { +#ifdef __GNUC__ + // builtin offered by GCC-compatible compilers including clang + __builtin_prefetch(ptr, /* 0 means read */ 0, /* 0 means no locality */ 0); +#else + (void)ptr; +#endif +} + +template +void optimized_ops_preload_l1_keep(const T* ptr) { +#ifdef __GNUC__ + // builtin offered by GCC-compatible compilers including clang + __builtin_prefetch(ptr, /* 0 means read */ 0, /* 3 means high locality */ 3); +#else + (void)ptr; +#endif +} + +template +void optimized_ops_prefetch_write_l1_keep(const T* ptr) { +#ifdef __GNUC__ + // builtin offered by GCC-compatible compilers including clang + __builtin_prefetch(ptr, /* 1 means write */ 1, /* 3 means high locality */ 3); +#else + (void)ptr; +#endif +} + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_COMMON_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h new file mode 100644 index 0000000..997b8fc --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h @@ -0,0 +1,115 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_ + +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" + +#ifndef TFLITE_DCHECK +#define TFLITE_DCHECK(condition) (condition) ? (void)0 : TFLITE_ASSERT_FALSE +#endif + +#ifndef TFLITE_DCHECK_EQ +#define TFLITE_DCHECK_EQ(x, y) ((x) == (y)) ? (void)0 : TFLITE_ASSERT_FALSE +#endif + +#ifndef TFLITE_DCHECK_NE +#define TFLITE_DCHECK_NE(x, y) ((x) != (y)) ? (void)0 : TFLITE_ASSERT_FALSE +#endif + +#ifndef TFLITE_DCHECK_GE +#define TFLITE_DCHECK_GE(x, y) ((x) >= (y)) ? (void)0 : TFLITE_ASSERT_FALSE +#endif + +#ifndef TFLITE_DCHECK_GT +#define TFLITE_DCHECK_GT(x, y) ((x) > (y)) ? (void)0 : TFLITE_ASSERT_FALSE +#endif + +#ifndef TFLITE_DCHECK_LE +#define TFLITE_DCHECK_LE(x, y) ((x) <= (y)) ? (void)0 : TFLITE_ASSERT_FALSE +#endif + +#ifndef TFLITE_DCHECK_LT +#define TFLITE_DCHECK_LT(x, y) ((x) < (y)) ? (void)0 : TFLITE_ASSERT_FALSE +#endif + +// TODO(ahentz): Clean up: We should stick to the DCHECK versions. +#ifndef TFLITE_CHECK +#define TFLITE_CHECK(condition) (condition) ? (void)0 : TFLITE_ABORT +#endif + +#ifndef TFLITE_CHECK_EQ +#define TFLITE_CHECK_EQ(x, y) ((x) == (y)) ? (void)0 : TFLITE_ABORT +#endif + +#ifndef TFLITE_CHECK_NE +#define TFLITE_CHECK_NE(x, y) ((x) != (y)) ? (void)0 : TFLITE_ABORT +#endif + +#ifndef TFLITE_CHECK_GE +#define TFLITE_CHECK_GE(x, y) ((x) >= (y)) ? (void)0 : TFLITE_ABORT +#endif + +#ifndef TFLITE_CHECK_GT +#define TFLITE_CHECK_GT(x, y) ((x) > (y)) ? (void)0 : TFLITE_ABORT +#endif + +#ifndef TFLITE_CHECK_LE +#define TFLITE_CHECK_LE(x, y) ((x) <= (y)) ? (void)0 : TFLITE_ABORT +#endif + +#ifndef TFLITE_CHECK_LT +#define TFLITE_CHECK_LT(x, y) ((x) < (y)) ? (void)0 : TFLITE_ABORT +#endif + +#ifndef TF_LITE_STATIC_MEMORY +// TODO(b/162019032): Consider removing these type-aliases. +using int8 = std::int8_t; +using uint8 = std::uint8_t; +using int16 = std::int16_t; +using uint16 = std::uint16_t; +using int32 = std::int32_t; +using uint32 = std::uint32_t; +#endif // !defined(TF_LITE_STATIC_MEMORY) + +// TFLITE_DEPRECATED() +// +// Duplicated from absl/base/macros.h to avoid pulling in that library. +// Marks a deprecated class, struct, enum, function, method and variable +// declarations. The macro argument is used as a custom diagnostic message (e.g. +// suggestion of a better alternative). +// +// Example: +// +// class TFLITE_DEPRECATED("Use Bar instead") Foo {...}; +// TFLITE_DEPRECATED("Use Baz instead") void Bar() {...} +// +// Every usage of a deprecated entity will trigger a warning when compiled with +// clang's `-Wdeprecated-declarations` option. This option is turned off by +// default, but the warnings will be reported by clang-tidy. +#if defined(__clang__) && __cplusplus >= 201103L +#define TFLITE_DEPRECATED(message) __attribute__((deprecated(message))) +#endif + +#ifndef TFLITE_DEPRECATED +#define TFLITE_DEPRECATED(message) +#endif + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h new file mode 100644 index 0000000..5b60967 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h @@ -0,0 +1,43 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_ + +#include + +namespace tflite { + +#if defined(TF_LITE_USE_GLOBAL_CMATH_FUNCTIONS) || \ + (defined(__ANDROID__) && !defined(__NDK_MAJOR__)) || defined(ARDUINO) || \ + defined(__ZEPHYR__) +#define TF_LITE_GLOBAL_STD_PREFIX +#else +#define TF_LITE_GLOBAL_STD_PREFIX std +#endif + +#define DECLARE_STD_GLOBAL_SWITCH1(tf_name, std_name) \ + template \ + inline T tf_name(const T x) { \ + return TF_LITE_GLOBAL_STD_PREFIX::std_name(x); \ + } + +DECLARE_STD_GLOBAL_SWITCH1(TfLiteRound, round); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/max.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/max.h new file mode 100644 index 0000000..afe5f08 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/max.h @@ -0,0 +1,38 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_ + +#include + +namespace tflite { + +#if defined(TF_LITE_USE_GLOBAL_MAX) || defined(__ZEPHYR__) +inline float TfLiteMax(const float& x, const float& y) { + return std::max(x, y); +} +#else +template +inline T TfLiteMax(const T& x, const T& y) { + return std::fmax(x, y); +} +#endif + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/min.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/min.h new file mode 100644 index 0000000..c92216c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/min.h @@ -0,0 +1,38 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_ + +#include + +namespace tflite { + +#if defined(TF_LITE_USE_GLOBAL_MIN) || defined(__ZEPHYR__) +inline float TfLiteMin(const float& x, const float& y) { + return std::min(x, y); +} +#else +template +inline T TfLiteMin(const T& x, const T& y) { + return std::fmin(x, y); +} +#endif + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/optimized/neon_check.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/optimized/neon_check.h new file mode 100644 index 0000000..825f366 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/optimized/neon_check.h @@ -0,0 +1,43 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_ + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) +#define USE_NEON +#include +#endif + +#if defined __GNUC__ && defined __SSE4_1__ && !defined TF_LITE_DISABLE_X86_NEON +#define USE_NEON +#include "NEON_2_SSE.h" +#endif + +// NEON_OR_PORTABLE(SomeFunc, args) calls NeonSomeFunc(args) if USE_NEON is +// defined, PortableSomeFunc(args) otherwise. +#ifdef USE_NEON +// Always use Neon code +#define NEON_OR_PORTABLE(funcname, ...) Neon##funcname(__VA_ARGS__) + +#else +// No NEON available: Use Portable code +#define NEON_OR_PORTABLE(funcname, ...) Portable##funcname(__VA_ARGS__) + +#endif // defined(USE_NEON) + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h new file mode 100644 index 0000000..a98109a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h @@ -0,0 +1,125 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_ + +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +inline RuntimeShape GetTensorShape(std::vector data) { + return RuntimeShape(data.size(), data.data()); +} + +// A list of tensors in a format that can be used by kernels like split and +// concatenation. +template +class VectorOfTensors { + public: + // Build with the tensors in 'tensor_list'. + VectorOfTensors(const TfLiteContext& context, + const TfLiteIntArray& tensor_list) { + int num_tensors = tensor_list.size; + + all_data_.reserve(num_tensors); + all_shape_.reserve(num_tensors); + all_shape_ptr_.reserve(num_tensors); + + for (int i = 0; i < num_tensors; ++i) { + TfLiteTensor* t = &context.tensors[tensor_list.data[i]]; + all_data_.push_back(GetTensorData(t)); + all_shape_.push_back(GetTensorShape(t)); + } + + // Taking the pointer from inside a std::vector is only OK if the vector is + // never modified, so we populate all_shape in the previous loop and then we + // are free to grab iterators here. + for (int i = 0; i < num_tensors; ++i) { + all_shape_ptr_.push_back(&all_shape_[i]); + } + } + // Return a pointer to the data pointers of all tensors in the list. For + // example: + // float* const* f = v.data(); + // f[0][1] is the second element of the first tensor. + T* const* data() const { return all_data_.data(); } + + // Return a pointer the shape pointers of all tensors in the list. For + // example: + // const RuntimeShape* const* d = v.dims(); + // dims[1] are the dimensions of the second tensor in the list. + const RuntimeShape* const* shapes() const { return all_shape_ptr_.data(); } + + private: + std::vector all_data_; + std::vector all_shape_; + std::vector all_shape_ptr_; +}; + +// A list of quantized tensors in a format that can be used by kernels like +// split and concatenation. +class VectorOfQuantizedTensors : public VectorOfTensors { + public: + // Build with the tensors in 'tensor_list'. + VectorOfQuantizedTensors(const TfLiteContext& context, + const TfLiteIntArray& tensor_list) + : VectorOfTensors(context, tensor_list) { + for (int i = 0; i < tensor_list.size; ++i) { + TfLiteTensor* t = &context.tensors[tensor_list.data[i]]; + zero_point_.push_back(t->params.zero_point); + scale_.push_back(t->params.scale); + } + } + + const float* scale() const { return scale_.data(); } + const int32_t* zero_point() const { return zero_point_.data(); } + + private: + std::vector zero_point_; + std::vector scale_; +}; + +// Writes randomly accessed values from `input` sequentially into `output`. +template +class SequentialTensorWriter { + public: + SequentialTensorWriter(const TfLiteTensor* input, TfLiteTensor* output) { + input_data_ = GetTensorData(input); + output_ptr_ = GetTensorData(output); + } + SequentialTensorWriter(const T* input_data, T* output_data) + : input_data_(input_data), output_ptr_(output_data) {} + + void Write(int position) { *output_ptr_++ = input_data_[position]; } + void WriteN(int position, int len) { + memcpy(output_ptr_, &input_data_[position], sizeof(T) * len); + output_ptr_ += len; + } + + private: + const T* input_data_; + T* output_ptr_; +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.cpp b/src/tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.cpp new file mode 100644 index 0000000..d92bf8f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.cpp @@ -0,0 +1,398 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" + +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" + +namespace tflite { + +namespace { +// These constants are used to manipulate the binary representation of doubles. +// Double-precision binary64 floating point format is: +// Bit | 63 | 62-52 | 51-0 | +// | Sign | Exponent | Fraction | +// To avoid 64-bit integers as much as possible, I break this into high and +// low 32-bit chunks. High is: +// Bit | 31 | 30-20 | 19-0 | +// | Sign | Exponent | High Fraction | +// Low is: +// Bit | 31-0 | +// | Low Fraction | +// We then access the components through logical bit-wise operations to +// extract the parts needed, with the positions and masks derived from the +// layout shown above. +constexpr uint64_t kSignMask = 0x8000000000000000LL; +constexpr uint64_t kExponentMask = 0x7ff0000000000000LL; +constexpr int32_t kExponentShift = 52; +constexpr int32_t kExponentBias = 1023; +constexpr uint32_t kExponentIsBadNum = 0x7ff; +constexpr uint64_t kFractionMask = 0x000fffffffc00000LL; +constexpr uint32_t kFractionShift = 22; +constexpr uint32_t kFractionRoundingMask = 0x003fffff; +constexpr uint32_t kFractionRoundingThreshold = 0x00200000; +} // namespace + +void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier, + int* shift) { + if (double_multiplier == 0.) { + *quantized_multiplier = 0; + *shift = 0; + return; + } +#ifdef TFLITE_EMULATE_FLOAT + // If we're trying to avoid the use of floating-point instructions (for + // example on microcontrollers) then use an alternative implementation + // that only requires integer and bitwise operations. To enable this, you + // need to set the define during the build process for your platform. + int64_t q_fixed = IntegerFrExp(double_multiplier, shift); +#else // TFLITE_EMULATE_FLOAT + const double q = std::frexp(double_multiplier, shift); + auto q_fixed = static_cast(TfLiteRound(q * (1ll << 31))); +#endif // TFLITE_EMULATE_FLOAT + TFLITE_CHECK(q_fixed <= (1ll << 31)); + if (q_fixed == (1ll << 31)) { + q_fixed /= 2; + ++*shift; + } + TFLITE_CHECK_LE(q_fixed, std::numeric_limits::max()); + // A shift amount smaller than -31 would cause all bits to be shifted out + // and thus all results would be zero. We implement that instead with + // q_fixed==0, so as to avoid hitting issues with right-shift + // operations with shift amounts greater than 31. Note that this happens + // roughly when abs(double_multiplier) < 2^-31 and the present handling means + // that we're effectively flushing tiny double_multiplier's to zero. + // We could conceivably handle values in the range (roughly) [32, 63] + // as 'denormals' i.e. (shift==0, q_fixed < 2^30). In that point of view + // the present handling is just doing 'flush denormals to zero'. We could + // reconsider and actually generate nonzero denormals if a need arises. + if (*shift < -31) { + *shift = 0; + q_fixed = 0; + } + *quantized_multiplier = static_cast(q_fixed); +} + +void QuantizeMultiplierGreaterThanOne(double double_multiplier, + int32_t* quantized_multiplier, + int* left_shift) { + TFLITE_CHECK_GT(double_multiplier, 1.); + QuantizeMultiplier(double_multiplier, quantized_multiplier, left_shift); + TFLITE_CHECK_GE(*left_shift, 0); +} + +void QuantizeMultiplierSmallerThanOneExp(double double_multiplier, + int32_t* quantized_multiplier, + int* left_shift) { + TFLITE_CHECK_LT(double_multiplier, 1.); + TFLITE_CHECK_GT(double_multiplier, 0.); + int shift; + QuantizeMultiplier(double_multiplier, quantized_multiplier, &shift); + TFLITE_CHECK_LE(shift, 0); + *left_shift = shift; +} + +int64_t IntegerFrExp(double input, int* shift) { + // Make sure our assumptions about the double layout hold. + TFLITE_CHECK_EQ(8, sizeof(double)); + + // We want to access the bits of the input double value directly, which is + // tricky to do safely, so use a union to handle the casting. + union { + double double_value; + uint64_t double_as_uint; + } cast_union; + cast_union.double_value = input; + const uint64_t u = cast_union.double_as_uint; + + // If the bitfield is all zeros apart from the sign bit, this is a normalized + // zero value, so return standard values for this special case. + if ((u & ~kSignMask) == 0) { + *shift = 0; + return 0; + } + + // Deal with NaNs and Infs, which are always indicated with a fixed pattern in + // the exponent, and distinguished by whether the fractions are zero or + // non-zero. + const uint32_t exponent_part = ((u & kExponentMask) >> kExponentShift); + if (exponent_part == kExponentIsBadNum) { + *shift = std::numeric_limits::max(); + if (u & kFractionMask) { + // NaN, so just return zero (with the exponent set to INT_MAX). + return 0; + } else { + // Infinity, so return +/- INT_MAX. + if (u & kSignMask) { + return std::numeric_limits::min(); + } else { + return std::numeric_limits::max(); + } + } + } + + // The shift is fairly easy to extract from the high bits of the double value, + // just by masking it out and applying a bias. The std::frexp() implementation + // always returns values between 0.5 and 1.0 though, whereas the exponent + // assumes 1.0 to 2.0 is the standard range, so I add on one to match that + // interface. + *shift = (exponent_part - kExponentBias) + 1; + + // There's an implicit high bit in the double format definition, so make sure + // we include that at the top, and then reconstruct the rest of the fractional + // value from the remaining fragments. + int64_t fraction = 0x40000000 + ((u & kFractionMask) >> kFractionShift); + + // We're cutting off some bits at the bottom, so to exactly match the standard + // frexp implementation here we'll apply rounding by adding one to the least + // significant bit of the result if the discarded portion is over half of the + // maximum. + if ((u & kFractionRoundingMask) > kFractionRoundingThreshold) { + fraction += 1; + } + // Negate the fraction if the sign bit was set. + if (u & kSignMask) { + fraction *= -1; + } + + return fraction; +} + +double DoubleFromFractionAndShift(int64_t fraction, int shift) { + union { + double double_value; + uint64_t double_as_uint; + } result; + + // Detect NaNs and infinities. + if (shift == std::numeric_limits::max()) { + if (fraction == 0) { + return std::numeric_limits::quiet_NaN(); + } else if (fraction > 0) { + return std::numeric_limits::infinity(); + } else { + return -std::numeric_limits::infinity(); + } + } + + // Return a normalized zero for a zero fraction. + if (fraction == 0) { + result.double_as_uint = 0; + return result.double_value; + } + + bool is_negative = (fraction < 0); + int64_t encoded_fraction = is_negative ? -fraction : fraction; + int64_t encoded_shift = (shift - 1); + while (encoded_fraction < 0x40000000) { + encoded_fraction *= 2; + encoded_shift -= 1; + } + while (encoded_fraction > 0x80000000) { + encoded_fraction /= 2; + encoded_shift += 1; + } + encoded_fraction -= 0x40000000; + if (encoded_shift < -1022) { + encoded_shift = -1023; + } else if (encoded_shift > 1022) { + encoded_shift = 1023; + } + encoded_shift += kExponentBias; + uint64_t encoded_sign = is_negative ? kSignMask : 0; + result.double_as_uint = encoded_sign | (encoded_shift << kExponentShift) | + (encoded_fraction << kFractionShift); + return result.double_value; +} + +double IntegerDoubleMultiply(double a, double b) { + int a_shift; + const int64_t a_fraction = IntegerFrExp(a, &a_shift); + int b_shift; + const int64_t b_fraction = IntegerFrExp(b, &b_shift); + // Detect NaNs and infinities. + if (a_shift == std::numeric_limits::max() || + (b_shift == std::numeric_limits::max())) { + return std::numeric_limits::quiet_NaN(); + } + const int result_shift = a_shift + b_shift + 1; + const int64_t result_fraction = (a_fraction * b_fraction) >> 32; + return DoubleFromFractionAndShift(result_fraction, result_shift); +} + +int IntegerDoubleCompare(double a, double b) { + int a_shift; + const int64_t a_fraction = IntegerFrExp(a, &a_shift); + int b_shift; + const int64_t b_fraction = IntegerFrExp(b, &b_shift); + + // Detect NaNs and infinities. + if (a_shift == std::numeric_limits::max() || + (b_shift == std::numeric_limits::max())) { + return 1; + } + + if ((a_fraction == 0) && (b_fraction < 0)) { + return 1; + } else if ((a_fraction < 0) && (b_fraction == 0)) { + return -1; + } else if (a_shift < b_shift) { + return -1; + } else if (a_shift > b_shift) { + return 1; + } else if (a_fraction < b_fraction) { + return -1; + } else if (a_fraction > b_fraction) { + return 1; + } else { + return 0; + } +} + +void PreprocessSoftmaxScaling(double beta, double input_scale, + int input_integer_bits, + int32_t* quantized_multiplier, int* left_shift) { + // If the overall multiplier (input and beta) is large, then exp() of an + // input difference of 1 scaled by this will be large. In other words, we + // can cap the multiplier and know that, when it is used, the output will be + // (round to) zero wherever the input is not at the maximum value. + + // If the overall scale is less than one, and input_integer_bits=0, then the + // result is double equivalent of Q0.31 (actually with more precision). Thus + // this generates a Q(input_integer_bits).(31-input_integer_bits) + // representation. +#ifdef TFLITE_EMULATE_FLOAT + const double input_beta = IntegerDoubleMultiply(beta, input_scale); + int shift; + int64_t fraction = IntegerFrExp(input_beta, &shift); + shift += (31 - input_integer_bits); + double input_beta_real_multiplier = + DoubleFromFractionAndShift(fraction, shift); + if (IntegerDoubleCompare(input_beta_real_multiplier, (1ll << 31) - 1.0) > 0) { + input_beta_real_multiplier = (1ll << 31) - 1.0; + } +#else // TFLITE_EMULATE_FLOAT + const double input_beta_real_multiplier = std::min( + beta * input_scale * (1 << (31 - input_integer_bits)), (1ll << 31) - 1.0); +#endif // TFLITE_EMULATE_FLOAT + + QuantizeMultiplierGreaterThanOne(input_beta_real_multiplier, + quantized_multiplier, left_shift); +} + +void PreprocessLogSoftmaxScalingExp(double beta, double input_scale, + int input_integer_bits, + int32_t* quantized_multiplier, + int* left_shift, + int32_t* reverse_scaling_divisor, + int* reverse_scaling_left_shift) { + PreprocessSoftmaxScaling(beta, input_scale, input_integer_bits, + quantized_multiplier, left_shift); + + // Also calculate what amounts to the inverse scaling factor for the input. + const double real_reverse_scaling_divisor = + (1 << (31 - *left_shift)) / static_cast(*quantized_multiplier); + tflite::QuantizeMultiplierSmallerThanOneExp(real_reverse_scaling_divisor, + reverse_scaling_divisor, + reverse_scaling_left_shift); +} + +int CalculateInputRadius(int input_integer_bits, int input_left_shift, + int total_signed_bits) { +#ifdef TFLITE_EMULATE_FLOAT + int64_t result = (1 << input_integer_bits) - 1; + result <<= (total_signed_bits - input_integer_bits); + result >>= input_left_shift; + return result; +#else // TFLITE_EMULATE_FLOAT + const double max_input_rescaled = + 1.0 * ((1 << input_integer_bits) - 1) * + (1ll << (total_signed_bits - input_integer_bits)) / + (1ll << input_left_shift); + // Tighten bound using floor. Suppose that we could use the exact value. + // After scaling the difference, the result would be at the maximum. Thus we + // must ensure that our value has lower magnitude. + return static_cast(std::floor(max_input_rescaled)); +#endif // TFLITE_EMULATE_FLOAT +} + +void NudgeQuantizationRange(const float min, const float max, + const int quant_min, const int quant_max, + float* nudged_min, float* nudged_max, + float* nudged_scale) { + // This code originates from tensorflow/core/kernels/fake_quant_ops_functor.h. + const float quant_min_float = static_cast(quant_min); + const float quant_max_float = static_cast(quant_max); + *nudged_scale = (max - min) / (quant_max_float - quant_min_float); + const float zero_point_from_min = quant_min_float - min / *nudged_scale; + uint16_t nudged_zero_point; + if (zero_point_from_min < quant_min_float) { + nudged_zero_point = static_cast(quant_min); + } else if (zero_point_from_min > quant_max_float) { + nudged_zero_point = static_cast(quant_max); + } else { + nudged_zero_point = static_cast(TfLiteRound(zero_point_from_min)); + } + *nudged_min = (quant_min_float - nudged_zero_point) * (*nudged_scale); + *nudged_max = (quant_max_float - nudged_zero_point) * (*nudged_scale); +} + +void FakeQuantizeArray(const float nudged_scale, const float nudged_min, + const float nudged_max, const float* input_data, + float* output_data, const float size) { + // This code originates from tensorflow/core/kernels/fake_quant_ops_functor.h. + const float inv_nudged_scale = 1.0f / nudged_scale; + + for (int i = 0; i < size; i++) { + const float src_val = input_data[i]; + const float clamped = std::min(nudged_max, std::max(nudged_min, src_val)); + const float clamped_shifted = clamped - nudged_min; + const float dst_val = + TfLiteRound(clamped_shifted * inv_nudged_scale) * nudged_scale + + nudged_min; + output_data[i] = dst_val; + } +} + +bool CheckedLog2(const float x, int* log2_result) { + // Using TfLiteRound instead of std::round and std::log instead of + // std::log2 to work around these functions being missing in a toolchain + // used in some TensorFlow tests as of May 2018. + const float x_log2 = std::log(x) * (1.0f / std::log(2.0f)); + const float x_log2_rounded = TfLiteRound(x_log2); + const float x_log2_fracpart = x_log2 - x_log2_rounded; + + *log2_result = static_cast(x_log2_rounded); + return std::abs(x_log2_fracpart) < 1e-3f; +} + +void QuantizeMultiplierArray(const double* effective_scales, size_t size, + int32_t* effective_scale_significand, + int* effective_shift) { + for (size_t i = 0; i < size; ++i) { + QuantizeMultiplier(effective_scales[i], &effective_scale_significand[i], + &effective_shift[i]); + } +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/quantization_util.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h similarity index 98% rename from src/tensorflow/lite/kernels/internal/quantization_util.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h index d380725..6aebd45 100644 --- a/src/tensorflow/lite/kernels/internal/quantization_util.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -19,9 +20,9 @@ limitations under the License. #include #include -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/round.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -290,3 +291,5 @@ void QuantizeMultiplierArray(const double* effective_scales, size_t size, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/add.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/add.h new file mode 100644 index 0000000..e9347af --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/add.h @@ -0,0 +1,449 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_ + +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { + +namespace reference_ops { + +template +inline void Add(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const T* input1_data, + const RuntimeShape& input2_shape, const T* input2_data, + const RuntimeShape& output_shape, T* output_data) { + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax( + input1_data[i] + input2_data[i], params.quantized_activation_min, + params.quantized_activation_max); + } +} + +inline void Add(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const float* input1_data, + const RuntimeShape& input2_shape, const float* input2_data, + const RuntimeShape& output_shape, float* output_data) { + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + for (int i = 0; i < flat_size; i++) { + auto x = input1_data[i] + input2_data[i]; + output_data[i] = ActivationFunctionWithMinMax( + x, params.float_activation_min, params.float_activation_max); + } +} + +// Element-wise add that can often be used for inner loop of broadcast add as +// well as the non-broadcast add. + +// This function is used for 8-bit as well as for 16-bit, but the accumulator +// is 32-bit for both cases. The overflow does not happen due to the +// choice of the shift (20 or 15, accordingly - see add.cc for more comments). +template +inline void AddElementwise(int size, const ArithmeticParams& params, + const T* input1_data, const T* input2_data, + T* output_data) { + TFLITE_DCHECK_GT(params.input1_offset, -std::numeric_limits::max()); + TFLITE_DCHECK_GT(params.input2_offset, -std::numeric_limits::max()); + TFLITE_DCHECK_LT(params.input1_offset, std::numeric_limits::max()); + TFLITE_DCHECK_LT(params.input2_offset, std::numeric_limits::max()); + + for (int i = 0; i < size; ++i) { + const int32_t input1_val = params.input1_offset + input1_data[i]; + const int32_t input2_val = params.input2_offset + input2_data[i]; + const int32_t shifted_input1_val = input1_val * (1 << params.left_shift); + const int32_t shifted_input2_val = input2_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, params.input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, params.input2_shift); + const int32_t raw_sum = scaled_input1_val + scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sum, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + output_data[i] = static_cast(clamped_output); + } +} + +// Scalar-broadcast add that can be used for inner loop of more general +// broadcast add, so that, for example, scalar-broadcast with batch will still +// be fast. +inline void AddScalarBroadcast(int size, const ArithmeticParams& params, + uint8_t input1_data, const uint8_t* input2_data, + uint8_t* output_data) { + TFLITE_DCHECK_GT(params.input1_offset, -256); + TFLITE_DCHECK_GT(params.input2_offset, -256); + TFLITE_DCHECK_LT(params.input1_offset, 256); + TFLITE_DCHECK_LT(params.input2_offset, 256); + + const int32_t input1_val = params.input1_offset + input1_data; + const int32_t shifted_input1_val = input1_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, params.input1_shift); + for (int i = 0; i < size; ++i) { + const int32_t input2_val = params.input2_offset + input2_data[i]; + const int32_t shifted_input2_val = input2_val * (1 << params.left_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, params.input2_shift); + const int32_t raw_sum = scaled_input1_val + scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sum, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + output_data[i] = static_cast(clamped_output); + } +} + +inline void Add(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const uint8_t* input1_data, + const RuntimeShape& input2_shape, const uint8_t* input2_data, + const RuntimeShape& output_shape, uint8_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + + TFLITE_DCHECK_GT(params.input1_offset, -256); + TFLITE_DCHECK_GT(params.input2_offset, -256); + TFLITE_DCHECK_LT(params.input1_offset, 256); + TFLITE_DCHECK_LT(params.input2_offset, 256); + AddElementwise(flat_size, params, input1_data, input2_data, output_data); +} + +inline void AddGeneralParamScale(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const int16_t* input1_data, + const RuntimeShape& input2_shape, + const int16_t* input2_data, + const RuntimeShape& output_shape, + int16_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + + int max_value = std::numeric_limits::max(); + + TFLITE_DCHECK_GT(params.input1_offset, -max_value); + TFLITE_DCHECK_GT(params.input2_offset, -max_value); + TFLITE_DCHECK_LT(params.input1_offset, max_value); + TFLITE_DCHECK_LT(params.input2_offset, max_value); + AddElementwise(flat_size, params, input1_data, input2_data, output_data); +} + +inline void Add(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const int16_t* input1_data, + const RuntimeShape& input2_shape, const int16_t* input2_data, + const RuntimeShape& output_shape, int16_t* output_data, + bool pot_scale = true) { + if (!pot_scale) { + AddGeneralParamScale(params, input1_shape, input1_data, input2_shape, + input2_data, output_shape, output_data); + return; + } + + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + + const int input1_shift = params.input1_shift; + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + const int16_t output_activation_min = params.quantized_activation_min; + const int16_t output_activation_max = params.quantized_activation_max; + + TFLITE_DCHECK(input1_shift == 0 || params.input2_shift == 0); + TFLITE_DCHECK_LE(input1_shift, 0); + TFLITE_DCHECK_LE(params.input2_shift, 0); + const int16_t* not_shift_input = + input1_shift == 0 ? input1_data : input2_data; + const int16_t* shift_input = input1_shift == 0 ? input2_data : input1_data; + const int input_right_shift = + input1_shift == 0 ? -params.input2_shift : -input1_shift; + + for (int i = 0; i < flat_size; i++) { + // F0 uses 0 integer bits, range [-1, 1]. + using F0 = gemmlowp::FixedPoint; + + F0 input_ready_scaled = F0::FromRaw(not_shift_input[i]); + F0 scaled_input = F0::FromRaw( + gemmlowp::RoundingDivideByPOT(shift_input[i], input_right_shift)); + F0 result = gemmlowp::SaturatingAdd(scaled_input, input_ready_scaled); + const int16_t raw_output = result.raw(); + const int16_t clamped_output = std::min( + output_activation_max, std::max(output_activation_min, raw_output)); + output_data[i] = clamped_output; + } +} + +inline void BroadcastAdd4DSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const float* input1_data, + const RuntimeShape& input2_shape, + const float* input2_data, + const RuntimeShape& output_shape, + float* output_data) { + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + const RuntimeShape extended_output_shape = + RuntimeShape::ExtendedShape(4, output_shape); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + for (int b = 0; b < extended_output_shape.Dims(0); ++b) { + for (int y = 0; y < extended_output_shape.Dims(1); ++y) { + for (int x = 0; x < extended_output_shape.Dims(2); ++x) { + for (int c = 0; c < extended_output_shape.Dims(3); ++c) { + output_data[Offset(extended_output_shape, b, y, x, c)] = + ActivationFunctionWithMinMax( + input1_data[SubscriptToIndex(desc1, b, y, x, c)] + + input2_data[SubscriptToIndex(desc2, b, y, x, c)], + params.float_activation_min, params.float_activation_max); + } + } + } + } +} + +inline void BroadcastAdd4DSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const int32_t* input1_data, + const RuntimeShape& input2_shape, + const int32_t* input2_data, + const RuntimeShape& output_shape, + int32_t* output_data) { + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + const RuntimeShape extended_output_shape = + RuntimeShape::ExtendedShape(4, output_shape); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + for (int b = 0; b < extended_output_shape.Dims(0); ++b) { + for (int y = 0; y < extended_output_shape.Dims(1); ++y) { + for (int x = 0; x < extended_output_shape.Dims(2); ++x) { + for (int c = 0; c < extended_output_shape.Dims(3); ++c) { + output_data[Offset(extended_output_shape, b, y, x, c)] = + ActivationFunctionWithMinMax( + input1_data[SubscriptToIndex(desc1, b, y, x, c)] + + input2_data[SubscriptToIndex(desc2, b, y, x, c)], + params.quantized_activation_min, + params.quantized_activation_max); + } + } + } + } +} + +// This function is used for 8-bit as well as for 16-bit, but the accumulator +// is 32-bit for both cases. The overflow does not happen due to the +// choice of the shift (20 or 15, accordingly - see add.cc for more comments). +template +inline void BroadcastAdd4DSlow( + const ArithmeticParams& params, const RuntimeShape& input1_shape, + const T* input1_data, const RuntimeShape& input2_shape, + const T* input2_data, const RuntimeShape& output_shape, T* output_data) { + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + const RuntimeShape extended_output_shape = + RuntimeShape::ExtendedShape(4, output_shape); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + for (int b = 0; b < extended_output_shape.Dims(0); ++b) { + for (int y = 0; y < extended_output_shape.Dims(1); ++y) { + for (int x = 0; x < extended_output_shape.Dims(2); ++x) { + for (int c = 0; c < extended_output_shape.Dims(3); ++c) { + const int32_t input1_val = + params.input1_offset + + input1_data[SubscriptToIndex(desc1, b, y, x, c)]; + const int32_t input2_val = + params.input2_offset + + input2_data[SubscriptToIndex(desc2, b, y, x, c)]; + const int32_t shifted_input1_val = + input1_val * (1 << params.left_shift); + const int32_t shifted_input2_val = + input2_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, + params.input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, + params.input2_shift); + const int32_t raw_sum = scaled_input1_val + scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sum, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + output_data[Offset(extended_output_shape, b, y, x, c)] = + static_cast(clamped_output); + } + } + } + } +} + +inline void BroadcastAddFivefold(const ArithmeticParams& unswitched_params, + const RuntimeShape& unswitched_input1_shape, + const uint8_t* unswitched_input1_data, + const RuntimeShape& unswitched_input2_shape, + const uint8_t* unswitched_input2_data, + const RuntimeShape& output_shape, + uint8_t* output_data) { + ArithmeticParams switched_params = unswitched_params; + switched_params.input1_offset = unswitched_params.input2_offset; + switched_params.input1_multiplier = unswitched_params.input2_multiplier; + switched_params.input1_shift = unswitched_params.input2_shift; + switched_params.input2_offset = unswitched_params.input1_offset; + switched_params.input2_multiplier = unswitched_params.input1_multiplier; + switched_params.input2_shift = unswitched_params.input1_shift; + + const bool use_unswitched = + unswitched_params.broadcast_category == + tflite::BroadcastableOpCategory::kFirstInputBroadcastsFast; + + const ArithmeticParams& params = + use_unswitched ? unswitched_params : switched_params; + const uint8_t* input1_data = + use_unswitched ? unswitched_input1_data : unswitched_input2_data; + const uint8_t* input2_data = + use_unswitched ? unswitched_input2_data : unswitched_input1_data; + + // Fivefold nested loops. The second input resets its position for each + // iteration of the second loop. The first input resets its position at the + // beginning of the fourth loop. The innermost loop is an elementwise add of + // sections of the arrays. + uint8_t* output_data_ptr = output_data; + const uint8_t* input1_data_ptr = input1_data; + const uint8_t* input2_data_reset = input2_data; + // In the fivefold pattern, y0, y2 and y4 are not broadcast, and so shared + // between input shapes. y3 for input 1 is always broadcast, and so the + // dimension there is 1, whereas optionally y1 might be broadcast for input 2. + // Put another way, + // input1.shape.FlatSize = y0 * y1 * y2 * y4, + // input2.shape.FlatSize = y0 * y2 * y3 * y4. + int y0 = params.broadcast_shape[0]; + int y1 = params.broadcast_shape[1]; + int y2 = params.broadcast_shape[2]; + int y3 = params.broadcast_shape[3]; + int y4 = params.broadcast_shape[4]; + if (y4 > 1) { + // General fivefold pattern, with y4 > 1 so there is a non-broadcast inner + // dimension. + for (int i0 = 0; i0 < y0; ++i0) { + const uint8_t* input2_data_ptr; + for (int i1 = 0; i1 < y1; ++i1) { + input2_data_ptr = input2_data_reset; + for (int i2 = 0; i2 < y2; ++i2) { + for (int i3 = 0; i3 < y3; ++i3) { + AddElementwise(y4, params, input1_data_ptr, input2_data_ptr, + output_data_ptr); + input2_data_ptr += y4; + output_data_ptr += y4; + } + // We have broadcast y4 of input1 data y3 times, and now move on. + input1_data_ptr += y4; + } + } + // We have broadcast y2*y3*y4 of input2 data y1 times, and now move on. + input2_data_reset = input2_data_ptr; + } + } else { + // Special case of y4 == 1, in which the innermost loop is a single element + // and can be combined with the next (y3) as an inner broadcast. + // + // Note that this handles the case of pure scalar broadcast when + // y0 == y1 == y2 == 1. With low overhead it handles cases such as scalar + // broadcast with batch (as y2 > 1). + // + // NOTE The process is the same as the above general case except simplified + // for y4 == 1 and the loop over y3 is contained within the + // AddScalarBroadcast function. + for (int i0 = 0; i0 < y0; ++i0) { + const uint8_t* input2_data_ptr; + for (int i1 = 0; i1 < y1; ++i1) { + input2_data_ptr = input2_data_reset; + for (int i2 = 0; i2 < y2; ++i2) { + AddScalarBroadcast(y3, params, *input1_data_ptr, input2_data_ptr, + output_data_ptr); + input2_data_ptr += y3; + output_data_ptr += y3; + input1_data_ptr += 1; + } + } + input2_data_reset = input2_data_ptr; + } + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/arg_min_max.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/arg_min_max.h similarity index 95% rename from src/tensorflow/lite/kernels/internal/reference/arg_min_max.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/arg_min_max.h index e6f34fd..c237246 100644 --- a/src/tensorflow/lite/kernels/internal/reference/arg_min_max.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/arg_min_max.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_ -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -66,3 +67,5 @@ void ArgMinMax(const RuntimeShape& input1_shape, const T1* input1_data, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/binary_function.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/binary_function.h new file mode 100644 index 0000000..fa537ec --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/binary_function.h @@ -0,0 +1,83 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +// Also appears to duplicate MinimumMaximum. +// +// R: Result type. T1: Input 1 type. T2: Input 2 type. +template +inline void BroadcastBinaryFunction4DSlow( + const RuntimeShape& unextended_input1_shape, const T1* input1_data, + const RuntimeShape& unextended_input2_shape, const T2* input2_data, + const RuntimeShape& unextended_output_shape, R* output_data, + R (*func)(T1, T2)) { + TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(4, unextended_output_shape); + + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(unextended_input1_shape, + unextended_input2_shape, &desc1, &desc2); + + for (int b = 0; b < output_shape.Dims(0); ++b) { + for (int y = 0; y < output_shape.Dims(1); ++y) { + for (int x = 0; x < output_shape.Dims(2); ++x) { + for (int c = 0; c < output_shape.Dims(3); ++c) { + auto out_idx = Offset(output_shape, b, y, x, c); + auto in1_idx = SubscriptToIndex(desc1, b, y, x, c); + auto in2_idx = SubscriptToIndex(desc2, b, y, x, c); + auto in1_val = input1_data[in1_idx]; + auto in2_val = input2_data[in2_idx]; + output_data[out_idx] = func(in1_val, in2_val); + } + } + } + } +} + +// R: Result type. T1: Input 1 type. T2: Input 2 type. +template +inline void BinaryFunction(const RuntimeShape& input1_shape, + const T1* input1_data, + const RuntimeShape& input2_shape, + const T2* input2_data, + const RuntimeShape& output_shape, R* output_data, + R (*func)(T1, T2)) { + const int flat_size = + MatchingFlatSize(input1_shape, input2_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = func(input1_data[i], input2_data[i]); + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/ceil.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/ceil.h similarity index 91% rename from src/tensorflow/lite/kernels/internal/reference/ceil.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/ceil.h index 66d1dc3..ebc26ab 100644 --- a/src/tensorflow/lite/kernels/internal/reference/ceil.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/ceil.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,7 +18,7 @@ limitations under the License. #include -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -35,3 +36,5 @@ inline void Ceil(const RuntimeShape& input_shape, const float* input_data, } // namespace reference_ops } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/comparisons.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/comparisons.h new file mode 100644 index 0000000..5ba7a7e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/comparisons.h @@ -0,0 +1,283 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +template +inline bool EqualFn(T lhs, T rhs) { + return lhs == rhs; +} + +template +inline bool NotEqualFn(T lhs, T rhs) { + return lhs != rhs; +} + +template +inline bool GreaterFn(T lhs, T rhs) { + return lhs > rhs; +} +template +inline bool GreaterEqualFn(T lhs, T rhs) { + return lhs >= rhs; +} +template +inline bool LessFn(T lhs, T rhs) { + return lhs < rhs; +} +template +inline bool LessEqualFn(T lhs, T rhs) { + return lhs <= rhs; +} + +template +using ComparisonFn = bool (*)(T, T); + +template F> +inline void ComparisonImpl( + const ComparisonParams& op_params, const RuntimeShape& input1_shape, + const T* input1_data, const RuntimeShape& input2_shape, + const T* input2_data, const RuntimeShape& output_shape, bool* output_data) { + const int64_t flatsize = + MatchingFlatSize(input1_shape, input2_shape, output_shape); + for (int64_t i = 0; i < flatsize; ++i) { + output_data[i] = F(input1_data[i], input2_data[i]); + } +} + +template F> +inline void Comparison(const ComparisonParams& op_params, + const RuntimeShape& input1_shape, + const float* input1_data, + const RuntimeShape& input2_shape, + const float* input2_data, + const RuntimeShape& output_shape, bool* output_data) { + ComparisonImpl(op_params, input1_shape, input1_data, input2_shape, + input2_data, output_shape, output_data); +} + +template F> +inline void ComparisonWithScaling( + const ComparisonParams& op_params, const RuntimeShape& input1_shape, + const T* input1_data, const RuntimeShape& input2_shape, + const T* input2_data, const RuntimeShape& output_shape, bool* output_data) { + int left_shift = op_params.left_shift; + int32_t input1_offset = op_params.input1_offset; + int32_t input1_multiplier = op_params.input1_multiplier; + int input1_shift = op_params.input1_shift; + int32_t input2_offset = op_params.input2_offset; + int32_t input2_multiplier = op_params.input2_multiplier; + int input2_shift = op_params.input2_shift; + + const int64_t flatsize = + MatchingFlatSize(input1_shape, input2_shape, output_shape); + for (int64_t i = 0; i < flatsize; ++i) { + const int32_t input1_val = input1_offset + input1_data[i]; + const int32_t input2_val = input2_offset + input2_data[i]; + const int32_t shifted_input1_val = input1_val * (1 << left_shift); + const int32_t shifted_input2_val = input2_val * (1 << left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, input1_multiplier, input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, input2_multiplier, input2_shift); + output_data[i] = F(scaled_input1_val, scaled_input2_val); + } +} + +struct BroadcastComparison4DSlowCommon { + const RuntimeShape output_shape; + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; +}; + +inline BroadcastComparison4DSlowCommon BroadcastComparison4DSlowPreprocess( + const RuntimeShape& unextended_input1_shape, + const RuntimeShape& unextended_input2_shape, + const RuntimeShape& unextended_output_shape) { + TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4); + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(unextended_input1_shape, + unextended_input2_shape, &desc1, &desc2); + return {RuntimeShape::ExtendedShape(4, unextended_output_shape), desc1, + desc2}; +} + +template F> +inline void BroadcastComparison4DSlowImpl( + const ComparisonParams& op_params, + const RuntimeShape& unextended_input1_shape, const T* input1_data, + const RuntimeShape& unextended_input2_shape, const T* input2_data, + const RuntimeShape& unextended_output_shape, bool* output_data) { + const BroadcastComparison4DSlowCommon dims = + BroadcastComparison4DSlowPreprocess(unextended_input1_shape, + unextended_input2_shape, + unextended_output_shape); + + for (int b = 0; b < dims.output_shape.Dims(0); ++b) { + for (int y = 0; y < dims.output_shape.Dims(1); ++y) { + for (int x = 0; x < dims.output_shape.Dims(2); ++x) { + for (int c = 0; c < dims.output_shape.Dims(3); ++c) { + output_data[Offset(dims.output_shape, b, y, x, c)] = + F(input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)], + input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)]); + } + } + } + } +} + +template F> +inline void BroadcastComparison4DSlow(const ComparisonParams& op_params, + const RuntimeShape& input1_shape, + const float* input1_data, + const RuntimeShape& input2_shape, + const float* input2_data, + const RuntimeShape& output_shape, + bool* output_data) { + BroadcastComparison4DSlowImpl(op_params, input1_shape, input1_data, + input2_shape, input2_data, + output_shape, output_data); +} + +template F> +inline void BroadcastComparison4DSlowWithScaling( + const ComparisonParams& op_params, + const RuntimeShape& unextended_input1_shape, const T* input1_data, + const RuntimeShape& unextended_input2_shape, const T* input2_data, + const RuntimeShape& unextended_output_shape, bool* output_data) { + const BroadcastComparison4DSlowCommon dims = + BroadcastComparison4DSlowPreprocess(unextended_input1_shape, + unextended_input2_shape, + unextended_output_shape); + + int left_shift = op_params.left_shift; + int32_t input1_offset = op_params.input1_offset; + int32_t input1_multiplier = op_params.input1_multiplier; + int input1_shift = op_params.input1_shift; + int32_t input2_offset = op_params.input2_offset; + int32_t input2_multiplier = op_params.input2_multiplier; + int input2_shift = op_params.input2_shift; + + for (int b = 0; b < dims.output_shape.Dims(0); ++b) { + for (int y = 0; y < dims.output_shape.Dims(1); ++y) { + for (int x = 0; x < dims.output_shape.Dims(2); ++x) { + for (int c = 0; c < dims.output_shape.Dims(3); ++c) { + const int32_t input1_val = + input1_offset + + input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)]; + const int32_t input2_val = + input2_offset + + input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)]; + const int32_t shifted_input1_val = input1_val * (1 << left_shift); + const int32_t shifted_input2_val = input2_val * (1 << left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, input1_multiplier, input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, input2_multiplier, input2_shift); + output_data[Offset(dims.output_shape, b, y, x, c)] = + F(scaled_input1_val, scaled_input2_val); + } + } + } + } +} + +#define TFLITE_COMPARISON_OP(name) \ + inline void name(const ComparisonParams& op_params, \ + const RuntimeShape& input1_shape, const float* input1_data, \ + const RuntimeShape& input2_shape, const float* input2_data, \ + const RuntimeShape& output_shape, bool* output_data) { \ + Comparison(op_params, input1_shape, input1_data, input2_shape, \ + input2_data, output_shape, output_data); \ + } \ + template \ + inline void name##NoScaling( \ + const ComparisonParams& op_params, const RuntimeShape& input1_shape, \ + const T* input1_data, const RuntimeShape& input2_shape, \ + const T* input2_data, const RuntimeShape& output_shape, \ + bool* output_data) { \ + ComparisonImpl(op_params, input1_shape, input1_data, \ + input2_shape, input2_data, output_shape, \ + output_data); \ + } \ + template \ + inline void name##WithScaling( \ + const ComparisonParams& op_params, const RuntimeShape& input1_shape, \ + const T* input1_data, const RuntimeShape& input2_shape, \ + const T* input2_data, const RuntimeShape& output_shape, \ + bool* output_data) { \ + ComparisonWithScaling(op_params, input1_shape, input1_data, \ + input2_shape, input2_data, \ + output_shape, output_data); \ + } \ + template \ + inline void Broadcast4DSlow##name##NoScaling( \ + const ComparisonParams& op_params, const RuntimeShape& input1_shape, \ + const T* input1_data, const RuntimeShape& input2_shape, \ + const T* input2_data, const RuntimeShape& output_shape, \ + bool* output_data) { \ + BroadcastComparison4DSlowImpl( \ + op_params, input1_shape, input1_data, input2_shape, input2_data, \ + output_shape, output_data); \ + } \ + inline void Broadcast4DSlow##name( \ + const ComparisonParams& op_params, const RuntimeShape& input1_shape, \ + const float* input1_data, const RuntimeShape& input2_shape, \ + const float* input2_data, const RuntimeShape& output_shape, \ + bool* output_data) { \ + BroadcastComparison4DSlow(op_params, input1_shape, input1_data, \ + input2_shape, input2_data, \ + output_shape, output_data); \ + } \ + template \ + inline void Broadcast4DSlow##name##WithScaling( \ + const ComparisonParams& op_params, const RuntimeShape& input1_shape, \ + const T* input1_data, const RuntimeShape& input2_shape, \ + const T* input2_data, const RuntimeShape& output_shape, \ + bool* output_data) { \ + BroadcastComparison4DSlowWithScaling( \ + op_params, input1_shape, input1_data, input2_shape, input2_data, \ + output_shape, output_data); \ + } +TFLITE_COMPARISON_OP(Equal); +TFLITE_COMPARISON_OP(NotEqual); +TFLITE_COMPARISON_OP(Greater); +TFLITE_COMPARISON_OP(GreaterEqual); +TFLITE_COMPARISON_OP(Less); +TFLITE_COMPARISON_OP(LessEqual); +#undef TFLITE_COMPARISON_OP + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/concatenation.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/concatenation.h similarity index 82% rename from src/tensorflow/lite/kernels/internal/reference/concatenation.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/concatenation.h index 82b95b6..4d78120 100644 --- a/src/tensorflow/lite/kernels/internal/reference/concatenation.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/concatenation.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -16,9 +17,10 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { namespace reference_ops { @@ -67,20 +69,19 @@ inline void Concatenation(const ConcatenationParams& params, } } -// TODO(prabhumk): This is the same as the optimized implementation. -// TODO(prabhumk): The quantized implementation of concatentation isn't fully +// TODO(b/174275780): The quantized implementation of concatentation isn't fully // quantized as it takes scale as a floating point value. This should be fixed // when optimizng this routine further. inline void ConcatenationWithScaling(const ConcatenationParams& params, const RuntimeShape* const* input_shapes, - const uint8* const* input_data, + const uint8_t* const* input_data, const RuntimeShape& output_shape, - uint8* output_data) { + uint8_t* output_data) { int axis = params.axis; - const int32* input_zeropoint = params.input_zeropoint; + const int32_t* input_zeropoint = params.input_zeropoint; const float* input_scale = params.input_scale; int inputs_count = params.inputs_count; - const int32 output_zeropoint = params.output_zeropoint; + const int32_t output_zeropoint = params.output_zeropoint; const float output_scale = params.output_scale; const int concat_dimensions = output_shape.DimensionsCount(); @@ -109,11 +110,11 @@ inline void ConcatenationWithScaling(const ConcatenationParams& params, } const float inverse_output_scale = 1.f / output_scale; - uint8* output_ptr = output_data; + uint8_t* output_ptr = output_data; for (int k = 0; k < outer_size; k++) { for (int i = 0; i < inputs_count; ++i) { const int copy_size = input_shapes[i]->Dims(axis) * base_inner_size; - const uint8* input_ptr = input_data[i] + k * copy_size; + const uint8_t* input_ptr = input_data[i] + k * copy_size; if (input_zeropoint[i] == output_zeropoint && input_scale[i] == output_scale) { memcpy(output_ptr, input_ptr, copy_size); @@ -121,9 +122,9 @@ inline void ConcatenationWithScaling(const ConcatenationParams& params, const float scale = input_scale[i] * inverse_output_scale; const float bias = -input_zeropoint[i] * scale; for (int j = 0; j < copy_size; ++j) { - const int32_t value = - static_cast(round(input_ptr[j] * scale + bias)) + - output_zeropoint; + const int32_t value = static_cast(tflite::TfLiteRound( + input_ptr[j] * scale + bias)) + + output_zeropoint; output_ptr[j] = static_cast( std::max(std::min(255, value), 0)); } @@ -137,3 +138,5 @@ inline void ConcatenationWithScaling(const ConcatenationParams& params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/conv.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/conv.h new file mode 100644 index 0000000..ee3dda2 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/conv.h @@ -0,0 +1,267 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline void Conv(const ConvParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& filter_shape, + const float* filter_data, const RuntimeShape& bias_shape, + const float* bias_data, const RuntimeShape& output_shape, + float* output_data, const RuntimeShape& im2col_shape, + float* im2col_data) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const float output_activation_min = params.float_activation_min; + const float output_activation_max = params.float_activation_max; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + (void)im2col_data; // only used in optimized code. + (void)im2col_shape; // only used in optimized code. + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + const int in_y_origin = (out_y * stride_height) - pad_height; + for (int out_x = 0; out_x < output_width; ++out_x) { + const int in_x_origin = (out_x * stride_width) - pad_width; + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + float total = 0.f; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + const int in_y = in_y_origin + dilation_height_factor * filter_y; + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + + if (!is_point_inside_image) { + continue; + } + + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + float input_value = input_data[Offset(input_shape, batch, in_y, + in_x, in_channel)]; + float filter_value = filter_data[Offset( + filter_shape, out_channel, filter_y, filter_x, in_channel)]; + total += (input_value * filter_value); + } + } + } + float bias_value = 0.0f; + if (bias_data) { + bias_value = bias_data[out_channel]; + } + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + ActivationFunctionWithMinMax(total + bias_value, + output_activation_min, + output_activation_max); + } + } + } + } +} + +inline void Conv(const ConvParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& filter_shape, + const uint8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + uint8_t* output_data, const RuntimeShape& im2col_shape, + uint8_t* im2col_data, void* cpu_backend_context) { + (void)cpu_backend_context; // only used in optimized code. + (void)im2col_data; // only used in optimized code. + (void)im2col_shape; // only used in optimized code. + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int32_t input_offset = params.input_offset; + const int32_t filter_offset = params.weights_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + const int in_y_origin = (out_y * stride_height) - pad_height; + for (int out_x = 0; out_x < output_width; ++out_x) { + const int in_x_origin = (out_x * stride_width) - pad_width; + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + int32_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + const int in_y = in_y_origin + dilation_height_factor * filter_y; + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + + if (!is_point_inside_image) { + continue; + } + + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + int32_t input_val = input_data[Offset(input_shape, batch, in_y, + in_x, in_channel)]; + int32_t filter_val = filter_data[Offset( + filter_shape, out_channel, filter_y, filter_x, in_channel)]; + acc += + (filter_val + filter_offset) * (input_val + input_offset); + } + } + } + if (bias_data) { + acc += bias_data[out_channel]; + } + acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, + output_shift); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + static_cast(acc); + } + } + } + } +} + +inline void HybridConvPerChannel( + const ConvParams& params, float* scaling_factors_ptr, + const RuntimeShape& input_shape, const int8_t* input_data, + const RuntimeShape& filter_shape, const int8_t* filter_data, + const RuntimeShape& bias_shape, const float* bias_data, + const RuntimeShape& output_shape, float* output_data, + const RuntimeShape& im2col_shape, int8_t* im2col_data, + const float* per_channel_scale, int32_t* input_offset) { + (void)im2col_data; // only used in optimized code. + (void)im2col_shape; // only used in optimized code. + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const float output_activation_min = params.float_activation_min; + const float output_activation_max = params.float_activation_max; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + const int in_x_origin = (out_x * stride_width) - pad_width; + const int in_y_origin = (out_y * stride_height) - pad_height; + int32_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + const int in_y = + in_y_origin + dilation_height_factor * filter_y; + // If the location is outside the bounds of the input image, + // use zero as a default value. + if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height)) { + int32_t input_val = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + int32_t filter_val = + filter_data[Offset(filter_shape, out_channel, filter_y, + filter_x, in_channel)]; + acc += filter_val * (input_val - input_offset[batch]); + } + } + } + } + float acc_float = + acc * per_channel_scale[out_channel] * scaling_factors_ptr[batch]; + if (bias_data) { + acc_float += bias_data[out_channel]; + } + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + ActivationFunctionWithMinMax(acc_float, output_activation_min, + output_activation_max); + } + } + } + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h similarity index 94% rename from src/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h index 0cecb16..0a864ae 100644 --- a/src/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,9 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { namespace reference_ops { @@ -98,3 +99,5 @@ inline void DepthwiseConv( } // end namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h new file mode 100644 index 0000000..4a392a6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h @@ -0,0 +1,300 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_ + +#include + +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +// Used in tests and template parameters to control which version of depthwise +// convolution is called. Primarily for reference code, and specializations +// forced in tests. +enum class DepthwiseConvImplementation { + // Run all tests against kUseStandardEntry even if also testing another + // kernel, since we need to be sure that the main DepthwiseConv() function in + // optimized_ops.h dispatches to a correctly-executing kernel. + kNone = 0, // The "default" option: use the normal + // DepthwiseConv kernel (entry) function. + kUseGenericKernel, // Forced use of generic kernel. + kUseNeon3x3, // 3x3 kernel that uses NEON when available. + kUseNeon3x3DotProduct, // 3x3 kernel that uses dot-product enabled NEON + // when available. + kUseCModel3x3DotProduct, // 3x3 kernel, reference C model that is intended + // to match overall design NEON code. + kUseUnwound3x3DotProduct, // 3x3 kernel, reference C model with unwound loops + // and some arrays. + kUseIntrinsics3x3DotProduct, // 3x3 kernel using NEON intrinsics. +}; + +// Category of depthwise convolution output rounding. +enum class DepthwiseConvOutputRounding { + kNone = 0, // Invalid: specific method must be specified. + kAwayFromZero, // Original method: exact halves rounded away from zero. + kUpward, // Halves towards +infinity: adds 0.5 before truncate. + // This is where a future kNearestEven would be placed. +}; + +// Category of depthwise convolution depth multiplication. +enum class DepthwiseConvDepthMultiplication { + kNoMultiplication = 0, // Depth multiplier = 1. + kUnitInputDepth, // Input depth = 1, output depth = depth multiplier. +}; + +namespace reference_ops { +namespace depthwise_conv { + +template +inline int32_t DepthwiseConvRound(int32_t x, int32_t quantized_multiplier, + int shift) { + TFLITE_DCHECK_NE(output_rounding, DepthwiseConvOutputRounding::kNone); + return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift); +} + +template <> +inline int32_t DepthwiseConvRound( + int32_t x, int32_t quantized_multiplier, int shift) { + return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift); +} + +template <> +inline int32_t DepthwiseConvRound( + int32_t x, int32_t quantized_multiplier, int shift) { + using gemmlowp::SaturatingRoundingDoublingHighMul; + const int left_shift = shift > 0 ? shift : 0; + const int right_shift = shift > 0 ? 0 : -shift; + const int rounding_offset = right_shift > 0 ? 1 << (right_shift - 1) : 0; + return (SaturatingRoundingDoublingHighMul(x * (1 << left_shift), + quantized_multiplier) + + rounding_offset) >> + right_shift; +} + +template +struct DepthwiseConvBasicKernel { + static inline void Run( + const DepthwiseParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& filter_shape, + const uint8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + uint8_t* output_data) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int depth_multiplier = params.depth_multiplier; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + const int32_t input_offset = params.input_offset; + const int32_t filter_offset = params.weights_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int input_depth = input_shape.Dims(3); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier); + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + + for (int b = 0; b < batches; ++b) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int ic = 0; ic < input_depth; ++ic) { + for (int m = 0; m < depth_multiplier; m++) { + const int oc = m + ic * depth_multiplier; + const int in_x_origin = (out_x * stride_width) - pad_width; + const int in_y_origin = (out_y * stride_height) - pad_height; + int32_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = + in_x_origin + dilation_width_factor * filter_x; + const int in_y = + in_y_origin + dilation_height_factor * filter_y; + // If the location is outside the bounds of the input image, + // use zero as a default value. + if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height)) { + int32_t input_val = + input_data[Offset(input_shape, b, in_y, in_x, ic)]; + int32_t filter_val = filter_data[Offset( + filter_shape, 0, filter_y, filter_x, oc)]; + acc += (filter_val + filter_offset) * + (input_val + input_offset); + } + } + } + if (bias_data) { + acc += bias_data[oc]; + } + acc = DepthwiseConvRound(acc, output_multiplier, + output_shift); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[Offset(output_shape, b, out_y, out_x, oc)] = + static_cast(acc); + } + } + } + } + } + } + + // TODO(b/148596273): Reconcile reference versions, perhaps with common + // MultiplyByQuantizedMultiplier or DepthwiseConvRound function. + static inline void RunPerChannel( + const DepthwiseParams& params, const RuntimeShape& input_shape, + const int8_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + int8_t* output_data) { + // Get parameters. + // TODO(b/141565753): Re-introduce ScopedProfilingLabel on Micro. + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int depth_multiplier = params.depth_multiplier; + const int32_t input_offset = params.input_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + const int32_t* output_multiplier = params.output_multiplier_per_channel; + const int32_t* output_shift = params.output_shift_per_channel; + + // Check dimensions of the tensors. + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int input_depth = input_shape.Dims(3); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier); + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + for (int m = 0; m < depth_multiplier; ++m) { + const int output_channel = m + in_channel * depth_multiplier; + const int in_x_origin = (out_x * stride_width) - pad_width; + const int in_y_origin = (out_y * stride_height) - pad_height; + int32_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = + in_x_origin + dilation_width_factor * filter_x; + const int in_y = + in_y_origin + dilation_height_factor * filter_y; + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + if (is_point_inside_image) { + int32_t input_val = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + int32_t filter_val = filter_data[Offset( + filter_shape, 0, filter_y, filter_x, output_channel)]; + // Accumulate with 32 bits accumulator. + // In the nudging process during model quantization, we + // force real value of 0.0 be represented by a quantized + // value. This guarantees that the input_offset is a int8_t, + // even though it is represented using int32_t. int32_t += + // int8_t + // * (int8_t - int8_t) so the highest value we can get from + // each accumulation is [-127, 127] * ([-128, 127] - + // [-128, 127]), which is [-32512, 32512]. log2(32512) + // = 14.98, which means we can accumulate at least 2^16 + // multiplications without overflow. The accumulator is + // applied to a filter so the accumulation logic will hold + // as long as the filter size (filter_y * filter_x * + // in_channel) does not exceed 2^16, which is the case in + // all the models we have seen so far. + acc += filter_val * (input_val + input_offset); + } + } + } + if (bias_data) { + acc += bias_data[output_channel]; + } + acc = DepthwiseConvRound( + acc, output_multiplier[output_channel], + output_shift[output_channel]); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, + output_channel)] = static_cast(acc); + } + } + } + } + } + } +}; + +} // namespace depthwise_conv + +inline void DepthwiseConv( + const DepthwiseParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& filter_shape, + const uint8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + uint8_t* output_data) { + return depthwise_conv::DepthwiseConvBasicKernel< + DepthwiseConvOutputRounding::kAwayFromZero>::Run(params, input_shape, + input_data, filter_shape, + filter_data, bias_shape, + bias_data, output_shape, + output_data); +} + +} // namespace reference_ops +} // end namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/dequantize.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/dequantize.h new file mode 100644 index 0000000..752da46 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/dequantize.h @@ -0,0 +1,81 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_ + +#include + +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +// Dequantizes into a float without rounding. +template +inline void Dequantize(const tflite::DequantizationParams& op_params, + const RuntimeShape& input_shape, + const InputT* input_data, + const RuntimeShape& output_shape, OutputT* output_data) { + int32_t zero_point = op_params.zero_point; + const double scale = op_params.scale; + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; i++) { + const int32_t val = input_data[i]; + const OutputT result = static_cast(scale * (val - zero_point)); + output_data[i] = result; + } +} + +// Dequantizes per-channel quantized tensor to float. +template +inline void PerChannelDequantize( + const tflite::PerChannelDequantizationParams& op_params, + const RuntimeShape& input_shape, const T* input_data, + const RuntimeShape& output_shape, float* output_data) { + // Ensure flat size is same. + MatchingFlatSize(input_shape, output_shape); + + const int32_t* zero_point = op_params.zero_point; + const float* scale = op_params.scale; + const int32_t quantized_dimension = op_params.quantized_dimension; + const int32_t num_dims = input_shape.DimensionsCount(); + const int32_t* dims_data = input_shape.DimsData(); + std::vector current_dim(num_dims, 0); + + do { + size_t offset = + ReducedOutputOffset(num_dims, reinterpret_cast(dims_data), + current_dim.data(), 0, nullptr); + const int channel = current_dim[quantized_dimension]; + const int32_t val = input_data[offset]; + const float result = + static_cast(scale[channel] * (val - zero_point[channel])); + output_data[offset] = result; + } while (NextIndex(num_dims, reinterpret_cast(dims_data), + current_dim.data())); +} + +} // namespace reference_ops + +} // namespace tflite +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/floor.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/floor.h similarity index 91% rename from src/tensorflow/lite/kernels/internal/reference/floor.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/floor.h index 0693fd4..1962ab8 100644 --- a/src/tensorflow/lite/kernels/internal/reference/floor.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/floor.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,7 +18,7 @@ limitations under the License. #include -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -37,3 +38,5 @@ inline void Floor(const RuntimeShape& input_shape, const float* input_data, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/fully_connected.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/fully_connected.h new file mode 100644 index 0000000..5a5f4c4 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/fully_connected.h @@ -0,0 +1,323 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace reference_ops { + +inline void FullyConnected( + const FullyConnectedParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& weights_shape, + const float* weights_data, const RuntimeShape& bias_shape, + const float* bias_data, const RuntimeShape& output_shape, + float* output_data) { + const float output_activation_min = params.float_activation_min; + const float output_activation_max = params.float_activation_max; + // TODO(b/62193649): This really should be: + // const int batches = ArraySize(output_dims, 1); + // but the current --variable_batch hack consists in overwriting the 3rd + // dimension with the runtime batch size, as we don't keep track for each + // array of which dimension is the batch dimension in it. + const int output_dims_count = output_shape.DimensionsCount(); + const int weights_dims_count = weights_shape.DimensionsCount(); + const int batches = FlatSizeSkipDim(output_shape, output_dims_count - 1); + const int output_depth = MatchingDim(weights_shape, weights_dims_count - 2, + output_shape, output_dims_count - 1); + const int accum_depth = weights_shape.Dims(weights_dims_count - 1); + for (int b = 0; b < batches; ++b) { + for (int out_c = 0; out_c < output_depth; ++out_c) { + float total = 0.f; + for (int d = 0; d < accum_depth; ++d) { + total += input_data[b * accum_depth + d] * + weights_data[out_c * accum_depth + d]; + } + float bias_value = 0.0f; + if (bias_data) { + bias_value = bias_data[out_c]; + } + output_data[out_c + output_depth * b] = ActivationFunctionWithMinMax( + total + bias_value, output_activation_min, output_activation_max); + } + } +} + +inline void FullyConnected( + const FullyConnectedParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& filter_shape, + const uint8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + uint8_t* output_data) { + const int32_t input_offset = params.input_offset; + const int32_t filter_offset = params.weights_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2); + TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1); + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + // TODO(b/62193649): This really should be: + // const int batches = ArraySize(output_dims, 1); + // but the current --variable_batch hack consists in overwriting the 3rd + // dimension with the runtime batch size, as we don't keep track for each + // array of which dimension is the batch dimension in it. + const int output_dim_count = output_shape.DimensionsCount(); + const int filter_dim_count = filter_shape.DimensionsCount(); + const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1); + const int output_depth = MatchingDim(filter_shape, filter_dim_count - 2, + output_shape, output_dim_count - 1); + const int accum_depth = filter_shape.Dims(filter_dim_count - 1); + for (int b = 0; b < batches; ++b) { + for (int out_c = 0; out_c < output_depth; ++out_c) { + int32_t acc = 0; + for (int d = 0; d < accum_depth; ++d) { + int32_t input_val = input_data[b * accum_depth + d]; + int32_t filter_val = filter_data[out_c * accum_depth + d]; + acc += (filter_val + filter_offset) * (input_val + input_offset); + } + if (bias_data) { + acc += bias_data[out_c]; + } + acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[out_c + output_depth * b] = static_cast(acc); + } + } +} + +inline void FullyConnected( + const FullyConnectedParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& filter_shape, + const uint8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + int16_t* output_data) { + const int32_t input_offset = params.input_offset; + const int32_t filter_offset = params.weights_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + TFLITE_DCHECK_EQ(output_offset, 0); + // TODO(b/62193649): This really should be: + // const int batches = ArraySize(output_dims, 1); + // but the current --variable_batch hack consists in overwriting the 3rd + // dimension with the runtime batch size, as we don't keep track for each + // array of which dimension is the batch dimension in it. + const int output_dim_count = output_shape.DimensionsCount(); + const int filter_dim_count = filter_shape.DimensionsCount(); + const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1); + const int output_depth = MatchingDim(filter_shape, filter_dim_count - 2, + output_shape, output_dim_count - 1); + const int accum_depth = filter_shape.Dims(filter_dim_count - 1); + for (int b = 0; b < batches; ++b) { + for (int out_c = 0; out_c < output_depth; ++out_c) { + // Internal accumulation. + // Initialize accumulator with the bias-value. + int32_t accum = bias_data[out_c]; + // Accumulation loop. + for (int d = 0; d < accum_depth; ++d) { + int16_t input_val = input_data[b * accum_depth + d] + input_offset; + int16_t filter_val = + filter_data[out_c * accum_depth + d] + filter_offset; + accum += filter_val * input_val; + } + // Down-scale the final int32_t accumulator to the scale used by our + // (16-bit, typically 3 integer bits) fixed-point format. The quantized + // multiplier and shift here have been pre-computed offline + // (e.g. by toco). + accum = + MultiplyByQuantizedMultiplier(accum, output_multiplier, output_shift); + // Saturate, cast to int16_t, and store to output array. + accum = std::max(accum, output_activation_min - output_offset); + accum = std::min(accum, output_activation_max - output_offset); + accum += output_offset; + output_data[out_c + output_depth * b] = accum; + } + } +} + +inline void ShuffledFullyConnected( + const FullyConnectedParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& weights_shape, + const uint8_t* shuffled_weights_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + int16_t* output_data, uint8_t* shuffled_input_workspace_data) { + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + + TFLITE_DCHECK_GE(input_shape.DimensionsCount(), 1); + TFLITE_DCHECK_GE(weights_shape.DimensionsCount(), 2); + TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1); + // TODO(b/62193649): This really should be: + // const int batches = ArraySize(output_dims, 1); + // but the current --variable_batch hack consists in overwriting the 3rd + // dimension with the runtime batch size, as we don't keep track for each + // array of which dimension is the batch dimension in it. + const int output_dim_count = output_shape.DimensionsCount(); + const int weights_dim_count = weights_shape.DimensionsCount(); + const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1); + const int output_depth = MatchingDim(weights_shape, weights_dim_count - 2, + output_shape, output_dim_count - 1); + const int accum_depth = weights_shape.Dims(weights_dim_count - 1); + TFLITE_DCHECK((accum_depth % 16) == 0); + TFLITE_DCHECK((output_depth % 4) == 0); + + // Shuffling and xoring of input activations into the workspace buffer + uint8_t* shuffled_input_workspace_ptr = shuffled_input_workspace_data; + if (batches == 1) { + for (int i = 0; i < accum_depth; i++) { + shuffled_input_workspace_data[i] = input_data[i] ^ 0x80; + } + } else if (batches == 4) { + for (int c = 0; c < accum_depth; c += 16) { + for (int b = 0; b < 4; b++) { + const uint8_t* src_data_ptr = input_data + b * accum_depth + c; + for (int j = 0; j < 16; j++) { + uint8_t src_val = *src_data_ptr++; + // Flip the sign bit, so that the kernel will only need to + // reinterpret these uint8_t values as int8_t, getting for free the + // subtraction of the zero_point value 128. + uint8_t dst_val = src_val ^ 0x80; + *shuffled_input_workspace_ptr++ = dst_val; + } + } + } + } else { + TFLITE_DCHECK(false); + return; + } + + // Actual computation + if (batches == 1) { + int16_t* output_ptr = output_data; + // Shuffled weights have had their sign bit (0x80) pre-flipped (xor'd) + // so that just reinterpreting them as int8_t values is equivalent to + // subtracting 128 from them, thus implementing for free the subtraction of + // the zero_point value 128. + const int8_t* shuffled_weights_ptr = + reinterpret_cast(shuffled_weights_data); + // Likewise, we preshuffled and pre-xored the input data above. + const int8_t* shuffled_input_data = + reinterpret_cast(shuffled_input_workspace_data); + for (int c = 0; c < output_depth; c += 4) { + // Internal accumulation. + // Initialize accumulator with the bias-value. + int32_t accum[4] = {0}; + // Accumulation loop. + for (int d = 0; d < accum_depth; d += 16) { + for (int i = 0; i < 4; i++) { + for (int j = 0; j < 16; j++) { + int8_t input_val = shuffled_input_data[d + j]; + int8_t weights_val = *shuffled_weights_ptr++; + accum[i] += weights_val * input_val; + } + } + } + for (int i = 0; i < 4; i++) { + // Add bias value + int32_t acc = accum[i] + bias_data[c + i]; + // Down-scale the final int32_t accumulator to the scale used by our + // (16-bit, typically 3 integer bits) fixed-point format. The quantized + // multiplier and shift here have been pre-computed offline + // (e.g. by toco). + acc = + MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift); + // Saturate, cast to int16_t, and store to output array. + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_ptr[c + i] = acc; + } + } + } else if (batches == 4) { + int16_t* output_ptr = output_data; + // Shuffled weights have had their sign bit (0x80) pre-flipped (xor'd) + // so that just reinterpreting them as int8_t values is equivalent to + // subtracting 128 from them, thus implementing for free the subtraction of + // the zero_point value 128. + const int8_t* shuffled_weights_ptr = + reinterpret_cast(shuffled_weights_data); + // Likewise, we preshuffled and pre-xored the input data above. + const int8_t* shuffled_input_data = + reinterpret_cast(shuffled_input_workspace_data); + for (int c = 0; c < output_depth; c += 4) { + const int8_t* shuffled_input_ptr = shuffled_input_data; + // Accumulation loop. + // Internal accumulation. + // Initialize accumulator with the bias-value. + int32_t accum[4][4]; + for (int i = 0; i < 4; i++) { + for (int b = 0; b < 4; b++) { + accum[i][b] = 0; + } + } + for (int d = 0; d < accum_depth; d += 16) { + for (int i = 0; i < 4; i++) { + for (int b = 0; b < 4; b++) { + for (int j = 0; j < 16; j++) { + int8_t input_val = shuffled_input_ptr[16 * b + j]; + int8_t weights_val = shuffled_weights_ptr[16 * i + j]; + accum[i][b] += weights_val * input_val; + } + } + } + shuffled_input_ptr += 64; + shuffled_weights_ptr += 64; + } + for (int i = 0; i < 4; i++) { + for (int b = 0; b < 4; b++) { + // Add bias value + int32_t acc = accum[i][b] + bias_data[c + i]; + // Down-scale the final int32_t accumulator to the scale used by our + // (16-bit, typically 3 integer bits) fixed-point format. The + // quantized multiplier and shift here have been pre-computed offline + // (e.g. by toco). + acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, + output_shift); + // Saturate, cast to int16_t, and store to output array. + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_ptr[b * output_depth + c + i] = acc; + } + } + } + } else { + TFLITE_DCHECK(false); + return; + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/hard_swish.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/hard_swish.h new file mode 100644 index 0000000..bba9f3a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/hard_swish.h @@ -0,0 +1,169 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_ + +#include "tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace reference_ops { + +inline int16_t SaturatingLeftShift(int16_t value, int amount) { + int32_t result = static_cast(value) * (1 << amount); + result = std::min(result, std::numeric_limits::max()); + result = std::max(result, std::numeric_limits::min()); + return result; +} + +// Similar to ARM instruction SQDMULH. +// Similar to gemmlowp::SaturatingRoundingDoublingHighMul except +// rounding to zero instead of to nearest (SQRDMULH). +inline std::int16_t SaturatingDoublingHighMul(std::int16_t a, std::int16_t b) { + bool overflow = a == b && a == std::numeric_limits::min(); + std::int32_t a_32(a); + std::int32_t b_32(b); + std::int32_t ab_32 = a_32 * b_32; + std::int16_t ab_x2_high16 = static_cast((ab_32) / (1 << 15)); + return overflow ? std::numeric_limits::max() : ab_x2_high16; +} + +template +inline void HardSwish(const RuntimeShape& input_shape, const T* input_data, + const RuntimeShape& output_shape, T* output_data) { + ruy::profiler::ScopeLabel label("ReferenceHardSwish/Float"); + auto matching_size = MatchingFlatSize(input_shape, output_shape); + const T* in_end = input_data + matching_size; + for (; input_data < in_end; input_data++, output_data++) { + const float in = *input_data; + *output_data = + in * std::min(static_cast(6), std::max(static_cast(0), in + 3)) / + 6; + } +} + +template +inline void HardSwish(const HardSwishParams& params, + const RuntimeShape& input_shape, const T* input_data, + const RuntimeShape& output_shape, T* output_data) { + ruy::profiler::ScopeLabel label("ReferenceHardSwish/Quantized"); + + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; i++) { + const int16_t input_value = input_data[i] - params.input_zero_point; + // Left-shift as much as we can without overflow/saturation to put + // significant bits in the high bits of our 16-bit fixedpoint values, so + // that fixed-point approximate computations below are as accurate as + // possible. + const int16_t input_value_on_hires_input_scale = input_value * (1 << 7); + // Compute the input value on essentially the output scale, just not + // right-shifted yet. This is the value that we'll use in the (x >= +3) + // case, and that in the general case we'll multiply against the "relu-ish" + // fixed-point multiplier in [0, 1]. + const int16_t input_value_on_preshift_output_scale = + gemmlowp::SaturatingRoundingDoublingHighMul( + input_value_on_hires_input_scale, + params.output_multiplier_fixedpoint_int16); + // Now compute the "relu-ish multiplier". In the (-3 <= x <= +3) case, that + // is just an affine rescaling of x from [-3, 3] to [0, 1]. In the general + // case, it is just that plus saturation at the boundaries of [-3, 3]. + // First, we rescale from [-3, 3] to [-1, 1], saturating. + // That is done by rescaling the input value with a fixed-point multiplier + // (reluish_multiplier_fixedpoint) and bit-shift such that we represent + // that input value on the scale where the real value 3.0f is represented + // by the quantized value 32768. (+32768 is actually not representable as + // int16_t, so this saturates at +32767, and that is seen empirically to be + // a negligible contribution to numerical error/bias). + // + // This code is careful to correctly implement any magnitude of multiplier, + // involving either a right shift or a left shift, with correct saturation + // behavior in the left-shift case. This forces this code to be more + // complicated, but is necessary for real applications: a partially + // trained quantized MobileNet v3-small model that motivated this code + // exhibits some large [min, max] range boundaries, of the order of + // magnitude of 10 or 100 depending on layers. + // + // The next few lines are basically just an ordinary + // MultiplyByQuantizedMultiplier, except that we are more careful here + // about the fine details of saturation when left-shifting, because here + // overflow in left-shift is a common case, not an anomaly as + // MultiplyByQuantizedMultiplier assumes. + int16_t reluish_value = input_value_on_hires_input_scale; + // Shift left, saturating, as much as we can while ensuring that this + // saturation will not contribute to the result. That is, left shift amount + // reduced by 1. + if (params.reluish_multiplier_exponent > 0) { + reluish_value = SaturatingLeftShift( + reluish_value, params.reluish_multiplier_exponent - 1); + } + // Apply the fixed-point multiplier, dividing the value by a divisor + // ranging in [1, 2]. + reluish_value = gemmlowp::SaturatingRoundingDoublingHighMul( + reluish_value, params.reluish_multiplier_fixedpoint_int16); + // Apply the last bit of left-shift. Thus, in the left-shifting case, if + // any saturation affects the result, it is happening here --- any + // saturation having occurred above is overwritten here, not affecting the + // result. + if (params.reluish_multiplier_exponent > 0) { + reluish_value = SaturatingLeftShift(reluish_value, 1); + } + // Shift right, in the right-shifting case. + if (params.reluish_multiplier_exponent < 0) { + reluish_value = gemmlowp::RoundingDivideByPOT( + reluish_value, -params.reluish_multiplier_exponent); + } + // At this point we have rescaled the value into a 16bit fixedpoint + // reluish_value in [-1, 1]. + // We now convert that to a 16bit fixedpoint value in [0, 1]. + reluish_value = (reluish_value + (1 << 15)) >> 1; + // Use of SaturatingDoublingHighMul here is important to cancel the biases + // from the above SaturatingRoundingDoublingHighMul. + // + // On a partially trained MobileNet-v3-small, + // + // | bias on | ImageNet + // | quantized | Top-1 + // Operation used here | values | accuracy (50k) + // --------------------------------------+------------+----------- + // SaturatingDoublingHighMul | -0.0024 | 58.920 + // SaturatingRoundingDoublingHighMul | -0.0067 | 58.064 + // + // In activations_test, this is covered by this testcase: + // QuantizedActivationsOpTest.HardSwishBias + // + const int16_t preshift_output_value = SaturatingDoublingHighMul( + reluish_value, input_value_on_preshift_output_scale); + // We were so far operating on the pre-shift output scale. Now we finally + // apply that output shift, arriving at the final output scale. + int16_t output_value = gemmlowp::RoundingDivideByPOT( + preshift_output_value, -params.output_multiplier_exponent); + output_value += params.output_zero_point; + output_value = + std::min(output_value, std::numeric_limits::max()); + output_value = + std::max(output_value, std::numeric_limits::min()); + output_data[i] = output_value; + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/add.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/add.h new file mode 100644 index 0000000..fb49d52 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/add.h @@ -0,0 +1,147 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_ + +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace reference_integer_ops { + +inline void CheckArithmeticParams(const ArithmeticParams& params) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + // Input offset is negative input zero point. Activation tensors are + // asymmetric quantized so they span the full int8 range. + TFLITE_DCHECK_GE(-params.input1_offset, std::numeric_limits::min()); + TFLITE_DCHECK_GE(-params.input2_offset, std::numeric_limits::min()); + TFLITE_DCHECK_LE(-params.input1_offset, std::numeric_limits::max()); + TFLITE_DCHECK_LE(-params.input2_offset, std::numeric_limits::max()); +} + +inline void ElementWise( + int size, const ArithmeticParams& params, const int8_t* input1_data, + const int8_t* input2_data, int8_t* output_data, + void (*check_arithmetic_params)(const ArithmeticParams&), + int8_t (*binary_func)(int8_t, int8_t, const ArithmeticParams&)) { + CheckArithmeticParams(params); + for (int i = 0; i < size; ++i) { + output_data[i] = binary_func(input1_data[i], input2_data[i], params); + } +} + +inline void BroadcastBinaryFunction4DSlow( + const ArithmeticParams& params, const RuntimeShape& input1_shape, + const int8_t* input1_data, const RuntimeShape& input2_shape, + const int8_t* input2_data, const RuntimeShape& output_shape, + int8_t* output_data, + void (*check_arithmetic_params)(const ArithmeticParams&), + int8_t (*binary_func)(int8_t, int8_t, const ArithmeticParams&)) { + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + const RuntimeShape extended_output_shape = + RuntimeShape::ExtendedShape(4, output_shape); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + for (int b = 0; b < extended_output_shape.Dims(0); ++b) { + for (int y = 0; y < extended_output_shape.Dims(1); ++y) { + for (int x = 0; x < extended_output_shape.Dims(2); ++x) { + for (int c = 0; c < extended_output_shape.Dims(3); ++c) { + output_data[Offset(extended_output_shape, b, y, x, c)] = binary_func( + input1_data[SubscriptToIndex(desc1, b, y, x, c)], + input2_data[SubscriptToIndex(desc2, b, y, x, c)], params); + } + } + } + } +} + +inline int8_t AddFunc(int8_t x, int8_t y, const ArithmeticParams& params) { + const int32_t input1_val = params.input1_offset + x; + const int32_t input2_val = params.input2_offset + y; + const int32_t shifted_input1_val = input1_val * (1 << params.left_shift); + const int32_t shifted_input2_val = input2_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, params.input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, params.input2_shift); + const int32_t raw_sum = scaled_input1_val + scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sum, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + return static_cast(clamped_output); +} + +// Element-wise add that can often be used for inner loop of broadcast add as +// well as the non-broadcast add. +inline void AddElementwise(int size, const ArithmeticParams& params, + const int8_t* input1_data, const int8_t* input2_data, + int8_t* output_data) { + ElementWise(size, params, input1_data, input2_data, output_data, + CheckArithmeticParams, AddFunc); +} + +inline void Add(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const int8_t* input1_data, + const RuntimeShape& input2_shape, const int8_t* input2_data, + const RuntimeShape& output_shape, int8_t* output_data) { + CheckArithmeticParams(params); + + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + + AddElementwise(flat_size, params, input1_data, input2_data, output_data); +} + +inline void BroadcastAdd4DSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const int8_t* input1_data, + const RuntimeShape& input2_shape, + const int8_t* input2_data, + const RuntimeShape& output_shape, + int8_t* output_data) { + BroadcastBinaryFunction4DSlow(params, input1_shape, input1_data, input2_shape, + input2_data, output_shape, output_data, + CheckArithmeticParams, AddFunc); +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h new file mode 100644 index 0000000..f2c7fb8 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h @@ -0,0 +1,224 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +// Fixed-point per-channel-quantization convolution reference kernel. +inline void ConvPerChannel( + const ConvParams& params, const int32_t* output_multiplier, + const int32_t* output_shift, const RuntimeShape& input_shape, + const int8_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + int8_t* output_data) { + // Get parameters. + const int32_t input_offset = params.input_offset; // r = s(q - Z) + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int32_t output_offset = params.output_offset; + + // Set min and max value of the output. + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + + // Consistency check. + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + + // Check dimensions of the tensors. + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + const int in_y_origin = (out_y * stride_height) - pad_height; + for (int out_x = 0; out_x < output_width; ++out_x) { + const int in_x_origin = (out_x * stride_width) - pad_width; + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + int32_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + const int in_y = in_y_origin + dilation_height_factor * filter_y; + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + + if (!is_point_inside_image) { + continue; + } + + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + int32_t input_val = input_data[Offset(input_shape, batch, in_y, + in_x, in_channel)]; + int32_t filter_val = filter_data[Offset( + filter_shape, out_channel, filter_y, filter_x, in_channel)]; + // Accumulate with 32 bits accumulator. + // In the nudging process during model quantization, we force + // real value of 0.0 be represented by a quantized value. This + // guarantees that the input_offset is a int8_t, even though + // it is represented using int32_t. int32_t += int8_t * + // (int8_t - int8_t) so the highest value we can get from each + // accumulation is [-127, 127] * ([-128, 127] - + // [-128, 127]), which is [-32512, 32512]. log2(32512) + // = 14.98, which means we can accumulate at least 2^16 + // multiplications without overflow. The accumulator is + // applied to a filter so the accumulation logic will hold as + // long as the filter size (filter_y * filter_x * in_channel) + // does not exceed 2^16, which is the case in all the models + // we have seen so far. + // TODO(b/174275578): Add a check to make sure the + // accumulator depth is smaller than 2^16. + acc += filter_val * (input_val + input_offset); + } + } + } + + if (bias_data) { + acc += bias_data[out_channel]; + } + acc = MultiplyByQuantizedMultiplier( + acc, output_multiplier[out_channel], output_shift[out_channel]); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + static_cast(acc); + } + } + } + } +} + +// Fixed-point per-channel-quantization convolution reference kernel. +// 16-bit data and 8-bit filter +inline void ConvPerChannel( + const ConvParams& params, const int32_t* output_multiplier, + const int32_t* output_shift, const RuntimeShape& input_shape, + const int16_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const std::int64_t* bias_data, const RuntimeShape& output_shape, + int16_t* output_data) { + // Get parameters. + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + + // Set min and max value of the output. + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + + // Consistency check. + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + + // Check dimensions of the tensors. + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + const int in_y_origin = (out_y * stride_height) - pad_height; + for (int out_x = 0; out_x < output_width; ++out_x) { + const int in_x_origin = (out_x * stride_width) - pad_width; + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + std::int64_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + const int in_y = in_y_origin + dilation_height_factor * filter_y; + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + + if (!is_point_inside_image) { + continue; + } + + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + int32_t input_val = input_data[Offset(input_shape, batch, in_y, + in_x, in_channel)]; + int32_t filter_val = filter_data[Offset( + filter_shape, out_channel, filter_y, filter_x, in_channel)]; + // Accumulate with 64 bits accumulator. + // int64_t += int8_t * int16_t so the highest value we can + // get from each accumulation is [-127, 127] * ([-32768, + // 32767] - + // [-32768, 32767]), which is [-8322945, 8322945]. + // log2(8322945) = 22.99. + acc += filter_val * input_val; + } + } + } + if (bias_data) { + acc += bias_data[out_channel]; + } + int32_t scaled_acc = MultiplyByQuantizedMultiplier( + acc, output_multiplier[out_channel], output_shift[out_channel]); + scaled_acc = std::max(scaled_acc, output_activation_min); + scaled_acc = std::min(scaled_acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + static_cast(scaled_acc); + } + } + } + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h new file mode 100644 index 0000000..b6c161f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h @@ -0,0 +1,292 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { +inline void DepthwiseConvPerChannel( + const DepthwiseParams& params, const int32_t* output_multiplier, + const int32_t* output_shift, const RuntimeShape& input_shape, + const int8_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + int8_t* output_data) { + // Get parameters. + // TODO(b/141565753): Re-introduce ScopedProfilingLabel on Micro. + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int depth_multiplier = params.depth_multiplier; + const int32_t input_offset = params.input_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + + // Check dimensions of the tensors. + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int input_depth = input_shape.Dims(3); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier); + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + for (int m = 0; m < depth_multiplier; ++m) { + const int output_channel = m + in_channel * depth_multiplier; + const int in_x_origin = (out_x * stride_width) - pad_width; + const int in_y_origin = (out_y * stride_height) - pad_height; + int32_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + const int in_y = + in_y_origin + dilation_height_factor * filter_y; + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + if (is_point_inside_image) { + int32_t input_val = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + int32_t filter_val = filter_data[Offset( + filter_shape, 0, filter_y, filter_x, output_channel)]; + // Accumulate with 32 bits accumulator. + // In the nudging process during model quantization, we force + // real value of 0.0 be represented by a quantized value. This + // guarantees that the input_offset is a int8_t, even though + // it is represented using int32_t. int32_t += int8_t * + // (int8_t - int8_t) so the highest value we can get from each + // accumulation is [-127, 127] * ([-128, 127] - + // [-128, 127]), which is [-32512, 32512]. log2(32512) + // = 14.98, which means we can accumulate at least 2^16 + // multiplications without overflow. The accumulator is + // applied to a filter so the accumulation logic will hold as + // long as the filter size (filter_y * filter_x * in_channel) + // does not exceed 2^16, which is the case in all the models + // we have seen so far. + // TODO(b/174275578): Add a check to make sure the + // accumulator depth is smaller than 2^16. + acc += filter_val * (input_val + input_offset); + } + } + } + if (bias_data) { + acc += bias_data[output_channel]; + } + acc = MultiplyByQuantizedMultiplier( + acc, output_multiplier[output_channel], + output_shift[output_channel]); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, + output_channel)] = static_cast(acc); + } + } + } + } + } +} + +inline void DepthwiseConvPerChannel( + const DepthwiseParams& params, const int32_t* output_multiplier, + const int32_t* output_shift, const RuntimeShape& input_shape, + const int16_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const std::int64_t* bias_data, const RuntimeShape& output_shape, + int16_t* output_data) { + // Get parameters. + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int depth_multiplier = params.depth_multiplier; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + + // Check dimensions of the tensors. + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int input_depth = input_shape.Dims(3); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier); + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + for (int m = 0; m < depth_multiplier; ++m) { + const int output_channel = m + in_channel * depth_multiplier; + const int in_x_origin = (out_x * stride_width) - pad_width; + const int in_y_origin = (out_y * stride_height) - pad_height; + std::int64_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + const int in_y = + in_y_origin + dilation_height_factor * filter_y; + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + if (is_point_inside_image) { + int32_t input_val = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + int32_t filter_val = filter_data[Offset( + filter_shape, 0, filter_y, filter_x, output_channel)]; + // Accumulate with 64 bits accumulator. + // We assume maximum of 2^16 accumulations as with the 8-bit + // case so actually the value in the accumulator should not + // exceed 40 bits + acc += static_cast(filter_val) * + static_cast(input_val); + } + } + } + if (bias_data) { + acc += bias_data[output_channel]; + } + int32_t scaled_acc = MultiplyByQuantizedMultiplier( + acc, output_multiplier[output_channel], + output_shift[output_channel]); + scaled_acc = std::max(scaled_acc, output_activation_min); + scaled_acc = std::min(scaled_acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, + output_channel)] = + static_cast(scaled_acc); + } + } + } + } + } +} + +inline void DepthwiseConvHybridPerChannel( + const DepthwiseParams& params, float* scaling_factors_ptr, + const RuntimeShape& input_shape, const int8_t* input_data, + const RuntimeShape& filter_shape, const int8_t* filter_data, + const RuntimeShape& bias_shape, const float* bias_data, + const RuntimeShape& output_shape, float* output_data, + const float* per_channel_scale, int32_t* input_offset) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int depth_multiplier = params.depth_multiplier; + const float output_activation_min = params.float_activation_min; + const float output_activation_max = params.float_activation_max; + // Check dimensions of the tensors. + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int input_depth = input_shape.Dims(3); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int bias_depth = bias_shape.FlatSize(); + TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier); + TFLITE_DCHECK_EQ(bias_depth, output_depth); + + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + for (int m = 0; m < depth_multiplier; ++m) { + const int output_channel = m + in_channel * depth_multiplier; + const int in_x_origin = (out_x * stride_width) - pad_width; + const int in_y_origin = (out_y * stride_height) - pad_height; + int32_t acc = 0; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + const int in_y = + in_y_origin + dilation_height_factor * filter_y; + // Zero padding by omitting the areas outside the image. + const bool is_point_inside_image = + (in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height); + if (is_point_inside_image) { + int32_t input_val = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + int32_t filter_val = filter_data[Offset( + filter_shape, 0, filter_y, filter_x, output_channel)]; + acc += filter_val * (input_val - input_offset[batch]); + } + } + } + float acc_float = static_cast(acc); + acc_float *= + per_channel_scale[output_channel] * scaling_factors_ptr[batch]; + if (bias_data && output_channel < bias_depth) { + acc_float += bias_data[output_channel]; + } + output_data[Offset(output_shape, batch, out_y, out_x, + output_channel)] = + ActivationFunctionWithMinMax(acc_float, output_activation_min, + output_activation_max); + } + } + } + } + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h new file mode 100644 index 0000000..efaad0a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h @@ -0,0 +1,111 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +inline void FullyConnected( + const FullyConnectedParams& params, const RuntimeShape& input_shape, + const int8_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + int8_t* output_data) { + const int32_t input_offset = params.input_offset; + const int32_t filter_offset = params.weights_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2); + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + const int filter_dim_count = filter_shape.DimensionsCount(); + const int batches = output_shape.Dims(0); + const int output_depth = output_shape.Dims(1); + TFLITE_DCHECK_LE(output_depth, filter_shape.Dims(filter_dim_count - 2)); + const int accum_depth = filter_shape.Dims(filter_dim_count - 1); + for (int b = 0; b < batches; ++b) { + for (int out_c = 0; out_c < output_depth; ++out_c) { + int32_t acc = 0; + for (int d = 0; d < accum_depth; ++d) { + int32_t input_val = input_data[b * accum_depth + d]; + int32_t filter_val = filter_data[out_c * accum_depth + d]; + acc += (filter_val + filter_offset) * (input_val + input_offset); + } + if (bias_data) { + acc += bias_data[out_c]; + } + acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[out_c + output_depth * b] = static_cast(acc); + } + } +} + +inline void FullyConnected( + const FullyConnectedParams& params, const RuntimeShape& input_shape, + const int16_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const int64_t* bias_data, const RuntimeShape& output_shape, + int16_t* output_data) { + const int32_t filter_offset = params.weights_offset; + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2); + + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + const int filter_dim_count = filter_shape.DimensionsCount(); + const int batches = output_shape.Dims(0); + const int output_depth = output_shape.Dims(1); + TFLITE_DCHECK_LE(output_depth, filter_shape.Dims(filter_dim_count - 2)); + const int accum_depth = filter_shape.Dims(filter_dim_count - 1); + for (int b = 0; b < batches; ++b) { + for (int out_c = 0; out_c < output_depth; ++out_c) { + int64_t acc = 0; + for (int d = 0; d < accum_depth; ++d) { + int32_t input_val = input_data[b * accum_depth + d]; + int32_t filter_val = filter_data[out_c * accum_depth + d]; + acc += (filter_val + filter_offset) * input_val; + } + if (bias_data) { + acc += bias_data[out_c]; + } + int32_t acc_scaled = + MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift); + acc_scaled = std::max(acc_scaled, output_activation_min); + acc_scaled = std::min(acc_scaled, output_activation_max); + output_data[out_c + output_depth * b] = static_cast(acc_scaled); + } + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h new file mode 100644 index 0000000..2a4a612 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h @@ -0,0 +1,68 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +inline void L2Normalization(int32_t input_zero_point, int32_t outer_size, + int32_t depth, const int8_t* input_data, + int8_t* output_data) { + static constexpr int8_t kMinInt8 = std::numeric_limits::min(); + static constexpr int8_t kMaxInt8 = std::numeric_limits::max(); + // The output scale must be in sync with Prepare(). + // Output is in 1/128 scale so the actual output range is nudged from [-1, 1] + // to [-1, 127/128]. + static constexpr int32_t kOutputScale = 7; + for (int outer_index = 0; outer_index < outer_size; ++outer_index) { + // int32_t = (int8_t - int8_t) ^ 2. + // ([-128, 127] - [-128, 127]) ^ 2 = [0, (2^8 - 1)^2] so the accumulator is + // safe from overflowing in at least 2^16 steps. + int32_t acc = 0; + for (int inner_index = 0; inner_index < depth; ++inner_index) { + int32_t input = + input_data[depth * outer_index + inner_index] - input_zero_point; + acc += input * input; + } + int32_t inv_l2norm_multiplier; + int inv_l2norm_shift; + GetInvSqrtQuantizedMultiplierExp(acc, kReverseShift, &inv_l2norm_multiplier, + &inv_l2norm_shift); + + for (int inner_index = 0; inner_index < depth; ++inner_index) { + int32_t input = + input_data[depth * outer_index + inner_index] - input_zero_point; + + // Rescale and downcast. Rescale is folded into the division. + int32_t output_in_q24 = MultiplyByQuantizedMultiplier( + input, inv_l2norm_multiplier, inv_l2norm_shift + kOutputScale); + output_in_q24 = + std::min(static_cast(kMaxInt8), + std::max(static_cast(kMinInt8), output_in_q24)); + output_data[depth * outer_index + inner_index] = + static_cast(output_in_q24); + } + } +} +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h new file mode 100644 index 0000000..2788939 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h @@ -0,0 +1,109 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_ + +#include +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +inline void Logistic(int32_t input_zero_point, int32_t input_range_radius, + int32_t input_multiplier, int32_t input_left_shift, + int32_t input_size, const int8_t* input_data, + int8_t* output_data) { + // Integer bits must be in sync with Prepare() function. + static constexpr int32_t kInputIntegerBits = 4; + static constexpr int32_t kOutputIntegerBits = 8; + static constexpr int8_t kMinInt8 = std::numeric_limits::min(); + static constexpr int8_t kMaxInt8 = std::numeric_limits::max(); + static constexpr int32_t kOutputZeroPoint = -128; + + for (int i = 0; i < input_size; ++i) { + const int32_t input = + static_cast(input_data[i]) - input_zero_point; + if (input <= -input_range_radius) { + output_data[i] = kMinInt8; + } else if (input >= input_range_radius) { + output_data[i] = kMaxInt8; + } else { + const int32_t input_in_q4 = MultiplyByQuantizedMultiplier( + input, input_multiplier, input_left_shift); + using FixedPoint4 = gemmlowp::FixedPoint; + const int32_t output_in_q0 = + gemmlowp::logistic(FixedPoint4::FromRaw(input_in_q4)).raw(); + + // Rescale and downcast. + using gemmlowp::RoundingDivideByPOT; + int32_t output_in_q23 = + RoundingDivideByPOT(output_in_q0, 31 - kOutputIntegerBits); + output_in_q23 = std::min(std::max(output_in_q23 + kOutputZeroPoint, + static_cast(kMinInt8)), + static_cast(kMaxInt8)); + output_data[i] = static_cast(output_in_q23); + } + } +} + +inline void Logistic(int32_t input_multiplier, int32_t input_size, + const int16_t* ptr_input_data, int16_t* ptr_output_data) { + // We use the LUT for sigmoid and take into account, that + // tanh(x) = 2*sigmoid(2*x) - 1 + + int32_t input_data_mul = (input_multiplier > 0) ? input_multiplier : 1; + + for (int i = 0; i < input_size; ++i, ptr_input_data++, ptr_output_data++) { + int32_t input_data = (*ptr_input_data) * input_data_mul; + + // Scale by 3/4 to expand range [-8,8]->[-10.7,10.7] and + // we do interpolation on unsigned values. + uint32_t abs_input_data = 3 * abs(input_data); + + // We divide by 2 power of 9, because + // we need to divide by 2 in power of 7 for + // the input conversion + 1/4 from the scale above. + // Define uh as uint32_t type not to make this function overflow. + uint32_t uh = abs_input_data >> 9; + uint32_t result; + + if (uh >= 255) { + // Saturate to maximum. + result = 0x7FFF << 10; + } else { + uint32_t ua = sigmoid_table_uint16[uh]; + uint32_t ub = sigmoid_table_uint16[uh + 1]; + uint32_t ut = abs_input_data & 0x1ff; + // Interpolation is done using the fractional bit. + result = (ua << 9) + ut * (ub - ua); + } + + result = (input_data >= 0) ? (result + (1 << 9)) + : ((1 << (16 + 9)) - result + (1 << 9) - 1); + + // Back to 16-bit. + result >>= 10; + + *ptr_output_data = result; + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mean.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mean.h new file mode 100644 index 0000000..1969d44 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mean.h @@ -0,0 +1,80 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +template +inline void Mean(const tflite::MeanParams& op_params, int32_t multiplier, + int32_t shift, const RuntimeShape& unextended_input_shape, + const integer_type* input_data, int32_t input_zero_point, + const RuntimeShape& unextended_output_shape, + integer_type* output_data, int32_t output_zero_point) { + // Current implementation only supports dimension equals 4 and simultaneous + // reduction over width and height. + TFLITE_CHECK_EQ(unextended_input_shape.DimensionsCount(), 4); + TFLITE_CHECK_LE(unextended_output_shape.DimensionsCount(), 4); + const RuntimeShape input_shape = + RuntimeShape::ExtendedShape(4, unextended_input_shape); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(4, unextended_output_shape); + const int output_batch = output_shape.Dims(0); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int output_depth = output_shape.Dims(3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int num_elements_in_axis = input_width * input_height; + + TFLITE_CHECK_EQ(op_params.axis_count, 2); + TFLITE_CHECK((op_params.axis[0] == 1 && op_params.axis[1] == 2) || + (op_params.axis[0] == 2 && op_params.axis[1] == 1)); + TFLITE_CHECK_EQ(output_height, 1); + TFLITE_CHECK_EQ(output_width, 1); + + static constexpr int32_t kMinInt = std::numeric_limits::min(); + static constexpr int32_t kMaxInt = std::numeric_limits::max(); + + for (int out_b = 0; out_b < output_batch; ++out_b) { + for (int out_d = 0; out_d < output_depth; ++out_d) { + int32_t acc = 0; + for (int in_h = 0; in_h < input_height; ++in_h) { + for (int in_w = 0; in_w < input_width; ++in_w) { + acc += input_data[Offset(input_shape, out_b, in_h, in_w, out_d)] - + input_zero_point; + } + } + acc = MultiplyByQuantizedMultiplier(acc, multiplier, shift); + acc = acc > 0 ? (acc + num_elements_in_axis / 2) / num_elements_in_axis + : (acc - num_elements_in_axis / 2) / num_elements_in_axis; + acc += output_zero_point; + acc = std::min(std::max(acc, kMinInt), kMaxInt); + output_data[Offset(output_shape, out_b, 0, 0, out_d)] = + static_cast(acc); + } + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h similarity index 65% rename from src/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h index 07c5963..33ca753 100644 --- a/src/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,35 +16,37 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_ -#include "fixedpoint/fixedpoint.h" -#include "tensorflow/lite/experimental/ruy/profiler/instrumentation.h" -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" namespace tflite { namespace reference_integer_ops { +template inline void MulElementwise(int size, const ArithmeticParams& params, - const int8_t* input1_data, const int8_t* input2_data, - int8_t* output_data) { + const T* input1_data, const T* input2_data, + T* output_data) { for (int i = 0; i < size; ++i) { - const int32 input1_val = params.input1_offset + input1_data[i]; - const int32 input2_val = params.input2_offset + input2_data[i]; - const int32 unclamped_result = + const int32_t input1_val = params.input1_offset + input1_data[i]; + const int32_t input2_val = params.input2_offset + input2_data[i]; + const int32_t unclamped_result = params.output_offset + MultiplyByQuantizedMultiplier(input1_val * input2_val, params.output_multiplier, params.output_shift); - const int32 clamped_output = + const int32_t clamped_output = std::min(params.quantized_activation_max, std::max(params.quantized_activation_min, unclamped_result)); - output_data[i] = static_cast(clamped_output); + output_data[i] = static_cast(clamped_output); } } +template inline void Mul(const ArithmeticParams& params, - const RuntimeShape& input1_shape, const int8_t* input1_data, - const RuntimeShape& input2_shape, const int8_t* input2_data, - const RuntimeShape& output_shape, int8_t* output_data) { + const RuntimeShape& input1_shape, const T* input1_data, + const RuntimeShape& input2_shape, const T* input2_data, + const RuntimeShape& output_shape, T* output_data) { TFLITE_DCHECK_LE(params.quantized_activation_min, params.quantized_activation_max); ruy::profiler::ScopeLabel label("Mul/8bit"); @@ -55,13 +58,13 @@ inline void Mul(const ArithmeticParams& params, // Mul with 16 bit inputs and int8_t outputs. inline void Mul(const ArithmeticParams& params, - const RuntimeShape& input1_shape, const int16* input1_data, - const RuntimeShape& input2_shape, const int16* input2_data, + const RuntimeShape& input1_shape, const int16_t* input1_data, + const RuntimeShape& input2_shape, const int16_t* input2_data, const RuntimeShape& output_shape, int8_t* output_data) { ruy::profiler::ScopeLabel label("Mul/Int16Int8"); - int32 output_offset = params.output_offset; - int32 output_activation_min = params.quantized_activation_min; - int32 output_activation_max = params.quantized_activation_max; + int32_t output_offset = params.output_offset; + int32_t output_activation_min = params.quantized_activation_min; + int32_t output_activation_max = params.quantized_activation_max; TFLITE_DCHECK_LE(output_activation_min, output_activation_max); const int flat_size = @@ -73,24 +76,23 @@ inline void Mul(const ArithmeticParams& params, F0 unclamped_result = F0::FromRaw(input1_data[i]) * F0::FromRaw(input2_data[i]); - int16 rescaled_result = + int16_t rescaled_result = gemmlowp::RoundingDivideByPOT(unclamped_result.raw(), 8); - int16 clamped_result = - std::min(output_activation_max - output_offset, rescaled_result); - clamped_result = - std::max(output_activation_min - output_offset, clamped_result); + int16_t clamped_result = std::min( + output_activation_max - output_offset, rescaled_result); + clamped_result = std::max(output_activation_min - output_offset, + clamped_result); output_data[i] = output_offset + clamped_result; } } -inline void BroadcastMul4DSlow(const ArithmeticParams& params, - const RuntimeShape& input1_shape, - const int8_t* input1_data, - const RuntimeShape& input2_shape, - const int8_t* input2_data, - const RuntimeShape& output_shape, - int8_t* output_data) { - ruy::profiler::ScopeLabel label("BroadcastMul4DSlow/8bit"); +template +inline void BroadcastMul4DSlow( + const ArithmeticParams& params, const RuntimeShape& input1_shape, + const T* input1_data, const RuntimeShape& input2_shape, + const T* input2_data, const RuntimeShape& output_shape, T* output_data) { + ruy::profiler::ScopeLabel label("BroadcastMul4DSlow"); + NdArrayDesc<4> desc1; NdArrayDesc<4> desc2; // The input shapes are extended as part of NdArrayDesc initialization. @@ -103,22 +105,22 @@ inline void BroadcastMul4DSlow(const ArithmeticParams& params, for (int y = 0; y < extended_output_shape.Dims(1); ++y) { for (int x = 0; x < extended_output_shape.Dims(2); ++x) { for (int c = 0; c < extended_output_shape.Dims(3); ++c) { - const int32 input1_val = + const int32_t input1_val = params.input1_offset + input1_data[SubscriptToIndex(desc1, b, y, x, c)]; - const int32 input2_val = + const int32_t input2_val = params.input2_offset + input2_data[SubscriptToIndex(desc2, b, y, x, c)]; - const int32 unclamped_result = + const int32_t unclamped_result = params.output_offset + MultiplyByQuantizedMultiplier(input1_val * input2_val, params.output_multiplier, params.output_shift); - const int32 clamped_output = std::min( + const int32_t clamped_output = std::min( params.quantized_activation_max, std::max(params.quantized_activation_min, unclamped_result)); output_data[Offset(extended_output_shape, b, y, x, c)] = - static_cast(clamped_output); + static_cast(clamped_output); } } } @@ -128,3 +130,5 @@ inline void BroadcastMul4DSlow(const ArithmeticParams& params, } // namespace reference_integer_ops } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h new file mode 100644 index 0000000..5bdca71 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h @@ -0,0 +1,261 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_ + +#include +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const int8_t* input_data, + const RuntimeShape& output_shape, int8_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + int32_t acc = 0; + int filter_count = 0; + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + acc += + input_data[Offset(input_shape, batch, in_y, in_x, channel)]; + filter_count++; + } + } + // Round to the closest integer value. + acc = acc > 0 ? (acc + filter_count / 2) / filter_count + : (acc - filter_count / 2) / filter_count; + acc = std::max(acc, params.quantized_activation_min); + acc = std::min(acc, params.quantized_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + static_cast(acc); + } + } + } + } +} + +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const int8_t* input_data, const RuntimeShape& output_shape, + int8_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + TFLITE_DCHECK_GE(params.quantized_activation_min, + std::numeric_limits::min()); + TFLITE_DCHECK_LE(params.quantized_activation_max, + std::numeric_limits::max()); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + int8_t max = std::numeric_limits::lowest(); + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + max = std::max( + max, + input_data[Offset(input_shape, batch, in_y, in_x, channel)]); + } + } + max = std::max(max, params.quantized_activation_min); + max = std::min(max, params.quantized_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + static_cast(max); + } + } + } + } +} + +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const int16_t* input_data, + const RuntimeShape& output_shape, + int16_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + int32_t acc = 0; + int filter_count = 0; + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + acc += + input_data[Offset(input_shape, batch, in_y, in_x, channel)]; + filter_count++; + } + } + // Round to the closest integer value. + acc = acc > 0 ? (acc + filter_count / 2) / filter_count + : (acc - filter_count / 2) / filter_count; + acc = std::max(acc, params.quantized_activation_min); + acc = std::min(acc, params.quantized_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + static_cast(acc); + } + } + } + } +} + +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const int16_t* input_data, const RuntimeShape& output_shape, + int16_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + TFLITE_DCHECK_GE(params.quantized_activation_min, + std::numeric_limits::min()); + TFLITE_DCHECK_LE(params.quantized_activation_max, + std::numeric_limits::max()); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + int16_t max = std::numeric_limits::lowest(); + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + max = std::max( + max, + input_data[Offset(input_shape, batch, in_y, in_x, channel)]); + } + } + max = std::max(max, params.quantized_activation_min); + max = std::min(max, params.quantized_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + static_cast(max); + } + } + } + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h new file mode 100644 index 0000000..4217116 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h @@ -0,0 +1,113 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_ + +#include + +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +inline void Tanh(int32_t input_zero_point, int32_t input_range_radius, + int32_t input_multiplier, int32_t input_shift, + const RuntimeShape& input_shape, const int8_t* input_data, + const RuntimeShape& output_shape, int8_t* output_data) { + // Integer bits must be in sync with Prepare() function. + static constexpr int32_t kInputIntegerBits = 4; + static constexpr int32_t kOutputScale = 7; + static constexpr int32_t kMinInt8 = std::numeric_limits::min(); + static constexpr int32_t kMaxInt8 = std::numeric_limits::max(); + using F4 = gemmlowp::FixedPoint; + + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; ++i) { + const int32_t input = + static_cast(input_data[i]) - input_zero_point; + if (input <= -input_range_radius) { + output_data[i] = kMinInt8; + } else if (input >= input_range_radius) { + output_data[i] = kMaxInt8; + } else { + const int32_t input_in_q4 = + MultiplyByQuantizedMultiplier(input, input_multiplier, input_shift); + const int32_t output_in_q0 = + gemmlowp::tanh(F4::FromRaw(input_in_q4)).raw(); + + // Rescale and downcast. + using gemmlowp::RoundingDivideByPOT; + int32_t output_in_q24 = + RoundingDivideByPOT(output_in_q0, 31 - kOutputScale); + output_in_q24 = std::min(std::max(output_in_q24, kMinInt8), kMaxInt8); + output_data[i] = static_cast(output_in_q24); + } + } +} + +inline void Tanh(int32_t input_multiplier, int32_t input_left_shift, + const RuntimeShape& input_shape, const int16_t* ptr_input_data, + const RuntimeShape& output_shape, int16_t* ptr_output_data) { + // We use the LUT for sigmoid and take into account, that + // tanh(x) = 2*sigmoid(2*x) - 1 + + int32_t input_data_mul = (input_multiplier > 0) ? input_multiplier : 1; + + int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; ++i, ptr_input_data++, ptr_output_data++) { + int32_t input_data = (*ptr_input_data) * input_data_mul; + + if (input_left_shift == 1) { + input_data <<= 1; + } + + // Scale by 3/4 to expand range [-8,8]->[-10.7,10.7]. + uint32_t abs_input_data = 3 * abs(input_data); + uint32_t uh = abs_input_data >> 8; + int32_t result; + + if (uh >= 255) { + // Saturate to maximum. + result = 0xFFFF << 8; + } else { + uint32_t ua = sigmoid_table_uint16[uh]; + uint32_t ub = sigmoid_table_uint16[uh + 1]; + + uint8_t ut = abs_input_data & 0xFF; + + result = (ua << 8) + ut * (ub - ua); + } + + result = (input_data >= 0) + ? (result - (1 << (14 + 9)) + (1 << (9 - 2))) + : (-result + (1 << (14 + 9)) + (1 << (9 - 2)) - 1); + + // Convert back to 16-bit. + result >>= (9 - 1); + + *ptr_output_data = result; + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h new file mode 100644 index 0000000..46f1660 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h @@ -0,0 +1,224 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" + +namespace tflite { +namespace reference_integer_ops { + +// Fixed-point per-channel-quantization transpose convolution reference kernel. +inline void TransposeConv( + const ConvParams& params, const int32_t* output_multiplier, + const int32_t* output_shift, const RuntimeShape& input_shape, + const int8_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + int8_t* output_data, const RuntimeShape& im2col_shape, int8_t* im2col_data, + int32_t* scratch_buffer) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + (void)im2col_data; // only used in optimized code. + (void)im2col_shape; // only used in optimized code. + + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int32_t input_offset = params.input_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_activation_min = std::numeric_limits::min(); + const int32_t output_activation_max = std::numeric_limits::max(); + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + + const int num_elements = output_shape.FlatSize(); + // We need to initialize scratch_buffer to all 0s, as we apply the same + // 'scatter' based trick as in float version. + memset(scratch_buffer, 0, num_elements * sizeof(int32_t)); + + // Loop through input elements one at a time. + for (int batch = 0; batch < batches; ++batch) { + for (int in_y = 0; in_y < input_height; ++in_y) { + for (int in_x = 0; in_x < input_width; ++in_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + // Loop through the output elements it will influence. + const int out_x_origin = (in_x * stride_width) - pad_width; + const int out_y_origin = (in_y * stride_height) - pad_height; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + for (int out_channel = 0; out_channel < output_depth; + ++out_channel) { + // Compute output element location. + const int out_x = out_x_origin + filter_x; + const int out_y = out_y_origin + filter_y; + // We cannot accumulate out of bounds. + if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) && + (out_y < output_height)) { + const int8_t input_value = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + const int8_t filter_value = + filter_data[Offset(filter_shape, out_channel, filter_y, + filter_x, in_channel)]; + scratch_buffer[Offset(output_shape, batch, out_y, out_x, + out_channel)] += + (input_value + input_offset) * filter_value; + } + } + } + } + } + } + } + } + + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + int32_t acc = scratch_buffer[Offset(output_shape, batch, out_y, out_x, + out_channel)]; + if (bias_data) { + acc += bias_data[out_channel]; + } + acc = MultiplyByQuantizedMultiplier( + acc, output_multiplier[out_channel], output_shift[out_channel]); + acc += output_offset; + acc = std::max(acc, output_activation_min); + acc = std::min(acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + static_cast(acc); + } + } + } + } +} + +// int16_t input (zero_point=0), int8_t filter, int64 accumulator +inline void TransposeConv( + const ConvParams& params, const int32_t* output_multiplier, + const int32_t* output_shift, const RuntimeShape& input_shape, + const int16_t* input_data, const RuntimeShape& filter_shape, + const int8_t* filter_data, const RuntimeShape& bias_shape, + const std::int64_t* bias_data, const RuntimeShape& output_shape, + int16_t* output_data, const RuntimeShape& im2col_shape, int8_t* im2col_data, + std::int64_t* scratch_buffer) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + (void)im2col_data; // only used in optimized code. + (void)im2col_shape; // only used in optimized code. + + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int32_t output_activation_min = std::numeric_limits::min(); + const int32_t output_activation_max = std::numeric_limits::max(); + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + + const int num_elements = output_shape.FlatSize(); + // We need to initialize scratch_buffer to all 0s, as we apply the same + // 'scatter' based trick as in float version. + memset(scratch_buffer, 0, num_elements * sizeof(std::int64_t)); + + // Loop through input elements one at a time. + for (int batch = 0; batch < batches; ++batch) { + for (int in_y = 0; in_y < input_height; ++in_y) { + for (int in_x = 0; in_x < input_width; ++in_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + // Loop through the output elements it will influence. + const int out_x_origin = (in_x * stride_width) - pad_width; + const int out_y_origin = (in_y * stride_height) - pad_height; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + for (int out_channel = 0; out_channel < output_depth; + ++out_channel) { + // Compute output element location. + const int out_x = out_x_origin + filter_x; + const int out_y = out_y_origin + filter_y; + // We cannot accumulate out of bounds. + if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) && + (out_y < output_height)) { + const int32_t input_value = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + const int32_t filter_value = + filter_data[Offset(filter_shape, out_channel, filter_y, + filter_x, in_channel)]; + scratch_buffer[Offset(output_shape, batch, out_y, out_x, + out_channel)] += + input_value * filter_value; + } + } + } + } + } + } + } + } + + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + std::int64_t acc = scratch_buffer[Offset(output_shape, batch, out_y, + out_x, out_channel)]; + if (bias_data) { + acc += bias_data[out_channel]; + } + int32_t scaled_acc = MultiplyByQuantizedMultiplier( + acc, output_multiplier[out_channel], output_shift[out_channel]); + scaled_acc = std::max(scaled_acc, output_activation_min); + scaled_acc = std::min(scaled_acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + static_cast(scaled_acc); + } + } + } + } +} + +} // namespace reference_integer_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/l2normalization.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/l2normalization.h new file mode 100644 index 0000000..28c54a6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/l2normalization.h @@ -0,0 +1,93 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline void L2Normalization(const tflite::L2NormalizationParams& op_params, + const RuntimeShape& input_shape, + const float* input_data, + const RuntimeShape& output_shape, + float* output_data, float epsilon = 1e-6) { + const int trailing_dim = input_shape.DimensionsCount() - 1; + const int outer_size = + MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); + const int depth = + MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); + for (int i = 0; i < outer_size; ++i) { + float squared_l2_norm = 0; + for (int c = 0; c < depth; ++c) { + const float val = input_data[depth * i + c]; + squared_l2_norm += val * val; + } + float l2_norm = std::sqrt(squared_l2_norm); + l2_norm = std::max(l2_norm, epsilon); + for (int c = 0; c < depth; ++c) { + output_data[depth * i + c] = input_data[depth * i + c] / l2_norm; + } + } +} + +inline void L2Normalization(const tflite::L2NormalizationParams& op_params, + const RuntimeShape& input_shape, + const uint8_t* input_data, + const RuntimeShape& output_shape, + uint8_t* output_data) { + const int trailing_dim = input_shape.DimensionsCount() - 1; + const int depth = + MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); + const int outer_size = + MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); + const int32_t input_zero_point = op_params.input_zero_point; + + for (int i = 0; i < outer_size; ++i) { + int32_t square_l2_norm = 0; + for (int c = 0; c < depth; c++) { + int32_t diff = input_data[depth * i + c] - input_zero_point; + square_l2_norm += diff * diff; + } + int32_t inv_l2norm_multiplier; + int inv_l2norm_shift; + GetInvSqrtQuantizedMultiplierExp(square_l2_norm, kReverseShift, + &inv_l2norm_multiplier, &inv_l2norm_shift); + for (int c = 0; c < depth; c++) { + int32_t diff = input_data[depth * i + c] - input_zero_point; + int32_t rescaled_diff = MultiplyByQuantizedMultiplierSmallerThanOneExp( + 128 * diff, inv_l2norm_multiplier, inv_l2norm_shift); + int32_t unclamped_output_val = 128 + rescaled_diff; + int32_t output_val = + std::min(static_cast(255), + std::max(static_cast(0), unclamped_output_val)); + output_data[depth * i + c] = static_cast(output_val); + } + } +} + +} // namespace reference_ops +} // namespace tflite +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/logistic.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/logistic.h new file mode 100644 index 0000000..fd88ddf --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/logistic.h @@ -0,0 +1,135 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_ + +#include + +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" + +namespace tflite { +namespace reference_ops { + +inline void Logistic(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const float cutoff_upper = 16.619047164916992188f; + const float cutoff_lower = -9.f; + + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + // Rational for using approximation in reference kernel. + // 0. This approximation gives enough precision for float. + // 1. This works around an issue on an embedded chipset where exp() does not + // return correctly as expected - exp(x) should return inf when overflown + // not 1.701417 IEEE 754 defines representation for inf. + // 2. This will speed up calculation and is matching the behavior in the + // optimized kernels. (check the definition of scalar_logistic_op) + + for (int i = 0; i < flat_size; i++) { + float val = input_data[i]; + float result; + if (val > cutoff_upper) { + result = 1.0f; + } else if (val < cutoff_lower) { + result = std::exp(val); + } else { + result = 1.f / (1.f + std::exp(-val)); + } + output_data[i] = result; + } +} + +// Convenience version that allows, for example, generated-code calls to be +// uniform between data types. +inline void Logistic(const LogisticParams&, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { + // Drop params: not needed. + Logistic(input_shape, input_data, output_shape, output_data); +} + +inline void Logistic(const LogisticParams& params, + const RuntimeShape& input_shape, const int16_t* input_data, + const RuntimeShape& output_shape, int16_t* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; i++) { + // F0 uses 0 integer bits, range [-1, 1]. + // This is the return type of math functions such as tanh, logistic, + // whose range is in [-1, 1]. + using F0 = gemmlowp::FixedPoint; + // F3 uses 3 integer bits, range [-8, 8], the input range expected here. + using F3 = gemmlowp::FixedPoint; + + const F3 input = F3::FromRaw(input_data[i]); + F0 output = gemmlowp::logistic(input); + output_data[i] = output.raw(); + } +} + +// Quantized int8_t logistic activation. Cheats by dequantizing and +// requantizing around the floating point logistic method. This implementation +// is slow on platforms without a floating point unit. + +// TODO(b/141211002): Delete this int8_t implementation once we can reuse the +// approach used in TFLite for int8_t Logistic. +inline void Logistic(const RuntimeShape& input_shape, const int8_t* input_data, + float input_scale, int input_zero_point, + const RuntimeShape& output_shape, int8_t* output_data, + float output_scale, int output_zero_point) { + const float cutoff_upper = 16.619047164916992188f; + const float cutoff_lower = -9.f; + + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + // Rational for using approximation in reference kernel. + // 0. This approximation gives enough precision for float. + // 1. This works around an issue on an embedded chipset where exp() does not + // return correctly as expected - exp(x) should return inf when overflown + // not 1.701417 IEEE 754 defines representation for inf. + // 2. This will speed up calculation and is matching the behavior in the + // optimized kernels. (check the definition of scalar_logistic_op) + + for (int i = 0; i < flat_size; i++) { + // Dequantize. + float val = + static_cast((input_data[i] - input_zero_point) * input_scale); + float result; + if (val > cutoff_upper) { + result = 1.0f; + } else if (val < cutoff_lower) { + result = std::exp(val); + } else { + result = 1.f / (1.f + std::exp(-val)); + } + // Requantize + int8_t output = + static_cast(result / output_scale + output_zero_point); + output_data[i] = output; + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/maximum_minimum.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/maximum_minimum.h new file mode 100644 index 0000000..646a336 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/maximum_minimum.h @@ -0,0 +1,67 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace reference_ops { + +template +void MaximumMinimumBroadcastSlow(const RuntimeShape& unextended_input1_shape, + const T* input1_data, + const RuntimeShape& unextended_input2_shape, + const T* input2_data, + const RuntimeShape& unextended_output_shape, + T* output_data, Op op) { + // Uses element-wise calculation if broadcast is not required. + if (unextended_input1_shape == unextended_input2_shape) { + const int flat_size = + MatchingElementsSize(unextended_input1_shape, unextended_input2_shape, + unextended_output_shape); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = op(input1_data[i], input2_data[i]); + } + } else { + TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), N); + + NdArrayDesc desc1; + NdArrayDesc desc2; + NdArrayDesc output_desc; + NdArrayDescsForElementwiseBroadcast( + unextended_input1_shape, unextended_input2_shape, &desc1, &desc2); + CopyDimsToDesc(RuntimeShape::ExtendedShape(N, unextended_output_shape), + &output_desc); + + auto maxmin_func = [&](int indexes[N]) { + output_data[SubscriptToIndex(output_desc, indexes)] = + op(input1_data[SubscriptToIndex(desc1, indexes)], + input2_data[SubscriptToIndex(desc2, indexes)]); + }; + NDOpsHelper(output_desc, maxmin_func); + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/mul.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/mul.h similarity index 84% rename from src/tensorflow/lite/kernels/internal/reference/mul.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/mul.h index 54e947d..db5fa65 100644 --- a/src/tensorflow/lite/kernels/internal/reference/mul.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/mul.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_ -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" namespace tflite { @@ -24,20 +25,20 @@ namespace reference_ops { // Element-wise mul that can often be used for inner loop of broadcast Mul as // well as the non-broadcast Mul. inline void MulElementwise(int size, const ArithmeticParams& params, - const uint8* input1_data, const uint8* input2_data, - uint8* output_data) { + const uint8_t* input1_data, + const uint8_t* input2_data, uint8_t* output_data) { for (int i = 0; i < size; ++i) { - const int32 input1_val = params.input1_offset + input1_data[i]; - const int32 input2_val = params.input2_offset + input2_data[i]; - const int32 unclamped_result = + const int32_t input1_val = params.input1_offset + input1_data[i]; + const int32_t input2_val = params.input2_offset + input2_data[i]; + const int32_t unclamped_result = params.output_offset + MultiplyByQuantizedMultiplier(input1_val * input2_val, params.output_multiplier, params.output_shift); - const int32 clamped_output = + const int32_t clamped_output = std::min(params.quantized_activation_max, std::max(params.quantized_activation_min, unclamped_result)); - output_data[i] = static_cast(clamped_output); + output_data[i] = static_cast(clamped_output); } } @@ -60,9 +61,9 @@ inline void Mul(const ArithmeticParams& params, } inline void Mul(const ArithmeticParams& params, - const RuntimeShape& input1_shape, const uint8* input1_data, - const RuntimeShape& input2_shape, const uint8* input2_data, - const RuntimeShape& output_shape, uint8* output_data) { + const RuntimeShape& input1_shape, const uint8_t* input1_data, + const RuntimeShape& input2_shape, const uint8_t* input2_data, + const RuntimeShape& output_shape, uint8_t* output_data) { TFLITE_DCHECK_LE(params.quantized_activation_min, params.quantized_activation_max); const int flat_size = @@ -73,11 +74,11 @@ inline void Mul(const ArithmeticParams& params, inline void BroadcastMul4DSlow(const ArithmeticParams& params, const RuntimeShape& input1_shape, - const uint8* input1_data, + const uint8_t* input1_data, const RuntimeShape& input2_shape, - const uint8* input2_data, + const uint8_t* input2_data, const RuntimeShape& output_shape, - uint8* output_data) { + uint8_t* output_data) { NdArrayDesc<4> desc1; NdArrayDesc<4> desc2; NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, @@ -89,22 +90,22 @@ inline void BroadcastMul4DSlow(const ArithmeticParams& params, for (int y = 0; y < extended_output_shape.Dims(1); ++y) { for (int x = 0; x < extended_output_shape.Dims(2); ++x) { for (int c = 0; c < extended_output_shape.Dims(3); ++c) { - const int32 input1_val = + const int32_t input1_val = params.input1_offset + input1_data[SubscriptToIndex(desc1, b, y, x, c)]; - const int32 input2_val = + const int32_t input2_val = params.input2_offset + input2_data[SubscriptToIndex(desc2, b, y, x, c)]; - const int32 unclamped_result = + const int32_t unclamped_result = params.output_offset + MultiplyByQuantizedMultiplier(input1_val * input2_val, params.output_multiplier, params.output_shift); - const int32 clamped_output = std::min( + const int32_t clamped_output = std::min( params.quantized_activation_max, std::max(params.quantized_activation_min, unclamped_result)); output_data[Offset(extended_output_shape, b, y, x, c)] = - static_cast(clamped_output); + static_cast(clamped_output); } } } @@ -164,3 +165,5 @@ void BroadcastMul4DSlow(const ArithmeticParams& params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/neg.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/neg.h similarity index 91% rename from src/tensorflow/lite/kernels/internal/reference/neg.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/neg.h index e127883..0a23c0a 100644 --- a/src/tensorflow/lite/kernels/internal/reference/neg.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/neg.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_ -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -35,3 +36,5 @@ inline void Negate(const RuntimeShape& input_shape, const T* input_data, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/pad.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/pad.h similarity index 80% rename from src/tensorflow/lite/kernels/internal/reference/pad.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/pad.h index e20aa5e..64fc186 100644 --- a/src/tensorflow/lite/kernels/internal/reference/pad.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/pad.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -18,7 +19,7 @@ limitations under the License. #include -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -32,8 +33,8 @@ constexpr int PadKernelMaxDimensionCount() { return 4; } // equivalent to a simple input1_data. For Pad, it should point to a zero // value. // -// Note that two typenames are required, so that T=P=int32 is considered a -// specialization distinct from P=int32. +// Note that two typenames are required, so that T=P=int32_t is considered a +// specialization distinct from P=int32_t. template inline void PadImpl(const tflite::PadParams& op_params, const RuntimeShape& input_shape, const T* input_data, @@ -116,11 +117,11 @@ inline void Pad(const tflite::PadParams& op_params, output_data); } -// The second (pad-value) input can be int32 when, say, the first is uint8. +// The second (pad-value) input can be int32_t when, say, the first is uint8_t. template inline void Pad(const tflite::PadParams& op_params, const RuntimeShape& input_shape, const T* input_data, - const int32* pad_value_ptr, const RuntimeShape& output_shape, + const int32_t* pad_value_ptr, const RuntimeShape& output_shape, T* output_data) { const T converted_pad_value = static_cast(*pad_value_ptr); PadImpl(op_params, input_shape, input_data, &converted_pad_value, @@ -130,40 +131,18 @@ inline void Pad(const tflite::PadParams& op_params, // This version avoids conflicting template matching. template <> inline void Pad(const tflite::PadParams& op_params, - const RuntimeShape& input_shape, const int32* input_data, - const int32* pad_value_ptr, const RuntimeShape& output_shape, - int32* output_data) { + const RuntimeShape& input_shape, const int32_t* input_data, + const int32_t* pad_value_ptr, const RuntimeShape& output_shape, + int32_t* output_data) { PadImpl(op_params, input_shape, input_data, pad_value_ptr, output_shape, output_data); } -// One could make all PadImageStyle calls simply delegate the work to the -// ordinary Pad. However, it is better that the reference code asserts false in -// similar cases. template inline void PadImageStyle(const tflite::PadParams& op_params, const RuntimeShape& input_shape, const T* input_data, const P* pad_value_ptr, const RuntimeShape& output_shape, T* output_data) { - TFLITE_ASSERT_FALSE; -} - -template -inline void PadImageStyle(const tflite::PadParams& op_params, - const RuntimeShape& input_shape, - const uint8* input_data, const P* pad_value_ptr, - const RuntimeShape& output_shape, - uint8* output_data) { - Pad(op_params, input_shape, input_data, pad_value_ptr, output_shape, - output_data); -} - -template -inline void PadImageStyle(const tflite::PadParams& op_params, - const RuntimeShape& input_shape, - const int8_t* input_data, const P* pad_value_ptr, - const RuntimeShape& output_shape, - int8_t* output_data) { Pad(op_params, input_shape, input_data, pad_value_ptr, output_shape, output_data); } @@ -182,3 +161,5 @@ inline void PadImageStyle(const tflite::PadParams& op_params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PAD_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/pooling.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/pooling.h new file mode 100644 index 0000000..19f1e84 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/pooling.h @@ -0,0 +1,300 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace reference_ops { + +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + float total = 0.f; + float filter_count = 0; + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + total += + input_data[Offset(input_shape, batch, in_y, in_x, channel)]; + filter_count++; + } + } + const float average = total / filter_count; + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + ActivationFunctionWithMinMax(average, params.float_activation_min, + params.float_activation_max); + } + } + } + } +} + +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const uint8_t* input_data, + const RuntimeShape& output_shape, + uint8_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + int32_t acc = 0; + int filter_count = 0; + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + acc += + input_data[Offset(input_shape, batch, in_y, in_x, channel)]; + filter_count++; + } + } + acc = (acc + filter_count / 2) / filter_count; + acc = std::max(acc, params.quantized_activation_min); + acc = std::min(acc, params.quantized_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + static_cast(acc); + } + } + } + } +} + +inline void L2Pool(const PoolParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + float sum_squares = 0.f; + int filter_count = 0; + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + const float val = + input_data[Offset(input_shape, batch, in_y, in_x, channel)]; + sum_squares += val * val; + filter_count++; + } + } + const float l2pool_result = std::sqrt(sum_squares / filter_count); + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + ActivationFunctionWithMinMax(l2pool_result, + params.float_activation_min, + params.float_activation_max); + } + } + } + } +} + +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + float max = std::numeric_limits::lowest(); + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + max = std::max( + max, + input_data[Offset(input_shape, batch, in_y, in_x, channel)]); + } + } + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + ActivationFunctionWithMinMax(max, params.float_activation_min, + params.float_activation_max); + } + } + } + } +} + +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& output_shape, + uint8_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + TFLITE_DCHECK_GE(params.quantized_activation_min, 0); + TFLITE_DCHECK_LE(params.quantized_activation_max, 255); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int channel = 0; channel < depth; ++channel) { + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; + // Compute the boundaries of the filter region clamped so as to + // ensure that the filter window fits in the input array. + const int filter_x_start = std::max(0, -in_x_origin); + const int filter_x_end = + std::min(params.filter_width, input_width - in_x_origin); + const int filter_y_start = std::max(0, -in_y_origin); + const int filter_y_end = + std::min(params.filter_height, input_height - in_y_origin); + uint8_t max = 0; + for (int filter_y = filter_y_start; filter_y < filter_y_end; + ++filter_y) { + for (int filter_x = filter_x_start; filter_x < filter_x_end; + ++filter_x) { + const int in_x = in_x_origin + filter_x; + const int in_y = in_y_origin + filter_y; + max = std::max( + max, + input_data[Offset(input_shape, batch, in_y, in_x, channel)]); + } + } + max = std::max(max, params.quantized_activation_min); + max = std::min(max, params.quantized_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, channel)] = + static_cast(max); + } + } + } + } +} +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/prelu.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/prelu.h new file mode 100644 index 0000000..e858430 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/prelu.h @@ -0,0 +1,112 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +// Broadcast prelu to output_shape for quantized uint8_t/int8_t data. +template +inline void BroadcastPrelu4DSlow( + const PreluParams& params, const RuntimeShape& input_shape, + const T* input_data, const RuntimeShape& alpha_shape, const T* alpha_data, + const RuntimeShape& output_shape, T* output_data) { + TFLITE_DCHECK_LE(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(alpha_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(output_shape.DimensionsCount(), 4); + const RuntimeShape extended_output_shape = + RuntimeShape::ExtendedShape(4, output_shape); + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(input_shape, alpha_shape, &desc1, &desc2); + + for (int b = 0; b < extended_output_shape.Dims(0); ++b) { + for (int y = 0; y < extended_output_shape.Dims(1); ++y) { + for (int x = 0; x < extended_output_shape.Dims(2); ++x) { + for (int c = 0; c < extended_output_shape.Dims(3); ++c) { + int output_index = Offset(extended_output_shape, b, y, x, c); + int input_index = SubscriptToIndex(desc1, b, y, x, c); + const int32_t input_value = + params.input_offset + input_data[input_index]; + int32_t output_value; + if (input_value >= 0) { + output_value = MultiplyByQuantizedMultiplier( + input_value, params.output_multiplier_1, params.output_shift_1); + } else { + auto alpha_index = SubscriptToIndex(desc2, b, y, x, c); + const int32_t alpha_value = + params.alpha_offset + alpha_data[alpha_index]; + + output_value = MultiplyByQuantizedMultiplier( + input_value * alpha_value, params.output_multiplier_2, + params.output_shift_2); + } + output_value += params.output_offset; + + const int32_t quantized_min = std::numeric_limits::min(); + const int32_t quantized_max = std::numeric_limits::max(); + const int32_t clamped_output = + std::min(quantized_max, std::max(quantized_min, output_value)); + output_data[output_index] = static_cast(clamped_output); + } + } + } + } +} + +template +inline void Prelu(const PreluParams& params, const RuntimeShape& input_shape, + const T* input_data, const RuntimeShape& alpha_shape, + const T* alpha_data, const RuntimeShape& output_shape, + T* output_data) { + const int32_t quantized_min = std::numeric_limits::min(); + const int32_t quantized_max = std::numeric_limits::max(); + + const int flat_size = + MatchingElementsSize(input_shape, alpha_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + const int32_t input_value = params.input_offset + input_data[i]; + int32_t output_value; + if (input_value >= 0) { + output_value = MultiplyByQuantizedMultiplier( + input_value, params.output_multiplier_1, params.output_shift_1); + } else { + const int32_t alpha_value = params.alpha_offset + alpha_data[i]; + + output_value = MultiplyByQuantizedMultiplier(input_value * alpha_value, + params.output_multiplier_2, + params.output_shift_2); + } + output_value += params.output_offset; + + const int32_t clamped_output = + std::min(quantized_max, std::max(quantized_min, output_value)); + output_data[i] = static_cast(clamped_output); + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h new file mode 100644 index 0000000..1a687dc --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h @@ -0,0 +1,141 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +// Consolidates dimensions in broadcast inputs, checks for five-fold pattern. +// +// For example, if sequence of dimensions of one input is +// ..., 1, 3, 1, 7, 9, 5,... and the other is ..., 2, 3, 1, 7, 1, 1, ... +// we can consolidate these as +// ..., 1, 3*7, 9*5, ... and 2, 3*7, 1. +// +// The category is updated in the less-frequent case of shapes that are +// not suited to a fivefold-loop broadcast. +// +// Falls back to generic pattern when it does not know how to process properly. +// +// Returns true iff there is some sort of broadcast, which includes five-fold +// patterns and falling back to generic broadcast. +inline bool ProcessBroadcastShapes(const RuntimeShape& shape0, + const RuntimeShape& shape1, + tflite::ArithmeticParams* params) { + const int dims_count = + std::max(shape0.DimensionsCount(), shape1.DimensionsCount()); + + params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast; + RuntimeShape scalar_shape(dims_count, 1); + + auto extended_shape0 = RuntimeShape::ExtendedShape(dims_count, shape0); + auto extended_shape1 = RuntimeShape::ExtendedShape(dims_count, shape1); + + // Check for "exact" match, implicitly accepting any scalar shapes. + if (extended_shape0 == extended_shape1) { + params->broadcast_category = BroadcastableOpCategory::kNonBroadcast; + return false; + } + + for (int i = dims_count - 1; i >= 0; --i) { + if (extended_shape0.Dims(i) == extended_shape1.Dims(i)) { + continue; + } else if (extended_shape0.Dims(i) == 1) { + params->broadcast_category = + BroadcastableOpCategory::kFirstInputBroadcastsFast; + break; + } else if (extended_shape1.Dims(i) == 1) { + params->broadcast_category = + BroadcastableOpCategory::kSecondInputBroadcastsFast; + break; + } else { + // This case is erroneous: there is a dimension that does not match and + // is not a broadcast from one shape to the other. + params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast; + return true; + } + } + + if (params->broadcast_category != + BroadcastableOpCategory::kFirstInputBroadcastsFast && + params->broadcast_category != + BroadcastableOpCategory::kSecondInputBroadcastsFast) { + // This is unreachable because at least one else clause in the above loop + // must be reached. + TFLITE_DCHECK(false); + params->broadcast_category = BroadcastableOpCategory::kNonBroadcast; + return false; + } + + // From this point it is assumed contractually that corresponding dimensions + // in shape0 and shape1 are either (a) equal or (b) one or other equals 1. + const bool swap_inputs = params->broadcast_category == + BroadcastableOpCategory::kSecondInputBroadcastsFast; + const RuntimeShape* shape_a = + swap_inputs ? &extended_shape1 : &extended_shape0; + const RuntimeShape* shape_b = + swap_inputs ? &extended_shape0 : &extended_shape1; + + int i = dims_count - 1; + params->broadcast_shape[0] = 1; + params->broadcast_shape[1] = 1; + params->broadcast_shape[2] = 1; + params->broadcast_shape[3] = 1; + params->broadcast_shape[4] = 1; + // y_0 is greedy: include dims if both or neither equal 1: in other words, + // test for equality rather than (shape_a->Dims(i) != 1). + while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) { + params->broadcast_shape[4] *= shape_b->Dims(i); + --i; + } + // Here either input_a or input_b has dim of 1 (if i >= 0). If it is input_b + // that has the unit dimension, the next two loops are not entered. + while (i >= 0 && shape_a->Dims(i) == 1) { + params->broadcast_shape[3] *= shape_b->Dims(i); + --i; + } + while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) { + params->broadcast_shape[2] *= shape_a->Dims(i); + --i; + } + // Here either input_a or input_b has dim of 1 (if i >= 0). + while (i >= 0 && shape_b->Dims(i) == 1) { + params->broadcast_shape[1] *= shape_a->Dims(i); + --i; + } + while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) { + params->broadcast_shape[0] *= shape_b->Dims(i); + --i; + } + + // Rarer case is when the broadcast dimensions cannot be handled by a fivefold + // loop. + if (i >= 0) { + params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast; + } + return true; +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/quantize.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/quantize.h new file mode 100644 index 0000000..b8d51a4 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/quantize.h @@ -0,0 +1,58 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +template +inline void AffineQuantize(const tflite::QuantizationParams& op_params, + const RuntimeShape& input_shape, + const InputT* input_data, + const RuntimeShape& output_shape, + OutputT* output_data) { + const int32_t zero_point = op_params.zero_point; + const double scale = op_params.scale; + const int flat_size = MatchingFlatSize(input_shape, output_shape); + static constexpr int32_t min_val = std::numeric_limits::min(); + static constexpr int32_t max_val = std::numeric_limits::max(); + + for (int i = 0; i < flat_size; i++) { + const InputT val = input_data[i]; + int32_t unclamped = + static_cast(TfLiteRound(val / static_cast(scale))) + + zero_point; + int32_t clamped = std::min(std::max(unclamped, min_val), max_val); + output_data[i] = clamped; + } +} + +} // namespace reference_ops + +} // namespace tflite +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/reduce.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/reduce.h new file mode 100644 index 0000000..00744ec --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/reduce.h @@ -0,0 +1,415 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REDUCE_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REDUCE_H_ + +#include "tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/max.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/min.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +// A generic reduce method that can be used for reduce_sum, reduce_mean, etc. +// This method iterates through input data and reduce elements along the +// dimensions given in axis. +template +inline bool Reduce(const In* input_data, const int* input_dims, + const int* output_dims, const int input_num_dims, + const int output_num_dims, const int* axis, + const int num_axis, int* input_iter, + Out reducer(const Out current, const In in), + Out* output_data) { + // Reset input iterator. + for (int idx = 0; idx < input_num_dims; ++idx) { + input_iter[idx] = 0; + } + // Iterate through input_data. + do { + size_t input_offset = + ReducedOutputOffset(input_num_dims, input_dims, input_iter, 0, nullptr); + size_t output_offset = ReducedOutputOffset(input_num_dims, input_dims, + input_iter, num_axis, axis); + output_data[output_offset] = + reducer(output_data[output_offset], input_data[input_offset]); + } while (NextIndex(input_num_dims, input_dims, input_iter)); + return true; +} + +// This method parses the input 'axis' to remove duplicates and handle negative +// values, and returns a valid 'out_axis' +inline bool ResolveAxis(const int num_dims, const int* axis, + const int64_t num_axis, int* out_axis, + int* out_num_axis) { + *out_num_axis = 0; // Just in case. + // Short-circuit axis resolution for scalars; the axis will go unused. + if (num_dims == 0) { + return true; + } + // o(n^2) is fine since out_num_axis should be really small, mostly <= 4 + for (int64_t idx = 0; idx < num_axis; ++idx) { + // Handle negative index. A positive index 'p_idx' can be represented as a + // negative index 'n_idx' as: n_idx = p_idx-num_dims + // eg: For num_dims=3, [0, 1, 2] is the same as [-3, -2, -1] */ + int current = axis[idx] < 0 ? (axis[idx] + num_dims) : axis[idx]; + TFLITE_DCHECK(current >= 0 && current < num_dims); + if (current < 0 || current >= num_dims) { + return false; + } + bool is_dup = false; + for (int j = 0; j < *out_num_axis; ++j) { + if (out_axis[j] == current) { + is_dup = true; + break; + } + } + if (!is_dup) { + out_axis[*out_num_axis] = current; + *out_num_axis += 1; + } + } + return true; +} + +// This method expects that output_data has been initialized. +template +inline bool ReduceSumImpl(const In* input_data, const int* input_dims, + const int* output_dims, const int input_num_dims, + const int output_num_dims, const int* axis, + const int num_axis, int* input_iter, + Out* output_data) { + auto reducer = [](const Out current, const In in) -> Out { + const Out actual_in = static_cast(in); + return current + actual_in; + }; + return Reduce(input_data, input_dims, output_dims, input_num_dims, + output_num_dims, axis, num_axis, input_iter, reducer, + output_data); +} + +template +inline bool InitTensorDataForReduce(const int* dims, const int num_dims, + const T init_value, T* data) { + size_t num_elements = 1; + for (int idx = 0; idx < num_dims; ++idx) { + size_t current = static_cast(dims[idx]); + // Overflow prevention. + if (num_elements > std::numeric_limits::max() / current) { + return false; + } + num_elements *= current; + } + for (size_t idx = 0; idx < num_elements; ++idx) { + data[idx] = init_value; + } + return true; +} + +// Computes the generic value (i.e., sum/max/min/prod) of elements across +// dimensions given in axis. It needs to pass in init_value and reducer. +template +inline bool ReduceGeneric(const T* input_data, const int* input_dims, + const int input_num_dims, T* output_data, + const int* output_dims, const int output_num_dims, + const int* axis, const int64_t num_axis_dimensions, + bool keep_dims, int* temp_index, int* resolved_axis, + T init_value, + T reducer(const T current, const T in)) { + // Return early when input shape has zero dim. + for (int i = 0; i < input_num_dims; ++i) { + if (input_dims[i] == 0) return true; + } + + // Reset output data. + if (!InitTensorDataForReduce(output_dims, output_num_dims, init_value, + output_data)) { + return false; + } + + // Resolve axis. + int num_resolved_axis = 0; + if (!ResolveAxis(input_num_dims, axis, num_axis_dimensions, resolved_axis, + &num_resolved_axis)) { + return false; + } + + return Reduce(input_data, input_dims, output_dims, input_num_dims, + output_num_dims, resolved_axis, num_resolved_axis, + temp_index, reducer, output_data); +} + +// Computes the mean of elements across dimensions given in axis. +// It does so in two stages, first calculates the sum of elements along the axis +// then divides it by the number of element in axis. +template +inline bool Mean(const T* input_data, const int* input_dims, + const int input_num_dims, T* output_data, + const int* output_dims, const int output_num_dims, + const int* axis, const int num_axis_dimensions, bool keep_dims, + int* temp_index, int* resolved_axis, U* temp_sum) { + ruy::profiler::ScopeLabel label("Mean"); + // Reset output data. + size_t num_outputs = 1; + for (int idx = 0; idx < output_num_dims; ++idx) { + size_t current = static_cast(output_dims[idx]); + // Overflow prevention. + if (num_outputs > std::numeric_limits::max() / current) { + return false; + } + num_outputs *= current; + } + for (size_t idx = 0; idx < num_outputs; ++idx) { + output_data[idx] = T(); + temp_sum[idx] = U(); + } + + // Resolve axis. + int num_resolved_axis = 0; + if (!ResolveAxis(input_num_dims, axis, num_axis_dimensions, resolved_axis, + &num_resolved_axis)) { + return false; + } + + if (!ReduceSumImpl(input_data, input_dims, output_dims, input_num_dims, + output_num_dims, resolved_axis, num_resolved_axis, + temp_index, temp_sum)) { + return false; + } + + // Calculate mean by dividing output_data by num of aggregated element. + size_t num_elements_in_axis = 1; + for (int idx = 0; idx < num_resolved_axis; ++idx) { + size_t current = static_cast(input_dims[resolved_axis[idx]]); + // Overflow prevention. + if (current > (std::numeric_limits::max() / num_elements_in_axis)) { + return false; + } + num_elements_in_axis *= current; + } + + if (num_elements_in_axis > 0) { + for (size_t idx = 0; idx < num_outputs; ++idx) { + output_data[idx] = + static_cast(temp_sum[idx] / static_cast(num_elements_in_axis)); + } + } + return true; +} + +template +inline void Mean(const tflite::MeanParams& op_params, + const RuntimeShape& unextended_input_shape, + const T* input_data, + const RuntimeShape& unextended_output_shape, T* output_data) { + ruy::profiler::ScopeLabel label("Mean4D"); + + // Current implementation only supports dimension equals 4 and simultaneous + // reduction over width and height. + TFLITE_CHECK_EQ(unextended_input_shape.DimensionsCount(), 4); + TFLITE_CHECK_LE(unextended_output_shape.DimensionsCount(), 4); + const RuntimeShape input_shape = + RuntimeShape::ExtendedShape(4, unextended_input_shape); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(4, unextended_output_shape); + + const int output_batch = output_shape.Dims(0); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int output_depth = output_shape.Dims(3); + + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + + TFLITE_CHECK_EQ(op_params.axis_count, 2); + TFLITE_CHECK((op_params.axis[0] == 1 && op_params.axis[1] == 2) || + (op_params.axis[0] == 2 && op_params.axis[1] == 1)); + TFLITE_CHECK_EQ(output_height, 1); + TFLITE_CHECK_EQ(output_width, 1); + + for (int out_b = 0; out_b < output_batch; ++out_b) { + for (int out_d = 0; out_d < output_depth; ++out_d) { + float value = 0; + for (int in_h = 0; in_h < input_height; ++in_h) { + for (int in_w = 0; in_w < input_width; ++in_w) { + value += input_data[Offset(input_shape, out_b, in_h, in_w, out_d)]; + } + } + output_data[Offset(output_shape, out_b, 0, 0, out_d)] = + value / (input_width * input_height); + } + } +} + +inline void Mean(const tflite::MeanParams& op_params, + const RuntimeShape& unextended_input_shape, + const uint8_t* input_data, int32_t input_zero_point, + float input_scale, const RuntimeShape& unextended_output_shape, + uint8_t* output_data, int32_t output_zero_point, + float output_scale) { + ruy::profiler::ScopeLabel label("Mean4D/Uint8"); + + // Current implementation only supports dimension equals 4 and simultaneous + // reduction over width and height. + TFLITE_CHECK_EQ(unextended_input_shape.DimensionsCount(), 4); + TFLITE_CHECK_LE(unextended_output_shape.DimensionsCount(), 4); + const RuntimeShape input_shape = + RuntimeShape::ExtendedShape(4, unextended_input_shape); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(4, unextended_output_shape); + const int output_batch = output_shape.Dims(0); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int output_depth = output_shape.Dims(3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const float num_elements_in_axis = input_width * input_height; + + TFLITE_CHECK_EQ(op_params.axis_count, 2); + TFLITE_CHECK((op_params.axis[0] == 1 && op_params.axis[1] == 2) || + (op_params.axis[0] == 2 && op_params.axis[1] == 1)); + TFLITE_CHECK_EQ(output_height, 1); + TFLITE_CHECK_EQ(output_width, 1); + + constexpr int32_t kMinValue = std::numeric_limits::min(); + constexpr int32_t kMaxValue = std::numeric_limits::max(); + + int32_t bias = + output_zero_point - + static_cast(input_zero_point * input_scale / output_scale); + double real_scale = + static_cast(input_scale / (num_elements_in_axis * output_scale)); + + int32_t multiplier; + int shift; + QuantizeMultiplier(real_scale, &multiplier, &shift); + for (int out_b = 0; out_b < output_batch; ++out_b) { + for (int out_d = 0; out_d < output_depth; ++out_d) { + int32_t acc = 0; + for (int in_h = 0; in_h < input_height; ++in_h) { + for (int in_w = 0; in_w < input_width; ++in_w) { + acc += input_data[Offset(input_shape, out_b, in_h, in_w, out_d)]; + } + } + acc = MultiplyByQuantizedMultiplier(acc, multiplier, shift); + acc += bias; + acc = std::min(std::max(acc, kMinValue), kMaxValue); + output_data[Offset(output_shape, out_b, 0, 0, out_d)] = + static_cast(acc); + } + } +} + +// Computes the mean of elements across dimensions given in axis. +// It does so in two stages, first calculates the sum of elements along the axis +// then divides it by the number of element in axis for quantized values. +template +inline bool QuantizedMeanOrSum(const T* input_data, int32_t input_zero_point, + float input_scale, const int* input_dims, + const int input_num_dims, T* output_data, + int32_t output_zero_point, float output_scale, + const int* output_dims, + const int output_num_dims, const int* axis, + const int num_axis_dimensions, bool keep_dims, + int* temp_index, int* resolved_axis, U* temp_sum, + bool compute_sum) { + const bool uint8_case = std::is_same::value; + const bool int16_case = std::is_same::value; + if (uint8_case) { + ruy::profiler::ScopeLabel label(compute_sum ? "Sum/Uint8" : "Mean/Uint8"); + } else if (int16_case) { + ruy::profiler::ScopeLabel label(compute_sum ? "Sum/Int16" : "Mean/Int16"); + } else { + ruy::profiler::ScopeLabel label(compute_sum ? "Sum/Int8" : "Mean/Int8"); + } + // Reset output data. + size_t num_outputs = 1; + for (int idx = 0; idx < output_num_dims; ++idx) { + size_t current = static_cast(output_dims[idx]); + // Overflow prevention. + if (num_outputs > std::numeric_limits::max() / current) { + return false; + } + num_outputs *= current; + } + for (size_t idx = 0; idx < num_outputs; ++idx) { + output_data[idx] = T(); + temp_sum[idx] = U(); + } + + // Resolve axis. + int num_resolved_axis = 0; + if (!ResolveAxis(input_num_dims, axis, num_axis_dimensions, resolved_axis, + &num_resolved_axis)) { + return false; + } + + if (!ReduceSumImpl(input_data, input_dims, output_dims, input_num_dims, + output_num_dims, resolved_axis, num_resolved_axis, + temp_index, temp_sum)) { + return false; + } + + // Calculate mean by dividing output_data by num of aggregated element. + size_t num_elements_in_axis = 1; + for (int idx = 0; idx < num_resolved_axis; ++idx) { + size_t current = static_cast(input_dims[resolved_axis[idx]]); + // Overflow prevention. + if (current > (std::numeric_limits::max() / num_elements_in_axis)) { + return false; + } + num_elements_in_axis *= current; + } + + if (num_elements_in_axis > 0) { + const float scale = input_scale / output_scale; + if (compute_sum) { + // TODO(b/116341117): Eliminate float and do this completely in 8bit. + const float bias = -input_zero_point * scale * num_elements_in_axis; + for (size_t idx = 0; idx < num_outputs; ++idx) { + const U value = + static_cast(TfLiteRound(temp_sum[idx] * scale + bias)) + + output_zero_point; + output_data[idx] = static_cast(value); + } + } else { + const float bias = -input_zero_point * scale; + for (size_t idx = 0; idx < num_outputs; ++idx) { + float float_mean = static_cast(temp_sum[idx]) / + static_cast(num_elements_in_axis); + float result = TfLiteMin( + TfLiteRound(float_mean * scale + bias) + output_zero_point, + static_cast(std::numeric_limits::max())); + result = TfLiteMax(result, + static_cast(std::numeric_limits::min())); + output_data[idx] = static_cast(result); + } + } + } + return true; +} + +} // namespace reference_ops + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REDUCE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/requantize.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/requantize.h new file mode 100644 index 0000000..8c20095 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/requantize.h @@ -0,0 +1,71 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_ + +#include "tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace reference_ops { + +template +inline void Requantize(const input_type* input_data, int32_t size, + int32_t effective_scale_multiplier, + int32_t effective_scale_shift, int32_t input_zeropoint, + int32_t output_zeropoint, output_type* output_data) { + ruy::profiler::ScopeLabel label("Requantize"); + const bool same_scale = + (effective_scale_multiplier == 1 << 30 && effective_scale_shift == 1); + if (same_scale) { + const bool mixed_type_int8_uint8 = + std::is_same::value && + std::is_same::value; + const bool mixed_type_uint8_int8 = + std::is_same::value && + std::is_same::value; + const int32_t zero_point_diff = input_zeropoint - output_zeropoint; + // Fast path to do requantization for the case when just a shift of 128 is + // needed. + if ((mixed_type_int8_uint8 && zero_point_diff == -128) || + (mixed_type_uint8_int8 && zero_point_diff == 128)) { + for (int i = 0; i < size; ++i) { + output_data[i] = input_data[i] ^ 0x80; + } + return; + } + } + static constexpr int32_t kMinOutput = std::numeric_limits::min(); + static constexpr int32_t kMaxOutput = std::numeric_limits::max(); + for (int i = 0; i < size; ++i) { + const int32_t input = input_data[i] - input_zeropoint; + const int32_t output = + MultiplyByQuantizedMultiplier(input, effective_scale_multiplier, + effective_scale_shift) + + output_zeropoint; + const int32_t clamped_output = + std::max(std::min(output, kMaxOutput), kMinOutput); + output_data[i] = static_cast(clamped_output); + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h new file mode 100644 index 0000000..b5935f8 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h @@ -0,0 +1,104 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_NEAREST_NEIGHBOR_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_NEAREST_NEIGHBOR_H_ + +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline int32_t GetNearestNeighbor(const int input_value, + const int32_t input_size, + const int32_t output_size, + const bool align_corners, + const bool half_pixel_centers) { + const float scale = + (align_corners && output_size > 1) + ? (input_size - 1) / static_cast(output_size - 1) + : input_size / static_cast(output_size); + const float offset = half_pixel_centers ? 0.5f : 0.0f; + int32_t output_value = std::min( + align_corners + ? static_cast(TfLiteRound((input_value + offset) * scale)) + : static_cast(std::floor((input_value + offset) * scale)), + input_size - 1); + if (half_pixel_centers) { + output_value = std::max(static_cast(0), output_value); + } + return output_value; +} + +template +inline void ResizeNearestNeighbor( + const tflite::ResizeNearestNeighborParams& op_params, + const RuntimeShape& unextended_input_shape, const T* input_data, + const RuntimeShape& output_size_shape, const int32_t* output_size_data, + const RuntimeShape& unextended_output_shape, T* output_data) { + TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4); + + const RuntimeShape input_shape = + RuntimeShape::ExtendedShape(4, unextended_input_shape); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(4, unextended_output_shape); + + int32_t batches = MatchingDim(input_shape, 0, output_shape, 0); + int32_t input_height = input_shape.Dims(1); + int32_t input_width = input_shape.Dims(2); + int32_t depth = MatchingDim(input_shape, 3, output_shape, 3); + + // The Tensorflow version of this op allows resize on the width and height + // axis only. + TFLITE_DCHECK_EQ(output_size_shape.FlatSize(), 2); + int32_t output_height = output_size_data[0]; + int32_t output_width = output_size_data[1]; + + const int col_offset = input_shape.Dims(3); + const int row_offset = input_shape.Dims(2) * col_offset; + const int batch_offset = input_shape.Dims(1) * row_offset; + + const T* input_ptr = input_data; + T* output_ptr = output_data; + for (int b = 0; b < batches; ++b) { + for (int y = 0; y < output_height; ++y) { + int32_t in_y = GetNearestNeighbor(y, input_height, output_height, + op_params.align_corners, + op_params.half_pixel_centers); + const T* y_input_ptr = input_ptr + in_y * row_offset; + for (int x = 0; x < output_width; ++x) { + int32_t in_x = GetNearestNeighbor(x, input_width, output_width, + op_params.align_corners, + op_params.half_pixel_centers); + const T* x_input_ptr = y_input_ptr + in_x * col_offset; + memcpy(output_ptr, x_input_ptr, depth * sizeof(T)); + output_ptr += depth; + } + } + input_ptr += batch_offset; + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_NEAREST_NEIGHBOR_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/round.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/round.h similarity index 93% rename from src/tensorflow/lite/kernels/internal/reference/round.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/round.h index 9bd8f3f..17334d5 100644 --- a/src/tensorflow/lite/kernels/internal/reference/round.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/round.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,7 +18,7 @@ limitations under the License. #include -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -49,3 +50,5 @@ inline void Round(const RuntimeShape& input_shape, const float* input_data, } // namespace reference_ops } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/softmax.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/softmax.h new file mode 100644 index 0000000..48f9f49 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/softmax.h @@ -0,0 +1,235 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_ + +#include + +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" + +namespace tflite { +namespace reference_ops { + +inline void Softmax(const SoftmaxParams& params, + const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const int trailing_dim = input_shape.DimensionsCount() - 1; + const int outer_size = + MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); + const int depth = + MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); + + for (int i = 0; i < outer_size; ++i) { + // Find max element value which we'll use to ensure numerical stability + // taking advantage of the following equality: + // exp(x[i])/sum(exp(x[i])) == exp(x[i]+C)/sum(exp(x[i]+C)) + float max = std::numeric_limits::lowest(); + for (int c = 0; c < depth; ++c) { + max = std::max(max, input_data[i * depth + c]); + } + + // Compute sum. + float sum = 0.f; + for (int c = 0; c < depth; ++c) { + const float exp_c = std::exp((input_data[i * depth + c] - max) * + static_cast(params.beta)); + output_data[i * depth + c] = exp_c; + sum += exp_c; + } + + // Compute result. + for (int c = 0; c < depth; ++c) { + output_data[i * depth + c] = output_data[i * depth + c] / sum; + } + } +} + +// Quantized softmax with int8_t/uint8_t input and int8_t/uint8_t/int16_t +// output. +template +inline void Softmax(const SoftmaxParams& params, + const RuntimeShape& input_shape, const InputT* input_data, + const RuntimeShape& output_shape, OutputT* output_data) { + const int32_t input_beta_multiplier = params.input_multiplier; + const int32_t input_beta_left_shift = params.input_left_shift; + const int diff_min = params.diff_min; + // The representation chosen for the input to the exp() function is Q5.26. + // We need to leave extra space since values that we skip might be as large as + // -32 before multiplying by input_beta_multiplier, and therefore as large as + // -16 afterwards. Note that exp(-8) is definitely not insignificant to + // accumulation, but exp(-16) definitely is. + static const int kScaledDiffIntegerBits = 5; + static const int kAccumulationIntegerBits = 12; + using FixedPointScaledDiff = + gemmlowp::FixedPoint; + using FixedPointAccum = + gemmlowp::FixedPoint; + using FixedPoint0 = gemmlowp::FixedPoint; + + const int trailing_dim = input_shape.DimensionsCount() - 1; + const int outer_size = + MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); + const int depth = + MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); + + for (int i = 0; i < outer_size; ++i) { + InputT max_in_row = std::numeric_limits::min(); + for (int c = 0; c < depth; ++c) { + max_in_row = std::max(max_in_row, input_data[i * depth + c]); + } + + FixedPointAccum sum_of_exps = FixedPointAccum::Zero(); + for (int c = 0; c < depth; ++c) { + int32_t input_diff = + static_cast(input_data[i * depth + c]) - max_in_row; + if (input_diff >= diff_min) { + const int32_t input_diff_rescaled = + MultiplyByQuantizedMultiplierGreaterThanOne( + input_diff, input_beta_multiplier, input_beta_left_shift); + const FixedPointScaledDiff scaled_diff_f8 = + FixedPointScaledDiff::FromRaw(input_diff_rescaled); + sum_of_exps = sum_of_exps + gemmlowp::Rescale( + exp_on_negative_values(scaled_diff_f8)); + } + } + + int num_bits_over_unit; + FixedPoint0 shifted_scale = FixedPoint0::FromRaw(GetReciprocal( + sum_of_exps.raw(), kAccumulationIntegerBits, &num_bits_over_unit)); + + for (int c = 0; c < depth; ++c) { + int32_t input_diff = + static_cast(input_data[i * depth + c]) - max_in_row; + if (input_diff >= diff_min) { + const int32_t input_diff_rescaled = + MultiplyByQuantizedMultiplierGreaterThanOne( + input_diff, input_beta_multiplier, input_beta_left_shift); + const FixedPointScaledDiff scaled_diff_f8 = + FixedPointScaledDiff::FromRaw(input_diff_rescaled); + + FixedPoint0 exp_in_0 = exp_on_negative_values(scaled_diff_f8); + int32_t unsat_output = gemmlowp::RoundingDivideByPOT( + (shifted_scale * exp_in_0).raw(), + num_bits_over_unit + 31 - (sizeof(OutputT) * 8)); + + const int32_t shifted_output = + unsat_output + + static_cast(std::numeric_limits::min()); + + output_data[i * depth + c] = static_cast(std::max( + std::min(shifted_output, + static_cast(std::numeric_limits::max())), + static_cast(std::numeric_limits::min()))); + } else { + output_data[i * depth + c] = std::numeric_limits::min(); + } + } + } +} + +// Computes exp(input - max_input) +inline int16_t SoftMaxCalculateExp(const SoftmaxParams& params, + const int16_t* input_data, const int depth, + int16_t max_in_row, int i, int c) { + int32_t input_diff = input_data[i * depth + c] - max_in_row; + // scale the input_diff such that [-65535, 0] correspond to [-10.0, 0.0] + // exp lut generated with range [-10, 0], as exp(-10) is negligible. + int32_t scaled_diff = MultiplyByQuantizedMultiplier( + input_diff, params.input_multiplier, params.input_left_shift); + // recenter to [-32768, 32767] + int32_t sym_scaled_diff = scaled_diff + 32767; + int16_t sat_sym_scaled_diff = + std::min(std::max(sym_scaled_diff, static_cast(-32768)), + static_cast(32767)); + // apply the exp() LUT activation function + return generic_int16_table_lookup(sat_sym_scaled_diff, params.exp_lut); +} +// Quantized softmax with int16_t input and int16_t output. +inline void SoftmaxInt16(const SoftmaxParams& params, + const RuntimeShape& input_shape, + const int16_t* input_data, + const RuntimeShape& output_shape, + int16_t* output_data) { + const int trailing_dim = input_shape.DimensionsCount() - 1; + const int outer_size = + MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); + const int depth = + MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); + + for (int i = 0; i < outer_size; ++i) { + // Find the largest element + int16_t max_in_row = std::numeric_limits::min(); + for (int c = 0; c < depth; ++c) { + max_in_row = std::max(max_in_row, input_data[i * depth + c]); + } + + // This loops computes the exp values and their sum. We will need the exp + // values later on in the function so we cache them in the output_data + // buffer. This is an optimization done to avoid calculating the exp values + // twice making use of the output_data buffer as scratch memory. + int32_t sum_of_exps = 0; // Q16.15 fixed point format. + int16_t* exp_results_Q015 = output_data + i * depth; + for (int c = 0; c < depth; ++c) { + exp_results_Q015[c] = + SoftMaxCalculateExp(params, input_data, depth, max_in_row, i, c); + sum_of_exps += exp_results_Q015[c]; + } + + // Compute the reciprocal 1/sum_of_exps + uint8_t headroom_plus_one = + CountLeadingZeros(static_cast(sum_of_exps)); + int32_t shifted_sum = + ((static_cast(sum_of_exps) << (headroom_plus_one - 1)) + + (1 << 13)) >> + 14; + // since the LUT computes 1/(1 + x) we need to first compute x = (sum - 1). + // also, the LUT expects a symmetrical input, so we must also recenter x + // from [0, 65535] to [-32768, 32767]. + int32_t sym_shifted_sum = shifted_sum + (-((1 << 15) + (1 << 16))); + int16_t sat_sym_shifted_sum = static_cast( + std::min(std::max(sym_shifted_sum, static_cast(-32768)), + static_cast(32767))); + // apply 1/(1 + x) LUT activation function + int16_t reciprocal_scale_Q015 = generic_int16_table_lookup( + sat_sym_shifted_sum, params.one_over_one_plus_x_lut); + + // Rescale the exp_result with reciprocal + // range of output is [0, 32767] correspond to [0.0, 1.0] + for (int c = 0; c < depth; ++c) { + uint8_t right_shift = 31 - headroom_plus_one; + int64_t round = 1 << (right_shift - 1); + int32_t result = (static_cast(exp_results_Q015[c]) * + static_cast(reciprocal_scale_Q015) + + round) >> + right_shift; + output_data[i * depth + c] = static_cast( + std::min(std::max(result, static_cast(0)), + static_cast(32767))); + } + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/strided_slice.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/strided_slice.h new file mode 100644 index 0000000..67f5845 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/strided_slice.h @@ -0,0 +1,124 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_ + +#include "tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/strided_slice_logic.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +template +inline void StridedSlice(const tflite::StridedSliceParams& op_params, + const RuntimeShape& unextended_input_shape, + const RuntimeShape& unextended_output_shape, + SequentialTensorWriter* writer) { + using strided_slice::LoopCondition; + using strided_slice::StartForAxis; + using strided_slice::StopForAxis; + + ruy::profiler::ScopeLabel label("StridedSlice"); + + // Note that the output_shape is not used herein. + tflite::StridedSliceParams params_copy = op_params; + + TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 5); + TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 5); + const RuntimeShape input_shape = + RuntimeShape::ExtendedShape(5, unextended_input_shape); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(5, unextended_output_shape); + + // Reverse and pad to 5 dimensions because that is what the runtime code + // requires (ie. all shapes must be 5D and are given backwards). + strided_slice::StridedSlicePadIndices(¶ms_copy, 5); + + const int start_0 = StartForAxis(params_copy, input_shape, 0); + const int stop_0 = StopForAxis(params_copy, input_shape, 0, start_0); + const int start_1 = StartForAxis(params_copy, input_shape, 1); + const int stop_1 = StopForAxis(params_copy, input_shape, 1, start_1); + const int start_2 = StartForAxis(params_copy, input_shape, 2); + const int stop_2 = StopForAxis(params_copy, input_shape, 2, start_2); + const int start_3 = StartForAxis(params_copy, input_shape, 3); + const int stop_3 = StopForAxis(params_copy, input_shape, 3, start_3); + const int start_4 = StartForAxis(params_copy, input_shape, 4); + const int stop_4 = StopForAxis(params_copy, input_shape, 4, start_4); + + for (int offset_0 = start_0 * input_shape.Dims(1), + end_0 = stop_0 * input_shape.Dims(1), + step_0 = params_copy.strides[0] * input_shape.Dims(1); + !LoopCondition(offset_0, end_0, params_copy.strides[0]); + offset_0 += step_0) { + for (int offset_1 = (offset_0 + start_1) * input_shape.Dims(2), + end_1 = (offset_0 + stop_1) * input_shape.Dims(2), + step_1 = params_copy.strides[1] * input_shape.Dims(2); + !LoopCondition(offset_1, end_1, params_copy.strides[1]); + offset_1 += step_1) { + for (int offset_2 = (offset_1 + start_2) * input_shape.Dims(3), + end_2 = (offset_1 + stop_2) * input_shape.Dims(3), + step_2 = params_copy.strides[2] * input_shape.Dims(3); + !LoopCondition(offset_2, end_2, params_copy.strides[2]); + offset_2 += step_2) { + for (int offset_3 = (offset_2 + start_3) * input_shape.Dims(4), + end_3 = (offset_2 + stop_3) * input_shape.Dims(4), + step_3 = params_copy.strides[3] * input_shape.Dims(4); + !LoopCondition(offset_3, end_3, params_copy.strides[3]); + offset_3 += step_3) { + for (int offset_4 = offset_3 + start_4, end_4 = offset_3 + stop_4; + !LoopCondition(offset_4, end_4, params_copy.strides[4]); + offset_4 += params_copy.strides[4]) { + writer->Write(offset_4); + } + } + } + } + } +} + +template +inline void StridedSlice(const tflite::StridedSliceParams& op_params, + const RuntimeShape& unextended_input_shape, + const T* input_data, + const RuntimeShape& unextended_output_shape, + T* output_data) { + SequentialTensorWriter writer(input_data, output_data); + StridedSlice(op_params, unextended_input_shape, unextended_output_shape, + &writer); +} + +template +inline void StridedSlice(const tflite::StridedSliceParams& op_params, + const RuntimeShape& unextended_input_shape, + const TfLiteTensor* input, + const RuntimeShape& unextended_output_shape, + TfLiteTensor* output) { + SequentialTensorWriter writer(input, output); + StridedSlice(op_params, unextended_input_shape, unextended_output_shape, + &writer); +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/sub.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/sub.h new file mode 100644 index 0000000..1719411 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/sub.h @@ -0,0 +1,515 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SUB_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SUB_H_ + +#include + +#include +#include + +#include "tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline void SubNonBroadcast(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const float* input1_data, + const RuntimeShape& input2_shape, + const float* input2_data, + const RuntimeShape& output_shape, + float* output_data) { + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax( + input1_data[i] - input2_data[i], params.float_activation_min, + params.float_activation_max); + } +} + +inline void SubNonBroadcast(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const int32_t* input1_data, + const RuntimeShape& input2_shape, + const int32_t* input2_data, + const RuntimeShape& output_shape, + int32_t* output_data) { + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax( + input1_data[i] - input2_data[i], params.quantized_activation_min, + params.quantized_activation_max); + } +} + +// TODO(b/151345304): We can implement BroadcastSub on buffers of arbitrary +// dimensionality if the runtime code does a single loop over one dimension +// that handles broadcasting as the base case. The code generator would then +// generate max(D1, D2) nested for loops. +template +inline void BroadcastSubSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const float* input1_data, + const RuntimeShape& input2_shape, + const float* input2_data, + const RuntimeShape& output_shape, + float* output_data) { + ruy::profiler::ScopeLabel label("BroadcastSubSlow/float"); + TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N); + NdArrayDesc desc1; + NdArrayDesc desc2; + NdArrayDesc output_desc; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + auto sub_func = [&](int indexes[N]) { + output_data[SubscriptToIndex(output_desc, indexes)] = + ActivationFunctionWithMinMax( + input1_data[SubscriptToIndex(desc1, indexes)] - + input2_data[SubscriptToIndex(desc2, indexes)], + params.float_activation_min, params.float_activation_max); + }; + NDOpsHelper(output_desc, sub_func); +} + +template +inline void BroadcastSubSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const uint8_t* input1_data, + const RuntimeShape& input2_shape, + const uint8_t* input2_data, + const RuntimeShape& output_shape, + uint8_t* output_data) { + ruy::profiler::ScopeLabel label("BroadcastSubSlow/uint8_t"); + TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N); + NdArrayDesc desc1; + NdArrayDesc desc2; + NdArrayDesc output_desc; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + auto sub_func = [&](int indexes[N]) { + const int32_t input1_val = + params.input1_offset + input1_data[SubscriptToIndex(desc1, indexes)]; + const int32_t input2_val = + params.input2_offset + input2_data[SubscriptToIndex(desc2, indexes)]; + const int32_t shifted_input1_val = input1_val * (1 << params.left_shift); + const int32_t shifted_input2_val = input2_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, params.input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, params.input2_shift); + const int32_t raw_sub = scaled_input1_val - scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sub, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + output_data[SubscriptToIndex(output_desc, indexes)] = + static_cast(clamped_output); + }; + NDOpsHelper(output_desc, sub_func); +} + +template +inline void BroadcastSubSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const int32_t* input1_data, + const RuntimeShape& input2_shape, + const int32_t* input2_data, + const RuntimeShape& output_shape, + int32_t* output_data) { + ruy::profiler::ScopeLabel label("BroadcastSubSlow/int32_t"); + TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N); + NdArrayDesc desc1; + NdArrayDesc desc2; + NdArrayDesc output_desc; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + auto sub_func = [&](int indexes[N]) { + output_data[SubscriptToIndex(output_desc, indexes)] = + ActivationFunctionWithMinMax( + input1_data[SubscriptToIndex(desc1, indexes)] - + input2_data[SubscriptToIndex(desc2, indexes)], + params.quantized_activation_min, params.quantized_activation_max); + }; + NDOpsHelper(output_desc, sub_func); +} + +template +inline void BroadcastSubSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const int8_t* input1_data, + const RuntimeShape& input2_shape, + const int8_t* input2_data, + const RuntimeShape& output_shape, + int8_t* output_data) { + ruy::profiler::ScopeLabel label("BroadcastSubSlow/int8_t"); + NdArrayDesc desc1; + NdArrayDesc desc2; + NdArrayDesc output_desc; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + auto sub_func = [&](int indexes[N]) { + const int32_t input1_val = + params.input1_offset + input1_data[SubscriptToIndex(desc1, indexes)]; + const int32_t input2_val = + params.input2_offset + input2_data[SubscriptToIndex(desc2, indexes)]; + const int32_t shifted_input1_val = input1_val * (1 << params.left_shift); + const int32_t shifted_input2_val = input2_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, params.input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, params.input2_shift); + const int32_t raw_sub = scaled_input1_val - scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sub, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + output_data[SubscriptToIndex(output_desc, indexes)] = + static_cast(clamped_output); + }; + NDOpsHelper(output_desc, sub_func); +} + +template +void BroadcastSubSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, + const int64_t* input1_data, + const RuntimeShape& input2_shape, + const int64_t* input2_data, + const RuntimeShape& output_shape, int64_t* output_data) { + ruy::profiler::ScopeLabel label("BroadcastSubSlow/int64_t"); + TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N); + NdArrayDesc desc1; + NdArrayDesc desc2; + NdArrayDesc output_desc; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + auto sub_func = [&](int indexes[N]) { + output_data[SubscriptToIndex(output_desc, indexes)] = + ActivationFunctionWithMinMax( + input1_data[SubscriptToIndex(desc1, indexes)] - + input2_data[SubscriptToIndex(desc2, indexes)], + params.int64_activation_min, params.int64_activation_max); + }; + NDOpsHelper(output_desc, sub_func); +} + +template +void BroadcastSubSlow(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const T* input1_data, + const RuntimeShape& input2_shape, const T* input2_data, + const RuntimeShape& output_shape, T* output_data) { + ruy::profiler::ScopeLabel label("BroadcastSubSlow/templated"); + TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N); + TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N); + NdArrayDesc desc1; + NdArrayDesc desc2; + NdArrayDesc output_desc; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + auto sub_func = [&](int indexes[N]) { + output_data[SubscriptToIndex(output_desc, indexes)] = + ActivationFunctionWithMinMax( + input1_data[SubscriptToIndex(desc1, indexes)] - + input2_data[SubscriptToIndex(desc2, indexes)], + params.quantized_activation_min, params.quantized_activation_max); + }; + NDOpsHelper(output_desc, sub_func); +} + +// Element-wise Sub that can often be used for inner loop of broadcast sub as +// well as the non-broadcast sub. +inline void SubElementwise(int size, const ArithmeticParams& params, + const uint8_t* input1_data, + const uint8_t* input2_data, uint8_t* output_data) { + TFLITE_DCHECK_GT(params.input1_offset, -256); + TFLITE_DCHECK_GT(params.input2_offset, -256); + TFLITE_DCHECK_LT(params.input1_offset, 256); + TFLITE_DCHECK_LT(params.input2_offset, 256); + + for (int i = 0; i < size; ++i) { + const int32_t input1_val = params.input1_offset + input1_data[i]; + const int32_t input2_val = params.input2_offset + input2_data[i]; + const int32_t shifted_input1_val = input1_val * (1 << params.left_shift); + const int32_t shifted_input2_val = input2_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, params.input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, params.input2_shift); + const int32_t raw_sub = scaled_input1_val - scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sub, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + output_data[i] = static_cast(clamped_output); + } +} + +// Element-wise add that can often be used for inner loop of broadcast add as +// well as the non-broadcast add. +inline void SubElementwise(int size, const ArithmeticParams& params, + const int8_t* input1_data, const int8_t* input2_data, + int8_t* output_data) { + const int32_t int8_max_value = std::numeric_limits::max(); + TFLITE_DCHECK_GE(params.input1_offset, -1 * int8_max_value); + TFLITE_DCHECK_GE(params.input2_offset, -1 * int8_max_value); + TFLITE_DCHECK_LE(params.input1_offset, int8_max_value); + TFLITE_DCHECK_LE(params.input2_offset, int8_max_value); + + for (int i = 0; i < size; ++i) { + const int32_t input1_val = params.input1_offset + input1_data[i]; + const int32_t input2_val = params.input2_offset + input2_data[i]; + const int32_t shifted_input1_val = input1_val * (1 << params.left_shift); + const int32_t shifted_input2_val = input2_val * (1 << params.left_shift); + const int32_t scaled_input1_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input1_val, params.input1_multiplier, params.input1_shift); + const int32_t scaled_input2_val = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + shifted_input2_val, params.input2_multiplier, params.input2_shift); + const int32_t raw_sub = scaled_input1_val - scaled_input2_val; + const int32_t raw_output = + MultiplyByQuantizedMultiplierSmallerThanOneExp( + raw_sub, params.output_multiplier, params.output_shift) + + params.output_offset; + const int32_t clamped_output = + std::min(params.quantized_activation_max, + std::max(params.quantized_activation_min, raw_output)); + output_data[i] = static_cast(clamped_output); + } +} + +inline void Sub(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const uint8_t* input1_data, + const RuntimeShape& input2_shape, const uint8_t* input2_data, + const RuntimeShape& output_shape, uint8_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + + TFLITE_DCHECK_GT(params.input1_offset, -256); + TFLITE_DCHECK_GT(params.input2_offset, -256); + TFLITE_DCHECK_LT(params.input1_offset, 256); + TFLITE_DCHECK_LT(params.input2_offset, 256); + SubElementwise(flat_size, params, input1_data, input2_data, output_data); +} + +inline void Sub(const ArithmeticParams& params, + const RuntimeShape& input1_shape, const int8_t* input1_data, + const RuntimeShape& input2_shape, const int8_t* input2_data, + const RuntimeShape& output_shape, int8_t* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + + const int32_t int8_max_value = std::numeric_limits::max(); + TFLITE_DCHECK_GE(params.input1_offset, -1 * int8_max_value); + TFLITE_DCHECK_GE(params.input2_offset, -1 * int8_max_value); + TFLITE_DCHECK_LE(params.input1_offset, int8_max_value); + TFLITE_DCHECK_LE(params.input2_offset, int8_max_value); + SubElementwise(flat_size, params, input1_data, input2_data, output_data); +} + +template +void Sub(const ArithmeticParams& params, const RuntimeShape& input1_shape, + const T* input1_data, const RuntimeShape& input2_shape, + const T* input2_data, const RuntimeShape& output_shape, + T* output_data) { + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1, + &desc2); + const RuntimeShape extended_output_shape = + RuntimeShape::ExtendedShape(4, output_shape); + + // In Tensorflow, the dimensions are canonically named (batch_number, row, + // col, channel), with extents (batches, height, width, depth), with the + // trailing dimension changing most rapidly (channels has the smallest stride, + // typically 1 element). + // + // In generated C code, we store arrays with the dimensions reversed. The + // first dimension has smallest stride. + // + // We name our variables by their Tensorflow convention, but generate C code + // nesting loops such that the innermost loop has the smallest stride for the + // best cache behavior. + for (int b = 0; b < extended_output_shape.Dims(0); ++b) { + for (int y = 0; y < extended_output_shape.Dims(1); ++y) { + for (int x = 0; x < extended_output_shape.Dims(2); ++x) { + for (int c = 0; c < extended_output_shape.Dims(3); ++c) { + output_data[Offset(extended_output_shape, b, y, x, c)] = + input1_data[SubscriptToIndex(desc1, b, y, x, c)] - + input2_data[SubscriptToIndex(desc2, b, y, x, c)]; + } + } + } + } +} + +inline void SetActivationMinMax(const ArithmeticParams& params, + int32_t* activation_min, + int32_t* activation_max) { + *activation_min = params.quantized_activation_min; + *activation_max = params.quantized_activation_max; +} + +inline void SetActivationMinMax(const ArithmeticParams& params, + float* activation_min, float* activation_max) { + *activation_min = params.float_activation_min; + *activation_max = params.float_activation_max; +} + +inline void SetActivationMinMax(const ArithmeticParams& params, + int64_t* activation_min, + int64_t* activation_max) { + *activation_min = params.int64_activation_min; + *activation_max = params.int64_activation_max; +} + +template +inline void SubWithActivation( + const ArithmeticParams& params, const RuntimeShape& input1_shape, + const T* input1_data, const RuntimeShape& input2_shape, + const T* input2_data, const RuntimeShape& output_shape, T* output_data) { + ruy::profiler::ScopeLabel label("SubWithActivation"); + const int flat_size = + MatchingElementsSize(input1_shape, input2_shape, output_shape); + T activation_min, activation_max; + SetActivationMinMax(params, &activation_min, &activation_max); + + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax( + input1_data[i] - input2_data[i], activation_min, activation_max); + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SUB_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/tanh.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/tanh.h new file mode 100644 index 0000000..50ac077 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/tanh.h @@ -0,0 +1,132 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TANH_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TANH_H_ + +#include + +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" + +namespace tflite { +namespace reference_ops { + +inline void Tanh(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; i++) { + float val = input_data[i]; + float result = std::tanh(val); + output_data[i] = result; + } +} + +// Convenience version that allows, for example, generated-code calls to be +// uniform between data types. +inline void Tanh(const TanhParams&, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { + // Drop params: not needed. + Tanh(input_shape, input_data, output_shape, output_data); +} + +inline void Tanh(const TanhParams& params, const RuntimeShape& input_shape, + const int16_t* input_data, const RuntimeShape& output_shape, + int16_t* output_data) { + const int input_left_shift = params.input_left_shift; + // Support for shifts is limited until we have a parameterized version of + // SaturatingRoundingMultiplyByPOT(). + TFLITE_DCHECK_GE(input_left_shift, 0); + TFLITE_DCHECK_LE(input_left_shift, 1); + + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + // F0 uses 0 integer bits, range [-1, 1]. + // This is the return type of math functions such as tanh, logistic, + // whose range is in [-1, 1]. + using F0 = gemmlowp::FixedPoint; + // F3 uses 3 integer bits, range [-8, 8], the input range expected here. + using F3 = gemmlowp::FixedPoint; + + if (input_left_shift == 0) { + for (int i = 0; i < flat_size; i++) { + F3 input = F3::FromRaw(input_data[i]); + F0 output = gemmlowp::tanh(input); + output_data[i] = output.raw(); + } + } else { + for (int i = 0; i < flat_size; i++) { + F3 input = F3::FromRaw( + gemmlowp::SaturatingRoundingMultiplyByPOT<1>(input_data[i])); + F0 output = gemmlowp::tanh(input); + output_data[i] = output.raw(); + } + } +} + +inline void Tanh(const TanhParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& output_shape, + uint8_t* output_data) { + const int32_t input_zero_point = params.input_zero_point; + const int32_t input_range_radius = params.input_range_radius; + const int32_t input_multiplier = params.input_multiplier; + const int input_left_shift = params.input_left_shift; + const int32_t output_zero_point = 128; + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; i++) { + const uint8_t input_val_u8 = input_data[i]; + const int32_t input_val_centered = + static_cast(input_val_u8) - input_zero_point; + uint8_t output_val; + if (input_val_centered <= -input_range_radius) { + output_val = 0; + } else if (input_val_centered >= input_range_radius) { + output_val = 255; + } else { + const int32_t input_val_rescaled = + MultiplyByQuantizedMultiplierGreaterThanOne( + input_val_centered, input_multiplier, input_left_shift); + using FixedPoint4 = gemmlowp::FixedPoint; + using FixedPoint0 = gemmlowp::FixedPoint; + const FixedPoint4 input_val_f4 = FixedPoint4::FromRaw(input_val_rescaled); + const FixedPoint0 output_val_f0 = gemmlowp::tanh(input_val_f4); + // Convert from Q0.31 to Q24.7. + using gemmlowp::RoundingDivideByPOT; + int32_t output_val_s32 = RoundingDivideByPOT(output_val_f0.raw(), 24); + output_val_s32 += output_zero_point; + if (output_val_s32 == 256) { + output_val_s32 = 255; + } + // Reinterpret as Q0.7, encoded in uint8_t. + TFLITE_DCHECK_GE(output_val_s32, 0); + TFLITE_DCHECK_LE(output_val_s32, 255); + output_val = static_cast(output_val_s32); + } + output_data[i] = output_val; + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TANH_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/transpose_conv.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/transpose_conv.h new file mode 100644 index 0000000..188b7b1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/reference/transpose_conv.h @@ -0,0 +1,220 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_CONV_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_CONV_H_ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline void TransposeConv( + const ConvParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& filter_shape, + const float* filter_data, const RuntimeShape& bias_shape, + const float* bias_data, const RuntimeShape& output_shape, + float* output_data, const RuntimeShape& im2col_shape, float* im2col_data) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + (void)im2col_data; // only used in optimized code. + (void)im2col_shape; // only used in optimized code. + + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + + // Although transpose convolution simplifies to convolution with transposed + // weights for strides of 1, non-unitary striding complicates matters. To + // keep this reference implementation as clear as possible, we use a + // "scatter" access pattern, where we loop through all the input elements, + // computing their influence on the output, rather than looping through the + // output elements in the typical "gather" access pattern of a conv. We + // therefore must initialize the output array to zero. + const int num_elements = output_shape.FlatSize(); + for (int i = 0; i < num_elements; i++) { + output_data[i] = 0.0f; + } + + // Loop through input elements one at a time. + for (int batch = 0; batch < batches; ++batch) { + for (int in_y = 0; in_y < input_height; ++in_y) { + for (int in_x = 0; in_x < input_width; ++in_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + // Loop through the output elements it will influence + const int out_x_origin = (in_x * stride_width) - pad_width; + const int out_y_origin = (in_y * stride_height) - pad_height; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + for (int out_channel = 0; out_channel < output_depth; + ++out_channel) { + // Compute output element location + const int out_x = out_x_origin + filter_x; + const int out_y = out_y_origin + filter_y; + // We cannot accumulate out of bounds + if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) && + (out_y < output_height)) { + float input_value = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + float filter_value = + filter_data[Offset(filter_shape, out_channel, filter_y, + filter_x, in_channel)]; + output_data[Offset(output_shape, batch, out_y, out_x, + out_channel)] += + input_value * filter_value; + } + } + } + } + } + } + } + } + if (bias_data) { + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + output_data[Offset(output_shape, batch, out_y, out_x, + out_channel)] += bias_data[out_channel]; + } + } + } + } + } +} + +inline void TransposeConv( + const ConvParams& params, const RuntimeShape& input_shape, + const uint8_t* input_data, const RuntimeShape& filter_shape, + const uint8_t* filter_data, const RuntimeShape& bias_shape, + const int32_t* bias_data, const RuntimeShape& output_shape, + uint8_t* output_data, const RuntimeShape& im2col_shape, + uint8_t* im2col_data, int32_t* scratch_buffer) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + (void)im2col_data; // only used in optimized code. + (void)im2col_shape; // only used in optimized code. + + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + const int32_t input_offset = params.input_offset; + const int32_t filter_offset = params.weights_offset; + const int32_t output_offset = params.output_offset; + const int32_t output_multiplier = params.output_multiplier; + const int output_shift = params.output_shift; + const int32_t output_activation_min = params.quantized_activation_min; + const int32_t output_activation_max = params.quantized_activation_max; + TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + if (bias_data) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + + const int num_elements = output_shape.FlatSize(); + // We need to initialize scratch_buffer to all 0s, as we apply the same + // 'scatter' based trick as in float version. + memset(scratch_buffer, 0, num_elements * sizeof(int32_t)); + + // Loop through input elements one at a time. + for (int batch = 0; batch < batches; ++batch) { + for (int in_y = 0; in_y < input_height; ++in_y) { + for (int in_x = 0; in_x < input_width; ++in_x) { + for (int in_channel = 0; in_channel < input_depth; ++in_channel) { + // Loop through the output elements it will influence. + const int out_x_origin = (in_x * stride_width) - pad_width; + const int out_y_origin = (in_y * stride_height) - pad_height; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + for (int out_channel = 0; out_channel < output_depth; + ++out_channel) { + // Compute output element location. + const int out_x = out_x_origin + filter_x; + const int out_y = out_y_origin + filter_y; + // We cannot accumulate out of bounds. + if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) && + (out_y < output_height)) { + uint8_t input_value = input_data[Offset( + input_shape, batch, in_y, in_x, in_channel)]; + uint8_t filter_value = + filter_data[Offset(filter_shape, out_channel, filter_y, + filter_x, in_channel)]; + scratch_buffer[Offset(output_shape, batch, out_y, out_x, + out_channel)] += + (input_value + input_offset) * + (filter_value + filter_offset); + } + } + } + } + } + } + } + } + for (int batch = 0; batch < batches; ++batch) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int out_channel = 0; out_channel < output_depth; ++out_channel) { + int32_t acc = scratch_buffer[Offset(output_shape, batch, out_y, out_x, + out_channel)]; + if (bias_data) { + acc += bias_data[out_channel]; + } + int32_t scaled_acc = MultiplyByQuantizedMultiplier( + acc, output_multiplier, output_shift); + scaled_acc += output_offset; + scaled_acc = std::max(scaled_acc, output_activation_min); + scaled_acc = std::min(scaled_acc, output_activation_max); + output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] = + static_cast(scaled_acc); + } + } + } + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_CONV_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/strided_slice_logic.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/strided_slice_logic.h new file mode 100644 index 0000000..e218629 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/strided_slice_logic.h @@ -0,0 +1,214 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_STRIDED_SLICE_LOGIC_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_STRIDED_SLICE_LOGIC_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace strided_slice { + +// Use until std::clamp() is available from C++17. +inline int Clamp(const int v, const int lo, const int hi) { + TFLITE_DCHECK(!(hi < lo)); + if (hi < v) return hi; + if (v < lo) return lo; + return v; +} + +inline void StridedSlicePadIndices(tflite::StridedSliceParams* p, + int dim_count) { + // Add indices and mask bits to fully include extra dimensions + TFLITE_CHECK_LE(dim_count, 5); + TFLITE_CHECK_GE(dim_count, p->start_indices_count); + TFLITE_CHECK_EQ(p->start_indices_count, p->stop_indices_count); + TFLITE_CHECK_EQ(p->stop_indices_count, p->strides_count); + + const int pad_count = dim_count - p->start_indices_count; + + // Pad indices at start, so move arrays by pad_count. + for (int i = p->start_indices_count - 1; i >= 0; --i) { + p->strides[i + pad_count] = p->strides[i]; + p->start_indices[i + pad_count] = p->start_indices[i]; + p->stop_indices[i + pad_count] = p->stop_indices[i]; + } + for (int i = 0; i < pad_count; ++i) { + p->start_indices[i] = 0; + p->stop_indices[i] = 1; + p->strides[i] = 1; + } + + // Pad masks with 0s or 1s as required. + p->shrink_axis_mask <<= pad_count; + p->ellipsis_mask <<= pad_count; + p->new_axis_mask <<= pad_count; + p->begin_mask <<= pad_count; + p->end_mask <<= pad_count; + p->begin_mask |= (1 << pad_count) - 1; + p->end_mask |= (1 << pad_count) - 1; + + p->start_indices_count = dim_count; + p->stop_indices_count = dim_count; + p->strides_count = dim_count; +} + +// Return the index for the first element along that axis. This index will be a +// positive integer between [0, axis_size] (or [-1, axis_size -1] if stride < 0) +// that can be used to index directly into the data. +inline int StartForAxis(const tflite::StridedSliceParams& params, + const RuntimeShape& input_shape, int axis) { + const auto begin_mask = params.begin_mask; + const auto* start_indices = params.start_indices; + const auto* strides = params.strides; + const int axis_size = input_shape.Dims(axis); + if (axis_size == 0) { + return 0; + } + // Begin with the specified index. + int start = start_indices[axis]; + + // begin_mask override + if (begin_mask & 1 << axis) { + if (strides[axis] > 0) { + // Forward iteration - use the first element. These values will get + // clamped below (Note: We could have set them to 0 and axis_size-1, but + // use lowest() and max() to maintain symmetry with StopForAxis()) + start = std::numeric_limits::lowest(); + } else { + // Backward iteration - use the last element. + start = std::numeric_limits::max(); + } + } + + // Handle negative indices + if (start < 0) { + start += axis_size; + } + + // Clamping + if (strides[axis] > 0) { + // Forward iteration + start = Clamp(start, 0, axis_size); + } else { + // Backward iteration + start = Clamp(start, -1, axis_size - 1); + } + + return start; +} + +// Return the "real" index for the end of iteration along that axis. This is an +// "end" in the traditional C sense, in that it points to one past the last +// element. ie. So if you were iterating through all elements of a 1D array of +// size 4, this function would return 4 as the stop, because it is one past the +// "real" indices of 0, 1, 2 & 3. +inline int StopForAxis(const tflite::StridedSliceParams& params, + const RuntimeShape& input_shape, int axis, + int start_for_axis) { + const auto end_mask = params.end_mask; + const auto shrink_axis_mask = params.shrink_axis_mask; + const auto* stop_indices = params.stop_indices; + const auto* strides = params.strides; + const int axis_size = input_shape.Dims(axis); + if (axis_size == 0) { + return 0; + } + + // Begin with the specified index + const bool shrink_axis = shrink_axis_mask & (1 << axis); + int stop = stop_indices[axis]; + + // When shrinking an axis, the end position does not matter (and can be + // incorrect when negative indexing is used, see Issue #19260). Always use + // start_for_axis + 1 to generate a length 1 slice, since start_for_axis has + // already been adjusted for negative indices. + if (shrink_axis) { + return start_for_axis + 1; + } + + // end_mask override + if (end_mask & (1 << axis)) { + if (strides[axis] > 0) { + // Forward iteration - use the last element. These values will get + // clamped below + stop = std::numeric_limits::max(); + } else { + // Backward iteration - use the first element. + stop = std::numeric_limits::lowest(); + } + } + + // Handle negative indices + if (stop < 0) { + stop += axis_size; + } + + // Clamping + // Because the end index points one past the last element, we need slightly + // different clamping ranges depending on the direction. + if (strides[axis] > 0) { + // Forward iteration + stop = Clamp(stop, 0, axis_size); + } else { + // Backward iteration + stop = Clamp(stop, -1, axis_size - 1); + } + + return stop; +} + +inline bool LoopCondition(int index, int stop, int stride) { + // True when we have reached the end of an axis and should loop. + return stride > 0 ? index >= stop : index <= stop; +} + +inline tflite::StridedSliceParams BuildStridedSliceParams( + int begin_mask, int end_mask, int shrink_axis_mask, + const std::vector& start_indices, const std::vector& stop_indices, + const std::vector& strides) { + tflite::StridedSliceParams op_params; + const int dims_count = start_indices.size(); + + op_params.start_indices_count = dims_count; + op_params.stop_indices_count = dims_count; + op_params.strides_count = dims_count; + for (int i = 0; i < dims_count; ++i) { + op_params.start_indices[i] = start_indices[i]; + op_params.stop_indices[i] = stop_indices[i]; + op_params.strides[i] = strides[i]; + } + + op_params.begin_mask = begin_mask; + op_params.ellipsis_mask = 0; + op_params.end_mask = end_mask; + op_params.new_axis_mask = 0; + op_params.shrink_axis_mask = shrink_axis_mask; + + return op_params; +} + +} // namespace strided_slice + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_STRIDED_SLICE_LOGIC_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/tensor_ctypes.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h similarity index 89% rename from src/tensorflow/lite/kernels/internal/tensor_ctypes.h rename to src/tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h index f1d3e17..f75781a 100644 --- a/src/tensorflow/lite/kernels/internal/tensor_ctypes.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,8 +16,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -45,3 +46,5 @@ inline RuntimeShape GetTensorShape(const TfLiteTensor* tensor) { } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/internal/types.h b/src/tensorflow_arm/tensorflow/lite/kernels/internal/types.h new file mode 100644 index 0000000..844aecd --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/internal/types.h @@ -0,0 +1,1172 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_TYPES_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_TYPES_H_ + +#include +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" + +namespace tflite { + +enum class FusedActivationFunctionType : uint8_t { + kNone, + kRelu6, + kRelu1, + kRelu +}; +enum class PaddingType : uint8_t { kNone, kSame, kValid }; + +struct PaddingValues { + int16_t width; + int16_t height; + // offset is used for calculating "remaining" padding, for example, `width` + // is 1 and `width_offset` is 1, so padding_left is 1 while padding_right is + // 1 + 1 = 2. + int16_t width_offset; + // Same as width_offset except it's over the height dimension. + int16_t height_offset; +}; + +// This enumeration allows for non-default formats for the weights array +// of a fully-connected operator, allowing the use of special optimized +// runtime paths. +enum class FullyConnectedWeightsFormat : uint8_t { + // Default format (flat 2D layout, the inner contiguous dimension + // is input_depth, the outer non-contiguous dimension is output_depth) + kDefault, + // Summary: optimized layout for fast CPU runtime implementation, + // aimed specifically at ARM CPUs at the moment, and specialized for + // 8-bit quantized layers. + // + // The use case we're concerned with here is: 8-bit quantization, + // large weights matrix that doesn't fit in cache (e.g. 4096x2048 in + // a key application that drove this), very small batch size (e.g. 1 -- 4). + // + // Even with 8-bit quantization of weights, the performance of memory + // accesses to the weights can become the dominant issue when + // the batch size is small, so each weight value is used in only a few + // arithmetic ops, i.e. the fully-connected node has a low arithmetic + // intensity. The specific issues that arise are of three kinds: + // (1) One may, ideally, max out DRAM bandwidth, i.e. be truly memory + // bound. That's the "good" issue to run into. + // (2) One may run into sub-optimal pre-fetching: the data hasn't been + // prefetched into the cache by the time we need it. + // (3) One may run into cache aliasing: multiple values that are + // pre-fetched, alias each other in the L1 cache (which typically + // has only 4-way set associativity in ARM CPUs) and thus evict + // each other before we get to using them. + // + // The point of this shuffling is to avoid issues (2) and (3) so that + // we get as fast as possible given only the hard constraint (1). + // This is achieved by turning the difficulty into a solution: the + // difficulty, that each value loaded from memory is used only in + // one kernel iteration, making this operation memory-intensive, hints at + // the solution, of shuffling the weights so that they are stored in the + // exact order as the kernel needs to load them, so that the memory + // accesses made by the kernel are trivial. This solves (2) because the + // trivial memory access pattern allows the CPU's automatic prefetching + // to perform very well (no need even for preload instructions), and this + // solves (3) because the values being loaded concurrently are now + // contiguous in the address space, thus don't alias each other in the cache. + // + // On ARM, we typically want our kernel to process a 4x16 block of weights + // at a time, because: + // - 16 is the number of bytes in a NEON register. + // - 4 is how many rows we need to handle concurrently in the kernel in + // order to have sufficient mutual independence of instructions to + // maximize arithmetic throughput. + // + // Finally, the 'Int8' part in the name refers to the fact that this + // weights format has each weights value encoded as a signed int8_t value, + // even if the data type of the weights buffer is uint8_t. This is intended + // to save runtime kernels the effort to have to XOR the top bit of these + // bytes before using them in signed arithmetic, see this file for more + // explanations on the 'signed int8_t trick' in matrix multiplication kernels: + // + // tensorflow/lite/toco/graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc + // + kShuffled4x16Int8, +}; + +// Quantization parameters, determining the mapping of quantized values +// to real values (i.e. determining how quantized values are mathematically +// interpreted). +// +// The correspondence is as follows: +// +// real_value = scale * (quantized_value - zero_point); +// +// In other words, zero_point designates which quantized value corresponds to +// the real 0 value, and scale designates the difference between the real values +// corresponding to consecutive quantized values differing by 1. +struct QuantizationParams { + int32_t zero_point = 0; + double scale = 0.0; +}; + +inline bool operator==(const QuantizationParams& qp1, + const QuantizationParams& qp2) { + return qp1.zero_point == qp2.zero_point && qp1.scale == qp2.scale; +} + +template +struct Dims { + int sizes[N]; + int strides[N]; +}; + +class RuntimeShape { + public: + // Shapes with dimensions up to 5 are stored directly in the structure, while + // larger shapes are separately allocated. + static constexpr int kMaxSmallSize = 5; + + RuntimeShape& operator=(RuntimeShape const&) = delete; + + RuntimeShape() : size_(0) {} + + explicit RuntimeShape(int dimensions_count) : size_(dimensions_count) { + if (dimensions_count > kMaxSmallSize) { +#ifdef TF_LITE_STATIC_MEMORY + TFLITE_CHECK(false && "No shape resizing supported on this platform"); +#else // TF_LITE_STATIC_MEMORY + dims_pointer_ = new int32_t[dimensions_count]; +#endif // TF_LITE_STATIC_MEMORY + } + } + + RuntimeShape(int shape_size, int32_t value) : size_(0) { + Resize(shape_size); + for (int i = 0; i < shape_size; ++i) { + SetDim(i, value); + } + } + + RuntimeShape(int dimensions_count, const int32_t* dims_data) : size_(0) { + ReplaceWith(dimensions_count, dims_data); + } + + RuntimeShape(const std::initializer_list init_list) : size_(0) { + BuildFrom(init_list); + } + + // Avoid using this constructor. We should be able to delete it when C++17 + // rolls out. + RuntimeShape(RuntimeShape const& other) : size_(other.DimensionsCount()) { + if (size_ > kMaxSmallSize) { + dims_pointer_ = new int32_t[size_]; + } + std::memcpy(DimsData(), other.DimsData(), sizeof(int32_t) * size_); + } + + bool operator==(const RuntimeShape& comp) const { + return this->size_ == comp.size_ && + std::memcmp(DimsData(), comp.DimsData(), size_ * sizeof(int32_t)) == + 0; + } + + ~RuntimeShape() { + if (size_ > kMaxSmallSize) { +#ifdef TF_LITE_STATIC_MEMORY + TFLITE_CHECK(false && "No shape resizing supported on this platform"); +#else // TF_LITE_STATIC_MEMORY + delete[] dims_pointer_; +#endif // TF_LITE_STATIC_MEMORY + } + } + + inline int32_t DimensionsCount() const { return size_; } + inline int32_t Dims(int i) const { + TFLITE_DCHECK_GE(i, 0); + TFLITE_DCHECK_LT(i, size_); + return size_ > kMaxSmallSize ? dims_pointer_[i] : dims_[i]; + } + inline void SetDim(int i, int32_t val) { + TFLITE_DCHECK_GE(i, 0); + TFLITE_DCHECK_LT(i, size_); + if (size_ > kMaxSmallSize) { + dims_pointer_[i] = val; + } else { + dims_[i] = val; + } + } + + inline int32_t* DimsData() { + return size_ > kMaxSmallSize ? dims_pointer_ : dims_; + } + inline const int32_t* DimsData() const { + return size_ > kMaxSmallSize ? dims_pointer_ : dims_; + } + // The caller must ensure that the shape is no bigger than 5-D. + inline const int32_t* DimsDataUpTo5D() const { return dims_; } + + inline void Resize(int dimensions_count) { + if (size_ > kMaxSmallSize) { +#ifdef TF_LITE_STATIC_MEMORY + TFLITE_CHECK(false && "No shape resizing supported on this platform"); +#else // TF_LITE_STATIC_MEMORY + delete[] dims_pointer_; +#endif // TF_LITE_STATIC_MEMORY + } + size_ = dimensions_count; + if (dimensions_count > kMaxSmallSize) { +#ifdef TF_LITE_STATIC_MEMORY + TFLITE_CHECK(false && "No shape resizing supported on this platform"); +#else // TF_LITE_STATIC_MEMORY + dims_pointer_ = new int32_t[dimensions_count]; +#endif // TF_LITE_STATIC_MEMORY + } + } + + inline void ReplaceWith(int dimensions_count, const int32_t* dims_data) { + Resize(dimensions_count); + int32_t* dst_dims = DimsData(); + std::memcpy(dst_dims, dims_data, dimensions_count * sizeof(int32_t)); + } + + template + inline void BuildFrom(const T& src_iterable) { + const int dimensions_count = + std::distance(src_iterable.begin(), src_iterable.end()); + Resize(dimensions_count); + int32_t* data = DimsData(); + for (auto it : src_iterable) { + *data = it; + ++data; + } + } + + // This will probably be factored out. Old code made substantial use of 4-D + // shapes, and so this function is used to extend smaller shapes. Note that + // (a) as Dims<4>-dependent code is eliminated, the reliance on this should be + // reduced, and (b) some kernels are stricly 4-D, but then the shapes of their + // inputs should already be 4-D, so this function should not be needed. + inline static RuntimeShape ExtendedShape(int new_shape_size, + const RuntimeShape& shape) { + return RuntimeShape(new_shape_size, shape, 1); + } + + inline void BuildFrom(const std::initializer_list init_list) { + BuildFrom>(init_list); + } + + // Returns the total count of elements, that is the size when flattened into a + // vector. + inline int FlatSize() const { + int buffer_size = 1; + const int* dims_data = reinterpret_cast(DimsData()); + for (int i = 0; i < size_; i++) { + buffer_size *= dims_data[i]; + } + return buffer_size; + } + + bool operator!=(const RuntimeShape& comp) const { return !((*this) == comp); } + + private: + // For use only by ExtendedShape(), written to guarantee (return-value) copy + // elision in C++17. + // This creates a shape padded to the desired size with the specified value. + RuntimeShape(int new_shape_size, const RuntimeShape& shape, int pad_value) + : size_(0) { + // If the following check fails, it is likely because a 4D-only kernel is + // being used with an array of larger dimension count. + TFLITE_CHECK_GE(new_shape_size, shape.DimensionsCount()); + Resize(new_shape_size); + const int size_increase = new_shape_size - shape.DimensionsCount(); + for (int i = 0; i < size_increase; ++i) { + SetDim(i, pad_value); + } + std::memcpy(DimsData() + size_increase, shape.DimsData(), + sizeof(int32_t) * shape.DimensionsCount()); + } + + int32_t size_; + union { + int32_t dims_[kMaxSmallSize]; + int32_t* dims_pointer_; + }; +}; + +// Converts inference-style shape to legacy tflite::Dims<4>. +inline tflite::Dims<4> ToRuntimeDims(const tflite::RuntimeShape& array_shape) { + tflite::Dims<4> result; + const int dimensions_count = array_shape.DimensionsCount(); + TFLITE_CHECK_LE(dimensions_count, 4); + int cum_prod = 1; + for (int i = 0; i < 4; i++) { + const int new_dim = + (i < dimensions_count) ? array_shape.Dims(dimensions_count - 1 - i) : 1; + result.sizes[i] = new_dim; + result.strides[i] = cum_prod; + cum_prod *= new_dim; + } + return result; +} + +// TODO(b/80418076): Move to legacy ops file, update invocations. +inline RuntimeShape DimsToShape(const tflite::Dims<4>& dims) { + return RuntimeShape( + {dims.sizes[3], dims.sizes[2], dims.sizes[1], dims.sizes[0]}); +} + +// Gets next index to iterate through a multidimensional array. +inline bool NextIndex(const int num_dims, const int* dims, int* current) { + if (num_dims == 0) { + return false; + } + TFLITE_DCHECK(dims != nullptr); + TFLITE_DCHECK(current != nullptr); + int carry = 1; + for (int idx = num_dims - 1; idx >= 0; --idx) { + int current_val = current[idx] + carry; + TFLITE_DCHECK_GE(dims[idx], current_val); + if (dims[idx] == current_val) { + current[idx] = 0; + } else { + current[idx] = current_val; + carry = 0; + break; + } + } + return (carry == 0); +} + +// Gets offset of index if reducing on axis. When reducing, the flattened offset +// will not change, if the input index changes on the given axis. For example, +// if you have a 3D tensor and you are reducing to 2D by eliminating axis 0, +// then index (0, 1, 2) and index (1, 1, 2) will map to the same flattened +// offset. +// TODO(kanlig): uses Dims to represent dimensions. +inline size_t ReducedOutputOffset(const int num_dims, const int* dims, + const int* index, const int num_axis, + const int* axis) { + if (num_dims == 0) { + return 0; + } + TFLITE_DCHECK(dims != nullptr); + TFLITE_DCHECK(index != nullptr); + size_t offset = 0; + for (int idx = 0; idx < num_dims; ++idx) { + // if we need to skip this axis + bool is_axis = false; + if (axis != nullptr) { + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) { + if (idx == axis[axis_idx]) { + is_axis = true; + break; + } + } + } + if (!is_axis) { + offset = offset * static_cast(dims[idx]) + + static_cast(index[idx]); + } + } + return offset; +} + +inline int Offset(const RuntimeShape& shape, int i0, int i1, int i2, int i3) { + TFLITE_DCHECK_EQ(shape.DimensionsCount(), 4); + const int* dims_data = reinterpret_cast(shape.DimsDataUpTo5D()); + TFLITE_DCHECK(i0 >= 0 && i0 < dims_data[0]); + TFLITE_DCHECK(i1 >= 0 && i1 < dims_data[1]); + TFLITE_DCHECK(i2 >= 0 && i2 < dims_data[2]); + TFLITE_DCHECK(i3 >= 0 && i3 < dims_data[3]); + return ((i0 * dims_data[1] + i1) * dims_data[2] + i2) * dims_data[3] + i3; +} + +inline int Offset(const RuntimeShape& shape, int i0, int i1, int i2, int i3, + int i4) { + TFLITE_DCHECK_EQ(shape.DimensionsCount(), 5); + const int* dims_data = reinterpret_cast(shape.DimsDataUpTo5D()); + TFLITE_DCHECK(i0 >= 0 && i0 < dims_data[0]); + TFLITE_DCHECK(i1 >= 0 && i1 < dims_data[1]); + TFLITE_DCHECK(i2 >= 0 && i2 < dims_data[2]); + TFLITE_DCHECK(i3 >= 0 && i3 < dims_data[3]); + TFLITE_DCHECK(i4 >= 0 && i4 < dims_data[4]); + return (((i0 * dims_data[1] + i1) * dims_data[2] + i2) * dims_data[3] + i3) * + dims_data[4] + + i4; +} + +inline int Offset(const Dims<4>& dims, int i0, int i1, int i2, int i3) { + TFLITE_DCHECK(i0 >= 0 && i0 < dims.sizes[0]); + TFLITE_DCHECK(i1 >= 0 && i1 < dims.sizes[1]); + TFLITE_DCHECK(i2 >= 0 && i2 < dims.sizes[2]); + TFLITE_DCHECK(i3 >= 0 && i3 < dims.sizes[3]); + return i0 * dims.strides[0] + i1 * dims.strides[1] + i2 * dims.strides[2] + + i3 * dims.strides[3]; +} + +inline int Offset(const Dims<4>& dims, int* index) { + return Offset(dims, index[0], index[1], index[2], index[3]); +} + +inline int Offset(const RuntimeShape& shape, int* index) { + return Offset(shape, index[0], index[1], index[2], index[3]); +} + +// Get array size, DCHECKing that the dim index is in range. +// +// Note that this will be phased out with Dims<4>, since RuntimeShape::Dims() +// already performs this check. +template +int ArraySize(const Dims& array, int index) { + TFLITE_DCHECK(index >= 0 && index < N); + return array.sizes[index]; +} + +// Get common array size, DCHECKing that they all agree. +template +int MatchingArraySize(const ArrayType1& array1, int index1, + const ArrayType2& array2, int index2) { + TFLITE_DCHECK_EQ(ArraySize(array1, index1), ArraySize(array2, index2)); + return ArraySize(array1, index1); +} + +template +int MatchingArraySize(const ArrayType1& array1, int index1, + const ArrayType2& array2, int index2, Args... args) { + TFLITE_DCHECK_EQ(ArraySize(array1, index1), ArraySize(array2, index2)); + return MatchingArraySize(array1, index1, args...); +} + +// Get common shape dim, DCHECKing that they all agree. +inline int MatchingDim(const RuntimeShape& shape1, int index1, + const RuntimeShape& shape2, int index2) { + TFLITE_DCHECK_EQ(shape1.Dims(index1), shape2.Dims(index2)); + return std::min(shape1.Dims(index1), shape2.Dims(index2)); +} + +template +int MatchingDim(const RuntimeShape& shape1, int index1, + const RuntimeShape& shape2, int index2, Args... args) { + TFLITE_DCHECK_EQ(shape1.Dims(index1), shape2.Dims(index2)); + return MatchingDim(shape1, index1, args...); +} + +// Will be phased out with Dims<4>, replaced by RuntimeShape::FlatSize(). +template +inline int FlatSize(const Dims& dims) { + int flat_size = 1; + for (int i = 0; i < N; ++i) { + flat_size *= dims.sizes[i]; + } + return flat_size; +} + +TFLITE_DEPRECATED("Prefer FlatSize.") +inline int RequiredBufferSizeForDims(const Dims<4>& dims) { + return FlatSize(dims); +} + +inline int MatchingElementsSize(const RuntimeShape& shape, + const RuntimeShape& check_shape_0) { + const int size_1 = shape.FlatSize(); + const int size_2 = check_shape_0.FlatSize(); + TFLITE_CHECK_EQ(size_1, size_2); + return size_1; +} + +inline int MatchingElementsSize(const RuntimeShape& shape, + const RuntimeShape& check_shape_0, + const RuntimeShape& check_shape_1) { + const int size_1 = shape.FlatSize(); + const int size_2 = check_shape_0.FlatSize(); + const int size_3 = check_shape_1.FlatSize(); + TFLITE_CHECK_EQ(size_1, size_2); + TFLITE_CHECK_EQ(size_2, size_3); + return size_1; +} + +// Flat size calculation, checking that dimensions match with one or more other +// arrays. +inline int MatchingFlatSize(const RuntimeShape& shape, + const RuntimeShape& check_shape_0) { + TFLITE_DCHECK_EQ(shape.DimensionsCount(), check_shape_0.DimensionsCount()); + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + return shape.FlatSize(); +} + +inline int MatchingFlatSize(const RuntimeShape& shape, + const RuntimeShape& check_shape_0, + const RuntimeShape& check_shape_1) { + TFLITE_DCHECK_EQ(shape.DimensionsCount(), check_shape_0.DimensionsCount()); + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + return MatchingFlatSize(shape, check_shape_1); +} + +inline int MatchingFlatSize(const RuntimeShape& shape, + const RuntimeShape& check_shape_0, + const RuntimeShape& check_shape_1, + const RuntimeShape& check_shape_2) { + TFLITE_DCHECK_EQ(shape.DimensionsCount(), check_shape_0.DimensionsCount()); + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + return MatchingFlatSize(shape, check_shape_1, check_shape_2); +} + +inline int MatchingFlatSize(const RuntimeShape& shape, + const RuntimeShape& check_shape_0, + const RuntimeShape& check_shape_1, + const RuntimeShape& check_shape_2, + const RuntimeShape& check_shape_3) { + TFLITE_DCHECK_EQ(shape.DimensionsCount(), check_shape_0.DimensionsCount()); + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + return MatchingFlatSize(shape, check_shape_1, check_shape_2, check_shape_3); +} + +// Flat size calculation, checking that dimensions match with one or more other +// arrays. +template +inline int MatchingFlatSize(const Dims& dims, const Dims& check_dims_0) { + for (int i = 0; i < N; ++i) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + return FlatSize(dims); +} + +template +inline int MatchingFlatSize(const Dims& dims, const Dims& check_dims_0, + const Dims& check_dims_1) { + for (int i = 0; i < N; ++i) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + return MatchingFlatSize(dims, check_dims_1); +} + +template +inline int MatchingFlatSize(const Dims& dims, const Dims& check_dims_0, + const Dims& check_dims_1, + const Dims& check_dims_2) { + for (int i = 0; i < N; ++i) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + return MatchingFlatSize(dims, check_dims_1, check_dims_2); +} + +template +inline int MatchingFlatSize(const Dims& dims, const Dims& check_dims_0, + const Dims& check_dims_1, + const Dims& check_dims_2, + const Dims& check_dims_3) { + for (int i = 0; i < N; ++i) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + return MatchingFlatSize(dims, check_dims_1, check_dims_2, check_dims_3); +} + +// Data is required to be contiguous, and so many operators can use either the +// full array flat size or the flat size with one dimension skipped (commonly +// the depth). +template +inline int FlatSizeSkipDim(const Dims& dims, int skip_dim) { + TFLITE_DCHECK(skip_dim >= 0 && skip_dim < N); + int flat_size = 1; + for (int i = 0; i < N; ++i) { + flat_size *= (i == skip_dim) ? 1 : dims.sizes[i]; + } + return flat_size; +} + +// A combination of MatchingFlatSize() and FlatSizeSkipDim(). +template +inline int MatchingFlatSizeSkipDim(const Dims& dims, int skip_dim, + const Dims& check_dims_0) { + for (int i = 0; i < N; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + } + return FlatSizeSkipDim(dims, skip_dim); +} + +template +inline int MatchingFlatSizeSkipDim(const Dims& dims, int skip_dim, + const Dims& check_dims_0, + const Dims& check_dims_1) { + for (int i = 0; i < N; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + } + return MatchingFlatSizeSkipDim(dims, skip_dim, check_dims_1); +} + +template +inline int MatchingFlatSizeSkipDim(const Dims& dims, int skip_dim, + const Dims& check_dims_0, + const Dims& check_dims_1, + const Dims& check_dims_2) { + for (int i = 0; i < N; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + } + return MatchingFlatSizeSkipDim(dims, skip_dim, check_dims_1, check_dims_2); +} + +template +inline int MatchingFlatSizeSkipDim(const Dims& dims, int skip_dim, + const Dims& check_dims_0, + const Dims& check_dims_1, + const Dims& check_dims_2, + const Dims& check_dims_3) { + for (int i = 0; i < N; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(ArraySize(dims, i), ArraySize(check_dims_0, i)); + } + } + return MatchingFlatSizeSkipDim(dims, skip_dim, check_dims_1, check_dims_2, + check_dims_3); +} + +// Data is required to be contiguous, and so many operators can use either the +// full array flat size or the flat size with one dimension skipped (commonly +// the depth). +inline int FlatSizeSkipDim(const RuntimeShape& shape, int skip_dim) { + const int dims_count = shape.DimensionsCount(); + TFLITE_DCHECK(skip_dim >= 0 && skip_dim < dims_count); + const auto* dims_data = shape.DimsData(); + int flat_size = 1; + for (int i = 0; i < dims_count; ++i) { + flat_size *= (i == skip_dim) ? 1 : dims_data[i]; + } + return flat_size; +} + +// A combination of MatchingFlatSize() and FlatSizeSkipDim(). +inline int MatchingFlatSizeSkipDim(const RuntimeShape& shape, int skip_dim, + const RuntimeShape& check_shape_0) { + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + } + return FlatSizeSkipDim(shape, skip_dim); +} + +inline int MatchingFlatSizeSkipDim(const RuntimeShape& shape, int skip_dim, + const RuntimeShape& check_shape_0, + const RuntimeShape& check_shape_1) { + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + } + return MatchingFlatSizeSkipDim(shape, skip_dim, check_shape_1); +} + +inline int MatchingFlatSizeSkipDim(const RuntimeShape& shape, int skip_dim, + const RuntimeShape& check_shape_0, + const RuntimeShape& check_shape_1, + const RuntimeShape& check_shape_2) { + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + } + return MatchingFlatSizeSkipDim(shape, skip_dim, check_shape_1, check_shape_2); +} + +inline int MatchingFlatSizeSkipDim(const RuntimeShape& shape, int skip_dim, + const RuntimeShape& check_shape_0, + const RuntimeShape& check_shape_1, + const RuntimeShape& check_shape_2, + const RuntimeShape& check_shape_3) { + const int dims_count = shape.DimensionsCount(); + for (int i = 0; i < dims_count; ++i) { + if (i != skip_dim) { + TFLITE_DCHECK_EQ(shape.Dims(i), check_shape_0.Dims(i)); + } + } + return MatchingFlatSizeSkipDim(shape, skip_dim, check_shape_1, check_shape_2, + check_shape_3); +} + +template +bool IsPackedWithoutStrides(const Dims& dims) { + int expected_stride = 1; + for (int d = 0; d < N; d++) { + if (dims.strides[d] != expected_stride) return false; + expected_stride *= dims.sizes[d]; + } + return true; +} + +template +void ComputeStrides(Dims* dims) { + dims->strides[0] = 1; + for (int d = 1; d < N; d++) { + dims->strides[d] = dims->strides[d - 1] * dims->sizes[d - 1]; + } +} + +enum class BroadcastableOpCategory : uint8_t { + kNone, + kNonBroadcast, // Matching input shapes. + kFirstInputBroadcastsFast, // Fivefold nested loops. + kSecondInputBroadcastsFast, // Fivefold nested loops. + kGenericBroadcast, // Fall-back. +}; + +struct MinMax { + float min; + float max; +}; +static_assert(sizeof(MinMax) == 8, ""); + +struct ActivationParams { + FusedActivationFunctionType activation_type; + // uint8_t, etc, activation params. + int32_t quantized_activation_min; + int32_t quantized_activation_max; +}; + +struct ReluParams : public ActivationParams { + int32_t input_offset; + int32_t output_offset; + int32_t output_multiplier; + int output_shift; +}; + +// Styles of resizing op usages. For example, kImageStyle can be used with a Pad +// op for pattern-specific optimization. +enum class ResizingCategory : uint8_t { + kNone, + kImageStyle, // 4D, operating on inner dimensions, say {0, a, b, 0}. + kGenericResize, +}; + +// For Add, Sub, Mul ops. +struct ArithmeticParams { + // Shape dependent / common to data / op types. + BroadcastableOpCategory broadcast_category; + // uint8_t inference params. + int32_t input1_offset; + int32_t input2_offset; + int32_t output_offset; + int32_t output_multiplier; + int output_shift; + // Add / Sub, not Mul, uint8_t inference params. + int left_shift; + int32_t input1_multiplier; + int input1_shift; + int32_t input2_multiplier; + int input2_shift; + + // TODO(b/158622529): Union the following activation params. + // uint8_t, etc, activation params. + int32_t quantized_activation_min; + int32_t quantized_activation_max; + // float activation params. + float float_activation_min; + float float_activation_max; + // int64_t activation params. + int64_t int64_activation_min; + int64_t int64_activation_max; + + // Processed output dimensions. + // Let input "a" be the one that broadcasts in the faster-changing dimension. + // Then, after coalescing, for shapes {a0, a1, a2, a3, a4} and + // {b0, b1, b2, b3, b4}, + // broadcast_shape[4] = b0 = a0. + // broadcast_shape[3] = b1; a1 = 1. + // broadcast_shape[2] = b2 = a2. + // broadcast_shape[1] = a3; b3 = 1. + // broadcast_shape[0] = b4 = a4. + int broadcast_shape[5]; +}; + +struct ConcatenationParams { + int8_t axis; + const int32_t* input_zeropoint; + const float* input_scale; + uint16_t inputs_count; + int32_t output_zeropoint; + float output_scale; +}; + +struct ComparisonParams { + // uint8_t inference params. + int left_shift; + int32_t input1_offset; + int32_t input1_multiplier; + int input1_shift; + int32_t input2_offset; + int32_t input2_multiplier; + int input2_shift; + // Shape dependent / common to inference types. + bool is_broadcast; +}; + +struct ConvParams { + PaddingType padding_type; + PaddingValues padding_values; + // TODO(starka): This was just "stride", so check that width+height is OK. + int16_t stride_width; + int16_t stride_height; + int16_t dilation_width_factor; + int16_t dilation_height_factor; + // uint8_t inference params. + // TODO(b/65838351): Use smaller types if appropriate. + int32_t input_offset; + int32_t weights_offset; + int32_t output_offset; + int32_t output_multiplier; + int output_shift; + // uint8_t, etc, activation params. + int32_t quantized_activation_min; + int32_t quantized_activation_max; + // float activation params. + float float_activation_min; + float float_activation_max; +}; + +struct DepthToSpaceParams { + int32_t block_size; +}; + +struct DepthwiseParams { + PaddingType padding_type; + PaddingValues padding_values; + int16_t stride_width; + int16_t stride_height; + int16_t dilation_width_factor; + int16_t dilation_height_factor; + int16_t depth_multiplier; + // uint8_t inference params. + // TODO(b/65838351): Use smaller types if appropriate. + int32_t input_offset; + int32_t weights_offset; + int32_t output_offset; + int32_t output_multiplier; + int output_shift; + // uint8_t, etc, activation params. + int32_t quantized_activation_min; + int32_t quantized_activation_max; + // float activation params. + float float_activation_min; + float float_activation_max; + const int32_t* output_multiplier_per_channel; + const int32_t* output_shift_per_channel; +}; + +struct DequantizationParams { + double scale; + int32_t zero_point; +}; + +struct PerChannelDequantizationParams { + const float* scale; + const int32_t* zero_point; + int32_t quantized_dimension; +}; + +struct FakeQuantParams { + MinMax minmax; + int32_t num_bits; +}; + +struct FullyConnectedParams { + // uint8_t inference params. + // TODO(b/65838351): Use smaller types if appropriate. + int32_t input_offset; + int32_t weights_offset; + int32_t output_offset; + int32_t output_multiplier; + int output_shift; + // uint8_t, etc, activation params. + int32_t quantized_activation_min; + int32_t quantized_activation_max; + // float activation params. + float float_activation_min; + float float_activation_max; + // Mark the operands as cacheable if they are unchanging, e.g. weights. + bool lhs_cacheable; + bool rhs_cacheable; + FullyConnectedWeightsFormat weights_format; +}; + +struct GatherParams { + int16_t axis; +}; + +struct L2NormalizationParams { + // uint8_t inference params. + int32_t input_zero_point; +}; + +struct LocalResponseNormalizationParams { + int32_t range; + double bias; + double alpha; + double beta; +}; + +struct HardSwishParams { + // zero_point of the input activations. + int16_t input_zero_point; + // zero_point of the output activations. + int16_t output_zero_point; + // 16bit fixed-point component of the multiplier to apply to go from the + // "high-res input scale", which is the input scale multiplied by 2^7, to the + // "relu-ish scale", which 3.0/32768. + // See the implementation of HardSwishPrepare. + int16_t reluish_multiplier_fixedpoint_int16; + // exponent/bit-shift component of the aforementioned multiplier. + int reluish_multiplier_exponent; + // 16bit fixed-point component of the multiplier to apply to go from the + // "high-res input scale", which is the input scale multiplied by 2^7, to the + // output scale. + // See the implementation of HardSwishPrepare. + int16_t output_multiplier_fixedpoint_int16; + // exponent/bit-shift component of the aforementioned multiplier. + int output_multiplier_exponent; +}; + +struct LogisticParams { + // uint8_t inference params. + int32_t input_zero_point; + int32_t input_range_radius; + int32_t input_multiplier; + int input_left_shift; +}; + +struct LstmCellParams { + int32_t weights_zero_point; + int32_t accum_multiplier; + int accum_shift; + int state_integer_bits; +}; + +struct MeanParams { + int8_t axis_count; + int16_t axis[4]; +}; + +struct PackParams { + int8_t axis; + const int32_t* input_zeropoint; + const float* input_scale; + uint16_t inputs_count; + int32_t output_zeropoint; + float output_scale; +}; + +struct PadParams { + int8_t left_padding_count; + int32_t left_padding[4]; + int8_t right_padding_count; + int32_t right_padding[4]; + ResizingCategory resizing_category; +}; + +struct PreluParams { + int32_t input_offset; + int32_t alpha_offset; + int32_t output_offset; + int32_t output_multiplier_1; + int output_shift_1; + int32_t output_multiplier_2; + int output_shift_2; +}; + +struct PoolParams { + FusedActivationFunctionType activation; + PaddingType padding_type; + PaddingValues padding_values; + int stride_height; + int stride_width; + int filter_height; + int filter_width; + // uint8_t, etc, activation params. + int32_t quantized_activation_min; + int32_t quantized_activation_max; + // float activation params. + float float_activation_min; + float float_activation_max; +}; + +struct ReshapeParams { + int8_t shape_count; + int32_t shape[4]; +}; + +struct ResizeBilinearParams { + bool align_corners; + // half_pixel_centers assumes pixels are of half the actual dimensions, and + // yields more accurate resizes. Corresponds to the same argument for the + // original TensorFlow op in TF2.0. + bool half_pixel_centers; +}; + +struct ResizeNearestNeighborParams { + bool align_corners; + bool half_pixel_centers; +}; + +struct SliceParams { + int8_t begin_count; + int32_t begin[5]; + int8_t size_count; + int32_t size[5]; +}; + +struct SoftmaxParams { + // beta is not really used (not a Tensorflow parameter) and not implemented + // for LogSoftmax. + double beta; + // uint8_t inference params. Used even when beta defaults to 1.0. + int32_t input_multiplier; + int32_t input_left_shift; + // Reverse scaling is only used by LogSoftmax. + int32_t reverse_scaling_divisor; + int32_t reverse_scaling_right_shift; + int diff_min; + int32_t zero_point; + float scale; + float* table; + // int16 LUT for exp(x), where x uniform distributed between [-10.0 , 0.0] + int16_t* exp_lut; + // int16 LUT for 1 / (1 + x), where x uniform distributed between [0.0 , 1.0] + int16_t* one_over_one_plus_x_lut; + uint8_t* uint8_table1; + uint8_t* uint8_table2; +}; + +struct SpaceToBatchParams { + // "Zero" padding for uint8_t means padding with the output offset. + int32_t output_offset; +}; + +struct SpaceToDepthParams { + int32_t block_size; +}; + +struct SplitParams { + // Graphs that split into, say, 2000 nodes are encountered. The indices in + // OperatorEdges are of type uint16_t. + uint16_t num_split; + int16_t axis; +}; + +struct SqueezeParams { + int8_t squeeze_dims_count; + int32_t squeeze_dims[4]; +}; + +struct StridedSliceParams { + int8_t start_indices_count; + int32_t start_indices[5]; + int8_t stop_indices_count; + int32_t stop_indices[5]; + int8_t strides_count; + int32_t strides[5]; + + int16_t begin_mask; + int16_t ellipsis_mask; + int16_t end_mask; + int16_t new_axis_mask; + int16_t shrink_axis_mask; +}; + +struct TanhParams { + int32_t input_zero_point; + int32_t input_range_radius; + int32_t input_multiplier; + int input_left_shift; +}; + +struct TransposeParams { + int8_t perm_count; + int32_t perm[5]; +}; + +struct UnpackParams { + uint16_t num_split; + int16_t axis; +}; + +struct LeakyReluParams { + float alpha; + int32_t input_offset; + int32_t output_offset; + int32_t output_multiplier_alpha; + int32_t output_shift_alpha; + int32_t output_multiplier_identity; + int32_t output_shift_identity; +}; + +template +inline void SetActivationParams(float min, float max, P* params) { + params->float_activation_min = min; + params->float_activation_max = max; +} + +template +inline void SetActivationParams(int32_t min, int32_t max, P* params) { + params->quantized_activation_min = min; + params->quantized_activation_max = max; +} + +template +inline void SetActivationParams(int64_t min, int64_t max, P* params) { + params->int64_activation_min = min; + params->int64_activation_max = max; +} + +template +inline void GetActivationParams(const P& params, int32_t* min, int32_t* max) { + *min = params.quantized_activation_min; + *max = params.quantized_activation_max; +} + +template +inline void GetActivationParams(const P& params, float* min, float* max) { + *min = params.float_activation_min; + *max = params.float_activation_max; +} + +template +inline void GetActivationParams(const P& params, int64_t* min, int64_t* max) { + *min = params.int64_activation_min; + *max = params.int64_activation_max; +} +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_TYPES_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.cpp b/src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.cpp new file mode 100644 index 0000000..05429da --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.cpp @@ -0,0 +1,483 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" + +#include +#include + +#include +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" + +namespace tflite { + +namespace { + +// Assumes tensor_index is a valid index (in bounds) +inline TfLiteTensor* GetTensorAtIndex(const TfLiteContext* context, + int tensor_index) { + if (context->tensors != nullptr) { + return &context->tensors[tensor_index]; + } else { + return context->GetTensor(context, tensor_index); + } +} + +// Validate in a single place to reduce binary size +inline TfLiteStatus ValidateTensorIndexingSafe(const TfLiteContext* context, + int index, int max_size, + const int* tensor_indices, + int* tensor_index) { + if (index < 0 || index >= max_size) { + TF_LITE_KERNEL_LOG(const_cast(context), + "Invalid tensor index %d (not in [0, %d))\n", index, + max_size); + return kTfLiteError; + } + if (tensor_indices[index] == kTfLiteOptionalTensor) { + TF_LITE_KERNEL_LOG(const_cast(context), + "Tensor at index %d was optional but was expected\n", + index); + return kTfLiteError; + } + + *tensor_index = tensor_indices[index]; + return kTfLiteOk; +} + +// Same as above but returns -1 for invalid inputs instead of status + logging +// error. +inline int ValidateTensorIndexing(const TfLiteContext* context, int index, + int max_size, const int* tensor_indices) { + if (index >= 0 && index < max_size) { + const int tensor_index = tensor_indices[index]; + if (tensor_index != kTfLiteOptionalTensor) { + return tensor_index; + } + } + return -1; +} + +inline TfLiteTensor* GetMutableInput(const TfLiteContext* context, + const TfLiteNode* node, int index) { + const int tensor_index = ValidateTensorIndexing( + context, index, node->inputs->size, node->inputs->data); + if (tensor_index < 0) { + return nullptr; + } + return GetTensorAtIndex(context, tensor_index); +} + +inline TfLiteStatus GetMutableInputSafe(const TfLiteContext* context, + const TfLiteNode* node, int index, + const TfLiteTensor** tensor) { + int tensor_index; + TF_LITE_ENSURE_OK( + context, ValidateTensorIndexingSafe(context, index, node->inputs->size, + node->inputs->data, &tensor_index)); + *tensor = GetTensorAtIndex(context, tensor_index); + return kTfLiteOk; +} + +} // anonymous namespace. + +const TfLiteTensor* GetInput(const TfLiteContext* context, + const TfLiteNode* node, int index) { + return GetMutableInput(context, node, index); +} + +TfLiteStatus GetInputSafe(const TfLiteContext* context, const TfLiteNode* node, + int index, const TfLiteTensor** tensor) { + return GetMutableInputSafe(context, node, index, tensor); +} + +TfLiteTensor* GetVariableInput(TfLiteContext* context, const TfLiteNode* node, + int index) { + TfLiteTensor* tensor = GetMutableInput(context, node, index); + return tensor->is_variable ? tensor : nullptr; +} + +TfLiteTensor* GetOutput(TfLiteContext* context, const TfLiteNode* node, + int index) { + const int tensor_index = ValidateTensorIndexing( + context, index, node->outputs->size, node->outputs->data); + if (tensor_index < 0) { + return nullptr; + } + return GetTensorAtIndex(context, tensor_index); +} + +TfLiteStatus GetOutputSafe(const TfLiteContext* context, const TfLiteNode* node, + int index, TfLiteTensor** tensor) { + int tensor_index; + TF_LITE_ENSURE_OK( + context, ValidateTensorIndexingSafe(context, index, node->outputs->size, + node->outputs->data, &tensor_index)); + *tensor = GetTensorAtIndex(context, tensor_index); + return kTfLiteOk; +} + +const TfLiteTensor* GetOptionalInputTensor(const TfLiteContext* context, + const TfLiteNode* node, int index) { + return GetInput(context, node, index); +} + +#ifndef TF_LITE_STATIC_MEMORY +TfLiteTensor* GetTemporary(TfLiteContext* context, const TfLiteNode* node, + int index) { + const int tensor_index = ValidateTensorIndexing( + context, index, node->temporaries->size, node->temporaries->data); + if (tensor_index < 0) { + return nullptr; + } + return GetTensorAtIndex(context, tensor_index); +} + +TfLiteStatus GetTemporarySafe(const TfLiteContext* context, + const TfLiteNode* node, int index, + TfLiteTensor** tensor) { + int tensor_index; + TF_LITE_ENSURE_OK(context, ValidateTensorIndexingSafe( + context, index, node->temporaries->size, + node->temporaries->data, &tensor_index)); + *tensor = GetTensorAtIndex(context, tensor_index); + return kTfLiteOk; +} + +const TfLiteTensor* GetIntermediates(TfLiteContext* context, + const TfLiteNode* node, int index) { + const int tensor_index = ValidateTensorIndexing( + context, index, node->intermediates->size, node->intermediates->data); + if (tensor_index < 0) { + return nullptr; + } + return GetTensorAtIndex(context, tensor_index); +} + +TfLiteStatus GetIntermediatesSafe(const TfLiteContext* context, + const TfLiteNode* node, int index, + TfLiteTensor** tensor) { + int tensor_index; + TF_LITE_ENSURE_OK(context, ValidateTensorIndexingSafe( + context, index, node->intermediates->size, + node->intermediates->data, &tensor_index)); + *tensor = GetTensorAtIndex(context, tensor_index); + return kTfLiteOk; +} +#endif // TF_LITE_STATIC_MEMORY + +// Per-axis +TfLiteStatus PopulateConvolutionQuantizationParams( + TfLiteContext* context, const TfLiteTensor* input, + const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output, + const TfLiteFusedActivation& activation, int32_t* multiplier, int* shift, + int32_t* output_activation_min, int32_t* output_activation_max, + int32_t* per_channel_multiplier, int* per_channel_shift) { + const auto* affine_quantization = + reinterpret_cast(filter->quantization.params); + return PopulateConvolutionQuantizationParams( + context, input, filter, bias, output, activation, multiplier, shift, + output_activation_min, output_activation_max, per_channel_multiplier, + per_channel_shift, affine_quantization->scale->size); +} + +// Per-axis & per-tensor +TfLiteStatus PopulateConvolutionQuantizationParams( + TfLiteContext* context, const TfLiteTensor* input, + const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output, + const TfLiteFusedActivation& activation, int32_t* multiplier, int* shift, + int32_t* output_activation_min, int32_t* output_activation_max, + int32_t* per_channel_multiplier, int* per_channel_shift, int num_channels) { + TF_LITE_ENSURE_EQ(context, input->quantization.type, + kTfLiteAffineQuantization); + TF_LITE_ENSURE_EQ(context, filter->quantization.type, + kTfLiteAffineQuantization); + // TODO(jianlijianli): Enable bias type check and bias scale == input scale + // * filter scale for each channel in affine quantization once bias + // quantization is properly populated. + // TF_LITE_ENSURE_EQ(context, bias->quantization.type, + // kTfLiteAffineQuantization); + + // Check data type. + const auto* affine_quantization = + reinterpret_cast(filter->quantization.params); + TF_LITE_ENSURE(context, affine_quantization); + TF_LITE_ENSURE(context, affine_quantization->scale); + const bool is_per_channel = affine_quantization->scale->size > 1; + if (is_per_channel) { + // Currently only Int8/Int16 is supported for per channel quantization. + TF_LITE_ENSURE(context, + input->type == kTfLiteInt8 || input->type == kTfLiteInt16); + TF_LITE_ENSURE_EQ(context, filter->type, kTfLiteInt8); + TF_LITE_ENSURE_EQ(context, affine_quantization->scale->size, num_channels); + TF_LITE_ENSURE_EQ( + context, num_channels, + filter->dims->data[affine_quantization->quantized_dimension]); + } + + // Populate multiplier and shift using affine quantization. + const float input_scale = input->params.scale; + const float output_scale = output->params.scale; + const float* filter_scales = affine_quantization->scale->data; + for (int i = 0; i < num_channels; ++i) { + // If per-tensor quantization parameter is specified, broadcast it along the + // quantization dimension (channels_out). + const float scale = is_per_channel ? filter_scales[i] : filter_scales[0]; + const double filter_scale = static_cast(scale); + const double effective_output_scale = static_cast(input_scale) * + filter_scale / + static_cast(output_scale); + int32_t significand; + int channel_shift; + QuantizeMultiplier(effective_output_scale, &significand, &channel_shift); + per_channel_multiplier[i] = significand; + per_channel_shift[i] = channel_shift; + } + + // Populate scalar quantization parameters. + // This check on legacy quantization parameters is kept only for backward + // compatibility. + if (input->type == kTfLiteUInt8) { + // Check bias scale == input scale * filter scale. + double real_multiplier = 0.0; + TF_LITE_ENSURE_STATUS(GetQuantizedConvolutionMultipler( + context, input, filter, bias, output, &real_multiplier)); + int exponent; + + // Populate quantization parameters with multiplier and shift. + QuantizeMultiplier(real_multiplier, multiplier, &exponent); + *shift = -exponent; + } + if (input->type == kTfLiteInt8 || input->type == kTfLiteUInt8 || + input->type == kTfLiteInt16) { + TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( + context, activation, output, output_activation_min, + output_activation_max)); + } + return kTfLiteOk; +} + +TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context, + const TfLiteTensor* input, + const TfLiteTensor* filter, + const TfLiteTensor* bias, + TfLiteTensor* output, + double* multiplier) { + const double input_product_scale = static_cast(input->params.scale) * + static_cast(filter->params.scale); + // The following conditions must be guaranteed by the training pipeline. + if (bias) { + const double bias_scale = static_cast(bias->params.scale); + // Here we're making sure the input_product_scale & bias_scale are about the + // same. Since we have: + // (output - output_zp) * output_scale = + // input_product_scale * input_product + bias * bias_scale ---- (0) + // + // (0) equals: + // (input_product + bias) * input_product_scale ----- (1) + // + + // bias * (bias_scale - input_product_scale) ------ (2) + // + // For the real kernel computation, we're doing (1), so we really need to + // make sure (2) has minimum impact on the output, so: + // bias * (bias_scale - input_product_scale) / output_scale should be + // a small number for an integer. + // Since normally bias should be within a small range. + // We should expect (bias_scale - input_product_scale) / output_scale to + // be a small number like 0.02. + const double scale_diff = std::abs(input_product_scale - bias_scale); + const double output_scale = static_cast(output->params.scale); + + TF_LITE_ENSURE(context, scale_diff / output_scale <= 0.02); + } + return GetQuantizedConvolutionMultipler(context, input, filter, output, + multiplier); +} + +TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context, + const TfLiteTensor* input, + const TfLiteTensor* filter, + TfLiteTensor* output, + double* multiplier) { + const double input_product_scale = + static_cast(input->params.scale * filter->params.scale); + TF_LITE_ENSURE(context, input_product_scale >= 0); + *multiplier = input_product_scale / static_cast(output->params.scale); + + return kTfLiteOk; +} + +namespace { +void CalculateActivationRangeQuantizedImpl(TfLiteFusedActivation activation, + int32_t qmin, int32_t qmax, + TfLiteTensor* output, + int32_t* act_min, int32_t* act_max) { + const auto scale = output->params.scale; + const auto zero_point = output->params.zero_point; + + auto quantize = [scale, zero_point](float f) { + return zero_point + static_cast(TfLiteRound(f / scale)); + }; + + if (activation == kTfLiteActRelu) { + *act_min = std::max(qmin, quantize(0.0)); + *act_max = qmax; + } else if (activation == kTfLiteActRelu6) { + *act_min = std::max(qmin, quantize(0.0)); + *act_max = std::min(qmax, quantize(6.0)); + } else if (activation == kTfLiteActReluN1To1) { + *act_min = std::max(qmin, quantize(-1.0)); + *act_max = std::min(qmax, quantize(1.0)); + } else { + *act_min = qmin; + *act_max = qmax; + } +} +} // namespace + +TfLiteStatus CalculateActivationRangeQuantized(TfLiteContext* context, + TfLiteFusedActivation activation, + TfLiteTensor* output, + int32_t* act_min, + int32_t* act_max) { + int32_t qmin = 0; + int32_t qmax = 0; + if (output->type == kTfLiteUInt8) { + qmin = std::numeric_limits::min(); + qmax = std::numeric_limits::max(); + } else if (output->type == kTfLiteInt8) { + qmin = std::numeric_limits::min(); + qmax = std::numeric_limits::max(); + } else if (output->type == kTfLiteInt16) { + qmin = std::numeric_limits::min(); + qmax = std::numeric_limits::max(); + } else { + TF_LITE_ENSURE(context, false); + } + + CalculateActivationRangeQuantizedImpl(activation, qmin, qmax, output, act_min, + act_max); + return kTfLiteOk; +} + +bool HaveSameShapes(const TfLiteTensor* input1, const TfLiteTensor* input2) { + return TfLiteIntArrayEqual(input1->dims, input2->dims); +} + +// TODO(petewarden): Having macros around this is ugly, look at other strategies +// before replicating this approach elsewhere. +#ifndef TF_LITE_STATIC_MEMORY +TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, + const TfLiteTensor* input1, + const TfLiteTensor* input2, + TfLiteIntArray** output_shape) { + int dims1 = NumDimensions(input1); + int dims2 = NumDimensions(input2); + int out_dims = std::max(dims1, dims2); + if (NumElements(input1) == 0) { + *output_shape = TfLiteIntArrayCopy(input1->dims); + return kTfLiteOk; + } + std::unique_ptr shape( + TfLiteIntArrayCreate(out_dims), TfLiteIntArrayFree); + for (int i = 0; i < out_dims; ++i) { + int d1 = i >= dims1 ? 1 : SizeOfDimension(input1, dims1 - i - 1); + int d2 = i >= dims2 ? 1 : SizeOfDimension(input2, dims2 - i - 1); + TF_LITE_ENSURE(context, d1 == d2 || d1 == 1 || d2 == 1); + shape->data[out_dims - i - 1] = std::max(d1, d2); + } + *output_shape = shape.release(); + return kTfLiteOk; +} + +TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, + const TfLiteTensor* input1, + const TfLiteTensor* input2, + const TfLiteTensor* input3, + TfLiteIntArray** output_shape) { + int dims1 = NumDimensions(input1); + int dims2 = NumDimensions(input2); + int dims3 = NumDimensions(input3); + int out_dims = std::max(std::max(dims1, dims2), dims3); + std::unique_ptr shape( + TfLiteIntArrayCreate(out_dims), TfLiteIntArrayFree); + for (int i = 0; i < out_dims; ++i) { + int d1 = i >= dims1 ? 1 : SizeOfDimension(input1, dims1 - i - 1); + int d2 = i >= dims2 ? 1 : SizeOfDimension(input2, dims2 - i - 1); + int d3 = i >= dims3 ? 1 : SizeOfDimension(input3, dims3 - i - 1); + int max_value = std::max(std::max(d1, d2), d3); + TF_LITE_ENSURE(context, d1 == 1 || d1 == max_value); + TF_LITE_ENSURE(context, d2 == 1 || d2 == max_value); + TF_LITE_ENSURE(context, d3 == 1 || d3 == max_value); + shape->data[out_dims - i - 1] = max_value; + } + *output_shape = shape.release(); + return kTfLiteOk; +} +#endif // TF_LITE_STATIC_MEMORY + +// Size of string is not constant, return 0 in such case. +int TfLiteTypeGetSize(TfLiteType type) { + switch (type) { + case kTfLiteUInt8: + TF_LITE_ASSERT_EQ(sizeof(uint8_t), 1); + return 1; + case kTfLiteInt8: + TF_LITE_ASSERT_EQ(sizeof(int8_t), 1); + return 1; + case kTfLiteBool: + return sizeof(bool); + case kTfLiteInt16: + TF_LITE_ASSERT_EQ(sizeof(int16_t), 2); + return 2; + case kTfLiteFloat16: + TF_LITE_ASSERT_EQ(sizeof(int16_t), 2); + return 2; + case kTfLiteFloat32: + TF_LITE_ASSERT_EQ(sizeof(float), 4); + return 4; + case kTfLiteInt32: + TF_LITE_ASSERT_EQ(sizeof(int32_t), 4); + return 4; + case kTfLiteInt64: + TF_LITE_ASSERT_EQ(sizeof(int64_t), 8); + return 8; + case kTfLiteUInt64: + TF_LITE_ASSERT_EQ(sizeof(uint64_t), 8); + return 8; + case kTfLiteFloat64: + TF_LITE_ASSERT_EQ(sizeof(double), 8); + return 8; + case kTfLiteComplex64: + TF_LITE_ASSERT_EQ(sizeof(std::complex), 8); + return 8; + case kTfLiteComplex128: + TF_LITE_ASSERT_EQ(sizeof(std::complex), 16); + return 16; + default: + return 0; + } +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.h b/src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.h new file mode 100644 index 0000000..f7612af --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/kernel_util.h @@ -0,0 +1,296 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_KERNEL_UTIL_H_ +#define TENSORFLOW_LITE_KERNELS_KERNEL_UTIL_H_ + +#include + +#include + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +// A fair number of functions in this header have historically been inline. +// It is ok to change functions to not be inline if the latency with +// benchmark_model for MobileNet + MobileBERT is unaffected. If such a change is +// made, move the newly non-inlined function declarations to the top of this +// header file. + +// Note: You must check if result is not null: +// +// TfLiteTensor* my_tensor = GetInput(context, node, kMyTensorIdx); +// TF_LITE_ENSURE(context, my_tensor != nullptr); +// +// This is because the index might point to the optional tensor constant +// (kTfLiteOptionalTensor) in which case there is no tensor to return. +const TfLiteTensor* GetInput(const TfLiteContext* context, + const TfLiteNode* node, int index); + +// Same as `GetInput` but returns boolean and uses output argument for tensor. +// +// TfLiteTensor* my_tensor; +// TF_LITE_ENSURE_OK(context, +// GetInputSafe(context, node, kMyTensorIdx, &my_tensor)); +// // can use my_tensor directly from here onwards, it is not nullptr +// +// Should be used in cases where the binary size is too large. +TfLiteStatus GetInputSafe(const TfLiteContext* context, const TfLiteNode* node, + int index, const TfLiteTensor** tensor); + +// Note: You must check if result is not null: +// +// TfLiteTensor* my_tensor = GetVariableInput(context, node, kMyTensorIdx); +// TF_LITE_ENSURE(context, my_tensor != nullptr); +// +// This is because the index might point to the optional tensor constant +// (kTfLiteOptionalTensor) in which case there is no tensor to return. +TfLiteTensor* GetVariableInput(TfLiteContext* context, const TfLiteNode* node, + int index); + +// Note: You must check if result is not null: +// +// TfLiteTensor* my_tensor = GetOutput(context, node, kMyTensorIdx); +// TF_LITE_ENSURE(context, my_tensor != nullptr); +// +// This is because the index might point to the optional tensor constant +// (kTfLiteOptionalTensor) in which case there is no tensor to return. +TfLiteTensor* GetOutput(TfLiteContext* context, const TfLiteNode* node, + int index); + +// Same as `GetOutput` but returns boolean and uses output argument for tensor. +// +// TfLiteTensor* my_tensor; +// TF_LITE_ENSURE_OK(context, +// GetOutputSafe(context, node, kMyTensorIdx, &my_tensor)); +// // can use my_tensor directly from here onwards, it is not nullptr +// +// Should be used in cases where the binary size is too large. +TfLiteStatus GetOutputSafe(const TfLiteContext* context, const TfLiteNode* node, + int index, TfLiteTensor** tensor); + +// Note: You must check if result is not null: +// +// TfLiteTensor* my_tensor = GetOptionalInputTensor(context, node, kIdx); +// TF_LITE_ENSURE(context, my_tensor != nullptr); +// +// This is because the index might point to the optional tensor constant +// (kTfLiteOptionalTensor) in which case there is no tensor to return. +// +// Deprecated. GetInput has the same functionality. +const TfLiteTensor* GetOptionalInputTensor(const TfLiteContext* context, + const TfLiteNode* node, int index); + +#ifndef TF_LITE_STATIC_MEMORY +// Note: You must check if result is not null: +// +// TfLiteTensor* my_tensor = GetTemporary(context, node, kMyTensorIdx); +// TF_LITE_ENSURE(context, my_tensor != nullptr); +// +// This is because the index might point to the optional tensor constant +// (kTfLiteOptionalTensor) in which case there is no tensor to return. +TfLiteTensor* GetTemporary(TfLiteContext* context, const TfLiteNode* node, + int index); + +// Same as `GetTemporary` but returns boolean and uses output argument for +// tensor. +// +// TfLiteTensor* my_tensor; +// TF_LITE_ENSURE_OK(context, +// GetTemporarySafe(context, node, kMyTensorIdx, +// &my_tensor)); +// // can use my_tensor directly from here onwards, it is not nullptr +// +// Should be used in cases where the binary size is too large. +TfLiteStatus GetTemporarySafe(const TfLiteContext* context, + const TfLiteNode* node, int index, + TfLiteTensor** tensor); + +// Note: You must check if result is not null: +// +// TfLiteTensor* my_tensor = GetIntermediates(context, node, kMyTensorIdx); +// TF_LITE_ENSURE(context, my_tensor != nullptr); +// +// This is because the index might point to the optional tensor constant +// (kTfLiteOptionalTensor) in which case there is no tensor to return. +const TfLiteTensor* GetIntermediates(TfLiteContext* context, + const TfLiteNode* node, int index); + +// Same as `GetIntermediates` but returns boolean and uses output argument for +// tensor. +// +// TfLiteTensor* my_tensor; +// TF_LITE_ENSURE_OK(context, +// GetIntermediatesSafe(context, node, kMyTensorIdx, +// &my_tensor)); +// // can use my_tensor directly from here onwards, it is not nullptr +// +// Should be used in cases where the binary size is too large. +TfLiteStatus GetIntermediatesSafe(const TfLiteContext* context, + const TfLiteNode* node, int index, + TfLiteTensor** tensor); +#endif // TF_LITE_STATIC_MEMORY + +inline int NumDimensions(const TfLiteTensor* t) { return t->dims->size; } +inline int SizeOfDimension(const TfLiteTensor* t, int dim) { + return t->dims->data[dim]; +} + +inline int NumInputs(const TfLiteNode* node) { return node->inputs->size; } +inline int NumOutputs(const TfLiteNode* node) { return node->outputs->size; } + +#ifndef TF_LITE_STATIC_MEMORY +inline int NumIntermediates(const TfLiteNode* node) { + return node->intermediates->size; +} +#endif // TF_LITE_STATIC_MEMORY + +inline int64_t NumElements(const TfLiteIntArray* dims) { + int64_t count = 1; + for (int i = 0; i < dims->size; ++i) { + count *= dims->data[i]; + } + return count; +} + +inline int64_t NumElements(const TfLiteTensor* t) { + return NumElements(t->dims); +} + +// Determines whether tensor is constant. +// TODO(b/138199592): Introduce new query which checks for constant OR +// persistent-read-only, which would be useful for most tensor kernels that +// are potentially dynamic based on the input tensor value availability at the +// time of prepare. +inline bool IsConstantTensor(const TfLiteTensor* tensor) { + return tensor->allocation_type == kTfLiteMmapRo; +} + +// Determines whether tensor is dynamic. Note that a tensor can be non-const and +// not dynamic. This function specifically checks for a dynamic tensor. +inline bool IsDynamicTensor(const TfLiteTensor* tensor) { + return tensor->allocation_type == kTfLiteDynamic; +} + +// Sets tensor to dynamic. +inline void SetTensorToDynamic(TfLiteTensor* tensor) { + if (tensor->allocation_type != kTfLiteDynamic) { + tensor->allocation_type = kTfLiteDynamic; + tensor->data.raw = nullptr; + } +} + +// Sets tensor to persistent and read-only. +inline void SetTensorToPersistentRo(TfLiteTensor* tensor) { + if (tensor->allocation_type != kTfLitePersistentRo) { + tensor->allocation_type = kTfLitePersistentRo; + tensor->data.raw = nullptr; + } +} + +// Determines whether it is a hybrid op - one that has float inputs and +// quantized weights. +inline bool IsHybridOp(const TfLiteTensor* input, const TfLiteTensor* weight) { + return ((weight->type == kTfLiteUInt8 || weight->type == kTfLiteInt8) && + input->type == kTfLiteFloat32); +} + +// Check dimensionality match and populate OpData for Conv and DepthwiseConv. +TfLiteStatus PopulateConvolutionQuantizationParams( + TfLiteContext* context, const TfLiteTensor* input, + const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output, + const TfLiteFusedActivation& activation, int32_t* multiplier, int* shift, + int32_t* output_activation_min, int32_t* output_activation_max, + int32_t* per_channel_multiplier, int* per_channel_shift); + +TfLiteStatus PopulateConvolutionQuantizationParams( + TfLiteContext* context, const TfLiteTensor* input, + const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output, + const TfLiteFusedActivation& activation, int32_t* multiplier, int* shift, + int32_t* output_activation_min, int32_t* output_activation_max, + int32_t* per_channel_multiplier, int* per_channel_shift, int num_channels); + +// Calculates the multiplication factor for a quantized convolution (or +// quantized depthwise convolution) involving the given tensors. Returns an +// error if the scales of the tensors are not compatible. +TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context, + const TfLiteTensor* input, + const TfLiteTensor* filter, + const TfLiteTensor* bias, + TfLiteTensor* output, + double* multiplier); + +TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context, + const TfLiteTensor* input, + const TfLiteTensor* filter, + TfLiteTensor* output, + double* multiplier); + +// Calculates the useful quantized range of an activation layer given its +// activation tensor. +TfLiteStatus CalculateActivationRangeQuantized(TfLiteContext* context, + TfLiteFusedActivation activation, + TfLiteTensor* output, + int32_t* act_min, + int32_t* act_max); + +// Calculates the useful range of an activation layer given its activation +// tensor.a +template +void CalculateActivationRange(TfLiteFusedActivation activation, + T* activation_min, T* activation_max) { + if (activation == kTfLiteActRelu) { + *activation_min = 0; + *activation_max = std::numeric_limits::max(); + } else if (activation == kTfLiteActRelu6) { + *activation_min = 0; + *activation_max = 6; + } else if (activation == kTfLiteActReluN1To1) { + *activation_min = -1; + *activation_max = 1; + } else { + *activation_min = std::numeric_limits::lowest(); + *activation_max = std::numeric_limits::max(); + } +} + +// Return true if the given tensors have the same shape. +bool HaveSameShapes(const TfLiteTensor* input1, const TfLiteTensor* input2); + +// Calculates the output_shape that is necessary for element-wise operations +// with broadcasting involving the two input tensors. +TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, + const TfLiteTensor* input1, + const TfLiteTensor* input2, + TfLiteIntArray** output_shape); + +// Calculates the output_shape that is necessary for element-wise operations +// with broadcasting involving the three input tensors. +TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, + const TfLiteTensor* input1, + const TfLiteTensor* input2, + const TfLiteTensor* input3, + TfLiteIntArray** output_shape); + +// Return the size of given type in bytes. Return 0 in in case of string. +int TfLiteTypeGetSize(TfLiteType type); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_KERNEL_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/kernels/op_macros.h b/src/tensorflow_arm/tensorflow/lite/kernels/op_macros.h new file mode 100644 index 0000000..41ec324 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/kernels/op_macros.h @@ -0,0 +1,86 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_OP_MACROS_H_ +#define TENSORFLOW_LITE_KERNELS_OP_MACROS_H_ + +// If we're on a platform without standard IO functions, fall back to a +// non-portable function. +#ifdef TF_LITE_MCU_DEBUG_LOG + +#include "tensorflow_arm/tensorflow/lite/micro/debug_log.h" + +#define DEBUG_LOG(x) \ + do { \ + DebugLog(x); \ + } while (0) + +inline void InfiniteLoop() { + DEBUG_LOG("HALTED\n"); + while (1) { + } +} + +#define TFLITE_ABORT InfiniteLoop(); + +#else // TF_LITE_MCU_DEBUG_LOG + +#include +#include + +#define DEBUG_LOG(x) \ + do { \ + fprintf(stderr, "%s", (x)); \ + } while (0) + +// Report Error for unsupported type by op 'op_name' and returns kTfLiteError. +#define TF_LITE_UNSUPPORTED_TYPE(context, type, op_name) \ + do { \ + TF_LITE_KERNEL_LOG((context), "%s:%d Type %s is unsupported by op %s.", \ + __FILE__, __LINE__, TfLiteTypeGetName(type), \ + (op_name)); \ + return kTfLiteError; \ + } while (0) + +#define TFLITE_ABORT abort() + +#endif // TF_LITE_MCU_DEBUG_LOG + +#if defined(NDEBUG) || defined(ARDUINO) +#define TFLITE_ASSERT_FALSE (static_cast(0)) +#else +#define TFLITE_ASSERT_FALSE TFLITE_ABORT +#endif + +#define TF_LITE_FATAL(msg) \ + do { \ + DEBUG_LOG(msg); \ + DEBUG_LOG("\nFATAL\n"); \ + TFLITE_ABORT; \ + } while (0) + +#define TF_LITE_ASSERT(x) \ + do { \ + if (!(x)) TF_LITE_FATAL(#x); \ + } while (0) + +#define TF_LITE_ASSERT_EQ(x, y) \ + do { \ + if ((x) != (y)) TF_LITE_FATAL(#x " didn't equal " #y); \ + } while (0) + +#endif // TENSORFLOW_LITE_KERNELS_OP_MACROS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/padding.h b/src/tensorflow_arm/tensorflow/lite/kernels/padding.h similarity index 96% rename from src/tensorflow/lite/kernels/padding.h rename to src/tensorflow_arm/tensorflow/lite/kernels/padding.h index 1116b1d..47c08ca 100644 --- a/src/tensorflow/lite/kernels/padding.h +++ b/src/tensorflow_arm/tensorflow/lite/kernels/padding.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_PADDING_H_ #define TENSORFLOW_LITE_KERNELS_PADDING_H_ -#include "tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" namespace tflite { @@ -78,3 +79,5 @@ inline TfLitePaddingValues ComputePaddingHeightWidth( } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_PADDING_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.cpp b/src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.cpp new file mode 100644 index 0000000..860ed34 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.cpp @@ -0,0 +1,87 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h" + +#include "tensorflow_arm/tensorflow/lite/micro/kernels/micro_ops.h" + +namespace tflite { + +AllOpsResolver::AllOpsResolver() { + // Please keep this list of Builtin Operators in alphabetical order. + AddAbs(); + AddAdd(); + AddArgMax(); + AddArgMin(); + AddAveragePool2D(); + AddCeil(); + AddConcatenation(); + AddConv2D(); + AddCos(); + AddDepthwiseConv2D(); + AddDequantize(); + AddDetectionPostprocess(); + AddEqual(); + AddEthosU(); + AddFloor(); + AddFullyConnected(); + AddGreater(); + AddGreaterEqual(); + AddHardSwish(); + AddL2Normalization(); + AddLess(); + AddLessEqual(); + AddLog(); + AddLogicalAnd(); + AddLogicalNot(); + AddLogicalOr(); + AddLogistic(); + AddMaximum(); + AddMaxPool2D(); + AddMean(); + AddMinimum(); + AddMul(); + AddNeg(); + AddNotEqual(); + AddPack(); + AddPad(); + AddPadV2(); + AddPrelu(); + AddQuantize(); + AddReduceMax(); + AddRelu(); + AddRelu6(); + AddReshape(); + AddResizeNearestNeighbor(); + AddRound(); + AddRsqrt(); + AddShape(); + AddSin(); + AddSoftmax(); + AddSplit(); + AddSplitV(); + AddSqrt(); + AddSquare(); + AddStridedSlice(); + AddSub(); + AddSvdf(); + AddTanh(); + AddUnpack(); +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h b/src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h new file mode 100644 index 0000000..d760b99 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h @@ -0,0 +1,41 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_ALL_OPS_RESOLVER_H_ +#define TENSORFLOW_LITE_MICRO_ALL_OPS_RESOLVER_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_mutable_op_resolver.h" + +namespace tflite { + +// The magic number in the template parameter is the maximum number of ops that +// can be added to AllOpsResolver. It can be increased if needed. And most +// applications that care about the memory footprint will want to directly use +// MicroMutableOpResolver and have an application specific template parameter. +// The examples directory has sample code for this. +class AllOpsResolver : public MicroMutableOpResolver<128> { + public: + AllOpsResolver(); + + private: + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_ALL_OPS_RESOLVER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/arduino/debug_log.cpp b/src/tensorflow_arm/tensorflow/lite/micro/arduino/debug_log.cpp new file mode 100644 index 0000000..61da33f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/arduino/debug_log.cpp @@ -0,0 +1,41 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/debug_log.h" + +#include "Arduino.h" + +// The Arduino DUE uses a different object for the default serial port shown in +// the monitor than most other models, so make sure we pick the right one. See +// https://github.com/arduino/Arduino/issues/3088#issuecomment-406655244 +#if defined(__SAM3X8E__) +#define DEBUG_SERIAL_OBJECT (SerialUSB) +#else +#define DEBUG_SERIAL_OBJECT (Serial) +#endif + +// On Arduino platforms, we set up a serial port and write to it for debug +// logging. +extern "C" void DebugLog(const char* s) { + static bool is_initialized = false; + if (!is_initialized) { + DEBUG_SERIAL_OBJECT.begin(9600); + is_initialized = true; + } + DEBUG_SERIAL_OBJECT.print(s); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.cpp b/src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.cpp new file mode 100644 index 0000000..25145ed --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.cpp @@ -0,0 +1,2848 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.h" + +// Keep model aligned to 8 bytes to guarantee aligned 64-bit accesses. +alignas(8) const unsigned char g_keyword_scrambled_model_data[] = { + 0x1c, 0x00, 0x00, 0x00, 0x54, 0x46, 0x4c, 0x33, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x0e, 0x00, 0x14, 0x00, 0x10, 0x00, 0x0c, 0x00, 0x08, 0x00, + 0x00, 0x00, 0x04, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0xd0, 0x6e, 0x00, 0x00, 0x60, 0x83, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0xbc, 0x6e, 0x00, 0x00, 0xac, 0x56, 0x00, 0x00, + 0x9c, 0x52, 0x00, 0x00, 0x8c, 0x51, 0x00, 0x00, 0x7c, 0x4d, 0x00, 0x00, + 0x2c, 0x4d, 0x00, 0x00, 0x1c, 0x49, 0x00, 0x00, 0x0c, 0x45, 0x00, 0x00, + 0xfc, 0x43, 0x00, 0x00, 0xec, 0x3f, 0x00, 0x00, 0x9c, 0x3f, 0x00, 0x00, + 0x8c, 0x3b, 0x00, 0x00, 0x7c, 0x37, 0x00, 0x00, 0x6c, 0x36, 0x00, 0x00, + 0x5c, 0x32, 0x00, 0x00, 0x0c, 0x32, 0x00, 0x00, 0xfc, 0x2d, 0x00, 0x00, + 0xec, 0x29, 0x00, 0x00, 0xdc, 0x28, 0x00, 0x00, 0xcc, 0x24, 0x00, 0x00, + 0x7c, 0x24, 0x00, 0x00, 0x6c, 0x22, 0x00, 0x00, 0x5c, 0x1a, 0x00, 0x00, + 0xcc, 0x19, 0x00, 0x00, 0xbc, 0x15, 0x00, 0x00, 0xac, 0x0d, 0x00, 0x00, + 0x1c, 0x0d, 0x00, 0x00, 0x0c, 0x09, 0x00, 0x00, 0xfc, 0x00, 0x00, 0x00, + 0x6c, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x2a, 0x91, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, + 0x34, 0xe1, 0x4f, 0xa1, 0x63, 0xa4, 0x62, 0xbf, 0x3e, 0x91, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0xa3, 0xb2, 0x8f, 0xee, + 0x35, 0xe6, 0xf2, 0xcc, 0x68, 0xa0, 0x33, 0xc4, 0x7d, 0x4e, 0xbb, 0xa9, + 0x10, 0x32, 0x8e, 0x3d, 0x76, 0x14, 0x1c, 0x33, 0x0e, 0x77, 0xf7, 0xc8, + 0x7b, 0x45, 0xc7, 0xdb, 0xcf, 0x87, 0xc7, 0x70, 0xa9, 0x29, 0xfd, 0x70, + 0x32, 0x96, 0x35, 0x7d, 0xe9, 0xac, 0x6d, 0x9b, 0xfd, 0xe4, 0xbc, 0x4a, + 0x57, 0xcd, 0x43, 0xcc, 0x73, 0x72, 0xdf, 0x07, 0x68, 0xc5, 0x67, 0xbd, + 0x8a, 0x91, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x80, 0x00, 0x00, 0x00, + 0xb0, 0xfb, 0x5f, 0xdf, 0x0e, 0xb9, 0xa2, 0xfd, 0x66, 0x86, 0x13, 0x1b, + 0x6d, 0x1d, 0x53, 0xdb, 0x83, 0xbf, 0x44, 0x29, 0x3f, 0x93, 0xee, 0x42, + 0x9a, 0xf4, 0x31, 0x6e, 0xc3, 0x15, 0x7e, 0x48, 0x72, 0x50, 0xc3, 0x53, + 0xef, 0x35, 0x1f, 0xc2, 0x29, 0x42, 0xb4, 0xd7, 0x4b, 0xd7, 0x98, 0x60, + 0xb9, 0x3e, 0xbb, 0x31, 0x35, 0xc3, 0xf6, 0x15, 0x7a, 0x9a, 0x2c, 0xfd, + 0xff, 0x04, 0xd9, 0x04, 0x57, 0x52, 0xae, 0x99, 0xa3, 0x95, 0xae, 0x6a, + 0x66, 0x52, 0x5f, 0x91, 0x17, 0x83, 0x0d, 0x27, 0x16, 0x02, 0x06, 0x64, + 0x80, 0x05, 0x99, 0x1c, 0x6c, 0xab, 0xb1, 0xa1, 0x0e, 0x44, 0x1f, 0x63, + 0xe9, 0xc1, 0xab, 0x8d, 0x08, 0x79, 0x56, 0xe0, 0x90, 0xa5, 0xb8, 0x3b, + 0xc4, 0x1e, 0xa5, 0x1f, 0x64, 0xe4, 0x0b, 0x72, 0x62, 0x19, 0x5f, 0x66, + 0xc0, 0x9b, 0x7b, 0xc4, 0xe5, 0x9f, 0x82, 0xa7, 0x16, 0x92, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x3e, 0x3d, 0xf4, 0x61, + 0x45, 0x2a, 0x48, 0x53, 0x1f, 0x22, 0x74, 0x65, 0xea, 0x5a, 0x00, 0x83, + 0x68, 0xf9, 0xbb, 0xa3, 0xc2, 0x1a, 0x8f, 0xe1, 0xfb, 0x76, 0x6a, 0xe9, + 0x1a, 0x0e, 0x4d, 0x32, 0xc6, 0xf3, 0x8d, 0x85, 0x54, 0xa1, 0xe9, 0xb8, + 0x35, 0xee, 0xba, 0x53, 0x40, 0xa2, 0xea, 0x7f, 0xc3, 0x99, 0x71, 0x17, + 0xdd, 0xd5, 0xfe, 0xdf, 0x5e, 0x15, 0xa0, 0x73, 0xf8, 0x78, 0x49, 0x73, + 0xcc, 0xf0, 0x18, 0x12, 0x06, 0x81, 0xd6, 0x19, 0x2c, 0xa8, 0xd7, 0x80, + 0x19, 0x19, 0xbf, 0x1e, 0x50, 0xb1, 0xfb, 0xb3, 0xa6, 0x56, 0x6f, 0x52, + 0xa6, 0xc0, 0xdd, 0x3f, 0xbb, 0x13, 0x6e, 0x04, 0xdf, 0x79, 0xca, 0x8b, + 0xa5, 0x9c, 0xa1, 0x78, 0x49, 0xca, 0xe5, 0x29, 0xbb, 0x29, 0x7c, 0x96, + 0xc6, 0x29, 0x06, 0x99, 0xec, 0x50, 0xd1, 0xe8, 0x9b, 0xb7, 0x53, 0xd2, + 0x36, 0x89, 0xb1, 0x5c, 0x38, 0xf4, 0x2f, 0xa1, 0xda, 0x6f, 0xd8, 0xd1, + 0x62, 0xd2, 0xd4, 0x97, 0xce, 0xf1, 0xbd, 0x73, 0x2d, 0x92, 0xdb, 0x62, + 0x0c, 0xb0, 0x77, 0xed, 0x32, 0x3a, 0xfc, 0x59, 0x94, 0xef, 0x2b, 0x48, + 0x60, 0xb2, 0x82, 0xa2, 0xb6, 0x51, 0xdb, 0x51, 0x47, 0x99, 0x4c, 0x50, + 0x93, 0x53, 0x9d, 0xa9, 0x3c, 0x94, 0x34, 0x9f, 0xa6, 0x3e, 0x4f, 0x87, + 0xd4, 0xa0, 0x40, 0xeb, 0x7b, 0xfa, 0x1b, 0x7d, 0x03, 0xa8, 0xf8, 0x8b, + 0xa5, 0x32, 0x3a, 0xaf, 0x7e, 0x6b, 0x25, 0x08, 0x97, 0x71, 0x8d, 0x0c, + 0x30, 0xc9, 0xa7, 0x23, 0xe3, 0x51, 0xb3, 0xf2, 0x86, 0xad, 0x12, 0xe2, + 0x79, 0x94, 0x7f, 0xf3, 0xf7, 0x88, 0x67, 0x3e, 0x8e, 0x8e, 0x04, 0x5e, + 0x4f, 0x01, 0x6f, 0x1d, 0x78, 0x42, 0x9e, 0x47, 0x81, 0xdf, 0x03, 0x39, + 0x3d, 0x9b, 0xbd, 0xb6, 0x06, 0x21, 0x82, 0xfe, 0xf2, 0x50, 0xe1, 0x14, + 0xbc, 0xe3, 0x5e, 0xe1, 0xbd, 0x8f, 0xfa, 0x35, 0x31, 0x4e, 0x66, 0xeb, + 0x67, 0x49, 0x1c, 0x07, 0x88, 0xb6, 0x22, 0x0c, 0xeb, 0xd9, 0x9f, 0x9b, + 0x8b, 0xe0, 0x9c, 0x3c, 0xf7, 0x91, 0xab, 0x98, 0x5b, 0x0e, 0x09, 0xdd, + 0xe3, 0x0b, 0x14, 0x55, 0xe9, 0xe4, 0x42, 0xd8, 0xce, 0xd7, 0xfd, 0x4c, + 0x20, 0x9f, 0x44, 0x93, 0xa6, 0x17, 0x8a, 0x68, 0x8f, 0xec, 0x62, 0xd1, + 0x97, 0x9c, 0xcc, 0xc4, 0xd9, 0x42, 0xda, 0xf1, 0x34, 0x04, 0xc6, 0xb6, + 0x0f, 0xc7, 0xe6, 0x2d, 0x26, 0x6e, 0x6f, 0x92, 0x7e, 0xd9, 0xd4, 0x40, + 0xc6, 0x70, 0xfa, 0x12, 0x2a, 0x1b, 0xbc, 0x50, 0xeb, 0x3b, 0x24, 0x96, + 0x8d, 0x7c, 0xae, 0xbe, 0xc3, 0x27, 0xce, 0x97, 0xcf, 0xcd, 0x10, 0x13, + 0x01, 0xc6, 0x48, 0x6a, 0x99, 0x38, 0x79, 0xb9, 0x1c, 0xc9, 0x09, 0xac, + 0x96, 0x8c, 0xf7, 0x82, 0x8f, 0xb8, 0x17, 0x94, 0x2c, 0x5f, 0x40, 0xcc, + 0x80, 0xf4, 0x9f, 0xaa, 0xcb, 0x83, 0x13, 0x7b, 0x3a, 0x78, 0x0a, 0x9f, + 0x79, 0x9e, 0xfc, 0x0e, 0x8f, 0x98, 0x60, 0x39, 0x86, 0x44, 0x8e, 0x4b, + 0xc4, 0xad, 0xe6, 0x98, 0x92, 0x08, 0x84, 0x48, 0x8f, 0x1d, 0x78, 0x10, + 0x9e, 0xf7, 0xb8, 0x61, 0x65, 0x46, 0xdb, 0x4a, 0xcf, 0xc5, 0x37, 0xe3, + 0x77, 0x76, 0xcf, 0x0a, 0x7e, 0x72, 0x3f, 0xe4, 0x51, 0x30, 0x28, 0x57, + 0x13, 0xfd, 0xdb, 0x7e, 0xd6, 0xa3, 0xdd, 0x64, 0xdd, 0x00, 0xd0, 0x7f, + 0xbc, 0x48, 0x1d, 0xaf, 0xde, 0x0e, 0x45, 0xc4, 0xc9, 0xfa, 0xf6, 0xb2, + 0xb7, 0x9a, 0x42, 0x8b, 0x18, 0x08, 0xed, 0xdb, 0xa9, 0xc3, 0x32, 0xf1, + 0x9c, 0xcf, 0x16, 0x74, 0x57, 0xce, 0xe9, 0x44, 0x21, 0xdb, 0x8a, 0x45, + 0x89, 0x70, 0x41, 0x5c, 0xbf, 0x10, 0xdf, 0x83, 0x4a, 0xe4, 0x4c, 0xd8, + 0xc9, 0x2e, 0x5b, 0xa3, 0x05, 0xed, 0x73, 0xb1, 0xb0, 0xb7, 0xc4, 0xd7, + 0x0d, 0xea, 0xf6, 0xb4, 0xc1, 0x5e, 0x12, 0x54, 0x30, 0x73, 0x5c, 0x93, + 0xd9, 0xf7, 0xc9, 0x24, 0x43, 0x8f, 0x4f, 0x8e, 0x94, 0x95, 0xb6, 0xfd, + 0xa3, 0x14, 0x42, 0x50, 0xb8, 0x66, 0xfb, 0xc4, 0xed, 0x72, 0xcf, 0x7b, + 0xa9, 0x73, 0xeb, 0xc4, 0x4a, 0x05, 0xea, 0xb4, 0x47, 0xca, 0x21, 0x56, + 0x28, 0xa8, 0x87, 0xb8, 0x87, 0x0b, 0xe3, 0x8d, 0xfd, 0x70, 0xf7, 0x33, + 0x76, 0xf0, 0x3d, 0xa4, 0x3b, 0x83, 0xab, 0x14, 0x01, 0xe1, 0xb0, 0xa9, + 0x44, 0xe8, 0xd7, 0x50, 0x26, 0x0b, 0xbb, 0x2d, 0x57, 0x39, 0x82, 0x7c, + 0x71, 0xd8, 0x12, 0xaf, 0xf3, 0x9f, 0x46, 0xbd, 0x62, 0xd6, 0x61, 0xf5, + 0xb7, 0x04, 0x94, 0xbf, 0x87, 0xea, 0xc4, 0xc4, 0x33, 0xcf, 0x36, 0x3b, + 0x4f, 0xc7, 0x71, 0xf1, 0x98, 0xe6, 0xb0, 0x96, 0x25, 0xd7, 0xac, 0x75, + 0xfc, 0x92, 0xe0, 0x69, 0x72, 0x37, 0x8d, 0x40, 0x31, 0xaa, 0x2c, 0x86, + 0xfb, 0x95, 0x3f, 0x9c, 0x23, 0xd4, 0x39, 0x99, 0xff, 0xea, 0x95, 0x79, + 0xb9, 0x2e, 0xb0, 0x33, 0xf1, 0xe8, 0xd0, 0x42, 0xb5, 0x70, 0x5c, 0xca, + 0x69, 0x48, 0x28, 0x23, 0x58, 0xb4, 0x07, 0xfc, 0x3e, 0x15, 0x29, 0x00, + 0xa9, 0x22, 0x44, 0x70, 0xd0, 0xc7, 0x01, 0x0d, 0x3e, 0xfc, 0x57, 0xb7, + 0x54, 0x3a, 0xc3, 0x43, 0xd6, 0x2f, 0x55, 0x09, 0x52, 0x4a, 0x6b, 0x8e, + 0x4c, 0x82, 0xbb, 0x4e, 0x3e, 0x38, 0xe1, 0x9e, 0x72, 0x83, 0xec, 0x40, + 0xf5, 0xf7, 0x0e, 0x3c, 0x24, 0xed, 0xda, 0xf2, 0x39, 0x6c, 0xad, 0xeb, + 0xff, 0xfb, 0x4a, 0x38, 0x50, 0x49, 0x28, 0x3d, 0x05, 0xb2, 0x98, 0x44, + 0x2b, 0x61, 0xa2, 0x9b, 0x3a, 0x3c, 0xad, 0xd9, 0x8c, 0xef, 0x3c, 0x72, + 0x50, 0x74, 0x13, 0x80, 0xc4, 0x7e, 0x6e, 0xf3, 0xc9, 0xdf, 0x63, 0xf6, + 0x41, 0xb2, 0x08, 0x78, 0x9b, 0x7c, 0xa9, 0x13, 0xd1, 0x21, 0xe7, 0x5e, + 0x6a, 0x0d, 0x64, 0xf7, 0x52, 0x75, 0xf2, 0x80, 0x69, 0xbe, 0x43, 0xf8, + 0xd4, 0xad, 0x49, 0xfc, 0x97, 0x76, 0x1c, 0xb6, 0x43, 0x9e, 0xcb, 0x45, + 0x4d, 0x75, 0x07, 0xae, 0xdb, 0xbf, 0xf5, 0x8a, 0xeb, 0xb9, 0x6b, 0x12, + 0x06, 0xbf, 0x94, 0xad, 0x77, 0x29, 0xb1, 0xae, 0x24, 0x9b, 0x4d, 0xdc, + 0xe1, 0x5e, 0xd7, 0x57, 0xec, 0xd1, 0xd8, 0xad, 0xf0, 0x06, 0x08, 0x43, + 0x33, 0x99, 0xd2, 0x04, 0xfc, 0xc8, 0xf6, 0x53, 0x3d, 0x73, 0xd4, 0x36, + 0xd3, 0x8e, 0x4a, 0xcd, 0xb1, 0xe9, 0xcb, 0x3a, 0x5f, 0x54, 0xbc, 0xde, + 0x16, 0xa2, 0x85, 0xde, 0x35, 0x27, 0x99, 0x32, 0x4f, 0xb9, 0x2c, 0x16, + 0xa2, 0x6e, 0xae, 0x75, 0x60, 0x77, 0xe9, 0x08, 0x0f, 0x08, 0xc4, 0xd0, + 0x62, 0xc7, 0xd2, 0x1f, 0x3b, 0x29, 0xdd, 0xb7, 0xea, 0xa3, 0x58, 0xaf, + 0x4c, 0x05, 0xd2, 0x82, 0x6a, 0xe0, 0xc4, 0xe9, 0x70, 0x7e, 0xf2, 0xca, + 0x82, 0x6a, 0xae, 0xc1, 0x9a, 0x42, 0x5d, 0x46, 0x4a, 0xb7, 0x8f, 0x4d, + 0x33, 0xfe, 0x6f, 0x47, 0xb5, 0x49, 0xb3, 0x89, 0x51, 0x31, 0x74, 0x68, + 0x14, 0xda, 0x0a, 0x41, 0x3d, 0x1f, 0x8e, 0x30, 0x8c, 0x77, 0xd1, 0xa9, + 0x36, 0x41, 0x78, 0x34, 0xb7, 0x7e, 0x4e, 0x7a, 0x77, 0x12, 0x43, 0x97, + 0x43, 0xba, 0xd6, 0x28, 0x14, 0x2a, 0x9f, 0x98, 0xb4, 0x39, 0x08, 0x5c, + 0xb7, 0xb8, 0x03, 0x63, 0x62, 0x68, 0xc6, 0x9a, 0x4d, 0xf5, 0xdc, 0x7c, + 0x0f, 0x7e, 0x77, 0xdc, 0x85, 0x53, 0x31, 0x8c, 0x53, 0x8b, 0x27, 0xc4, + 0xb7, 0x3d, 0xd0, 0x94, 0x9b, 0x7e, 0x59, 0x59, 0x03, 0x09, 0x8c, 0x30, + 0x70, 0x7d, 0x9c, 0x73, 0x89, 0x6c, 0x5f, 0xbf, 0xf9, 0xc7, 0x72, 0x76, + 0x12, 0x98, 0xe3, 0xbe, 0xc3, 0x67, 0xdf, 0xa1, 0x76, 0xa3, 0xec, 0x44, + 0x30, 0x70, 0x2f, 0x6a, 0x86, 0x28, 0xb9, 0x9d, 0x7f, 0x93, 0xf2, 0x4a, + 0x34, 0x48, 0x1f, 0x2e, 0x2e, 0x95, 0x88, 0xdb, 0x1f, 0x2c, 0x19, 0x46, + 0x2e, 0x91, 0x5f, 0x81, 0x0d, 0x08, 0x9d, 0x03, 0x0b, 0xaf, 0x59, 0x0a, + 0x41, 0xad, 0x4d, 0x6c, 0x09, 0x0e, 0x9f, 0xd1, 0xc4, 0xdb, 0xac, 0x59, + 0x27, 0x04, 0x1c, 0x73, 0xe9, 0xf3, 0xe8, 0x54, 0xd9, 0x11, 0x31, 0xb2, + 0xed, 0x2d, 0x8c, 0xeb, 0x99, 0x26, 0x48, 0x9e, 0xac, 0x88, 0x96, 0xcb, + 0x19, 0x49, 0xfa, 0x4a, 0x82, 0xd5, 0x5d, 0xb8, 0x0f, 0x22, 0x3f, 0xb6, + 0x5c, 0x02, 0x2a, 0xb9, 0xd9, 0xfe, 0x4d, 0x9d, 0xdb, 0x85, 0x90, 0x19, + 0x7f, 0x1a, 0x44, 0xa3, 0x74, 0x68, 0xbf, 0xa2, 0x3b, 0xb4, 0x3b, 0xeb, + 0xab, 0x99, 0xc2, 0x46, 0x50, 0x7e, 0xec, 0xa9, 0xb4, 0x86, 0xfa, 0x50, + 0xcb, 0x71, 0x7e, 0x75, 0xa5, 0xca, 0xa6, 0x2f, 0x40, 0x1d, 0xa1, 0x4a, + 0x5c, 0x91, 0xd7, 0x2a, 0xa6, 0x17, 0x11, 0x4d, 0x19, 0x2b, 0xb3, 0x0f, + 0xf0, 0xb3, 0x06, 0x70, 0x51, 0x5c, 0x52, 0x8c, 0xdf, 0xe3, 0x19, 0x92, + 0x08, 0x40, 0xa2, 0xb4, 0xc0, 0xf2, 0xe8, 0x44, 0xcc, 0x36, 0xaa, 0xf9, + 0xf8, 0xfc, 0x2d, 0x83, 0x79, 0xc6, 0x58, 0xc1, 0xdf, 0x32, 0xb7, 0xde, + 0x0f, 0x3e, 0xc0, 0xa8, 0x7e, 0xeb, 0xf2, 0x30, 0x16, 0xdf, 0x38, 0xcb, + 0x69, 0xd9, 0x44, 0x0d, 0x44, 0xf4, 0x45, 0x9c, 0x81, 0xc8, 0xe7, 0x06, + 0xae, 0x95, 0xaf, 0xff, 0x17, 0x3b, 0x1c, 0x3f, 0xda, 0xa5, 0xf8, 0xfd, + 0x9c, 0xf1, 0x0a, 0xca, 0xda, 0xc0, 0xfa, 0x02, 0xc4, 0xce, 0x78, 0xfb, + 0x35, 0x8c, 0xfe, 0x55, 0xad, 0x0d, 0x9b, 0xeb, 0x10, 0xf1, 0x7b, 0xb1, + 0x09, 0xf8, 0xef, 0xfc, 0xde, 0x7a, 0x69, 0x74, 0x76, 0xef, 0x91, 0x64, + 0x33, 0xc4, 0x08, 0x15, 0x73, 0x85, 0x56, 0xae, 0x9c, 0xf6, 0xdd, 0x55, + 0x19, 0x96, 0xe6, 0x41, 0x12, 0xc9, 0x87, 0x91, 0x9e, 0xc6, 0x18, 0xe8, + 0xbf, 0xa0, 0x59, 0xfd, 0x20, 0xab, 0xb5, 0xcf, 0x0f, 0x6e, 0x30, 0xd3, + 0xc5, 0x70, 0xf2, 0x50, 0xa4, 0x2a, 0xdf, 0xb0, 0x45, 0xfc, 0x82, 0x1a, + 0x3b, 0xfe, 0x0c, 0xad, 0x41, 0x95, 0xf1, 0xd6, 0x85, 0xa2, 0xc9, 0xff, + 0xbe, 0x3a, 0x64, 0x70, 0x43, 0xc0, 0xc5, 0xc8, 0x80, 0x11, 0x0d, 0x20, + 0xcd, 0xf2, 0xa2, 0xbb, 0x43, 0x68, 0x0e, 0xf4, 0x01, 0xb3, 0x73, 0x79, + 0x9f, 0x68, 0x41, 0x63, 0x3e, 0xda, 0xf9, 0xf4, 0x23, 0x57, 0x97, 0x84, + 0x99, 0xe8, 0x5e, 0xdb, 0xaa, 0x24, 0xab, 0x9c, 0x40, 0x83, 0xf9, 0x3f, + 0x4f, 0x5a, 0x53, 0xa6, 0xf1, 0xe8, 0x95, 0xcf, 0xcb, 0x50, 0x13, 0x51, + 0xa7, 0x8c, 0x71, 0x1d, 0xff, 0xcc, 0x66, 0xab, 0xff, 0xca, 0xc5, 0xc3, + 0x73, 0x45, 0xb7, 0x21, 0x1d, 0x65, 0x7a, 0xe5, 0x1f, 0x3f, 0x1a, 0x58, + 0x23, 0x28, 0xc8, 0xf3, 0xbf, 0x98, 0x25, 0xc0, 0x83, 0x68, 0xf0, 0x62, + 0x63, 0x90, 0xcf, 0x1f, 0x20, 0xb8, 0x04, 0x5c, 0xc4, 0x80, 0x5b, 0xf4, + 0x6d, 0xdc, 0xe9, 0xac, 0xd8, 0x13, 0x3b, 0x42, 0xf8, 0x4e, 0xa2, 0x1c, + 0xce, 0x3f, 0x8d, 0x15, 0xd3, 0x87, 0x1b, 0x44, 0x79, 0x52, 0x34, 0x4b, + 0x63, 0x4d, 0xbf, 0x95, 0xec, 0xae, 0xf9, 0xc6, 0x7b, 0x7b, 0x85, 0x8c, + 0x4f, 0x20, 0x58, 0x9d, 0x48, 0x03, 0x2f, 0x77, 0x2e, 0x8b, 0x6f, 0x66, + 0x76, 0xb9, 0xb8, 0xb7, 0x34, 0x5a, 0x63, 0x06, 0x85, 0x82, 0x5f, 0x23, + 0x8f, 0x8d, 0x0c, 0x92, 0x3b, 0xd2, 0x8a, 0x1b, 0x39, 0xee, 0x6a, 0xbc, + 0xf6, 0x94, 0x2a, 0xc6, 0x73, 0xa6, 0x99, 0x98, 0xdc, 0x96, 0xd7, 0xc1, + 0xfe, 0x9b, 0xc8, 0xfb, 0x86, 0x5a, 0xad, 0xce, 0xf8, 0xd5, 0x32, 0x62, + 0x96, 0x63, 0xaf, 0x4c, 0x4a, 0xae, 0xec, 0x26, 0x3d, 0x84, 0x69, 0x50, + 0x5f, 0x37, 0x9b, 0x29, 0xac, 0x15, 0x76, 0x3d, 0x33, 0x96, 0x06, 0xde, + 0xc1, 0x6d, 0xa2, 0xc7, 0xc3, 0x8a, 0x20, 0x2e, 0xf7, 0x08, 0x55, 0x83, + 0x23, 0x9c, 0x23, 0x2d, 0x3a, 0xa1, 0x32, 0xbc, 0x47, 0x48, 0xd5, 0x6a, + 0x71, 0xb9, 0xcc, 0x2d, 0x99, 0xa0, 0x37, 0x07, 0x46, 0x45, 0xbe, 0xf0, + 0x27, 0x5a, 0x25, 0x72, 0x58, 0x47, 0x6d, 0xbf, 0x23, 0xdc, 0x48, 0x44, + 0x45, 0x95, 0xb1, 0x62, 0xf1, 0x7e, 0x4c, 0x95, 0x1c, 0xb4, 0x17, 0x8b, + 0x59, 0x2e, 0xf3, 0x4f, 0x45, 0x3b, 0x5d, 0x67, 0x92, 0x52, 0xd8, 0xc1, + 0x91, 0xfa, 0x53, 0xaa, 0x87, 0xc0, 0xa7, 0xb0, 0x9f, 0x10, 0xe8, 0xac, + 0x45, 0x52, 0xbb, 0x17, 0xee, 0xf6, 0x18, 0xbe, 0x02, 0x70, 0xce, 0x79, + 0x66, 0x72, 0xf9, 0xf6, 0xca, 0x66, 0xff, 0xa4, 0x9a, 0xd9, 0xb7, 0x07, + 0xa9, 0xc1, 0x23, 0x7e, 0x7b, 0x9c, 0xe3, 0x02, 0x7a, 0xcc, 0xa3, 0x67, + 0xb7, 0xb0, 0x37, 0xba, 0xae, 0x12, 0xda, 0x48, 0x6e, 0x7f, 0xde, 0x5f, + 0x75, 0x15, 0xca, 0xd2, 0x46, 0xdd, 0xb0, 0x82, 0xbf, 0x6d, 0xe9, 0x51, + 0x66, 0xa5, 0x9e, 0x0c, 0xd5, 0x03, 0xbd, 0x97, 0x0e, 0x1b, 0x88, 0xf6, + 0x61, 0x5a, 0x8b, 0xe0, 0xdd, 0x3e, 0x59, 0x4c, 0x35, 0xfd, 0xb0, 0x3b, + 0x79, 0x8c, 0x1c, 0x96, 0x97, 0x35, 0x62, 0x36, 0x62, 0x4c, 0x4b, 0x46, + 0xb1, 0x21, 0xf7, 0xf0, 0x34, 0xdc, 0xd9, 0x9f, 0xf8, 0x53, 0x7d, 0xca, + 0xbc, 0x4d, 0xaf, 0xf4, 0xb7, 0x2f, 0xa7, 0x5d, 0x18, 0xf9, 0x3b, 0xa9, + 0xb0, 0xbb, 0xdf, 0xfa, 0x28, 0x2b, 0x58, 0xce, 0x46, 0x01, 0x3f, 0x76, + 0xf2, 0x39, 0x45, 0x8b, 0x3c, 0xda, 0x62, 0x2b, 0x6b, 0xe1, 0x5f, 0x14, + 0xfc, 0x79, 0x17, 0x2d, 0xe2, 0xe5, 0x8c, 0xc5, 0xde, 0x91, 0xfd, 0xf5, + 0x6d, 0x9b, 0x6b, 0xbb, 0xb0, 0x13, 0xae, 0xbe, 0x1e, 0xa8, 0x8f, 0x3c, + 0xfd, 0x24, 0xbe, 0xb8, 0x39, 0x80, 0x03, 0x06, 0x8b, 0xff, 0xca, 0x90, + 0x88, 0x0f, 0x45, 0xc4, 0xeb, 0x50, 0x52, 0xf5, 0x00, 0x8c, 0x16, 0x9d, + 0x26, 0xaa, 0xec, 0xb1, 0x44, 0xd6, 0xfe, 0x67, 0xa3, 0xc1, 0xec, 0x4a, + 0x12, 0xa6, 0x7c, 0x7c, 0xc3, 0x46, 0x1c, 0x64, 0x61, 0x67, 0xec, 0xce, + 0x1e, 0xa2, 0xb4, 0xdd, 0x6e, 0x7f, 0x02, 0x14, 0xf4, 0x1c, 0x17, 0xa7, + 0x31, 0x9f, 0xc2, 0xc6, 0xc0, 0x21, 0x41, 0x88, 0x61, 0xd8, 0xca, 0x06, + 0xa5, 0xe4, 0xef, 0xa4, 0xaa, 0x4d, 0xa3, 0xad, 0x5f, 0xd4, 0x0c, 0x6b, + 0x14, 0x38, 0x2e, 0xe8, 0x87, 0x5a, 0x68, 0x10, 0x51, 0xd8, 0xbb, 0xa6, + 0xd9, 0xdc, 0xd3, 0x7f, 0x1f, 0xea, 0xa8, 0xcc, 0x3f, 0x43, 0xa4, 0x04, + 0x95, 0xb4, 0xde, 0x2f, 0x07, 0x5d, 0x91, 0x1c, 0x8e, 0xc3, 0xbc, 0xaa, + 0x46, 0x8a, 0xa8, 0x42, 0xa7, 0x2c, 0x0f, 0x1f, 0xb3, 0xe2, 0x8a, 0x0b, + 0xa0, 0x3f, 0xfb, 0x87, 0x9e, 0x42, 0xa5, 0x60, 0xce, 0x5a, 0x54, 0x91, + 0x26, 0x51, 0xea, 0x81, 0x6f, 0xf1, 0x54, 0x93, 0xe7, 0xa0, 0xf8, 0x64, + 0xab, 0x1d, 0x0d, 0x9d, 0x64, 0x6a, 0xd5, 0x19, 0x03, 0xbb, 0x94, 0x7f, + 0x0a, 0xb8, 0x6b, 0x87, 0xc3, 0x1a, 0x38, 0xe5, 0xe8, 0xba, 0x13, 0x17, + 0xeb, 0x13, 0xcc, 0xac, 0xcb, 0x1f, 0x96, 0x4c, 0x3b, 0x18, 0xfb, 0xe8, + 0x5c, 0x54, 0xce, 0x1a, 0x91, 0x44, 0xf5, 0x49, 0x6c, 0x38, 0x2a, 0x92, + 0x8a, 0x0d, 0x3d, 0x08, 0xc2, 0x5f, 0x6c, 0xac, 0x48, 0xb3, 0xdc, 0x2e, + 0xa6, 0x5a, 0xa8, 0xee, 0x22, 0x9a, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x04, 0x00, 0x00, 0x96, 0xc5, 0x3a, 0x4e, 0x42, 0x7d, 0x27, 0xce, + 0x44, 0x84, 0xf1, 0x67, 0x8c, 0xc5, 0xdd, 0x75, 0x3b, 0x8a, 0xed, 0x2e, + 0x29, 0x62, 0x7b, 0xb0, 0xe6, 0xa3, 0xb4, 0x61, 0x73, 0x10, 0xff, 0x0e, + 0x0c, 0x98, 0x74, 0xef, 0xbb, 0xc4, 0xca, 0x03, 0x88, 0xa4, 0x96, 0x61, + 0xef, 0x36, 0x6d, 0xa2, 0xb1, 0xc8, 0xf0, 0xac, 0xf1, 0xb2, 0x08, 0x56, + 0xc7, 0x99, 0xcf, 0xae, 0x0a, 0x37, 0x85, 0x60, 0x78, 0x2d, 0x14, 0xda, + 0xb1, 0xa7, 0x00, 0xb6, 0x00, 0x04, 0x76, 0x80, 0x0e, 0x9f, 0x2a, 0x30, + 0x8b, 0x85, 0xd9, 0xc1, 0xaf, 0xee, 0x27, 0x80, 0x20, 0xed, 0xef, 0x25, + 0x5c, 0x98, 0x6b, 0xcc, 0xf8, 0x72, 0xfb, 0x3f, 0x13, 0xe6, 0x9b, 0x47, + 0xee, 0xa1, 0x18, 0x55, 0xa0, 0x68, 0xbe, 0xd4, 0x21, 0x59, 0x72, 0xa8, + 0xa4, 0xd2, 0x33, 0x57, 0x50, 0xfc, 0x6b, 0xa8, 0x49, 0x1b, 0x74, 0xdb, + 0x5a, 0x16, 0xb8, 0x52, 0x0c, 0xda, 0xa0, 0xa3, 0xff, 0x33, 0x56, 0x82, + 0x0f, 0x0a, 0x90, 0x82, 0xee, 0xf1, 0x1b, 0xb3, 0x05, 0x44, 0x39, 0x01, + 0xf7, 0x1e, 0xff, 0xcb, 0xea, 0xd0, 0xb6, 0x20, 0xbc, 0x84, 0xb1, 0xf9, + 0xa2, 0xc1, 0x56, 0xe6, 0xfa, 0x47, 0xc9, 0xfd, 0x45, 0x77, 0x51, 0x8e, + 0x01, 0xe4, 0x17, 0x20, 0x6f, 0x99, 0xe3, 0x90, 0x2f, 0xcc, 0xaf, 0xd9, + 0x61, 0x32, 0x91, 0x62, 0x58, 0xf4, 0x98, 0xf5, 0xf4, 0xeb, 0x13, 0xeb, + 0xdc, 0x8a, 0xac, 0xb2, 0x9e, 0xcf, 0xe7, 0xa7, 0xd4, 0x97, 0x22, 0x12, + 0x08, 0x10, 0x6d, 0x40, 0xea, 0x26, 0xea, 0x42, 0x29, 0x6e, 0x75, 0x62, + 0x47, 0x08, 0x17, 0xa8, 0x69, 0x0f, 0xf7, 0x35, 0x59, 0x23, 0x86, 0x83, + 0xfd, 0xb5, 0x61, 0x98, 0x9c, 0x4d, 0x37, 0xda, 0x9f, 0xfc, 0xfb, 0x16, + 0xb7, 0x6c, 0x52, 0xee, 0xa8, 0x9c, 0x3e, 0x93, 0x43, 0xc5, 0x2b, 0xd4, + 0xd0, 0x9f, 0x69, 0x2c, 0xc9, 0x1f, 0x2e, 0xdf, 0x5b, 0xe6, 0xc6, 0x5f, + 0x71, 0xd1, 0xd7, 0xb2, 0x8f, 0x3a, 0xba, 0x60, 0x75, 0x3d, 0x34, 0x41, + 0x43, 0x9b, 0x13, 0xc0, 0x3b, 0x30, 0xc5, 0xe9, 0x84, 0x81, 0xde, 0x85, + 0x4e, 0x65, 0x7b, 0x21, 0x37, 0xb8, 0xef, 0x24, 0x19, 0xaa, 0x26, 0x0c, + 0x27, 0xa7, 0xd9, 0x29, 0x47, 0x1a, 0x15, 0x42, 0x1e, 0x30, 0x79, 0x79, + 0x96, 0x09, 0x62, 0x26, 0xad, 0x98, 0x8b, 0xcb, 0x3d, 0xeb, 0x66, 0x83, + 0x77, 0xd9, 0x79, 0x4d, 0x05, 0x81, 0x72, 0xe9, 0xe0, 0x6f, 0x13, 0x00, + 0x7e, 0xa3, 0x92, 0x82, 0x1c, 0x90, 0x83, 0x4b, 0x15, 0x97, 0x0f, 0x92, + 0xe2, 0xd3, 0x3d, 0xd7, 0x6c, 0xb9, 0x60, 0x9a, 0x23, 0x52, 0xbe, 0x59, + 0xc9, 0x36, 0x9e, 0xf7, 0x77, 0x09, 0x79, 0x01, 0xcc, 0xec, 0x17, 0xd1, + 0x74, 0xbc, 0x58, 0x65, 0x45, 0x3c, 0x86, 0xf1, 0xbc, 0xbd, 0x95, 0x54, + 0x46, 0x45, 0x7b, 0x4c, 0xa2, 0xea, 0x2a, 0x6e, 0xa8, 0xd1, 0x66, 0x03, + 0xb2, 0x6a, 0xe0, 0xd3, 0x07, 0x8d, 0xe0, 0x09, 0x81, 0x42, 0xe3, 0x97, + 0xc4, 0xe7, 0x37, 0xc5, 0x82, 0xcf, 0xb1, 0xec, 0xba, 0xbd, 0xf4, 0xb6, + 0x41, 0xb2, 0xb8, 0xa6, 0x3a, 0x85, 0x4b, 0x4f, 0x46, 0x48, 0xe9, 0x9b, + 0x72, 0xf5, 0xb0, 0x64, 0x66, 0x75, 0x42, 0xb4, 0x00, 0xbe, 0x11, 0x6d, + 0x86, 0x93, 0x07, 0x50, 0xa7, 0xef, 0x55, 0x42, 0xcf, 0xe8, 0x61, 0xd0, + 0x9b, 0x11, 0x84, 0x8c, 0x74, 0xe4, 0xb8, 0x3f, 0x48, 0xb3, 0x61, 0xe3, + 0xea, 0x66, 0x86, 0x94, 0x95, 0x12, 0x77, 0x26, 0x75, 0x30, 0xb5, 0xd3, + 0x7a, 0xad, 0x2d, 0x58, 0x46, 0x1b, 0x4b, 0xd9, 0x2d, 0x1e, 0x0b, 0xff, + 0xd7, 0x03, 0x56, 0x3b, 0xbd, 0x65, 0xb0, 0xf9, 0xfe, 0x43, 0x1c, 0x9c, + 0x18, 0x82, 0x78, 0x5e, 0x06, 0x02, 0x21, 0x70, 0xb2, 0x7f, 0xb5, 0x63, + 0x71, 0x85, 0x95, 0x79, 0xae, 0x1e, 0xc6, 0x62, 0x7a, 0x7c, 0x63, 0x46, + 0x70, 0x1c, 0x58, 0x72, 0x1d, 0xde, 0xca, 0xb4, 0xfc, 0xc8, 0x56, 0x38, + 0x32, 0xf4, 0x0b, 0x56, 0x87, 0x6b, 0x5b, 0x53, 0xd2, 0x2c, 0x35, 0xef, + 0x5b, 0x33, 0x59, 0x13, 0x76, 0x82, 0x30, 0x80, 0x23, 0x10, 0x07, 0x4c, + 0x3f, 0xac, 0x9c, 0x58, 0x2d, 0x04, 0xe6, 0x6a, 0xd3, 0x5c, 0xf9, 0xb6, + 0x59, 0x4e, 0x85, 0xfe, 0x01, 0x71, 0xf0, 0xf7, 0xf2, 0x1f, 0x46, 0xd5, + 0x20, 0x3c, 0x9b, 0xc2, 0x1e, 0x73, 0x1c, 0x56, 0x9c, 0x76, 0x8c, 0x12, + 0x95, 0x51, 0xd4, 0x6f, 0x5b, 0x3a, 0xa7, 0x5f, 0xa7, 0xe4, 0xfa, 0xb7, + 0x1a, 0xdd, 0xb6, 0x4c, 0x01, 0x02, 0xae, 0x9c, 0x02, 0x0d, 0x66, 0x2f, + 0x40, 0x87, 0xa1, 0xbc, 0xf3, 0xde, 0xf4, 0xdb, 0x65, 0xee, 0xcc, 0xca, + 0xe1, 0x7a, 0xa2, 0xf4, 0xf7, 0xf5, 0x7c, 0x2a, 0x3f, 0xa4, 0x67, 0xbb, + 0x07, 0x50, 0x7a, 0x29, 0x8a, 0xcf, 0x2c, 0x7a, 0x0e, 0x0d, 0xc7, 0x95, + 0x8b, 0xf4, 0xe2, 0x50, 0xe1, 0xc1, 0x40, 0x16, 0x99, 0x5c, 0x72, 0xe7, + 0xe4, 0x01, 0xeb, 0x29, 0x6a, 0x99, 0xf2, 0x67, 0x23, 0x46, 0x1f, 0xaa, + 0xea, 0xc1, 0x51, 0x30, 0xeb, 0x7d, 0x34, 0x52, 0x91, 0x37, 0x2d, 0xc6, + 0x5c, 0x3a, 0x7c, 0x54, 0xc0, 0x79, 0xdc, 0xf9, 0xbf, 0x08, 0x2a, 0xf6, + 0xe1, 0x1e, 0xee, 0xc6, 0xd2, 0xe9, 0x30, 0x27, 0x60, 0x0c, 0xa2, 0x63, + 0x16, 0x06, 0x3d, 0xe2, 0xf5, 0x6f, 0xea, 0xe4, 0x4d, 0x9f, 0x2d, 0x36, + 0x62, 0x95, 0x47, 0x5d, 0x00, 0x22, 0x9f, 0x0c, 0xbb, 0x71, 0xad, 0xea, + 0xe7, 0x62, 0x59, 0x21, 0xd1, 0xaf, 0x04, 0x5a, 0xfc, 0x1f, 0x28, 0x6b, + 0x6f, 0x71, 0xec, 0xd4, 0xbd, 0x9c, 0x88, 0xfb, 0x3f, 0x04, 0xea, 0xd6, + 0xb2, 0x24, 0xe5, 0x28, 0xfe, 0xc5, 0x3e, 0x15, 0x00, 0x8c, 0xa2, 0xdf, + 0x18, 0x3d, 0x10, 0x9a, 0xb1, 0xcd, 0x64, 0xda, 0x87, 0x41, 0xc8, 0xa1, + 0x1c, 0x97, 0xd5, 0x44, 0xd9, 0x51, 0xd2, 0x96, 0xed, 0xad, 0x28, 0x1f, + 0x03, 0x89, 0x21, 0xbd, 0x79, 0x91, 0x48, 0x9c, 0x8e, 0x17, 0xfd, 0x36, + 0x72, 0xf6, 0x69, 0x4f, 0x3f, 0x02, 0x57, 0xcc, 0x3f, 0x1c, 0x49, 0x82, + 0x00, 0x45, 0x9e, 0x29, 0x83, 0x14, 0x12, 0xbb, 0xd2, 0xd0, 0x1a, 0x66, + 0x0f, 0x57, 0x24, 0xd4, 0x9f, 0x46, 0x0c, 0xf4, 0xb8, 0x28, 0x85, 0x52, + 0xe2, 0xa1, 0xc2, 0x3a, 0x8c, 0x34, 0x4a, 0x81, 0xe3, 0xbc, 0xa2, 0x67, + 0x67, 0x12, 0x13, 0xc4, 0xe7, 0xd7, 0x2c, 0x4e, 0xa9, 0xf5, 0xed, 0x63, + 0xf2, 0x18, 0x9c, 0x0c, 0xe2, 0x4d, 0x25, 0x23, 0x30, 0x3e, 0x49, 0x29, + 0xa6, 0x37, 0xdf, 0xc2, 0xdc, 0xf6, 0x5e, 0xae, 0x45, 0xd7, 0x8d, 0x56, + 0xba, 0x29, 0x4f, 0xee, 0xc9, 0x26, 0xd7, 0xbf, 0x10, 0x4d, 0x0a, 0x3b, + 0x3d, 0x1f, 0xd5, 0x72, 0xe1, 0xe6, 0xf5, 0x23, 0x4a, 0x17, 0x2d, 0xe4, + 0x40, 0x55, 0x9b, 0x39, 0x66, 0x36, 0xe4, 0x6d, 0x6d, 0xb6, 0x8d, 0x2a, + 0x7e, 0x76, 0x73, 0xa5, 0x86, 0x20, 0x3d, 0x18, 0xa0, 0x6c, 0x35, 0x59, + 0xc8, 0x1c, 0xef, 0x0f, 0x36, 0x1d, 0x6f, 0xba, 0x89, 0xb9, 0x9e, 0x7a, + 0x58, 0x1d, 0x43, 0xad, 0x85, 0x8b, 0x6b, 0xcc, 0x25, 0xb8, 0xe4, 0xdd, + 0xa1, 0x35, 0xd9, 0xef, 0xc4, 0xb1, 0xf6, 0x99, 0x27, 0x17, 0xb7, 0xbe, + 0xd1, 0x4f, 0xa1, 0x81, 0x4e, 0xb6, 0x19, 0xcd, 0xa0, 0x92, 0xeb, 0x56, + 0x41, 0x4f, 0x37, 0xca, 0x3b, 0x43, 0x85, 0x86, 0xdf, 0x5d, 0x5a, 0x8c, + 0xd4, 0x5b, 0xc4, 0x28, 0xdb, 0x16, 0xea, 0x3a, 0x2e, 0x9e, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x80, 0x00, 0x00, 0x00, 0xea, 0x59, 0x40, 0xc4, + 0x40, 0x8b, 0x6a, 0x8a, 0xb8, 0x7f, 0x1e, 0x0b, 0xfe, 0xab, 0xa4, 0xac, + 0x42, 0x91, 0xc5, 0xfa, 0x2c, 0x7e, 0xb4, 0xf9, 0x5c, 0xd5, 0x4c, 0x6a, + 0x74, 0x82, 0x90, 0x81, 0x96, 0xb0, 0xf4, 0xd4, 0xba, 0xc9, 0xa3, 0x2e, + 0x26, 0x0a, 0xc9, 0x55, 0x65, 0xac, 0xde, 0x83, 0x37, 0xec, 0x0e, 0xf6, + 0xdc, 0x8c, 0x34, 0xe6, 0x57, 0xde, 0x32, 0x0a, 0x02, 0x62, 0x4f, 0x6a, + 0x92, 0xa5, 0xb4, 0x40, 0xde, 0x57, 0xf4, 0xd1, 0xa3, 0x1c, 0xd3, 0xf7, + 0x4a, 0x15, 0xcc, 0x27, 0x26, 0x00, 0xba, 0xf3, 0xfa, 0x4e, 0xc6, 0xe9, + 0xc3, 0x05, 0x3d, 0x3a, 0x89, 0x96, 0x7d, 0x41, 0xac, 0xca, 0x28, 0x7f, + 0x69, 0x02, 0x40, 0x03, 0x93, 0x86, 0x85, 0x85, 0x73, 0x00, 0x09, 0x5a, + 0xcf, 0x5f, 0x1d, 0xaa, 0x46, 0x41, 0x9d, 0x08, 0xbf, 0xea, 0x45, 0x9b, + 0x93, 0xda, 0x9e, 0x81, 0xba, 0x9e, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x08, 0x00, 0x00, 0x6a, 0x1f, 0x9b, 0x03, 0xdd, 0xe4, 0x16, 0x07, + 0x7f, 0x5b, 0xb0, 0xee, 0xac, 0x55, 0xc4, 0x50, 0xe6, 0x2b, 0x17, 0xed, + 0x7f, 0x50, 0x4d, 0x71, 0x73, 0xae, 0xe0, 0x4d, 0xce, 0x08, 0xd9, 0x8b, + 0x83, 0x2c, 0x01, 0x48, 0x02, 0xd3, 0xbb, 0xca, 0x86, 0xd7, 0xca, 0x5f, + 0xc7, 0xce, 0x59, 0xdf, 0xc1, 0xcc, 0xf7, 0x7b, 0x54, 0xf8, 0x0d, 0x4f, + 0x81, 0x9e, 0x50, 0x6a, 0x65, 0x66, 0x4a, 0xec, 0x7a, 0x1b, 0x92, 0xb2, + 0x39, 0x8f, 0x5d, 0x41, 0x33, 0xcf, 0xe6, 0x1b, 0x34, 0x5d, 0xe1, 0xf6, + 0xef, 0xcb, 0xa0, 0x55, 0x7e, 0x1f, 0x45, 0x38, 0xb9, 0x56, 0x15, 0x3b, + 0x70, 0xab, 0xc8, 0x2f, 0x1c, 0xb9, 0x7d, 0x37, 0xe1, 0xb4, 0x03, 0x44, + 0x5a, 0xf6, 0x57, 0x97, 0x03, 0x54, 0x4c, 0x22, 0x88, 0xc3, 0x82, 0xfd, + 0x91, 0xc1, 0xf1, 0x63, 0xb4, 0x50, 0x46, 0x11, 0x64, 0x07, 0xfd, 0x85, + 0xe5, 0x78, 0x57, 0xdd, 0x19, 0x2a, 0x6b, 0x64, 0x3e, 0xec, 0xb8, 0xf3, + 0xb5, 0x95, 0x29, 0x72, 0xf1, 0x9d, 0xdd, 0xb9, 0xad, 0xd0, 0x78, 0x26, + 0x86, 0x10, 0x10, 0x19, 0xe4, 0x79, 0xae, 0xdc, 0x56, 0xb7, 0x54, 0x4f, + 0x94, 0xc6, 0x26, 0x9a, 0x93, 0xa8, 0x2e, 0x1b, 0x1c, 0xda, 0x87, 0x3a, + 0xa2, 0x44, 0xb9, 0x0b, 0x0f, 0xab, 0x70, 0x3b, 0xb7, 0x6c, 0xbf, 0x58, + 0x67, 0x32, 0x7d, 0xa3, 0x2a, 0xcb, 0x4e, 0x02, 0x92, 0xa1, 0x26, 0x0e, + 0x20, 0x5e, 0xb3, 0xec, 0xc4, 0x04, 0x5b, 0x7f, 0xe5, 0xbd, 0x30, 0xeb, + 0xc8, 0xdd, 0xf1, 0x72, 0x5a, 0x7e, 0xcb, 0x93, 0x22, 0xa0, 0x01, 0x9f, + 0xbb, 0x24, 0x9f, 0x50, 0x01, 0x1f, 0x24, 0x02, 0x85, 0x6d, 0xe6, 0x4d, + 0x55, 0xc4, 0x07, 0xe9, 0x87, 0x38, 0xbf, 0x1a, 0x3b, 0x05, 0x82, 0xc4, + 0x73, 0x4b, 0x87, 0x3c, 0xb4, 0x0a, 0x48, 0x8c, 0x06, 0x67, 0xe7, 0xbf, + 0xcc, 0xe7, 0xe5, 0xc3, 0xb2, 0x81, 0x60, 0xe2, 0xd1, 0xb1, 0x8f, 0x98, + 0xbd, 0x7d, 0xbd, 0x4e, 0x9a, 0xca, 0xbe, 0xcb, 0x81, 0x47, 0x25, 0xaa, + 0xfa, 0x91, 0xcf, 0x78, 0xce, 0xcb, 0x1a, 0x11, 0x79, 0xcf, 0x97, 0xa3, + 0x95, 0x95, 0x6f, 0xd7, 0xae, 0x80, 0xc9, 0xd5, 0x95, 0xb7, 0xcf, 0xe2, + 0x9d, 0x98, 0x65, 0x80, 0xfd, 0x2e, 0xee, 0x46, 0x5e, 0x46, 0x8c, 0xde, + 0x52, 0xb4, 0xdc, 0xce, 0xa8, 0xab, 0x4e, 0x0c, 0x12, 0x9f, 0x89, 0x9c, + 0x84, 0x80, 0xfe, 0x08, 0x64, 0x12, 0x12, 0x95, 0x62, 0xea, 0x65, 0xcc, + 0x34, 0x80, 0xcf, 0x92, 0x5f, 0xc2, 0xae, 0x76, 0xe7, 0x2f, 0xbb, 0xa8, + 0xdb, 0x6a, 0x66, 0x60, 0xaf, 0x88, 0xba, 0x65, 0x32, 0xcf, 0xf7, 0x6e, + 0xd8, 0xd0, 0x69, 0xb0, 0x12, 0x23, 0xd6, 0xc2, 0x32, 0xe5, 0x8e, 0x51, + 0xc5, 0x61, 0x28, 0x45, 0xf7, 0xf9, 0xea, 0x73, 0xce, 0x04, 0x2d, 0x56, + 0x43, 0x10, 0x8b, 0x4f, 0x6b, 0xfa, 0x32, 0xa8, 0x92, 0x8f, 0xd9, 0xb4, + 0xfd, 0xa4, 0x74, 0xa8, 0xea, 0xca, 0xd3, 0x84, 0xbb, 0x5a, 0x34, 0x57, + 0xf9, 0xda, 0x25, 0x40, 0x1f, 0x5e, 0xc2, 0x66, 0x43, 0x05, 0xdd, 0x13, + 0x88, 0x91, 0x60, 0xa1, 0x75, 0xd3, 0xc4, 0x27, 0xff, 0xda, 0x24, 0x3d, + 0xd9, 0xd7, 0x47, 0x46, 0x30, 0xd0, 0x76, 0xc4, 0x9e, 0x97, 0xe3, 0x43, + 0xd7, 0x45, 0xaf, 0x49, 0x36, 0xf2, 0x18, 0xdd, 0x3f, 0x86, 0x9a, 0xec, + 0x9a, 0x70, 0xeb, 0x5a, 0xe2, 0xa0, 0x4b, 0x45, 0x21, 0xb3, 0x32, 0x3d, + 0x0c, 0x8c, 0x03, 0x13, 0xae, 0x46, 0xb5, 0x1a, 0x0a, 0x03, 0x36, 0xfe, + 0xfe, 0xfa, 0xc9, 0x4d, 0x46, 0xf8, 0xfe, 0x6f, 0x99, 0x8c, 0xe4, 0x77, + 0x0c, 0x27, 0x59, 0xf7, 0xc3, 0xfc, 0x32, 0xb3, 0xa5, 0xae, 0xdc, 0x49, + 0xac, 0x31, 0x27, 0xa6, 0x14, 0x92, 0xfb, 0xe3, 0x69, 0x35, 0x8d, 0xa0, + 0x50, 0x55, 0x09, 0x90, 0xdf, 0x67, 0x08, 0x4c, 0x0e, 0xaf, 0x71, 0xc2, + 0xe8, 0xb8, 0xdc, 0x45, 0xe3, 0x6d, 0x58, 0x3f, 0x19, 0x8d, 0xcd, 0xeb, + 0xe3, 0x02, 0x49, 0xd8, 0xc8, 0x8b, 0x29, 0xb3, 0xef, 0x2b, 0xf0, 0x39, + 0x5c, 0x11, 0xaa, 0x52, 0x44, 0x0d, 0x1a, 0x3a, 0x7a, 0x62, 0xda, 0x6d, + 0xe3, 0xdd, 0x03, 0x30, 0x6d, 0x3e, 0x18, 0x30, 0x1d, 0xc0, 0xd0, 0x05, + 0x67, 0x98, 0xf5, 0x2a, 0xc7, 0xa1, 0x58, 0xd7, 0xf8, 0x6f, 0x7d, 0x07, + 0x59, 0x27, 0x95, 0xb9, 0x8d, 0x4d, 0xd7, 0xc8, 0x5e, 0x8b, 0x89, 0x14, + 0xb7, 0x1b, 0x35, 0xaa, 0x72, 0x02, 0x39, 0x3c, 0x41, 0x7c, 0x91, 0x93, + 0x81, 0xe1, 0xad, 0xbe, 0x77, 0x28, 0x80, 0xa2, 0x9c, 0xa8, 0x00, 0x18, + 0xa5, 0x70, 0xec, 0xec, 0x96, 0x95, 0x37, 0xa3, 0xee, 0x15, 0xa0, 0x69, + 0x0e, 0x05, 0xb5, 0xb4, 0xb6, 0xa7, 0x8b, 0xb9, 0x41, 0x88, 0x4f, 0x56, + 0x39, 0xa7, 0xbe, 0x24, 0xce, 0x4c, 0xe0, 0x9c, 0x24, 0x5a, 0xa1, 0xab, + 0xcd, 0x82, 0xf1, 0x16, 0x3f, 0xc0, 0xaf, 0xe1, 0x42, 0xe0, 0x7d, 0x1b, + 0xd9, 0x8f, 0xb8, 0x04, 0xa1, 0x88, 0xd9, 0xc3, 0xaf, 0x4f, 0xda, 0xfd, + 0x0b, 0x5c, 0xc3, 0x04, 0xf3, 0xdb, 0xe6, 0x76, 0x6e, 0xe9, 0xdc, 0xea, + 0x6f, 0xa2, 0xa5, 0x75, 0x2c, 0xc7, 0x91, 0x7d, 0x4b, 0xd5, 0x68, 0x55, + 0xbb, 0x2d, 0x14, 0xdb, 0x06, 0x76, 0xf7, 0xcc, 0x0a, 0x88, 0x6c, 0x2b, + 0xa1, 0x57, 0xd6, 0x15, 0x9c, 0x46, 0xcf, 0x5b, 0x6f, 0x9e, 0x7e, 0xc5, + 0x39, 0xda, 0x97, 0x26, 0x5e, 0xf5, 0x25, 0x06, 0xed, 0x8e, 0x9b, 0x1d, + 0x1b, 0x91, 0x07, 0x89, 0x08, 0xce, 0xd7, 0x38, 0x43, 0x64, 0x8e, 0xf5, + 0x3a, 0x52, 0x4a, 0xfb, 0x3e, 0xff, 0x2c, 0xb3, 0x78, 0x40, 0xb5, 0xdd, + 0xb2, 0x8a, 0xd3, 0x6a, 0xc5, 0xb0, 0xa3, 0x4a, 0xb8, 0xe7, 0x27, 0xa0, + 0x5a, 0x8f, 0x0f, 0xda, 0x53, 0x49, 0xc9, 0x77, 0x2a, 0xef, 0x78, 0xc6, + 0xec, 0xaf, 0x10, 0xe5, 0x71, 0xc5, 0x7a, 0x85, 0xdf, 0xb2, 0x85, 0x02, + 0xe3, 0x55, 0x7a, 0x91, 0x3a, 0x68, 0xb2, 0x9d, 0x3d, 0xd9, 0x01, 0xc5, + 0x5f, 0x3c, 0xa8, 0x1d, 0x99, 0xc6, 0xe7, 0xad, 0x09, 0xd1, 0x39, 0x3a, + 0x92, 0xc5, 0x77, 0x9c, 0xdf, 0x99, 0x56, 0x9f, 0xfe, 0xf8, 0xfd, 0xc8, + 0x4f, 0x19, 0xa3, 0xa0, 0xdf, 0xff, 0x17, 0xac, 0xa9, 0x03, 0x32, 0x85, + 0x4c, 0x29, 0xca, 0x89, 0x58, 0xdc, 0x88, 0xdd, 0xeb, 0x79, 0x68, 0x5e, + 0x0f, 0x37, 0x1a, 0xf7, 0x05, 0xfd, 0x39, 0x91, 0x25, 0x61, 0xf3, 0x04, + 0xda, 0x97, 0xfc, 0x7b, 0xcc, 0x40, 0x63, 0xfd, 0x5b, 0x3b, 0x27, 0x8e, + 0x92, 0x6d, 0x98, 0x0f, 0xcc, 0x9c, 0x9b, 0xda, 0xb2, 0xc6, 0xca, 0x56, + 0xff, 0x7e, 0xcc, 0xa2, 0xc0, 0x45, 0x3e, 0xf6, 0xdf, 0xa7, 0xe8, 0x2a, + 0xef, 0x0c, 0xde, 0xec, 0xa4, 0x1d, 0x2c, 0x3e, 0x03, 0xfd, 0xa4, 0x44, + 0x60, 0x4a, 0xf5, 0x83, 0x8f, 0x09, 0x2d, 0xe8, 0xd5, 0x46, 0xf6, 0x1c, + 0x2d, 0x39, 0x28, 0x0c, 0xdf, 0xa1, 0x2b, 0x05, 0x6e, 0x3c, 0x36, 0xdd, + 0x91, 0x81, 0x52, 0xf1, 0x56, 0xdc, 0xbb, 0x79, 0x62, 0xd8, 0x2e, 0x27, + 0x5d, 0x9f, 0x3c, 0xce, 0x81, 0x5c, 0x70, 0xe5, 0x4d, 0x33, 0x06, 0xd5, + 0x14, 0x04, 0xb7, 0xbc, 0x7b, 0x7a, 0xb4, 0xf7, 0x4a, 0x48, 0x8f, 0x97, + 0x85, 0x96, 0x69, 0xc9, 0x40, 0x52, 0xb1, 0x1c, 0x28, 0x82, 0xb3, 0x63, + 0xee, 0x94, 0x2f, 0xcb, 0x40, 0xad, 0xd7, 0x78, 0xb1, 0xc4, 0x21, 0x05, + 0x36, 0xd9, 0x46, 0xf0, 0x83, 0xcd, 0xee, 0x52, 0x7a, 0xa6, 0xa4, 0x40, + 0xb0, 0x2f, 0xf0, 0x1c, 0xfa, 0x42, 0x98, 0x54, 0x5b, 0xfe, 0x5e, 0xd6, + 0x84, 0x73, 0xca, 0x39, 0xbe, 0x87, 0xf2, 0x92, 0xee, 0x3d, 0x21, 0xcc, + 0x69, 0x81, 0xe5, 0xe8, 0x8a, 0xc3, 0x23, 0x64, 0x98, 0xd5, 0x1d, 0xcd, + 0x5c, 0x6c, 0x37, 0xc8, 0x8b, 0x08, 0x22, 0x12, 0x9f, 0x85, 0xc9, 0xed, + 0xb4, 0xa6, 0x07, 0xe1, 0x62, 0x79, 0x35, 0x5d, 0x26, 0x11, 0x4a, 0x6b, + 0x33, 0x37, 0x91, 0x78, 0xe8, 0xe2, 0xba, 0x8b, 0x8a, 0xb7, 0xbb, 0x0f, + 0xd2, 0xb3, 0xa2, 0x02, 0x0c, 0x57, 0x35, 0x99, 0x88, 0x6b, 0x9b, 0x64, + 0x79, 0x1f, 0x4a, 0x48, 0xd4, 0x3b, 0x5c, 0xeb, 0xb4, 0x83, 0xc3, 0xad, + 0x9c, 0x6a, 0xb0, 0xcf, 0x7f, 0x70, 0xe8, 0x22, 0x46, 0x25, 0xfe, 0x7e, + 0x02, 0x44, 0x83, 0x02, 0xb3, 0x08, 0x2e, 0x34, 0x08, 0x4b, 0xff, 0xa2, + 0xc1, 0x60, 0xbb, 0xd8, 0x89, 0x16, 0xf8, 0xaa, 0xab, 0xea, 0xf7, 0xa0, + 0x10, 0x9a, 0xc9, 0xe9, 0xa4, 0x81, 0xa7, 0x87, 0x32, 0x5b, 0xc1, 0xd0, + 0xd9, 0x70, 0x6f, 0xb6, 0x7c, 0x65, 0xd5, 0x0e, 0x65, 0x93, 0xfe, 0x6d, + 0x66, 0xaa, 0xab, 0xd0, 0x03, 0x07, 0xf2, 0xbe, 0x39, 0xd6, 0xc8, 0xac, + 0xf2, 0x06, 0x58, 0x58, 0x46, 0xc0, 0x1a, 0xbd, 0xa4, 0x96, 0x38, 0x31, + 0x32, 0x89, 0x04, 0xdf, 0xcd, 0x3c, 0x2e, 0x98, 0xb8, 0x39, 0xba, 0xe2, + 0xca, 0x6b, 0xd0, 0x53, 0xce, 0x4a, 0xc8, 0x95, 0x81, 0x84, 0x17, 0xce, + 0x7f, 0x1d, 0xc1, 0x5a, 0xc4, 0xc2, 0x73, 0x30, 0x6d, 0x0b, 0x8c, 0xf8, + 0x66, 0x38, 0x4e, 0xa3, 0x14, 0x84, 0x15, 0x36, 0x9e, 0x0d, 0x56, 0x6b, + 0xa6, 0x77, 0x65, 0xa4, 0x2c, 0x77, 0x00, 0x8b, 0x43, 0x57, 0xc6, 0x25, + 0xc5, 0xd0, 0x17, 0x79, 0x6b, 0x5d, 0xbc, 0xcd, 0xc8, 0x25, 0x8f, 0x20, + 0x09, 0xcc, 0xbd, 0x80, 0x10, 0xdf, 0x35, 0xf6, 0x9c, 0x04, 0x80, 0x23, + 0xdc, 0x97, 0xe0, 0xba, 0x29, 0x48, 0x2e, 0x95, 0x0f, 0xb1, 0x9b, 0xc7, + 0xe6, 0x0b, 0x89, 0x16, 0xe2, 0x81, 0x3b, 0x32, 0x69, 0xc4, 0xde, 0xc6, + 0x12, 0x09, 0x47, 0xff, 0x50, 0xe4, 0x45, 0xb7, 0x35, 0xd2, 0x61, 0x9b, + 0x52, 0x6e, 0xbe, 0xaf, 0xd2, 0xeb, 0x0c, 0x50, 0xf1, 0x57, 0x9f, 0x59, + 0xe1, 0xc1, 0x4f, 0x8c, 0x79, 0x07, 0x05, 0xce, 0x8d, 0x64, 0xb2, 0xf0, + 0xd3, 0x4f, 0xe1, 0x7b, 0xfa, 0x30, 0x0a, 0xc2, 0x5d, 0x0c, 0x47, 0x6c, + 0x17, 0x77, 0x1f, 0xe5, 0xd8, 0x14, 0xfd, 0xc1, 0x01, 0x70, 0x51, 0x60, + 0xb2, 0x20, 0xfd, 0x86, 0xbc, 0x19, 0x5e, 0x01, 0xa6, 0x19, 0x3a, 0x21, + 0xa5, 0x0a, 0x1c, 0xd9, 0xa9, 0x78, 0xbb, 0xc9, 0x01, 0x65, 0xe4, 0xb3, + 0x48, 0xb8, 0xe1, 0xe7, 0xb5, 0xf4, 0x4e, 0xa9, 0xb6, 0xe2, 0x5b, 0xeb, + 0xf5, 0x76, 0x06, 0x1a, 0xd9, 0x08, 0x40, 0xff, 0x72, 0xb2, 0xe3, 0x01, + 0x50, 0xb1, 0xad, 0xb3, 0xa3, 0xf6, 0xef, 0x72, 0x05, 0x0c, 0xf4, 0xce, + 0x24, 0x2c, 0x63, 0x89, 0x63, 0x9e, 0x21, 0xb8, 0xb0, 0xbe, 0xc7, 0x45, + 0xae, 0x47, 0x2b, 0x9e, 0x61, 0x81, 0x4c, 0x76, 0x96, 0x7b, 0x18, 0x37, + 0x74, 0xcb, 0x00, 0xef, 0x38, 0x72, 0x24, 0x0a, 0x63, 0xc1, 0x64, 0xd6, + 0x41, 0xc8, 0x6a, 0xf1, 0xe7, 0x11, 0x20, 0x4b, 0xc2, 0x95, 0x70, 0xb8, + 0xf8, 0x8f, 0xd9, 0xae, 0x8c, 0x12, 0xd8, 0x6f, 0x63, 0x30, 0xca, 0x56, + 0x46, 0x11, 0xda, 0x49, 0x1f, 0x84, 0x3d, 0xae, 0xab, 0x78, 0x29, 0x02, + 0x6c, 0x43, 0xa3, 0xef, 0x9d, 0x97, 0x59, 0x15, 0x53, 0xcd, 0xc7, 0x47, + 0x65, 0x30, 0xc7, 0xae, 0x31, 0x4a, 0x41, 0xb4, 0x66, 0x9c, 0xbb, 0x51, + 0x0b, 0xbd, 0xe2, 0x7d, 0x41, 0x2c, 0xd0, 0x75, 0x57, 0x93, 0xce, 0x2e, + 0xeb, 0x31, 0x7f, 0x56, 0xb2, 0xa4, 0x2b, 0x9f, 0xcc, 0xef, 0x6f, 0xf0, + 0x77, 0x19, 0xad, 0x4d, 0x2e, 0x37, 0x00, 0x75, 0x53, 0xae, 0x22, 0x44, + 0x69, 0x1c, 0x8a, 0x90, 0xf2, 0xcd, 0x0f, 0x6b, 0x37, 0xdb, 0xfd, 0x71, + 0x64, 0x80, 0xd8, 0x57, 0x1b, 0x8f, 0xff, 0x14, 0xd4, 0x5f, 0xe1, 0xd1, + 0x0f, 0x06, 0x13, 0x61, 0x29, 0xa9, 0x80, 0x9d, 0xc7, 0x8a, 0xa0, 0xb5, + 0xaa, 0xfc, 0xe0, 0xb4, 0xb4, 0xf0, 0x31, 0xf0, 0xec, 0x78, 0x03, 0x28, + 0xb9, 0xf7, 0xd9, 0xa7, 0xc8, 0xad, 0x2e, 0x16, 0xb8, 0x18, 0x82, 0x43, + 0x66, 0x8b, 0xae, 0xb2, 0x45, 0x2b, 0x0c, 0x9d, 0x69, 0xbd, 0x1b, 0xc5, + 0x20, 0xc6, 0x41, 0xe7, 0x4f, 0x4b, 0x7b, 0x46, 0x3d, 0x7a, 0x6d, 0x9f, + 0x13, 0x2e, 0x0f, 0xf3, 0x85, 0x3e, 0x5b, 0x12, 0xe5, 0xbf, 0x1b, 0x20, + 0xc3, 0x5f, 0x6b, 0xf7, 0xf7, 0xa3, 0xd7, 0x33, 0xd2, 0xcb, 0x18, 0xa5, + 0xa4, 0xa2, 0xd3, 0x59, 0x91, 0x9a, 0x04, 0xfa, 0x9d, 0xa5, 0x55, 0xad, + 0x09, 0x5a, 0x1e, 0x0b, 0x10, 0xd0, 0x46, 0x18, 0xe4, 0x09, 0xe8, 0x1b, + 0x44, 0xd3, 0x78, 0x45, 0xc0, 0xdf, 0xa2, 0xef, 0xfc, 0x59, 0x8a, 0x1b, + 0x22, 0x60, 0xc9, 0x58, 0x7d, 0x65, 0x45, 0xa9, 0xac, 0xd5, 0xd4, 0xc4, + 0x44, 0xd3, 0x08, 0x44, 0x40, 0x4d, 0x3d, 0x7e, 0x39, 0x81, 0x72, 0x15, + 0x49, 0xd7, 0x2c, 0xda, 0x33, 0xaf, 0xc5, 0xb5, 0x8a, 0x3c, 0xbf, 0x81, + 0x88, 0x4f, 0x12, 0xe4, 0xe8, 0xe6, 0x00, 0xb6, 0xd9, 0xcd, 0xb2, 0x70, + 0x08, 0x15, 0x72, 0xf6, 0x46, 0xc7, 0x98, 0x7c, 0x1d, 0x54, 0xd0, 0x66, + 0x2d, 0xa1, 0xd8, 0xda, 0xb0, 0xe5, 0x9f, 0xa3, 0x2f, 0x2c, 0xfb, 0x34, + 0xb3, 0x21, 0x8b, 0x61, 0xf4, 0xce, 0x60, 0x2b, 0xb5, 0x5e, 0x3d, 0x14, + 0x2c, 0xbe, 0x19, 0x9d, 0x5f, 0x01, 0xe1, 0x21, 0x34, 0x11, 0x6b, 0x10, + 0xd4, 0x17, 0x58, 0xb3, 0x0a, 0x30, 0xe4, 0x17, 0x51, 0x0b, 0xf2, 0xbb, + 0xa6, 0xb7, 0x00, 0xa2, 0xe8, 0xa5, 0xa3, 0x41, 0x1d, 0x65, 0x2d, 0x26, + 0x93, 0x26, 0x7d, 0xdc, 0xad, 0x6f, 0x83, 0xeb, 0x66, 0x55, 0xde, 0x60, + 0x21, 0x56, 0x19, 0x4f, 0x9b, 0x7b, 0x26, 0x4a, 0x80, 0xf5, 0xab, 0x8b, + 0xbf, 0xe4, 0xb1, 0xa1, 0xd6, 0x33, 0x32, 0xbf, 0x86, 0x8c, 0x3c, 0xd0, + 0x12, 0x03, 0xd4, 0xb9, 0x23, 0x54, 0x1b, 0x94, 0x2f, 0xa5, 0x34, 0x4d, + 0x59, 0x18, 0x33, 0x8e, 0x8c, 0xf7, 0x1f, 0xc9, 0x6d, 0x75, 0xfb, 0x2a, + 0x22, 0x6c, 0x64, 0xb7, 0x79, 0xd8, 0x3b, 0xf6, 0x4e, 0x98, 0xd8, 0xa8, + 0x2c, 0x06, 0xd1, 0x92, 0x32, 0x44, 0xec, 0x38, 0x40, 0x3b, 0x53, 0x16, + 0x40, 0x8f, 0x92, 0x72, 0x87, 0xa8, 0xb8, 0xc0, 0x8f, 0x25, 0x4c, 0x4f, + 0x24, 0xfc, 0x8d, 0xc6, 0xa6, 0xeb, 0x2f, 0xdf, 0x2f, 0x0d, 0x2f, 0xd3, + 0x6e, 0x70, 0x71, 0xfe, 0xf0, 0x2e, 0xe9, 0x84, 0xd3, 0xc1, 0xd1, 0x70, + 0x4b, 0x8f, 0x7b, 0x60, 0xb0, 0xb7, 0xe3, 0x79, 0x52, 0x6a, 0x6b, 0x26, + 0x03, 0x8f, 0x6a, 0x0f, 0x8d, 0x85, 0xd7, 0x5f, 0xf7, 0x39, 0x31, 0x0e, + 0x26, 0x73, 0x84, 0x3f, 0x9b, 0x10, 0x6f, 0x29, 0x63, 0x14, 0x36, 0xa2, + 0xec, 0x44, 0x7d, 0x84, 0xc6, 0x4a, 0xec, 0xfe, 0xac, 0xcb, 0xe4, 0xfa, + 0xf6, 0x68, 0x83, 0x68, 0xe0, 0x8f, 0xd3, 0x8a, 0x60, 0x73, 0xf1, 0x5c, + 0x71, 0x02, 0x0c, 0xa2, 0x88, 0x2c, 0xa2, 0x35, 0x35, 0x5c, 0x3f, 0xb1, + 0xbe, 0xb3, 0x6b, 0x5c, 0xe1, 0x78, 0x75, 0x40, 0x20, 0x87, 0x67, 0xca, + 0x07, 0x1c, 0x9c, 0x02, 0xc7, 0xf2, 0x9d, 0x1c, 0xda, 0x1b, 0x86, 0x1b, + 0xc6, 0xa6, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, + 0x93, 0xca, 0x30, 0xae, 0xea, 0x26, 0x6a, 0x1b, 0x15, 0x46, 0x0a, 0xe3, + 0x57, 0x23, 0x4c, 0x0c, 0x98, 0x8e, 0x3e, 0xbb, 0x43, 0x14, 0x73, 0xdf, + 0x17, 0x91, 0xe2, 0xee, 0x39, 0xf9, 0xc2, 0x2f, 0xdc, 0xad, 0x0e, 0x00, + 0xf5, 0xdd, 0xe3, 0x97, 0xba, 0x8c, 0xee, 0x53, 0xc4, 0x70, 0x37, 0x46, + 0xcf, 0x04, 0xc3, 0xc8, 0x56, 0x38, 0x2e, 0x39, 0x75, 0x32, 0x6d, 0x98, + 0xc4, 0x14, 0xae, 0xa4, 0x29, 0xa3, 0xc6, 0xb6, 0x66, 0x45, 0x48, 0xdf, + 0xc0, 0xa9, 0x4b, 0x4f, 0xef, 0xb9, 0xb4, 0x89, 0x0d, 0x64, 0x00, 0x5c, + 0xd1, 0xc8, 0x2b, 0xf7, 0xc5, 0x1a, 0x1b, 0x06, 0xb7, 0x49, 0xb1, 0xe3, + 0x4d, 0x87, 0xf9, 0x3f, 0xba, 0x39, 0xa3, 0x56, 0x7f, 0x43, 0xcc, 0x15, + 0x9c, 0x3d, 0xba, 0x71, 0x7b, 0xeb, 0x45, 0x0f, 0x15, 0x1b, 0x6c, 0x84, + 0x75, 0x6d, 0x43, 0x0b, 0x27, 0x12, 0x6b, 0xbc, 0x0a, 0x6d, 0xe4, 0xf6, + 0x4f, 0xc7, 0xbb, 0x9e, 0x91, 0xb5, 0x09, 0x5f, 0x79, 0x2a, 0xbf, 0xda, + 0x34, 0x91, 0x44, 0x47, 0x52, 0x64, 0x00, 0x89, 0x27, 0x17, 0x5c, 0xe9, + 0x90, 0x8b, 0xcb, 0xbe, 0x21, 0x47, 0x65, 0x1c, 0x54, 0x61, 0x48, 0x17, + 0x66, 0xb7, 0xa1, 0x60, 0x27, 0x31, 0x04, 0x42, 0x3b, 0x33, 0x3d, 0xda, + 0xf7, 0x61, 0x3d, 0x4b, 0x91, 0xa5, 0x74, 0x4b, 0xde, 0x16, 0xf2, 0x79, + 0x3e, 0xf7, 0x89, 0x87, 0xb3, 0xdd, 0xa2, 0x49, 0xd7, 0x54, 0x1b, 0x39, + 0xff, 0xb5, 0xec, 0x9d, 0x1d, 0x09, 0x7e, 0x5a, 0x3c, 0xd1, 0xdc, 0x0e, + 0x2a, 0x0e, 0x2c, 0x40, 0x4e, 0xa5, 0x8c, 0x9d, 0xc8, 0x9b, 0xa5, 0xb2, + 0x40, 0xa4, 0xaa, 0x3b, 0xac, 0x93, 0x19, 0xf7, 0xa1, 0x8b, 0xf8, 0x4a, + 0x40, 0x08, 0x5d, 0x1d, 0xb0, 0xae, 0x0f, 0x67, 0xa7, 0x21, 0xaf, 0xe3, + 0xb1, 0xfc, 0xff, 0xa0, 0x95, 0x66, 0x2b, 0xf7, 0x82, 0x2d, 0x8a, 0x26, + 0x0f, 0xc3, 0xed, 0x62, 0xb6, 0xcb, 0x4c, 0x86, 0xe9, 0x20, 0x78, 0x3f, + 0x08, 0x53, 0x8f, 0x41, 0xf1, 0xa1, 0x04, 0x77, 0xd9, 0xe6, 0xea, 0x26, + 0x6d, 0x33, 0x48, 0xb3, 0xbb, 0xed, 0xfc, 0xd7, 0xa3, 0x2b, 0xe2, 0x39, + 0xcf, 0x78, 0x4e, 0x11, 0x26, 0xad, 0x39, 0x83, 0x6e, 0x72, 0xbf, 0xc6, + 0x34, 0x23, 0x97, 0x5d, 0x7b, 0x64, 0x1e, 0x78, 0x00, 0x34, 0x92, 0x5d, + 0x3f, 0x23, 0x28, 0x60, 0x7f, 0x88, 0xf0, 0xca, 0x96, 0x4a, 0x15, 0xbf, + 0x8a, 0xb7, 0xd0, 0xd9, 0x99, 0x8b, 0xdb, 0x26, 0xdc, 0x7e, 0x8d, 0x35, + 0x53, 0x60, 0x07, 0x85, 0x80, 0xc4, 0x9c, 0x0d, 0x81, 0xe2, 0x93, 0x85, + 0x76, 0x2d, 0x85, 0x21, 0x6e, 0xda, 0x29, 0xe5, 0xb1, 0x08, 0x46, 0x09, + 0x1b, 0x8a, 0xd9, 0xd2, 0xd7, 0x16, 0x74, 0xee, 0x26, 0x3e, 0xc4, 0x8c, + 0x2e, 0x6b, 0x0c, 0xbc, 0x95, 0xea, 0x4a, 0xb2, 0xd6, 0x6f, 0x43, 0xd1, + 0x3a, 0x8f, 0xbd, 0x77, 0xb4, 0x67, 0x63, 0x6b, 0xd2, 0xe0, 0xf0, 0x81, + 0x74, 0xb7, 0xc5, 0x11, 0x60, 0x10, 0x6b, 0xc6, 0x0f, 0xfd, 0x84, 0x2e, + 0x5c, 0x8f, 0x3b, 0xf5, 0x68, 0xa7, 0x62, 0xc6, 0x4f, 0xa6, 0xee, 0x19, + 0x44, 0xea, 0xc0, 0xe4, 0x64, 0x12, 0x71, 0x2f, 0xfb, 0xa3, 0x4d, 0xb0, + 0x8e, 0x5e, 0xe1, 0x79, 0x65, 0xd4, 0xf3, 0xed, 0x73, 0x04, 0xf1, 0x6d, + 0xc6, 0x75, 0x54, 0x28, 0x13, 0xe2, 0xd6, 0xa1, 0x26, 0xf9, 0xa4, 0x29, + 0x20, 0x5b, 0xd0, 0x3c, 0x3d, 0xf3, 0x7a, 0x18, 0x9a, 0x3d, 0xec, 0x6a, + 0x4c, 0xfd, 0xa5, 0x00, 0xdf, 0xec, 0xfd, 0x64, 0x38, 0x66, 0xa7, 0xba, + 0x59, 0xb3, 0x9b, 0x9c, 0x44, 0xfb, 0x10, 0x08, 0xb8, 0x79, 0xea, 0x85, + 0xbf, 0xa4, 0x14, 0xce, 0xce, 0x85, 0x22, 0x3f, 0x16, 0x00, 0x1c, 0x57, + 0xc8, 0x5a, 0x1b, 0xf5, 0xff, 0xde, 0x7e, 0xa9, 0xcc, 0xf3, 0xb5, 0x1d, + 0x57, 0x06, 0xda, 0xbb, 0x6c, 0x0a, 0x1e, 0xd4, 0x09, 0x74, 0x84, 0x1d, + 0xfa, 0xdf, 0x33, 0x1e, 0xe2, 0x8f, 0x10, 0xf7, 0x73, 0xab, 0x71, 0xb8, + 0x64, 0xce, 0xc0, 0x49, 0xc0, 0x36, 0xd3, 0x39, 0x31, 0x4c, 0x12, 0x5b, + 0xf3, 0xf9, 0xb4, 0x2c, 0x88, 0xba, 0xd4, 0x1a, 0xbd, 0x0c, 0x99, 0xbd, + 0x0e, 0xad, 0x51, 0xe0, 0xca, 0xdb, 0x25, 0x66, 0x83, 0xe0, 0x55, 0x18, + 0xeb, 0xa6, 0x4e, 0x56, 0xcb, 0x2f, 0xa5, 0xf2, 0x42, 0x7a, 0xa1, 0x05, + 0xf0, 0x3a, 0x71, 0x5a, 0x78, 0x3a, 0x7a, 0x6d, 0x12, 0x9f, 0x43, 0xc5, + 0xcc, 0xb3, 0xfd, 0xf2, 0xbf, 0x05, 0x16, 0xef, 0x07, 0xf9, 0xde, 0x0d, + 0x51, 0xf0, 0x33, 0x86, 0x43, 0x57, 0x40, 0xbc, 0xa9, 0xbd, 0xa0, 0x23, + 0xff, 0xbb, 0xe6, 0x15, 0xa1, 0xeb, 0xe9, 0x78, 0x0d, 0x72, 0x76, 0xf2, + 0xb6, 0x6e, 0x46, 0xe2, 0x86, 0xab, 0x3c, 0x52, 0x2c, 0xc6, 0x77, 0xdd, + 0x57, 0xf7, 0x4d, 0x36, 0xbb, 0x41, 0x08, 0x21, 0xaa, 0xe6, 0x44, 0x50, + 0xed, 0xaf, 0x18, 0xb3, 0xdd, 0x6b, 0x57, 0x46, 0x9e, 0x44, 0x93, 0x20, + 0xe0, 0x62, 0x95, 0xcd, 0xcf, 0xe4, 0x96, 0x92, 0xc3, 0x0d, 0x16, 0xb2, + 0xc3, 0xf4, 0x0f, 0x3f, 0x87, 0x17, 0xb9, 0x7b, 0x60, 0x60, 0xfa, 0xfb, + 0x81, 0x5c, 0xb3, 0xb7, 0x89, 0x73, 0xf7, 0x35, 0xf7, 0x27, 0xf1, 0x0e, + 0xa4, 0xa1, 0xba, 0xea, 0x6a, 0xe3, 0x5c, 0x0f, 0xf7, 0x15, 0xbc, 0x28, + 0x57, 0x27, 0x8f, 0xd8, 0xca, 0x82, 0x19, 0xd0, 0xa3, 0x9d, 0xe5, 0xe0, + 0x44, 0xbf, 0x78, 0xa4, 0x09, 0x69, 0x27, 0xa0, 0x69, 0xb5, 0xd4, 0xbe, + 0x00, 0xe6, 0x03, 0x97, 0xbc, 0x8b, 0xfc, 0x25, 0x70, 0xb3, 0x49, 0x30, + 0xe3, 0x24, 0x19, 0x77, 0xb4, 0x93, 0x46, 0x03, 0xe6, 0x22, 0xaf, 0x76, + 0xd2, 0x90, 0x00, 0x05, 0x46, 0xb8, 0xa4, 0xf5, 0x4c, 0xaa, 0x04, 0x63, + 0xa0, 0x57, 0xe0, 0x20, 0x6e, 0x1a, 0xed, 0x21, 0x86, 0xd0, 0x38, 0x5b, + 0xe6, 0xa7, 0xb0, 0xe7, 0x75, 0xe3, 0x76, 0xb3, 0x15, 0x8b, 0xdc, 0x10, + 0x52, 0x15, 0x21, 0x7b, 0xd0, 0xc4, 0x75, 0x26, 0x1d, 0x6e, 0x0d, 0x4c, + 0x08, 0x5b, 0x95, 0x9a, 0xd0, 0xda, 0xbe, 0x23, 0x98, 0xde, 0x60, 0x2a, + 0xe9, 0xa4, 0x92, 0xf0, 0x92, 0x84, 0xdc, 0x86, 0x60, 0xf5, 0x23, 0x31, + 0xf5, 0xe9, 0xd6, 0x00, 0xc1, 0x78, 0xab, 0x05, 0x94, 0xd3, 0x47, 0x4d, + 0x32, 0x0f, 0x82, 0xa0, 0x99, 0x0b, 0xfe, 0x6b, 0x58, 0xf9, 0x24, 0xf6, + 0x17, 0xa0, 0x5f, 0x24, 0x6a, 0xc6, 0x01, 0xa8, 0xfa, 0xca, 0xdc, 0xb6, + 0x83, 0xcb, 0xd2, 0x3b, 0xb7, 0x0b, 0x04, 0x3e, 0x6a, 0xaf, 0x23, 0x17, + 0x3e, 0x14, 0xce, 0x52, 0x1c, 0xe3, 0x06, 0x66, 0x29, 0x17, 0x6f, 0x7e, + 0x66, 0x06, 0xa9, 0x68, 0x7f, 0xca, 0xad, 0xa8, 0xb7, 0x2d, 0xa4, 0x5d, + 0xa6, 0x16, 0xcd, 0xed, 0xee, 0x14, 0x96, 0xc8, 0x12, 0x69, 0x4e, 0x70, + 0x72, 0x2a, 0x75, 0x82, 0x08, 0x3f, 0x3e, 0x27, 0xa0, 0xea, 0x43, 0x84, + 0xa9, 0x9a, 0x91, 0x87, 0x4f, 0x20, 0x61, 0x55, 0x8d, 0x70, 0xad, 0x6c, + 0x59, 0x5d, 0x13, 0x80, 0xbb, 0x52, 0x55, 0x81, 0x8b, 0x59, 0x94, 0x0f, + 0xc2, 0x54, 0x79, 0x59, 0xe8, 0x9d, 0x58, 0xe5, 0x91, 0x10, 0xb3, 0xef, + 0x1c, 0xda, 0xaa, 0xdd, 0x91, 0x0b, 0xb0, 0x14, 0x3b, 0xad, 0x02, 0x98, + 0x40, 0x3c, 0x54, 0xc4, 0x23, 0xb9, 0x40, 0x54, 0x7e, 0x88, 0x10, 0x3e, + 0x24, 0xe5, 0xf6, 0xdf, 0x5c, 0x9e, 0x7a, 0x9f, 0xd0, 0xff, 0x5e, 0x9c, + 0xb6, 0x30, 0x17, 0x94, 0xd2, 0xaa, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x80, 0x00, 0x00, 0x00, 0x96, 0xff, 0x2f, 0x01, 0x60, 0x2c, 0x1b, 0xe3, + 0xc6, 0xcb, 0xa4, 0x41, 0xa1, 0x44, 0x13, 0x14, 0xe2, 0x44, 0x77, 0x1c, + 0x96, 0xe8, 0xe6, 0x4f, 0x70, 0x99, 0x3a, 0xef, 0xa1, 0x6f, 0x1f, 0x7f, + 0xb9, 0xe9, 0x1e, 0x35, 0x37, 0x5b, 0x94, 0x90, 0x78, 0xcc, 0x8d, 0xcd, + 0x6c, 0x9f, 0xf6, 0x73, 0xed, 0x23, 0xa2, 0x28, 0x64, 0x58, 0x50, 0x64, + 0x05, 0xbc, 0xc9, 0x9b, 0x5a, 0xec, 0x3f, 0x2b, 0x61, 0xcf, 0xa7, 0x35, + 0x56, 0x8c, 0x77, 0x68, 0xd6, 0xcf, 0x9b, 0xc5, 0x62, 0xee, 0x3a, 0xb2, + 0xfe, 0x78, 0xba, 0x02, 0xe7, 0x26, 0x8a, 0x89, 0x30, 0x19, 0xcc, 0xb0, + 0x98, 0xbf, 0x30, 0x2c, 0xae, 0x13, 0x6c, 0x93, 0x86, 0x19, 0x84, 0x13, + 0x01, 0x2f, 0x39, 0x4e, 0x33, 0xd1, 0x15, 0x99, 0xf7, 0x1e, 0xb8, 0x86, + 0xdb, 0xb6, 0xf9, 0x56, 0x42, 0x0e, 0x4a, 0xb1, 0x5e, 0xf0, 0x9a, 0x06, + 0x5e, 0xab, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, + 0xcd, 0xde, 0xad, 0x40, 0x34, 0xcd, 0x79, 0x0a, 0x29, 0x84, 0x05, 0x3f, + 0xb5, 0xbe, 0x49, 0x84, 0x43, 0xcc, 0xa6, 0xe3, 0xe9, 0xdc, 0x84, 0x14, + 0xe7, 0xb3, 0x1b, 0x96, 0xe8, 0xda, 0x35, 0x15, 0x38, 0xf5, 0xb3, 0xb5, + 0x91, 0xc3, 0xc3, 0x94, 0xc6, 0x79, 0xeb, 0xf5, 0x22, 0x78, 0xf0, 0x0b, + 0xda, 0xb0, 0x91, 0xa7, 0x43, 0x71, 0x8e, 0xa6, 0x52, 0x0f, 0x81, 0x06, + 0xc8, 0xdf, 0xb5, 0x1f, 0x92, 0xb0, 0xfe, 0x93, 0x38, 0x4c, 0xf4, 0x17, + 0x66, 0x31, 0xea, 0x08, 0x72, 0xb9, 0xaa, 0xfd, 0x40, 0x8d, 0xbf, 0x56, + 0x19, 0xb1, 0xb5, 0x8e, 0x4e, 0x4e, 0x73, 0x7f, 0x4b, 0x0c, 0x70, 0x94, + 0x7c, 0x9f, 0xfc, 0x23, 0x35, 0xba, 0xd2, 0x23, 0x88, 0x1d, 0x83, 0x28, + 0x45, 0xd7, 0x1b, 0x63, 0xfb, 0x36, 0x86, 0x06, 0xf3, 0x99, 0x81, 0x6e, + 0xd7, 0xf1, 0xd4, 0x53, 0x6d, 0x30, 0x3c, 0x8d, 0xac, 0xc6, 0x9a, 0xd5, + 0xe8, 0x4f, 0x11, 0x58, 0xba, 0xfd, 0x67, 0x06, 0xe7, 0x1a, 0xb4, 0xa1, + 0x45, 0x13, 0xf2, 0x3b, 0xdc, 0x71, 0xf0, 0xc6, 0x53, 0xfc, 0x8b, 0x2f, + 0x14, 0xe4, 0xe0, 0xd6, 0x8c, 0x96, 0x4c, 0x48, 0xc0, 0x30, 0x6e, 0x00, + 0x0f, 0x42, 0xfe, 0xa7, 0x9d, 0x0f, 0xf2, 0x52, 0x58, 0xf9, 0x35, 0x33, + 0x99, 0xda, 0xd5, 0x9d, 0x61, 0x26, 0x6b, 0x80, 0xff, 0x08, 0x51, 0x54, + 0x26, 0xfa, 0x8d, 0xfc, 0x67, 0x60, 0x93, 0x0e, 0xcd, 0x78, 0x41, 0x5a, + 0x31, 0x47, 0x14, 0xb0, 0x65, 0x89, 0x30, 0xcb, 0x0c, 0xc5, 0xa0, 0x37, + 0xa8, 0xe0, 0xcf, 0x24, 0xa4, 0x2f, 0xad, 0xa7, 0x9c, 0xa2, 0xe8, 0x81, + 0x17, 0xbe, 0x2f, 0xd5, 0xd1, 0xa8, 0xff, 0x9d, 0x5e, 0x7f, 0xd9, 0x6c, + 0x56, 0xe6, 0xc4, 0x60, 0x8d, 0xa5, 0x47, 0x5e, 0x43, 0x1e, 0x34, 0x23, + 0xb3, 0x6a, 0xdf, 0x6c, 0xf8, 0xd1, 0x85, 0x11, 0xaa, 0x74, 0x85, 0x71, + 0x27, 0xc5, 0x80, 0x37, 0x60, 0xb4, 0x2b, 0x53, 0x5a, 0xc4, 0x35, 0xd1, + 0xe8, 0x4b, 0x01, 0x58, 0x1f, 0xdb, 0x73, 0xf3, 0x2c, 0x8b, 0xbb, 0x17, + 0x36, 0x76, 0x35, 0x6b, 0xa0, 0x82, 0x47, 0xf5, 0x16, 0x21, 0x41, 0x43, + 0xc9, 0x1f, 0x53, 0xf9, 0xe9, 0x47, 0xf0, 0x9c, 0x6d, 0xe3, 0x23, 0x59, + 0x74, 0xdc, 0x1a, 0x8f, 0x4e, 0x6c, 0x71, 0x83, 0x7e, 0xd0, 0x2b, 0x50, + 0x44, 0x86, 0x5f, 0xbf, 0x60, 0x92, 0xeb, 0x9a, 0x9b, 0xa2, 0xc9, 0x2b, + 0xa8, 0xc4, 0x77, 0x4e, 0x3f, 0xf8, 0xa6, 0x39, 0x50, 0x5c, 0x7e, 0x2a, + 0x70, 0xb0, 0x5d, 0x28, 0xb2, 0x81, 0xa9, 0xaf, 0x16, 0x5e, 0x27, 0xeb, + 0x03, 0x0e, 0x82, 0xad, 0x28, 0x51, 0x16, 0xd1, 0xf4, 0x58, 0x75, 0x1a, + 0xf9, 0x6a, 0xbf, 0x73, 0xd7, 0x84, 0x07, 0x7f, 0x4c, 0x4e, 0x29, 0x02, + 0x9b, 0x60, 0x81, 0x85, 0xa9, 0xbf, 0xc7, 0xa0, 0x8f, 0x8a, 0xdc, 0xa4, + 0xc5, 0x17, 0x51, 0x24, 0x15, 0x28, 0x9e, 0x5e, 0x78, 0x84, 0x21, 0x02, + 0xca, 0x26, 0x61, 0x4e, 0x95, 0xa6, 0x8d, 0xa6, 0x98, 0x7d, 0x1f, 0x84, + 0x19, 0x24, 0x8b, 0x31, 0x76, 0x89, 0x2a, 0x5f, 0xa9, 0xfb, 0xaa, 0x8a, + 0x8c, 0xce, 0xe4, 0x30, 0xd6, 0xec, 0x5b, 0x39, 0xb7, 0x09, 0x80, 0x23, + 0x4c, 0xe1, 0x6e, 0x8f, 0x7c, 0x10, 0xe8, 0x8a, 0x60, 0x35, 0xd7, 0xa3, + 0xe0, 0x5f, 0xcd, 0xfa, 0x3d, 0x8f, 0xd8, 0x5d, 0xec, 0xc9, 0xc5, 0xa0, + 0x73, 0x41, 0x89, 0xe5, 0x39, 0xf2, 0x42, 0xff, 0x08, 0xa0, 0x12, 0xb7, + 0x4a, 0x5e, 0x46, 0x06, 0x31, 0xbd, 0x88, 0x5e, 0x9e, 0x05, 0x17, 0x51, + 0xb3, 0xe7, 0x88, 0x10, 0x19, 0x32, 0xff, 0x8a, 0x1e, 0xce, 0x66, 0xbc, + 0x84, 0x1f, 0xed, 0x52, 0x52, 0x77, 0xe1, 0x5e, 0xa6, 0x21, 0xe4, 0xad, + 0x59, 0xca, 0xa3, 0x77, 0xea, 0x66, 0x28, 0x15, 0x73, 0x3a, 0xfd, 0xe4, + 0x75, 0x46, 0x99, 0x59, 0x5c, 0x7a, 0x9b, 0x9d, 0x11, 0xb4, 0x76, 0x45, + 0x06, 0x45, 0x41, 0x1e, 0x94, 0xb7, 0xd9, 0xb8, 0xcb, 0xbf, 0x71, 0xec, + 0xba, 0x9f, 0x4a, 0x1b, 0xbc, 0xfd, 0x5c, 0x06, 0x64, 0xfd, 0x31, 0x52, + 0xc0, 0xe4, 0xa7, 0x21, 0x2f, 0x22, 0x92, 0xf0, 0x51, 0x33, 0x92, 0x1d, + 0x40, 0x3c, 0x01, 0x81, 0x3b, 0xa8, 0x2e, 0x4e, 0xb6, 0x60, 0xcd, 0xd4, + 0x36, 0x3b, 0x2e, 0x1d, 0x5e, 0x43, 0xd9, 0x94, 0xf1, 0x51, 0xd3, 0x59, + 0x94, 0x6a, 0xd5, 0x5f, 0x1f, 0xd3, 0xa6, 0x55, 0xda, 0x15, 0xf1, 0x3e, + 0x2c, 0x60, 0xb8, 0xc3, 0xda, 0x0e, 0x56, 0x53, 0xea, 0xcd, 0x39, 0x27, + 0x94, 0x86, 0x94, 0xb2, 0x5b, 0xd8, 0x9a, 0x12, 0x94, 0xb0, 0xb6, 0x77, + 0x28, 0xba, 0xde, 0xb6, 0x60, 0x4d, 0x2b, 0x6e, 0x3d, 0xf6, 0xf1, 0x48, + 0xf7, 0x77, 0xa1, 0x49, 0xe0, 0x9f, 0x1e, 0xc9, 0xe6, 0xcb, 0x95, 0x26, + 0x61, 0x5a, 0xc9, 0xed, 0x49, 0x40, 0x17, 0x57, 0x15, 0xfc, 0x3c, 0xb8, + 0x28, 0x79, 0xb8, 0x42, 0x2a, 0xf9, 0xd4, 0x19, 0xb9, 0x5f, 0x41, 0xc2, + 0x25, 0xd7, 0x88, 0x34, 0xb3, 0x25, 0x4e, 0xca, 0xff, 0x9e, 0x59, 0x9a, + 0x33, 0xc8, 0x12, 0xf9, 0xd5, 0x70, 0xc0, 0x8b, 0x43, 0x13, 0xc4, 0x8d, + 0x45, 0x99, 0xaa, 0xd7, 0xeb, 0xb1, 0xe9, 0xb7, 0x5b, 0xab, 0x48, 0xd1, + 0x26, 0x60, 0x8c, 0x13, 0x55, 0x8a, 0x41, 0xd3, 0x68, 0x58, 0xd4, 0xa6, + 0x30, 0x6e, 0x88, 0x3e, 0x81, 0x6e, 0x61, 0x06, 0x13, 0x66, 0xd5, 0x8e, + 0x5d, 0x87, 0x4f, 0xd9, 0xb1, 0x66, 0xb3, 0xc5, 0x88, 0xa9, 0xc0, 0x73, + 0xcb, 0x7f, 0x42, 0xec, 0x96, 0x64, 0xad, 0x72, 0x85, 0x72, 0xaf, 0xeb, + 0xa9, 0xc4, 0x17, 0x86, 0xab, 0xe7, 0x23, 0xd7, 0x96, 0xf7, 0xb2, 0xb3, + 0x51, 0xe1, 0x9a, 0x3b, 0x0e, 0xaf, 0x89, 0xca, 0x7b, 0xf1, 0x70, 0x7b, + 0xc7, 0x82, 0xfc, 0xc7, 0x6c, 0x37, 0xd9, 0x7b, 0x82, 0x0f, 0x94, 0xcf, + 0xd1, 0xa9, 0x33, 0xc2, 0xa4, 0xab, 0xed, 0xad, 0xee, 0x64, 0x5d, 0x04, + 0xf2, 0xcb, 0x8e, 0x99, 0x22, 0x33, 0x69, 0x85, 0x85, 0xb6, 0x1a, 0x9b, + 0x09, 0x18, 0xbe, 0xcd, 0x63, 0xf6, 0x5d, 0x52, 0xbc, 0x26, 0x99, 0x3e, + 0x52, 0xe5, 0x0c, 0xc5, 0xee, 0xdd, 0xbb, 0x07, 0xbc, 0x38, 0xc1, 0x67, + 0x96, 0x8c, 0xe6, 0xe4, 0x18, 0xfa, 0x07, 0x91, 0x48, 0xef, 0x9c, 0x70, + 0x9d, 0x5b, 0x1c, 0x0e, 0xd5, 0xd3, 0x59, 0xee, 0x44, 0x13, 0xf7, 0x00, + 0xa6, 0x20, 0xad, 0x65, 0x1d, 0xb7, 0x96, 0x2f, 0x79, 0x7b, 0x04, 0xa3, + 0x10, 0x90, 0x29, 0x8c, 0xa3, 0x2e, 0x14, 0x39, 0xd3, 0xe4, 0x6e, 0x46, + 0xf7, 0x6e, 0x96, 0x68, 0xd9, 0xef, 0x45, 0xf7, 0x3c, 0xcd, 0xc7, 0xca, + 0x33, 0x64, 0x8e, 0x31, 0x80, 0x48, 0x7b, 0x7c, 0x81, 0x9a, 0x48, 0xff, + 0xd5, 0x0d, 0x74, 0xe7, 0x77, 0x46, 0x61, 0x9b, 0xde, 0xed, 0x83, 0xe9, + 0x4f, 0x92, 0xc1, 0x16, 0xad, 0x44, 0x40, 0x23, 0xce, 0x04, 0x31, 0xbf, + 0xcf, 0xe2, 0x5a, 0x68, 0x5a, 0xf4, 0x0f, 0xe1, 0x87, 0x79, 0xb0, 0x32, + 0x0b, 0x09, 0x6b, 0x72, 0x2b, 0x16, 0x06, 0x67, 0x82, 0x0b, 0x92, 0x35, + 0xdb, 0x4c, 0xe2, 0x4a, 0x60, 0x99, 0xaf, 0x52, 0x10, 0x4b, 0xa5, 0xcf, + 0xac, 0x66, 0x49, 0x56, 0x04, 0xc0, 0xd6, 0x6f, 0x62, 0x53, 0x6f, 0xcb, + 0x62, 0xe9, 0xa5, 0xca, 0x18, 0x8e, 0x86, 0x3f, 0x36, 0xfd, 0xea, 0x55, + 0x16, 0x6d, 0x6c, 0x6a, 0x8f, 0xa7, 0x9c, 0x70, 0x15, 0xd7, 0xf4, 0x57, + 0x68, 0x04, 0x84, 0x60, 0x3b, 0xb0, 0x32, 0xc4, 0xea, 0x9d, 0x70, 0xb9, + 0xa6, 0x34, 0xe5, 0xfa, 0xa1, 0x24, 0x54, 0x7f, 0xef, 0xac, 0xb4, 0x5f, + 0xa0, 0xc0, 0x40, 0x3f, 0x73, 0xdf, 0x56, 0xa6, 0xd9, 0x17, 0xf4, 0xff, + 0x50, 0xae, 0x21, 0x0d, 0x5a, 0xe0, 0xb0, 0xf9, 0x5b, 0x7a, 0x61, 0x6e, + 0xa6, 0x85, 0x85, 0xbf, 0x19, 0x03, 0xe2, 0x74, 0x1f, 0x03, 0x70, 0x76, + 0x3c, 0xed, 0x02, 0x7d, 0xfa, 0xf9, 0x1e, 0x17, 0xdd, 0x42, 0x30, 0xf0, + 0x32, 0x47, 0x46, 0xae, 0xf5, 0x64, 0xe6, 0x5e, 0x2b, 0x40, 0x86, 0x97, + 0xb1, 0x24, 0x52, 0x69, 0x67, 0x79, 0x8e, 0x0d, 0xcc, 0x07, 0xcb, 0x72, + 0x29, 0xe9, 0xba, 0x2d, 0xf7, 0xcb, 0xe3, 0x86, 0x06, 0xaa, 0x6d, 0x79, + 0xf8, 0xb6, 0x93, 0x0a, 0x9c, 0x97, 0xef, 0x47, 0x37, 0x13, 0x2e, 0x6b, + 0xfd, 0x59, 0x0c, 0xc9, 0x5e, 0x5e, 0xcd, 0x71, 0x6f, 0x99, 0x0d, 0x88, + 0x9d, 0xbb, 0x7c, 0x2b, 0x22, 0xd5, 0xbe, 0xee, 0x26, 0x1c, 0xe1, 0xad, + 0xc8, 0x4d, 0x5f, 0x6b, 0xd1, 0xf4, 0x30, 0x4d, 0x46, 0x1d, 0x54, 0x11, + 0x4b, 0xa0, 0x7f, 0x94, 0x71, 0xc0, 0x44, 0x4a, 0x42, 0x11, 0xf5, 0x89, + 0xec, 0xb5, 0x24, 0x45, 0xf1, 0xf0, 0x30, 0x54, 0xf8, 0x62, 0xdb, 0x58, + 0x3d, 0x7c, 0x2a, 0x82, 0xe5, 0xbe, 0x13, 0xcf, 0xdc, 0x88, 0xfb, 0xd3, + 0x1e, 0x4d, 0xa5, 0x3e, 0xad, 0x95, 0xa2, 0xe6, 0x48, 0x73, 0xb2, 0xbe, + 0x96, 0xef, 0x8e, 0x0b, 0x28, 0xf9, 0xbe, 0x2a, 0xd6, 0x68, 0x9e, 0x9c, + 0x7b, 0x5a, 0xaf, 0x20, 0xf6, 0xa5, 0x3f, 0x99, 0x61, 0x57, 0xe8, 0x1c, + 0xb2, 0xc3, 0xd0, 0x7f, 0x2c, 0xb5, 0xe9, 0x66, 0x8e, 0x88, 0xec, 0x13, + 0x51, 0xbc, 0x8e, 0xb6, 0xe2, 0x91, 0xbf, 0x5e, 0x8c, 0x1c, 0xdd, 0x0e, + 0x0a, 0x13, 0x06, 0xc6, 0x62, 0x1c, 0x41, 0x8d, 0xa1, 0xc0, 0xf2, 0xfa, + 0x76, 0x35, 0xaa, 0x77, 0x06, 0x3f, 0x76, 0x50, 0xf6, 0x43, 0xf2, 0x25, + 0x00, 0x79, 0xde, 0xca, 0xa1, 0x06, 0x6f, 0xb4, 0x17, 0x4b, 0x99, 0x5a, + 0x00, 0x32, 0xd6, 0xb0, 0x1f, 0x80, 0x53, 0x16, 0xaa, 0x87, 0x72, 0xa2, + 0x34, 0xaf, 0x90, 0x3d, 0x60, 0xde, 0x0e, 0x6d, 0x83, 0xda, 0xb2, 0x11, + 0x2f, 0x39, 0xdc, 0x1a, 0xfe, 0x51, 0x74, 0x10, 0x3c, 0x41, 0xd5, 0x41, + 0x65, 0x4a, 0xa0, 0x11, 0xde, 0x95, 0x34, 0xef, 0xa0, 0xc9, 0xa8, 0xd3, + 0xcb, 0xb9, 0x7d, 0x51, 0x7d, 0xff, 0x26, 0x88, 0xd8, 0x29, 0x0e, 0xa0, + 0xd4, 0xa7, 0x07, 0x33, 0xe7, 0x7d, 0x59, 0x9f, 0x35, 0xc1, 0xb5, 0xf7, + 0x78, 0x78, 0x84, 0xf0, 0x20, 0x41, 0x3f, 0x02, 0x7d, 0x41, 0x90, 0x01, + 0x8d, 0xa4, 0xd8, 0xd7, 0xeb, 0x56, 0x7f, 0x38, 0xbc, 0x1e, 0x15, 0xdf, + 0xfc, 0x34, 0xe7, 0x99, 0xd4, 0x92, 0xd5, 0xf3, 0x9e, 0x16, 0x0b, 0x5c, + 0xeb, 0xb6, 0x78, 0xac, 0x84, 0x06, 0x8e, 0xfe, 0xd0, 0x7c, 0xce, 0x4a, + 0x43, 0x49, 0x3b, 0xe1, 0xab, 0x57, 0xc0, 0x12, 0xd6, 0x9d, 0xa4, 0xee, + 0x91, 0x10, 0x81, 0xe2, 0xfc, 0x02, 0x26, 0x7a, 0xca, 0x81, 0x5b, 0x2f, + 0x34, 0x51, 0xdd, 0x25, 0x4d, 0xc8, 0xf9, 0x3e, 0x59, 0x0f, 0x3d, 0x64, + 0x51, 0xbf, 0x42, 0xc4, 0x92, 0x9d, 0x8f, 0x39, 0x8a, 0x31, 0x09, 0x24, + 0x19, 0x44, 0xc0, 0xf4, 0xea, 0xca, 0x59, 0xcb, 0x86, 0x6c, 0x02, 0x7a, + 0xe5, 0x30, 0x79, 0xe2, 0x2c, 0x76, 0x08, 0x8f, 0x98, 0x0d, 0x4d, 0x12, + 0xc3, 0x98, 0xb4, 0x24, 0x04, 0x4f, 0x51, 0xec, 0x4e, 0xec, 0xbd, 0x8c, + 0xc4, 0x79, 0x51, 0x7f, 0xe1, 0xce, 0x76, 0x28, 0x0b, 0x7b, 0xc5, 0x3f, + 0x5b, 0x48, 0x19, 0x76, 0x68, 0x31, 0x8e, 0x28, 0xff, 0x18, 0x24, 0xe3, + 0x91, 0xe7, 0x49, 0x0d, 0x10, 0xbd, 0x00, 0xc6, 0x58, 0xfd, 0xb6, 0x88, + 0x63, 0xbd, 0xb4, 0x4b, 0xb8, 0xed, 0xdd, 0xb7, 0x53, 0xce, 0x89, 0xdb, + 0x7f, 0xf4, 0xc3, 0x21, 0x31, 0xad, 0x20, 0x78, 0x06, 0x71, 0xaf, 0xc0, + 0xe3, 0xdc, 0xb8, 0xf4, 0x80, 0xc8, 0x33, 0x1d, 0x8b, 0xff, 0x5a, 0x92, + 0x68, 0x4d, 0xc1, 0x5b, 0x58, 0x3e, 0xf6, 0x7f, 0xba, 0x42, 0xa5, 0x6d, + 0xec, 0x03, 0x36, 0xc9, 0x3f, 0x83, 0x1f, 0x0c, 0x33, 0x57, 0x6a, 0x43, + 0x5f, 0x11, 0x72, 0x19, 0x2c, 0xda, 0x71, 0x58, 0xf2, 0x50, 0x50, 0x06, + 0x97, 0xd0, 0xdf, 0xd1, 0x4f, 0x0b, 0x00, 0x1a, 0xea, 0x85, 0x3b, 0x37, + 0x2f, 0xf0, 0x40, 0x52, 0xd9, 0x2a, 0xe8, 0x54, 0xa5, 0xee, 0x0f, 0x49, + 0x74, 0x39, 0x96, 0x5d, 0x60, 0x8f, 0x14, 0x59, 0x86, 0x59, 0x86, 0xfb, + 0x67, 0x71, 0x5c, 0x26, 0x5f, 0xe9, 0xab, 0x32, 0x77, 0x83, 0xdf, 0x02, + 0x19, 0x85, 0xae, 0x4d, 0x7d, 0x9c, 0x8d, 0x4f, 0x61, 0x05, 0x3c, 0x0c, + 0xc6, 0x74, 0x9e, 0x36, 0x33, 0xb8, 0x14, 0x85, 0xab, 0xa2, 0x0b, 0x5d, + 0x22, 0xf2, 0x50, 0x3e, 0xa4, 0x88, 0xac, 0x67, 0xf9, 0x06, 0xe5, 0x30, + 0x8e, 0xf9, 0x67, 0x34, 0xd5, 0x94, 0x5b, 0x35, 0xb7, 0x3d, 0x39, 0x5f, + 0x4e, 0xae, 0xfe, 0xf7, 0x57, 0xd3, 0x95, 0x7b, 0x0a, 0xd9, 0x92, 0x4a, + 0x66, 0x29, 0xa0, 0x18, 0x35, 0x54, 0x14, 0x44, 0x79, 0x72, 0xc3, 0xbc, + 0xa8, 0x1a, 0xd3, 0xa3, 0xbe, 0x6f, 0x9e, 0xcc, 0x68, 0xb6, 0x5f, 0xd4, + 0x42, 0xab, 0xe8, 0x09, 0x60, 0x57, 0x2e, 0xb2, 0x9a, 0x5b, 0x62, 0x38, + 0xfb, 0x0a, 0x35, 0x9c, 0x4f, 0xf7, 0xe0, 0xd2, 0x06, 0x04, 0x1f, 0x79, + 0x7f, 0xa7, 0x7b, 0xd3, 0x63, 0xc9, 0xbd, 0x16, 0x58, 0x38, 0x7b, 0xaa, + 0x08, 0xf3, 0x14, 0x6c, 0x25, 0xf8, 0xa5, 0xe9, 0x4b, 0x45, 0x34, 0x89, + 0x76, 0x74, 0xcb, 0x41, 0x9c, 0x2a, 0xd9, 0xca, 0xb3, 0x12, 0x46, 0x6d, + 0x85, 0x4d, 0x63, 0x2d, 0x24, 0x1b, 0x19, 0x6b, 0x3f, 0x61, 0x6b, 0x4b, + 0x15, 0x83, 0x2d, 0x8f, 0x61, 0xab, 0xd1, 0x55, 0x93, 0x4e, 0x26, 0xd6, + 0x7a, 0x0a, 0x8a, 0xff, 0x58, 0x44, 0xf7, 0x39, 0x31, 0x1a, 0xab, 0xa6, + 0x98, 0x31, 0x41, 0x03, 0xb6, 0xc9, 0xf5, 0x50, 0xe3, 0x7b, 0xc0, 0x59, + 0x74, 0x60, 0x91, 0xb4, 0x79, 0x02, 0x25, 0xc1, 0xb5, 0xbd, 0xcb, 0x6e, + 0x40, 0x61, 0xfe, 0x68, 0x29, 0x83, 0x1b, 0xd2, 0x49, 0xe1, 0x31, 0xde, + 0xdd, 0x53, 0xb0, 0xb8, 0x96, 0xa2, 0xce, 0xea, 0x8b, 0x66, 0x2c, 0x5a, + 0x80, 0x51, 0x0b, 0xc1, 0x2d, 0x9a, 0xfa, 0x9d, 0xc6, 0xcc, 0x2b, 0xbb, + 0xaa, 0xce, 0x98, 0xaa, 0x26, 0x15, 0x8f, 0x4a, 0xe7, 0xdb, 0x17, 0x6c, + 0xe5, 0x58, 0xc9, 0xae, 0xe4, 0x9c, 0x1d, 0xab, 0x59, 0x84, 0x3e, 0x27, + 0x76, 0x03, 0xe3, 0x82, 0x64, 0x6f, 0x6e, 0x6f, 0x63, 0xd2, 0x12, 0x84, + 0xe3, 0x9b, 0x9d, 0x7e, 0x53, 0x1a, 0x54, 0x8d, 0xc1, 0xf0, 0x94, 0xae, + 0xad, 0x8f, 0x6a, 0x12, 0x4e, 0xa7, 0x30, 0xdb, 0x55, 0xbe, 0x09, 0xe2, + 0x56, 0x08, 0xc4, 0x3a, 0xb0, 0x55, 0xb0, 0x24, 0x96, 0xa6, 0x3e, 0x28, + 0xd0, 0x35, 0xfb, 0x58, 0x47, 0xba, 0x2d, 0x51, 0xbb, 0x72, 0x20, 0x59, + 0xd2, 0xdd, 0x9c, 0xe2, 0xb5, 0x31, 0x90, 0xac, 0x74, 0x5d, 0x9f, 0x3d, + 0x8c, 0x1c, 0x96, 0xc0, 0x60, 0x61, 0xa8, 0xbb, 0x3c, 0xb3, 0x6d, 0x6d, + 0x92, 0x4a, 0xca, 0xbb, 0x60, 0x5e, 0x82, 0x0d, 0x7f, 0xab, 0x4b, 0x36, + 0x4c, 0x93, 0x0d, 0x88, 0x71, 0xaf, 0xb6, 0x53, 0xb0, 0x38, 0xb4, 0x1c, + 0xb4, 0x7b, 0xd4, 0x13, 0x32, 0x6c, 0xe4, 0xee, 0x6a, 0xb3, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x88, 0x83, 0x91, 0x4c, + 0x2e, 0x1e, 0xbe, 0xa4, 0xb5, 0x96, 0xff, 0x67, 0x50, 0xe9, 0x81, 0x0e, + 0x5d, 0x0e, 0xad, 0xc4, 0x1f, 0xeb, 0x98, 0x38, 0xcc, 0x54, 0x9d, 0x27, + 0xa6, 0xf1, 0x37, 0x23, 0xce, 0xb4, 0x5b, 0xff, 0x12, 0xb1, 0xb8, 0x35, + 0x5e, 0x03, 0x02, 0x04, 0xad, 0xa6, 0x6f, 0x43, 0xfc, 0xe4, 0xbe, 0x0c, + 0xe0, 0x93, 0xd5, 0xef, 0x09, 0xfa, 0x04, 0xe9, 0x5a, 0x22, 0xd4, 0x81, + 0xc1, 0x27, 0x4f, 0x5f, 0x6e, 0x83, 0x5a, 0x8a, 0x2d, 0xbb, 0x8f, 0xa4, + 0x91, 0xcc, 0x82, 0x37, 0x3b, 0x14, 0x98, 0x58, 0x86, 0x44, 0xb7, 0xa9, + 0x58, 0xf3, 0x3d, 0x49, 0x71, 0x7a, 0x37, 0xcd, 0xc5, 0xb9, 0xc9, 0x46, + 0xd5, 0xd4, 0x17, 0x60, 0x1a, 0xbf, 0x93, 0xa9, 0xe9, 0x08, 0x25, 0x40, + 0xd1, 0x65, 0xae, 0xdd, 0x85, 0xa6, 0xcc, 0x06, 0xca, 0x91, 0xe1, 0x63, + 0xf9, 0x6b, 0x15, 0xa8, 0x04, 0x61, 0xd2, 0xa6, 0x59, 0x21, 0x1a, 0x1c, + 0xc9, 0xa9, 0xa9, 0xc8, 0x54, 0x86, 0xac, 0xa5, 0xd6, 0x95, 0x39, 0x83, + 0x4b, 0x6b, 0x69, 0xa6, 0x94, 0xd8, 0xc0, 0xfb, 0x66, 0x0f, 0x3a, 0xbe, + 0xc7, 0xf3, 0xcc, 0xd5, 0xb7, 0x1b, 0x60, 0x02, 0x95, 0x45, 0x4a, 0x12, + 0xc9, 0xfe, 0x75, 0x7c, 0x1b, 0xb2, 0x86, 0x96, 0x28, 0x07, 0xa2, 0x18, + 0x7a, 0x6c, 0x90, 0x6f, 0x32, 0x0c, 0xc8, 0x34, 0xbc, 0x75, 0x4d, 0x96, + 0x03, 0xa6, 0x0f, 0x3d, 0x35, 0x1b, 0x64, 0x76, 0x95, 0x55, 0xff, 0x25, + 0xd4, 0x71, 0xcf, 0x8a, 0x73, 0x6d, 0x9b, 0x74, 0xfe, 0xff, 0x9e, 0x31, + 0x9e, 0x5e, 0x89, 0x5a, 0x1a, 0xeb, 0x8d, 0x06, 0x3b, 0xf2, 0xf6, 0x06, + 0x5d, 0xc3, 0xba, 0x04, 0xca, 0x0f, 0x07, 0x2c, 0xbd, 0x54, 0x52, 0xd9, + 0x1c, 0x2f, 0x0e, 0x13, 0x5e, 0x25, 0x13, 0xe5, 0xd7, 0x8e, 0x19, 0x42, + 0x1b, 0x52, 0x2e, 0xd2, 0x8f, 0xc5, 0x8e, 0x1c, 0x34, 0x2e, 0x4d, 0xd5, + 0x51, 0x7d, 0x91, 0x64, 0xbc, 0xb4, 0x0d, 0xc9, 0xe7, 0x1c, 0x6c, 0x47, + 0xe9, 0xbb, 0x67, 0x9a, 0x96, 0xde, 0xad, 0xff, 0xba, 0x35, 0x25, 0x6d, + 0x57, 0xa1, 0x93, 0xfe, 0xe2, 0x8d, 0x02, 0xeb, 0xf0, 0x2f, 0x54, 0xfd, + 0x46, 0xc0, 0x8f, 0xea, 0x32, 0x7b, 0x57, 0xda, 0xe0, 0x29, 0x1c, 0x19, + 0xba, 0xa4, 0xa6, 0x1c, 0x6e, 0xeb, 0x7a, 0xa8, 0x8a, 0xe1, 0xc6, 0x12, + 0xf5, 0xa3, 0x24, 0x1a, 0x96, 0xe1, 0x02, 0xc0, 0xf4, 0x7d, 0x14, 0x72, + 0xd6, 0x12, 0x8e, 0x6c, 0x8c, 0xd2, 0xfd, 0x88, 0x78, 0x48, 0xf3, 0x74, + 0x38, 0x86, 0x04, 0x68, 0x6d, 0x7c, 0xf4, 0x4c, 0x40, 0x17, 0xf6, 0x8f, + 0xb2, 0x6c, 0xd7, 0x66, 0x66, 0x3b, 0x38, 0xa1, 0xbb, 0x1e, 0xff, 0x72, + 0x1f, 0x64, 0x56, 0xc2, 0x53, 0x1c, 0x6f, 0x84, 0x2b, 0xbd, 0x23, 0xd9, + 0xb4, 0x6b, 0x87, 0x79, 0x99, 0xec, 0x81, 0x8d, 0x1a, 0x58, 0x00, 0xf0, + 0x2c, 0xc1, 0xc4, 0x57, 0x74, 0x0f, 0xce, 0x32, 0xe2, 0x5e, 0xae, 0x02, + 0x1c, 0xe8, 0x94, 0xc6, 0x44, 0xaa, 0x7b, 0x9a, 0x32, 0xb5, 0x33, 0xac, + 0xfc, 0x41, 0x65, 0xf2, 0xca, 0xcc, 0xc6, 0x74, 0x36, 0xb2, 0xc9, 0x0e, + 0x26, 0x73, 0xae, 0x68, 0x98, 0xa4, 0x36, 0xe8, 0x98, 0x39, 0xad, 0x05, + 0x3f, 0xca, 0x12, 0xcc, 0x86, 0xfd, 0xc6, 0x57, 0xf0, 0x02, 0x4e, 0x45, + 0xcb, 0x54, 0x34, 0xdd, 0x66, 0x26, 0xab, 0xda, 0x95, 0xa5, 0x85, 0xec, + 0x02, 0x03, 0xb6, 0x29, 0x30, 0x11, 0x40, 0x54, 0x9a, 0x6a, 0x87, 0x2e, + 0x97, 0xa1, 0x7e, 0xeb, 0x34, 0x39, 0x78, 0x3b, 0xbc, 0x5f, 0x8e, 0xc5, + 0x0e, 0x21, 0x29, 0x4b, 0xb7, 0x1b, 0xe7, 0x14, 0x08, 0x34, 0xb7, 0x9a, + 0x0a, 0xb2, 0x6c, 0x25, 0x76, 0xb5, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0xe2, 0x7d, 0x48, 0xdd, 0x1a, 0xcb, 0xb6, 0x5c, + 0x6f, 0xbe, 0x32, 0x9d, 0xd2, 0x2b, 0x9e, 0x10, 0x65, 0xd7, 0x1e, 0xec, + 0xc8, 0xb5, 0x10, 0x64, 0x8f, 0x5d, 0xef, 0xfe, 0x9b, 0x6c, 0x9b, 0x02, + 0x6a, 0x6d, 0xf7, 0x98, 0x7b, 0xf7, 0x17, 0xfd, 0x49, 0x1b, 0x6a, 0xc5, + 0x3c, 0xa0, 0xfc, 0xa8, 0x94, 0x95, 0xed, 0x48, 0x81, 0x04, 0x53, 0x8c, + 0xbe, 0xe4, 0x4e, 0xaf, 0xc1, 0x9d, 0xc3, 0xdf, 0xc2, 0xb5, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0xae, 0xb0, 0x67, 0x5b, + 0x99, 0x26, 0x07, 0xfb, 0x6c, 0x98, 0xfe, 0xbb, 0x35, 0xf1, 0x5b, 0x02, + 0xc6, 0x03, 0xfc, 0x97, 0x21, 0x16, 0x8d, 0x48, 0xd4, 0x4f, 0x03, 0xd9, + 0x7c, 0x9f, 0xa6, 0x1e, 0x6f, 0x5a, 0x58, 0x17, 0x6d, 0x26, 0xb4, 0xc5, + 0x4c, 0xe5, 0x93, 0x0a, 0x9c, 0xb2, 0x40, 0xbc, 0x60, 0xc7, 0x2b, 0xdb, + 0x3b, 0xc0, 0x3c, 0x5c, 0x44, 0x4b, 0xdd, 0x58, 0xbe, 0xdc, 0xc5, 0xb5, + 0x6a, 0xf9, 0x5e, 0x73, 0x07, 0x58, 0x8f, 0x45, 0x7b, 0xac, 0xba, 0x82, + 0x96, 0x49, 0x4d, 0x22, 0x70, 0x7a, 0x3d, 0x69, 0x26, 0x8b, 0x88, 0x13, + 0xf1, 0x8d, 0xfc, 0xdf, 0x73, 0xd5, 0x20, 0x3c, 0x52, 0x92, 0x16, 0xb1, + 0x6e, 0xb7, 0x41, 0xbe, 0x23, 0x9b, 0x51, 0xf7, 0xc9, 0x38, 0x8a, 0xc7, + 0x6e, 0x68, 0x82, 0xd1, 0x59, 0x50, 0x09, 0x4b, 0x44, 0x3b, 0x28, 0x06, + 0x60, 0x75, 0x7a, 0xe5, 0xa1, 0x36, 0xbb, 0x62, 0x44, 0xe3, 0xd0, 0x68, + 0x14, 0xea, 0xad, 0xf9, 0x18, 0xcc, 0xd5, 0x42, 0x5d, 0x18, 0x53, 0xe6, + 0x4a, 0xfe, 0xde, 0x32, 0xe1, 0xe7, 0xf8, 0x8c, 0x9d, 0x35, 0xf4, 0x4a, + 0xcb, 0x23, 0x2f, 0x91, 0xb5, 0xb0, 0xb2, 0x01, 0x5c, 0x22, 0x8c, 0x42, + 0x42, 0xd5, 0xf0, 0x82, 0x6f, 0x9f, 0x64, 0xe5, 0x99, 0x4d, 0x36, 0x0b, + 0xfc, 0x78, 0x38, 0x30, 0x47, 0x8f, 0x0b, 0x57, 0x86, 0x4f, 0x1b, 0xc9, + 0x05, 0x0e, 0x08, 0xc4, 0xf4, 0xab, 0x9e, 0x90, 0xb4, 0x4f, 0x36, 0x54, + 0xe8, 0xa1, 0x3f, 0x90, 0xd2, 0xf3, 0xb4, 0xb4, 0xdd, 0xf3, 0x43, 0x2f, + 0xc4, 0x43, 0xbb, 0x99, 0x8e, 0xb8, 0x61, 0x59, 0x5e, 0xfa, 0x1b, 0x3c, + 0xc1, 0xeb, 0x9d, 0x35, 0x62, 0x34, 0x82, 0x45, 0xef, 0x41, 0xe9, 0xfc, + 0x35, 0xae, 0xb4, 0x0b, 0xce, 0x52, 0x5b, 0x40, 0x7d, 0xdd, 0x86, 0x83, + 0x52, 0x74, 0x77, 0x11, 0xc2, 0x9b, 0x8c, 0xa3, 0x63, 0xc2, 0x2d, 0xdd, + 0x8c, 0x76, 0x13, 0xc5, 0xc0, 0xde, 0x3e, 0x6b, 0xe1, 0x0f, 0xeb, 0x0f, + 0x0a, 0x25, 0x41, 0x2f, 0x8b, 0x4a, 0x98, 0x30, 0xcb, 0x1a, 0x43, 0xa3, + 0xc1, 0xcc, 0x44, 0x9a, 0x6c, 0xdc, 0x92, 0x40, 0xc4, 0x7a, 0x1f, 0x8a, + 0x6f, 0x74, 0xf3, 0xf5, 0x52, 0x72, 0xf7, 0x81, 0x6e, 0x74, 0x75, 0xe6, + 0xea, 0xd9, 0x57, 0x91, 0xae, 0xf2, 0x3f, 0x35, 0x4b, 0x99, 0xd9, 0x3f, + 0x85, 0xe0, 0x92, 0xaa, 0x35, 0xac, 0x28, 0xbf, 0x43, 0xb8, 0xad, 0xc7, + 0xc5, 0xf6, 0x15, 0x2f, 0x7c, 0xfb, 0x34, 0x48, 0xf3, 0x04, 0x12, 0xf4, + 0x2f, 0x92, 0x74, 0xc8, 0xea, 0xbc, 0x24, 0x6e, 0x3b, 0x0e, 0x9e, 0xf0, + 0xaf, 0x02, 0x97, 0x95, 0xbc, 0x90, 0x7f, 0xc4, 0xf8, 0xe2, 0x04, 0x9a, + 0x8f, 0xfc, 0xbc, 0x50, 0xfe, 0xf7, 0x89, 0x17, 0x2c, 0xdb, 0xd6, 0x5e, + 0xbf, 0xd9, 0x8e, 0x89, 0x8b, 0x06, 0x1d, 0x0b, 0x81, 0x2a, 0x55, 0x5c, + 0x5f, 0xb6, 0xa6, 0xa5, 0xd2, 0xaa, 0x79, 0x9c, 0x39, 0x31, 0x76, 0x03, + 0x98, 0x42, 0xd6, 0xb7, 0x37, 0x1f, 0xc8, 0x51, 0x8a, 0x1c, 0x5d, 0xcd, + 0x9c, 0x78, 0xa4, 0x22, 0x6e, 0x12, 0x10, 0x0a, 0x33, 0xc9, 0xe0, 0xfe, + 0xfc, 0xe8, 0x15, 0xe7, 0xef, 0xd8, 0x6d, 0xc7, 0xc9, 0xc2, 0x8e, 0x18, + 0x82, 0x2f, 0xa6, 0x09, 0x8a, 0xdc, 0x41, 0x6b, 0x89, 0xea, 0xd9, 0xd6, + 0x96, 0xfd, 0xba, 0x6e, 0xae, 0x2d, 0x0c, 0xf9, 0x3c, 0x4c, 0x1a, 0xfa, + 0x98, 0x83, 0x51, 0x45, 0x9d, 0x1e, 0xa5, 0xc1, 0x81, 0x54, 0x37, 0x5d, + 0x28, 0xca, 0xa6, 0xfe, 0x48, 0xf4, 0x77, 0x17, 0x92, 0x1d, 0x0c, 0xb3, + 0x39, 0x77, 0x22, 0xd9, 0xc7, 0xc2, 0xaf, 0x70, 0x0a, 0xd3, 0xa6, 0x57, + 0x69, 0xfb, 0xb9, 0xe0, 0xc4, 0x73, 0x7a, 0x68, 0xee, 0x27, 0x6e, 0x3a, + 0x6e, 0xae, 0x32, 0xf6, 0x09, 0xb3, 0x0b, 0x40, 0x72, 0xc6, 0x26, 0x6e, + 0xc5, 0x88, 0x6b, 0xce, 0x99, 0x88, 0x60, 0x6f, 0x6e, 0xa9, 0xe6, 0xd7, + 0x35, 0x5e, 0x3b, 0x36, 0x0d, 0x14, 0xb8, 0x2f, 0xde, 0x67, 0xc8, 0x2e, + 0x52, 0xc1, 0xf1, 0x58, 0x87, 0x32, 0x2a, 0x52, 0x21, 0x27, 0x1e, 0x04, + 0xed, 0xc4, 0x82, 0xd7, 0xeb, 0x85, 0x12, 0x3e, 0xea, 0xd0, 0x07, 0xa0, + 0x80, 0x48, 0xe9, 0xbd, 0x9b, 0x3a, 0x8e, 0x8b, 0xa0, 0xfc, 0x07, 0xf0, + 0x69, 0x4e, 0xc7, 0x1d, 0xd9, 0x9a, 0x73, 0x18, 0x63, 0xb8, 0xe6, 0x4a, + 0xa0, 0x81, 0xf0, 0xdb, 0xb9, 0x88, 0xf4, 0x2b, 0x1f, 0x0d, 0xda, 0x31, + 0xc0, 0xb0, 0x55, 0x79, 0x56, 0x48, 0x22, 0xbb, 0x49, 0x7f, 0xb1, 0xf1, + 0xf6, 0x6f, 0x42, 0xd3, 0xba, 0x68, 0x3a, 0x8f, 0xe7, 0xac, 0x53, 0x30, + 0x96, 0xec, 0x51, 0x7d, 0xfc, 0xc0, 0x35, 0xe9, 0x59, 0xe7, 0x0e, 0xed, + 0x29, 0x46, 0x50, 0x3c, 0x4b, 0x36, 0xc6, 0x2a, 0xaa, 0x3b, 0xbe, 0xce, + 0xd3, 0xda, 0x4d, 0x65, 0xb0, 0xe8, 0x52, 0x68, 0xf0, 0x23, 0xde, 0x02, + 0x77, 0xb3, 0xcc, 0xce, 0x78, 0xdd, 0x8c, 0xf8, 0xbe, 0x5d, 0x0d, 0xa9, + 0xb6, 0x96, 0x85, 0xbf, 0x92, 0x2a, 0x6b, 0x1b, 0xe8, 0x76, 0x05, 0x13, + 0x30, 0xd8, 0x3d, 0x80, 0xaa, 0xa2, 0xa3, 0xbc, 0x07, 0xba, 0x9c, 0x75, + 0x5b, 0x42, 0x03, 0xd8, 0xde, 0x42, 0x44, 0xf7, 0x29, 0x43, 0x29, 0x0d, + 0x48, 0x2b, 0x02, 0xd0, 0xcc, 0xe9, 0x17, 0x47, 0x23, 0x73, 0x6d, 0xc5, + 0x91, 0x6d, 0x4e, 0xc5, 0xcf, 0xc3, 0x58, 0xaf, 0x6e, 0xa2, 0x9e, 0xe7, + 0xe1, 0x88, 0xac, 0x62, 0xff, 0xbc, 0x69, 0x57, 0xad, 0x0f, 0x08, 0xf8, + 0x32, 0xfd, 0x79, 0xcb, 0x30, 0xbc, 0xd2, 0xe5, 0x20, 0xd9, 0x0f, 0xd1, + 0x33, 0xbf, 0xe4, 0x49, 0x7a, 0x2b, 0x5c, 0xb3, 0x63, 0x13, 0x4d, 0xed, + 0x17, 0xe7, 0x5b, 0xf4, 0x36, 0x9d, 0x3c, 0x4e, 0x51, 0xb2, 0xf7, 0xf2, + 0xcd, 0xfb, 0xec, 0x42, 0x79, 0x46, 0xae, 0x18, 0x50, 0xdf, 0xbf, 0x5b, + 0xb1, 0x9a, 0x49, 0x22, 0xae, 0xe9, 0xf3, 0x86, 0x3f, 0xe0, 0xb4, 0xc6, + 0x9c, 0x08, 0xd6, 0xd9, 0xf4, 0x68, 0xbb, 0x33, 0x0e, 0x59, 0x3d, 0x76, + 0xf0, 0xd7, 0x54, 0x04, 0x19, 0x66, 0xee, 0x61, 0x11, 0x0d, 0x48, 0x10, + 0x21, 0x16, 0x7c, 0xac, 0x49, 0xab, 0xe0, 0x19, 0x85, 0x93, 0x48, 0x65, + 0x7c, 0x5e, 0x6c, 0x1a, 0xf5, 0xb0, 0xc6, 0x80, 0xa1, 0x2a, 0xd5, 0x71, + 0x42, 0xec, 0x2f, 0x25, 0xf7, 0xb8, 0x84, 0xcd, 0xf0, 0x5c, 0xcd, 0xee, + 0x44, 0xcb, 0xeb, 0x74, 0x96, 0x3c, 0xb0, 0x56, 0xcb, 0xaf, 0x7e, 0x9e, + 0x4a, 0x12, 0x06, 0xae, 0x57, 0x43, 0x2d, 0xb2, 0x11, 0x96, 0x05, 0xdb, + 0xb3, 0x1a, 0x01, 0xa7, 0x1d, 0x02, 0x81, 0x1c, 0x36, 0x41, 0x65, 0xf0, + 0x67, 0xd6, 0xd0, 0x0f, 0xec, 0x34, 0x7d, 0xd3, 0x89, 0xac, 0x60, 0x67, + 0x95, 0x81, 0x84, 0xe7, 0xbb, 0x9a, 0x59, 0x36, 0x3b, 0xde, 0xa4, 0x88, + 0xda, 0xf2, 0xd2, 0xa2, 0x0c, 0xba, 0xfb, 0x93, 0xbf, 0xc8, 0xad, 0xe8, + 0x57, 0xa0, 0x2b, 0xbb, 0x4e, 0xa9, 0x38, 0xe7, 0x86, 0x6b, 0x95, 0x34, + 0x24, 0x96, 0xc0, 0x09, 0xd9, 0xfd, 0x5f, 0x1c, 0x93, 0xd9, 0x72, 0xfa, + 0xc4, 0x14, 0x72, 0x9c, 0x19, 0x6f, 0xee, 0x12, 0x17, 0xee, 0x65, 0xb4, + 0x8c, 0x83, 0x39, 0x3c, 0x0f, 0xbf, 0x25, 0xcf, 0xee, 0x05, 0x8c, 0x6a, + 0x56, 0x18, 0xf0, 0x20, 0x72, 0xc1, 0xbf, 0xe4, 0xce, 0x37, 0xbf, 0x2b, + 0xba, 0x70, 0x1e, 0xc2, 0xc8, 0xcd, 0x58, 0xb9, 0x60, 0xc7, 0xfb, 0xd0, + 0xce, 0xb9, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, + 0x7c, 0x63, 0x50, 0x90, 0xcb, 0x9c, 0xce, 0x59, 0xb1, 0x47, 0xb0, 0x49, + 0x9b, 0xfc, 0xfb, 0x3d, 0x3d, 0x62, 0xcf, 0x58, 0x4c, 0x2a, 0x79, 0xf0, + 0x72, 0x7f, 0x81, 0x41, 0xac, 0x82, 0x2d, 0xa9, 0xf0, 0x0e, 0x4d, 0xd2, + 0xe0, 0xbd, 0xca, 0x17, 0xb7, 0x59, 0x9f, 0xdb, 0xfe, 0x51, 0x90, 0x88, + 0xb9, 0xeb, 0x4e, 0xac, 0x80, 0x30, 0x64, 0xc4, 0x49, 0xd1, 0xb6, 0x65, + 0x67, 0xef, 0x9d, 0x5c, 0x04, 0xe0, 0x9d, 0xbe, 0x47, 0x75, 0x9b, 0x6e, + 0x30, 0x76, 0xad, 0x37, 0x9a, 0x56, 0xff, 0xcd, 0x40, 0x26, 0x3e, 0xe2, + 0x7d, 0x30, 0x55, 0x09, 0x92, 0x25, 0x36, 0x2f, 0xf8, 0x55, 0xb8, 0x9b, + 0x66, 0x49, 0x41, 0x9d, 0x78, 0x6d, 0x3f, 0x54, 0x41, 0x01, 0x93, 0x9c, + 0x5e, 0x0c, 0x4a, 0x38, 0x79, 0x76, 0xb4, 0x98, 0xae, 0xf9, 0x99, 0x21, + 0x05, 0x6a, 0xfb, 0xbc, 0x44, 0xf7, 0xdc, 0x85, 0x5e, 0x5f, 0x18, 0x49, + 0x22, 0x11, 0x6d, 0xa5, 0x9e, 0x6b, 0x59, 0x60, 0xf8, 0x73, 0x8b, 0xcb, + 0x38, 0xbb, 0xc9, 0xbf, 0x49, 0x0e, 0x57, 0x65, 0x48, 0x41, 0x41, 0xa2, + 0x40, 0x67, 0x91, 0x1d, 0x54, 0xac, 0xa7, 0xef, 0x16, 0x8b, 0xc7, 0xd1, + 0xe6, 0xdb, 0xc5, 0x9c, 0xd4, 0x04, 0x67, 0xd8, 0x75, 0x21, 0x2b, 0x1d, + 0x11, 0xc1, 0x79, 0x45, 0xb4, 0x91, 0x7a, 0x97, 0x00, 0xde, 0xc6, 0xc5, + 0x8a, 0xd1, 0xd7, 0xea, 0xc1, 0x22, 0xe1, 0x58, 0x61, 0xf2, 0x89, 0x3d, + 0xdb, 0x04, 0x3d, 0xe4, 0xe9, 0xe7, 0xbf, 0x4b, 0x99, 0x8a, 0xc6, 0xf2, + 0x09, 0xc4, 0xe2, 0x6d, 0x0b, 0xda, 0x13, 0xfb, 0xff, 0xbf, 0x0b, 0xfc, + 0x78, 0x33, 0xb8, 0x7b, 0x3e, 0xd8, 0xba, 0x27, 0xba, 0xae, 0xdf, 0xce, + 0xea, 0x80, 0x08, 0x38, 0xd8, 0x33, 0x00, 0xa9, 0xb6, 0x88, 0x48, 0xa9, + 0x3b, 0x54, 0xf0, 0x95, 0xda, 0xba, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x04, 0x00, 0x00, 0xb1, 0xd7, 0x8d, 0x6c, 0xb9, 0x96, 0xdc, 0x64, + 0x9b, 0x0c, 0x74, 0x54, 0x59, 0x82, 0xf6, 0x6e, 0x7c, 0x4e, 0x23, 0x83, + 0x04, 0x2e, 0x49, 0xfb, 0x56, 0x4b, 0xcd, 0x0d, 0x76, 0x29, 0xb1, 0xce, + 0x40, 0xa3, 0xd0, 0x02, 0x16, 0x8e, 0x1c, 0x0a, 0x00, 0x5b, 0x8c, 0x06, + 0xf9, 0x07, 0x97, 0x12, 0x0c, 0x33, 0xd5, 0x48, 0x6d, 0xae, 0x7d, 0x2c, + 0x8f, 0x74, 0x32, 0x24, 0xcf, 0x91, 0xd7, 0xbe, 0xb2, 0x05, 0xcf, 0x2f, + 0x93, 0xd5, 0x43, 0x90, 0xce, 0x02, 0x97, 0xf8, 0x51, 0xb3, 0xba, 0x56, + 0x5d, 0x94, 0x41, 0xa4, 0x11, 0xf3, 0x21, 0xc0, 0xcc, 0x28, 0xf8, 0x5a, + 0x00, 0x0a, 0xd4, 0x53, 0xdd, 0xac, 0xfe, 0x25, 0x03, 0xea, 0x2b, 0x6b, + 0x9d, 0x7e, 0x1a, 0xe1, 0x5f, 0x5c, 0xa7, 0x47, 0xa2, 0x72, 0x4f, 0x92, + 0x60, 0x25, 0x7c, 0x1c, 0xa5, 0x34, 0xa6, 0x86, 0x0e, 0xda, 0x8f, 0x3f, + 0xec, 0xe2, 0xe4, 0xad, 0xa9, 0x41, 0xcc, 0x3d, 0x94, 0x43, 0xfd, 0x28, + 0xd8, 0xb0, 0x0f, 0x05, 0x9e, 0x2b, 0x27, 0x3f, 0xe0, 0x84, 0xbc, 0x9e, + 0x7a, 0xa5, 0x83, 0x3d, 0x3b, 0xac, 0x83, 0xd3, 0x16, 0x92, 0x8c, 0xd2, + 0x4a, 0x81, 0xdd, 0xba, 0x0a, 0xb7, 0xc5, 0x9f, 0x83, 0x0f, 0x78, 0xb8, + 0xab, 0x2d, 0xca, 0xf8, 0x6c, 0x06, 0xd7, 0x82, 0xb8, 0x61, 0x7d, 0x2a, + 0x31, 0x3a, 0x39, 0x97, 0x5f, 0xc7, 0x00, 0x6e, 0x46, 0xf2, 0xc5, 0x12, + 0x71, 0x55, 0x5b, 0x10, 0xaf, 0xbb, 0x07, 0x4c, 0x2f, 0xa3, 0x51, 0x53, + 0x22, 0x20, 0xab, 0xed, 0x02, 0x95, 0xc6, 0x5f, 0xaa, 0xb8, 0xc0, 0xcb, + 0xe5, 0xe0, 0x25, 0x97, 0xf7, 0xda, 0x1d, 0xd8, 0x5a, 0xff, 0x76, 0x0c, + 0x3e, 0x33, 0x1b, 0x7a, 0x15, 0xb8, 0x34, 0x75, 0xcf, 0xe9, 0xf3, 0x53, + 0x61, 0x03, 0x2d, 0x52, 0x29, 0x69, 0x3a, 0xc3, 0xd9, 0x22, 0xc0, 0x2d, + 0x80, 0xed, 0x66, 0xc4, 0xf4, 0x89, 0x60, 0x14, 0xdb, 0xec, 0x7d, 0xcc, + 0x99, 0x5c, 0x94, 0x27, 0xab, 0xed, 0xd2, 0x17, 0xf4, 0x36, 0xfc, 0x7e, + 0x99, 0x98, 0xb6, 0x86, 0xb6, 0x7c, 0x54, 0xd6, 0xec, 0xb5, 0xad, 0x62, + 0xcc, 0xb0, 0xf7, 0x8c, 0x52, 0x99, 0xf2, 0x44, 0x27, 0x3a, 0xb0, 0xff, + 0x8f, 0x09, 0xae, 0xe1, 0x61, 0xd8, 0x9f, 0xdd, 0x2f, 0x6b, 0xea, 0xd0, + 0x12, 0x70, 0x8c, 0x9d, 0x8f, 0x4c, 0x36, 0x98, 0x1e, 0x2e, 0xb5, 0x50, + 0x63, 0x33, 0x9c, 0x4b, 0xc3, 0xd4, 0xa0, 0xe6, 0x96, 0x96, 0x75, 0xfd, + 0x8a, 0xc4, 0x0c, 0xa7, 0xea, 0x9d, 0xf1, 0x23, 0x9e, 0x38, 0xff, 0x1a, + 0x67, 0x36, 0x5f, 0x5f, 0x17, 0x88, 0x1a, 0x43, 0x25, 0xea, 0x76, 0xb5, + 0xcd, 0xce, 0x43, 0xf8, 0x71, 0x2b, 0xdb, 0xf0, 0xcd, 0x76, 0xbd, 0x94, + 0x57, 0xdb, 0x77, 0xcd, 0xb2, 0x8f, 0xd1, 0xc0, 0xeb, 0x00, 0x61, 0x7f, + 0x66, 0xb0, 0x43, 0x6e, 0xe0, 0x9f, 0x11, 0x0e, 0x65, 0xf7, 0x4e, 0x00, + 0x74, 0xc3, 0xeb, 0xb1, 0xeb, 0x0c, 0x24, 0x5d, 0x15, 0x56, 0x16, 0x47, + 0x87, 0xcf, 0x34, 0xbe, 0x2a, 0xdd, 0x77, 0x55, 0xa4, 0x09, 0x15, 0x79, + 0x8c, 0xaa, 0xce, 0x32, 0x90, 0x9b, 0x16, 0x40, 0x94, 0x7f, 0x19, 0x27, + 0xbc, 0xbf, 0x45, 0x4b, 0xa5, 0xf0, 0xd0, 0x9e, 0x5b, 0xb9, 0x46, 0x6e, + 0x72, 0x8f, 0x49, 0x3b, 0x7a, 0xc1, 0x92, 0xb0, 0xd5, 0x25, 0x1b, 0x0b, + 0xf3, 0xd0, 0x8a, 0x47, 0x8b, 0xbe, 0xa4, 0xf9, 0x6a, 0x09, 0x84, 0x9a, + 0x5b, 0x5b, 0xea, 0xbb, 0x6f, 0xd8, 0xaf, 0xcd, 0x67, 0x9b, 0x79, 0x7c, + 0x8f, 0xcc, 0xd7, 0x5f, 0x3a, 0xc3, 0xd0, 0xb7, 0xba, 0x28, 0x83, 0x81, + 0x4a, 0x05, 0x51, 0xaf, 0xa0, 0x52, 0x34, 0xe3, 0x4f, 0xec, 0x82, 0xdc, + 0x97, 0xd8, 0x69, 0xb2, 0x0d, 0x68, 0x35, 0x87, 0x58, 0xc0, 0xcf, 0x58, + 0x0d, 0xf6, 0x6b, 0x6d, 0x2a, 0xc0, 0x72, 0xe4, 0x90, 0x8c, 0x7b, 0x45, + 0xba, 0xf1, 0x13, 0x6f, 0x8c, 0xd2, 0xdd, 0xc5, 0x8e, 0xc8, 0xec, 0xf9, + 0xfb, 0xde, 0xe5, 0xaa, 0xcb, 0xc0, 0xff, 0x77, 0x2d, 0x99, 0xb1, 0x69, + 0x7f, 0xe3, 0x38, 0x61, 0x35, 0xb6, 0x45, 0xdd, 0x73, 0x45, 0x84, 0x89, + 0x1b, 0x96, 0x7e, 0x6a, 0x1d, 0xd9, 0xe6, 0x76, 0xa8, 0x16, 0x0f, 0x42, + 0xc9, 0x41, 0xec, 0x5d, 0x25, 0x01, 0xb0, 0x45, 0xa6, 0xaa, 0x69, 0x87, + 0x11, 0xa1, 0xb8, 0x9e, 0x68, 0x48, 0x68, 0xe9, 0xb5, 0xc2, 0xff, 0x83, + 0x8f, 0x71, 0xb9, 0xd7, 0xbb, 0xae, 0x59, 0x8b, 0x1b, 0x4c, 0x44, 0xd8, + 0xe3, 0xce, 0xab, 0x88, 0xfb, 0x64, 0xd9, 0x61, 0x5a, 0x7d, 0xce, 0x3a, + 0x27, 0xb5, 0xa3, 0xfd, 0x5d, 0xa3, 0xb8, 0xa1, 0x15, 0x63, 0x0b, 0x75, + 0x39, 0xc3, 0xa4, 0xfb, 0x60, 0x53, 0xfd, 0x11, 0x21, 0x35, 0x0f, 0x19, + 0x28, 0x14, 0xcd, 0x8a, 0xcf, 0x33, 0xaa, 0x4f, 0x6a, 0x1e, 0x56, 0x87, + 0xd5, 0x6e, 0x43, 0x9b, 0xa3, 0x72, 0x95, 0x8c, 0x34, 0xa2, 0xac, 0x11, + 0x76, 0x95, 0xd7, 0xdd, 0xbf, 0x10, 0xf4, 0x0f, 0x2a, 0x64, 0xd2, 0x4d, + 0x7b, 0xc6, 0x9b, 0x7d, 0xf7, 0xa5, 0xb3, 0x84, 0x9a, 0x9a, 0x5e, 0xcf, + 0x7f, 0x95, 0x6d, 0x44, 0xd1, 0xb2, 0x19, 0xbb, 0xed, 0x37, 0x42, 0x4b, + 0x4b, 0x6d, 0xb7, 0x10, 0x02, 0x5f, 0x00, 0x1f, 0x24, 0xce, 0xb2, 0x8b, + 0x3e, 0x7d, 0xc6, 0x6e, 0x6c, 0x90, 0x75, 0xad, 0x3f, 0x9d, 0x63, 0x04, + 0x76, 0x20, 0x7a, 0x56, 0x48, 0xa1, 0x6a, 0x37, 0x74, 0xd2, 0xb7, 0x4f, + 0xa3, 0x64, 0x62, 0xaa, 0xce, 0x75, 0x8c, 0x15, 0x75, 0x79, 0xa0, 0xbd, + 0xdd, 0x01, 0x46, 0xca, 0xa0, 0x31, 0x1a, 0x16, 0x1f, 0xef, 0x8b, 0xc6, + 0x54, 0x57, 0xfa, 0x6e, 0x43, 0xdf, 0xb0, 0x99, 0xed, 0xa4, 0xcb, 0xeb, + 0x91, 0x35, 0x14, 0x0c, 0xa9, 0x1d, 0xb5, 0xa9, 0x32, 0x99, 0xe3, 0x89, + 0x74, 0xaa, 0xa4, 0x65, 0x1e, 0x82, 0x47, 0xfa, 0x37, 0x23, 0xe5, 0x86, + 0xb6, 0xc0, 0xb6, 0x89, 0x9a, 0xd9, 0xae, 0x29, 0x39, 0x7b, 0x66, 0xc7, + 0x5b, 0x02, 0x08, 0x86, 0xd4, 0xf0, 0x75, 0xc2, 0x05, 0x86, 0xc3, 0x75, + 0xd2, 0x2a, 0x1e, 0xec, 0x6e, 0x75, 0x29, 0x58, 0x8c, 0x25, 0x3b, 0x95, + 0x21, 0xde, 0x42, 0xd5, 0xb7, 0x15, 0x30, 0x09, 0x49, 0x78, 0x55, 0xd5, + 0xf2, 0x30, 0x80, 0x93, 0x8a, 0xce, 0x84, 0x27, 0xdb, 0x4a, 0x09, 0x30, + 0x0c, 0x7f, 0x4d, 0xd1, 0x0f, 0xda, 0x66, 0x58, 0xe1, 0x01, 0xfd, 0x75, + 0x83, 0xf5, 0x39, 0x2e, 0xe2, 0x6b, 0xde, 0xff, 0x20, 0x8a, 0xf7, 0xcc, + 0x81, 0x8e, 0x99, 0xb4, 0xeb, 0x76, 0x74, 0x38, 0x2b, 0xe0, 0x6d, 0x61, + 0x8f, 0x39, 0x59, 0x10, 0x7d, 0xb5, 0xd3, 0x14, 0x96, 0x04, 0x1d, 0x22, + 0x89, 0xef, 0x15, 0x7c, 0x28, 0x5a, 0xd6, 0x8d, 0xf3, 0xb7, 0x6a, 0x9a, + 0xce, 0x21, 0x77, 0xfd, 0x4f, 0x22, 0x26, 0x28, 0xb8, 0xb5, 0xb3, 0x73, + 0xfd, 0x2a, 0x7b, 0x42, 0x26, 0x77, 0x41, 0x93, 0xed, 0xf9, 0x8f, 0xa9, + 0x92, 0xd5, 0x9f, 0x2e, 0x60, 0xec, 0x60, 0x98, 0xf1, 0xd5, 0x11, 0xe2, + 0xe0, 0xd7, 0x45, 0xa7, 0xe4, 0xf2, 0x82, 0x61, 0x2f, 0x41, 0x1b, 0xd9, + 0x8e, 0x78, 0xd5, 0x6b, 0x68, 0x74, 0xf0, 0xc3, 0x83, 0x01, 0x16, 0x60, + 0x6e, 0x34, 0x88, 0x45, 0x8a, 0x86, 0x44, 0x5b, 0xa5, 0xa8, 0x55, 0xbc, + 0xfa, 0x8f, 0xbd, 0x93, 0x95, 0x3f, 0xab, 0x19, 0x54, 0x8f, 0x06, 0x8e, + 0xca, 0x0b, 0x4a, 0x18, 0x3f, 0x7a, 0x9c, 0x3f, 0xe6, 0xbe, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x81, 0x32, 0x41, 0x46, + 0x59, 0x26, 0xf4, 0xef, 0x93, 0x9f, 0x04, 0xc2, 0x67, 0x13, 0x32, 0x45, + 0xc0, 0x79, 0x70, 0x27, 0x21, 0x2b, 0xaf, 0x35, 0xf3, 0xc4, 0x88, 0x52, + 0x28, 0xea, 0xca, 0x8a, 0x08, 0x01, 0x6f, 0x61, 0xab, 0x10, 0xa3, 0xf0, + 0x6b, 0x3b, 0x54, 0x64, 0xf1, 0x63, 0x83, 0x38, 0x2b, 0x26, 0x18, 0x5a, + 0x67, 0xc4, 0x67, 0x38, 0x3f, 0x2c, 0x9a, 0xc9, 0x48, 0x33, 0x77, 0xb4, + 0xb2, 0xc2, 0xc7, 0x08, 0x21, 0x5e, 0xc4, 0x19, 0x59, 0xe1, 0xfa, 0x32, + 0xa4, 0x4c, 0x3e, 0xba, 0x65, 0x92, 0x98, 0x39, 0x71, 0x2f, 0x99, 0x08, + 0xf8, 0xb3, 0x7a, 0x03, 0x53, 0xd7, 0x68, 0xb2, 0x5e, 0xb0, 0xef, 0xe0, + 0x1e, 0x7d, 0xb2, 0x23, 0x5d, 0x2b, 0xd7, 0x09, 0xa6, 0x78, 0xa4, 0x7c, + 0x08, 0xed, 0x8a, 0xf6, 0x96, 0xa0, 0x10, 0x17, 0x62, 0x8b, 0x8a, 0xa0, + 0xac, 0x22, 0x67, 0x02, 0xa8, 0x66, 0x1a, 0xb5, 0x02, 0xde, 0xa5, 0xfa, + 0x69, 0x29, 0x5f, 0x24, 0x89, 0x46, 0x68, 0xd6, 0x51, 0x2a, 0xfe, 0x88, + 0xf0, 0x40, 0xde, 0xd1, 0x12, 0x2e, 0xed, 0x13, 0x7b, 0x49, 0xf6, 0xe1, + 0x7a, 0xcf, 0x61, 0xcb, 0x70, 0x9d, 0xaa, 0x51, 0x07, 0xc2, 0x54, 0x76, + 0x89, 0x29, 0x94, 0x29, 0x8b, 0x0e, 0xf5, 0xe8, 0x81, 0xc7, 0xdb, 0x59, + 0x1e, 0x75, 0xda, 0x6a, 0x94, 0x18, 0x16, 0xae, 0xbb, 0x43, 0x87, 0x56, + 0x66, 0x8b, 0x84, 0xe9, 0xa9, 0xd0, 0xd2, 0x8f, 0x5b, 0xbf, 0x1d, 0x24, + 0x3a, 0xb7, 0x64, 0xff, 0xe9, 0x22, 0x21, 0x65, 0xaf, 0x2b, 0x45, 0x8d, + 0x28, 0xea, 0xbc, 0x07, 0x10, 0x6e, 0xfb, 0x4d, 0x6f, 0x35, 0xe5, 0xeb, + 0x5d, 0x29, 0x72, 0xe1, 0x94, 0xad, 0xed, 0x25, 0xd7, 0x39, 0x63, 0x32, + 0x37, 0x0b, 0xb2, 0xd7, 0x54, 0x1f, 0xe4, 0x0d, 0xe7, 0xb3, 0xd1, 0xa6, + 0x2a, 0xcf, 0x8e, 0x97, 0xf1, 0xa8, 0xfc, 0xb1, 0x61, 0xdc, 0xb4, 0x8f, + 0x29, 0xa2, 0x68, 0x4a, 0xe6, 0x2f, 0x8a, 0x69, 0x2c, 0xa1, 0x1d, 0xe2, + 0x9e, 0x65, 0x71, 0xb7, 0x83, 0xef, 0x63, 0xf5, 0x36, 0xdc, 0xa0, 0x94, + 0x5a, 0x45, 0x8a, 0x85, 0x5e, 0x28, 0x86, 0x21, 0xd2, 0xbf, 0x7a, 0x2f, + 0x76, 0x1c, 0x2a, 0x15, 0xb2, 0xe8, 0xaf, 0x63, 0x37, 0xbe, 0xd8, 0x0a, + 0xef, 0x54, 0xee, 0xe6, 0xd9, 0xb3, 0xdb, 0x41, 0x55, 0xba, 0xd8, 0x14, + 0x7c, 0x10, 0x61, 0x06, 0x40, 0x45, 0x69, 0x37, 0x60, 0xf7, 0x6a, 0x7a, + 0x23, 0x70, 0x30, 0x57, 0x3e, 0xe5, 0x12, 0x24, 0xbc, 0x5e, 0x82, 0x89, + 0xd8, 0x37, 0xc9, 0x33, 0xb9, 0x38, 0xa5, 0xba, 0xed, 0xdd, 0x93, 0x58, + 0x81, 0x15, 0xec, 0x15, 0x70, 0x2f, 0x30, 0xfa, 0xaf, 0xf7, 0xf5, 0xcb, + 0x41, 0x74, 0xea, 0xc0, 0x91, 0xbe, 0x53, 0x4c, 0xc2, 0x74, 0x1b, 0x5b, + 0x8c, 0x74, 0xd8, 0xc3, 0x4a, 0x12, 0xaa, 0x57, 0xd6, 0x61, 0xb1, 0xb8, + 0x81, 0x5d, 0x81, 0x37, 0x1e, 0x5b, 0x3d, 0x5a, 0xbc, 0xa6, 0xb2, 0x27, + 0xe3, 0x01, 0x4c, 0xf0, 0xad, 0x7b, 0xdf, 0x50, 0xf9, 0xd7, 0xb7, 0xcc, + 0xa8, 0x5c, 0x3d, 0x9a, 0xb7, 0x60, 0x3e, 0x63, 0x3f, 0x6a, 0x08, 0x0b, + 0x82, 0xdc, 0x3e, 0xfa, 0x24, 0x33, 0xd3, 0x01, 0xbf, 0xef, 0xeb, 0x52, + 0x3f, 0x91, 0x61, 0xda, 0xe2, 0x26, 0x10, 0xdf, 0xe4, 0x9b, 0x77, 0x91, + 0x22, 0xc5, 0x4e, 0x9c, 0x0b, 0x32, 0xff, 0x27, 0x85, 0x85, 0x0c, 0x99, + 0x50, 0x8f, 0xad, 0x5d, 0x06, 0x18, 0x52, 0xb4, 0x64, 0x09, 0xc4, 0xa4, + 0x84, 0xd4, 0x81, 0x07, 0x0a, 0x97, 0x55, 0xf8, 0x96, 0x52, 0xb2, 0x9a, + 0xf4, 0x06, 0x2c, 0x9a, 0x3b, 0x8b, 0xaa, 0x67, 0x18, 0x3a, 0xee, 0xbc, + 0xca, 0x8f, 0x46, 0xf6, 0x4a, 0x33, 0x5b, 0x56, 0x09, 0xb2, 0x72, 0x87, + 0xdb, 0xbb, 0x57, 0x67, 0x53, 0x82, 0x77, 0x31, 0x66, 0xbb, 0xf1, 0x33, + 0x6d, 0x55, 0x82, 0xaa, 0x80, 0xd4, 0x4d, 0xb8, 0xab, 0xbd, 0x2a, 0xda, + 0x10, 0x3a, 0xc8, 0xf0, 0x14, 0x1e, 0xcb, 0x8e, 0x76, 0x6c, 0xc8, 0x74, + 0x05, 0xb3, 0x51, 0xbd, 0x63, 0x06, 0x69, 0x05, 0x2a, 0x21, 0xd6, 0x2f, + 0xe4, 0x38, 0xae, 0xf8, 0xd4, 0xe9, 0xa7, 0xe8, 0xc8, 0x5a, 0x65, 0x7d, + 0x54, 0x34, 0x33, 0x0d, 0xf6, 0x07, 0xd6, 0x8c, 0xe5, 0x72, 0x9b, 0xfb, + 0x60, 0x49, 0xd2, 0xaf, 0xb4, 0x17, 0xc4, 0x74, 0x8d, 0xe5, 0x54, 0xda, + 0x96, 0x56, 0x7d, 0x97, 0x62, 0xe8, 0xec, 0x0d, 0x2b, 0x02, 0x2e, 0x59, + 0xf8, 0xa1, 0x06, 0x6a, 0xb6, 0x3e, 0x15, 0xeb, 0x64, 0x1a, 0x48, 0x3d, + 0x53, 0x2c, 0x42, 0x3b, 0x97, 0xa1, 0x3f, 0x47, 0x8b, 0x74, 0x87, 0x8b, + 0x96, 0x63, 0x08, 0x4c, 0x99, 0x38, 0x5a, 0xb6, 0x93, 0xa8, 0xcc, 0xee, + 0x62, 0x3a, 0x00, 0x6d, 0x5c, 0xab, 0x77, 0x3c, 0x46, 0xae, 0x6e, 0xeb, + 0xf1, 0xf9, 0x63, 0xf1, 0xa2, 0x31, 0x21, 0x38, 0xc3, 0x4f, 0xe2, 0x3a, + 0x33, 0x7f, 0xe7, 0xc6, 0x69, 0xd5, 0x1c, 0x7e, 0x5b, 0x4f, 0xb1, 0x50, + 0x3b, 0xbe, 0x31, 0xa7, 0x42, 0xa3, 0x97, 0x7b, 0xe3, 0x90, 0xd0, 0x07, + 0xfd, 0x05, 0xb9, 0xf2, 0x47, 0xc4, 0xc8, 0xdd, 0x1c, 0x3c, 0xa4, 0x22, + 0x96, 0x04, 0xca, 0x28, 0x17, 0xcc, 0x5c, 0x49, 0x7e, 0xc6, 0x93, 0x98, + 0xd3, 0x8b, 0xd2, 0xf6, 0x4a, 0xb6, 0xbe, 0x8d, 0xa2, 0xdd, 0xb6, 0x7c, + 0x66, 0x0c, 0x29, 0xcb, 0x1d, 0x98, 0xf6, 0xe4, 0xe5, 0x30, 0x4c, 0x84, + 0xbf, 0x6f, 0x71, 0x4e, 0xc2, 0x12, 0x9f, 0x35, 0xd6, 0xf8, 0xc6, 0x30, + 0xe9, 0x9e, 0x1a, 0x8a, 0x2f, 0xd1, 0x96, 0xb3, 0x3c, 0x0f, 0xf5, 0x78, + 0xa7, 0xe0, 0xbd, 0x4b, 0xe0, 0xd8, 0x3d, 0x57, 0xa5, 0x44, 0xa0, 0xd9, + 0x10, 0x79, 0xd2, 0x10, 0x50, 0xc7, 0x77, 0x73, 0x09, 0xf8, 0xb4, 0xcf, + 0x66, 0xe3, 0x0c, 0xfb, 0x96, 0xf8, 0x52, 0xb3, 0x7e, 0x44, 0xf0, 0x03, + 0x54, 0xd4, 0xa2, 0x57, 0x38, 0x8a, 0x96, 0xfc, 0x7c, 0x4c, 0x9f, 0x3a, + 0xf2, 0xa2, 0x48, 0xbb, 0x3e, 0xd1, 0x11, 0x2c, 0xab, 0xdf, 0x53, 0x96, + 0xac, 0x58, 0x33, 0xb9, 0xdd, 0xd2, 0x4f, 0x8a, 0x0a, 0x89, 0x0e, 0xd3, + 0x6f, 0x58, 0x8c, 0xa1, 0x0a, 0x0b, 0xa7, 0xd7, 0x1f, 0x0a, 0x70, 0xe3, + 0x43, 0x12, 0x56, 0xb8, 0x6c, 0xf8, 0x75, 0x4e, 0x2b, 0xb0, 0x17, 0x29, + 0xe4, 0x95, 0x85, 0xd8, 0x85, 0x95, 0x63, 0x55, 0xa8, 0x82, 0xf0, 0xe7, + 0x7d, 0xf3, 0xf1, 0x78, 0x66, 0xd1, 0x92, 0x71, 0x99, 0xad, 0x30, 0x94, + 0xe9, 0x54, 0x2c, 0xe1, 0x57, 0xf3, 0x6a, 0xe6, 0x0c, 0x5e, 0xc7, 0x58, + 0xba, 0xb7, 0x61, 0xd3, 0x74, 0x72, 0x96, 0x06, 0x0b, 0x01, 0x3d, 0xc2, + 0xa1, 0xb4, 0x38, 0x81, 0x19, 0x44, 0xbc, 0x84, 0x52, 0x22, 0xc9, 0x67, + 0x81, 0x99, 0xfb, 0x0a, 0xc2, 0xff, 0x50, 0x67, 0xbe, 0x38, 0x5e, 0x13, + 0x16, 0x60, 0x83, 0x35, 0xb9, 0x2f, 0xa9, 0x55, 0xbb, 0x30, 0x6b, 0x19, + 0xfc, 0x2a, 0x40, 0x24, 0x74, 0x20, 0x57, 0x78, 0xb9, 0x55, 0xb7, 0x70, + 0x86, 0x65, 0x43, 0x1c, 0x76, 0x2e, 0x91, 0x83, 0x5e, 0x33, 0xc2, 0xd4, + 0xcc, 0xb5, 0x1c, 0x45, 0xaf, 0xa3, 0x87, 0x95, 0x9b, 0x77, 0x50, 0x44, + 0x7e, 0xdd, 0xca, 0x3f, 0x51, 0x21, 0xae, 0xf2, 0x15, 0xa9, 0x32, 0x94, + 0xca, 0xde, 0x3b, 0x97, 0x13, 0x6b, 0xff, 0xe0, 0x79, 0x39, 0x40, 0xf0, + 0x66, 0x7d, 0x5e, 0xef, 0xec, 0x0a, 0x35, 0xd2, 0x0d, 0x09, 0x19, 0x13, + 0xf2, 0xc2, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0xdc, 0x07, 0x2e, 0x46, 0xab, 0x4d, 0x6d, 0xf7, 0x24, 0xba, 0x02, 0xe3, + 0xc5, 0xe3, 0xed, 0x64, 0xc6, 0x77, 0x5a, 0x14, 0xae, 0x38, 0x52, 0x8c, + 0x16, 0x2c, 0x52, 0x0e, 0xf6, 0x65, 0x99, 0xcc, 0xf6, 0x9f, 0x77, 0xcc, + 0x2e, 0xaf, 0x14, 0xd1, 0xf0, 0x0f, 0xa7, 0x3e, 0x5b, 0x74, 0xff, 0xb9, + 0xd3, 0x30, 0x02, 0x5e, 0x52, 0xc8, 0x6f, 0x57, 0xef, 0x28, 0xf5, 0xfa, + 0x9e, 0x70, 0x00, 0xfc, 0x3e, 0xc3, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x04, 0x00, 0x00, 0xaa, 0x9f, 0x86, 0xb0, 0x6d, 0xa1, 0x0c, 0xfa, + 0xef, 0xb3, 0x6a, 0x50, 0xa6, 0xfe, 0xff, 0xa9, 0x61, 0x0b, 0x18, 0x72, + 0xee, 0xc6, 0xcd, 0x3a, 0x34, 0x5e, 0xa8, 0x81, 0x31, 0x54, 0x25, 0x05, + 0xc1, 0xd9, 0x66, 0x3d, 0x17, 0xbb, 0x03, 0x21, 0x07, 0x69, 0x3a, 0x37, + 0xe8, 0xd4, 0x6a, 0x68, 0xe1, 0xa3, 0x19, 0x5a, 0x8d, 0x14, 0x11, 0x09, + 0xef, 0xae, 0xfe, 0x94, 0x19, 0x8a, 0xe4, 0xb9, 0x6e, 0xe8, 0xfa, 0x12, + 0x2a, 0x5d, 0x00, 0x29, 0x27, 0x6d, 0x5a, 0xa5, 0x09, 0x34, 0x79, 0x2b, + 0xa8, 0xcc, 0x42, 0xb4, 0xde, 0xe0, 0x91, 0xb9, 0x06, 0x0c, 0x11, 0x17, + 0x25, 0x7a, 0x35, 0x57, 0x51, 0x40, 0xf3, 0xc7, 0xc6, 0x4a, 0x69, 0x98, + 0x2b, 0x2b, 0x3e, 0x5d, 0x32, 0xd8, 0x8f, 0xb0, 0x1d, 0xee, 0x77, 0xe3, + 0xaf, 0x4f, 0x71, 0x05, 0x04, 0xd2, 0xff, 0x51, 0xed, 0xa4, 0x69, 0x50, + 0x24, 0x2a, 0xe5, 0xaa, 0xbb, 0xc6, 0x7a, 0x7f, 0xb2, 0xdf, 0x1d, 0xc2, + 0x02, 0x2e, 0x52, 0xd1, 0xd9, 0x5b, 0xe7, 0x6c, 0x50, 0x31, 0x4e, 0xdf, + 0x8e, 0x3f, 0x37, 0xfc, 0xf5, 0x34, 0x0e, 0xdb, 0x4c, 0x5d, 0x7d, 0xc8, + 0xe4, 0x72, 0x40, 0xcb, 0x95, 0xa5, 0x41, 0xeb, 0x78, 0x5f, 0x64, 0x20, + 0x55, 0x19, 0xc7, 0xf9, 0x9c, 0x71, 0x40, 0x8f, 0xcc, 0x2d, 0x86, 0xc0, + 0xf4, 0x36, 0x2b, 0x0e, 0x28, 0xb4, 0xad, 0x1b, 0xde, 0x60, 0x67, 0x03, + 0x0f, 0x7c, 0x18, 0xd9, 0xc3, 0x73, 0x67, 0x0d, 0x44, 0x3d, 0xbe, 0x7c, + 0xcf, 0x96, 0x22, 0x0b, 0x0e, 0x3a, 0x0b, 0xcf, 0x04, 0x95, 0x92, 0x7d, + 0x4b, 0xa2, 0x6a, 0x0b, 0x47, 0x72, 0x73, 0xa8, 0x9b, 0x96, 0x3d, 0xc6, + 0x03, 0x34, 0xb1, 0x69, 0xc2, 0x50, 0x60, 0x89, 0x8c, 0x55, 0x8f, 0x8e, + 0x74, 0xa8, 0x9e, 0x25, 0xe4, 0x0e, 0x73, 0xef, 0x4f, 0x51, 0xbe, 0xed, + 0x5c, 0x14, 0xd3, 0xfa, 0x94, 0x58, 0x8d, 0x5c, 0xa0, 0xb1, 0xfc, 0x37, + 0x6e, 0x9c, 0x9e, 0x61, 0xe5, 0x12, 0x13, 0xb2, 0x88, 0xc6, 0xcf, 0x60, + 0x3f, 0x0d, 0x51, 0x33, 0x22, 0xfa, 0xfb, 0x2d, 0x2b, 0x8d, 0x43, 0x9b, + 0x3d, 0x1e, 0x88, 0x24, 0x50, 0x78, 0xf7, 0x7e, 0x45, 0xb1, 0x0f, 0xa9, + 0xe6, 0x77, 0xf8, 0x78, 0xff, 0x57, 0x6a, 0x05, 0x06, 0x0c, 0x7e, 0x1e, + 0x7f, 0xe9, 0x90, 0xe8, 0x61, 0x68, 0xbc, 0x9e, 0xc4, 0xe5, 0x06, 0x04, + 0x76, 0xcc, 0x01, 0x57, 0x1a, 0x55, 0x9e, 0x45, 0x26, 0xd6, 0xd8, 0xc2, + 0x50, 0x25, 0xfc, 0x72, 0x4e, 0x18, 0xbe, 0xf2, 0x2f, 0xc0, 0x1b, 0xc8, + 0x14, 0xeb, 0x24, 0xda, 0x15, 0x0a, 0x83, 0x38, 0xc5, 0xdd, 0xc9, 0xd7, + 0x12, 0x35, 0x55, 0xdf, 0x2c, 0x23, 0xea, 0x17, 0xca, 0xbf, 0x18, 0xc9, + 0x80, 0x63, 0x4b, 0x77, 0x8b, 0x17, 0x01, 0x05, 0x1b, 0xa3, 0x0b, 0x0f, + 0xdd, 0xc6, 0xe0, 0xdf, 0xc9, 0xa6, 0x8c, 0x50, 0x95, 0x8d, 0x6c, 0x96, + 0x67, 0xff, 0x88, 0x38, 0x3b, 0x76, 0x72, 0x11, 0x35, 0xa0, 0x1c, 0xc8, + 0x96, 0x9c, 0xe5, 0x90, 0x79, 0x0e, 0x62, 0x57, 0x00, 0xd9, 0x57, 0xf8, + 0xa4, 0xc2, 0xc2, 0x0a, 0x17, 0x8e, 0xd7, 0x03, 0x6d, 0x4d, 0x14, 0xb6, + 0x96, 0x8a, 0x76, 0x67, 0x58, 0xce, 0x9c, 0xb3, 0x10, 0x49, 0x06, 0xeb, + 0x56, 0x43, 0x40, 0xcb, 0xd4, 0xd7, 0x59, 0x42, 0xa4, 0xd7, 0x21, 0x6a, + 0x51, 0x3d, 0x1c, 0x54, 0xd7, 0xd6, 0xa2, 0xcf, 0xf8, 0xf6, 0x72, 0x35, + 0x04, 0xa6, 0xe3, 0x53, 0xca, 0xc5, 0x62, 0xee, 0xa9, 0xc3, 0x6d, 0x1b, + 0xc4, 0xc5, 0xd9, 0xa7, 0x37, 0xc2, 0x04, 0x01, 0xc9, 0x4a, 0x2e, 0x26, + 0xdd, 0x12, 0x6e, 0x41, 0x64, 0xb4, 0xe8, 0xe8, 0xc7, 0xf8, 0xab, 0x8a, + 0xab, 0x1d, 0x7f, 0x2d, 0x58, 0xc2, 0xc4, 0xf0, 0x5d, 0x11, 0x35, 0x52, + 0x88, 0xbc, 0x0f, 0x44, 0x6e, 0x91, 0x1e, 0x87, 0xb4, 0xb1, 0x91, 0x52, + 0x32, 0xe4, 0x38, 0x6d, 0x5e, 0x8d, 0x30, 0xf0, 0xbc, 0xc3, 0x15, 0x80, + 0x47, 0x36, 0x35, 0xb0, 0x93, 0xf3, 0xc4, 0x82, 0xc7, 0x73, 0xc1, 0x67, + 0x0c, 0x7a, 0x31, 0x36, 0xbc, 0x73, 0x67, 0x66, 0xae, 0x48, 0x82, 0x27, + 0x6e, 0x14, 0xd0, 0xd5, 0x12, 0x10, 0xce, 0x5e, 0x37, 0xcd, 0x7e, 0xa5, + 0xcb, 0xff, 0x91, 0xf0, 0x62, 0xdb, 0x95, 0x74, 0x0c, 0x8c, 0x1e, 0x78, + 0x11, 0x02, 0xb3, 0x02, 0x0b, 0x31, 0xe7, 0x4e, 0x8b, 0x58, 0x6a, 0xde, + 0x20, 0x93, 0x8b, 0x8e, 0x62, 0x03, 0x24, 0xc9, 0xca, 0xf8, 0x44, 0x1d, + 0x0c, 0x1b, 0xd8, 0x5d, 0xcc, 0xe2, 0x8e, 0x02, 0xc6, 0x5c, 0x06, 0x45, + 0xe6, 0x94, 0x8f, 0xa2, 0x3e, 0xf5, 0xe9, 0xf5, 0x88, 0x87, 0xb2, 0x84, + 0x1e, 0xb6, 0xb6, 0xfc, 0x9f, 0x8e, 0x79, 0xf5, 0x4b, 0x24, 0x81, 0x3e, + 0x5d, 0xf4, 0x10, 0x6e, 0xdd, 0x8c, 0x8c, 0xae, 0xc6, 0x2c, 0x26, 0xb2, + 0xfc, 0xf3, 0x99, 0xe8, 0x8c, 0x65, 0x5d, 0x6c, 0xa8, 0x1d, 0x6f, 0x1e, + 0x32, 0x0a, 0xee, 0x87, 0xf6, 0xe1, 0xdd, 0x5e, 0x7f, 0x7a, 0x90, 0x8c, + 0x3f, 0xe8, 0x47, 0x95, 0x9b, 0xc8, 0x2c, 0x49, 0xc9, 0xe4, 0x2d, 0xea, + 0x58, 0xfc, 0x29, 0x1a, 0xb7, 0xa1, 0xf9, 0xb8, 0x84, 0x41, 0xa0, 0xf1, + 0x77, 0x83, 0x56, 0x73, 0x86, 0xea, 0xf4, 0xf5, 0x2a, 0xa6, 0x6b, 0x00, + 0x64, 0x39, 0x08, 0x8f, 0xf0, 0x22, 0x1a, 0x4c, 0xf2, 0x5a, 0xd0, 0xaa, + 0x39, 0xae, 0x8a, 0xbc, 0x03, 0x99, 0xf7, 0xcc, 0x80, 0xdf, 0x2b, 0x85, + 0xbe, 0x1a, 0x97, 0x28, 0x63, 0x04, 0x72, 0x75, 0x75, 0xb4, 0x9c, 0xd3, + 0x17, 0xcc, 0x1e, 0xa1, 0xd2, 0x47, 0x18, 0x45, 0xad, 0xb4, 0x0a, 0x32, + 0x31, 0x36, 0x64, 0x48, 0x3f, 0x7b, 0x4b, 0xc0, 0xd6, 0x78, 0x46, 0xaa, + 0x90, 0x89, 0xf9, 0x36, 0x3d, 0xb4, 0xb3, 0x50, 0x51, 0xd9, 0x55, 0x6f, + 0xa9, 0xe7, 0x25, 0xaf, 0xa0, 0xca, 0x9d, 0x45, 0x83, 0xc3, 0x0b, 0x2a, + 0x0c, 0xf9, 0x3f, 0xe4, 0x08, 0xf4, 0xbd, 0x23, 0x45, 0x85, 0xcf, 0x41, + 0x93, 0xd3, 0x21, 0x5f, 0x53, 0xa2, 0x5b, 0xa9, 0xf5, 0xe9, 0x8f, 0x2a, + 0x2d, 0x53, 0x3c, 0x36, 0x17, 0xce, 0x37, 0x35, 0x3e, 0x9e, 0x6b, 0xbc, + 0xba, 0xaa, 0xa5, 0x61, 0x79, 0x98, 0x8e, 0xbd, 0x19, 0xf4, 0x5f, 0xa9, + 0xb8, 0x96, 0xa2, 0xce, 0x32, 0x00, 0xab, 0x51, 0xcb, 0xfa, 0x30, 0x3a, + 0x83, 0x92, 0x91, 0xad, 0x08, 0x61, 0x62, 0x51, 0x7f, 0x19, 0xa9, 0x2a, + 0x84, 0xf2, 0xab, 0x7e, 0x5e, 0xa7, 0x5a, 0x54, 0x7f, 0x68, 0x2a, 0x7b, + 0x4f, 0xde, 0x45, 0x1d, 0xef, 0x73, 0x5f, 0xc0, 0x40, 0x6e, 0xec, 0x6c, + 0xe9, 0xa5, 0x6b, 0x46, 0x54, 0x7c, 0x24, 0x8b, 0xa4, 0xe5, 0xb4, 0x82, + 0x31, 0x1f, 0x3e, 0x79, 0x2e, 0x21, 0x8c, 0xf1, 0xbd, 0xad, 0x7c, 0x28, + 0xcc, 0xbd, 0x58, 0x72, 0xe9, 0x6a, 0x04, 0x56, 0x67, 0x0f, 0x62, 0x98, + 0x5a, 0x97, 0x4b, 0xe2, 0x67, 0x70, 0xbb, 0x17, 0xb1, 0x84, 0x5b, 0xd4, + 0x6e, 0xab, 0x90, 0x29, 0x20, 0x93, 0x34, 0xa8, 0x03, 0x0f, 0xed, 0x1a, + 0xf0, 0x1b, 0x92, 0x87, 0x43, 0xa5, 0x6a, 0x1c, 0xdc, 0xd7, 0x22, 0x68, + 0x83, 0x98, 0x74, 0x2a, 0x4c, 0x51, 0xef, 0x71, 0x19, 0xd5, 0x3d, 0x05, + 0x19, 0x61, 0xb2, 0x52, 0xa8, 0x6e, 0xda, 0x72, 0x51, 0x66, 0x9f, 0xf0, + 0x12, 0xf6, 0x18, 0x60, 0xcc, 0xd7, 0x2f, 0x2e, 0x83, 0x14, 0x09, 0xdb, + 0x55, 0x1c, 0xf2, 0xaf, 0xfd, 0xa4, 0x40, 0xf1, 0x4a, 0xc7, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x9c, 0x52, 0xff, 0x48, + 0x06, 0x61, 0x76, 0x6d, 0xd7, 0x44, 0xb1, 0x0c, 0x32, 0x62, 0x15, 0xa1, + 0xc3, 0x97, 0x03, 0xdd, 0xed, 0x20, 0x3c, 0x3a, 0x09, 0x16, 0xe5, 0x7d, + 0x8c, 0xf9, 0x7b, 0x22, 0x5e, 0x3a, 0xdd, 0xf0, 0xc6, 0xf0, 0x3a, 0xd4, + 0x94, 0x85, 0x1c, 0x60, 0x74, 0x91, 0xa3, 0xe2, 0x8a, 0xe5, 0x3e, 0xd4, + 0x95, 0x28, 0x8b, 0x1a, 0x7b, 0xbe, 0x07, 0xc0, 0xe3, 0x6b, 0xb9, 0x85, + 0x82, 0x0b, 0x24, 0xba, 0x1c, 0xfc, 0xc0, 0x0a, 0x21, 0x33, 0xad, 0x00, + 0x19, 0xce, 0xb5, 0x8f, 0x73, 0x05, 0xf1, 0xac, 0x03, 0xbe, 0x1f, 0x22, + 0xd5, 0x32, 0x5e, 0x50, 0xe3, 0xe0, 0x62, 0x26, 0xf4, 0xb0, 0x85, 0xd8, + 0xf7, 0xa7, 0xf4, 0xa7, 0xff, 0x10, 0xb8, 0xbc, 0xe0, 0x3e, 0x4d, 0xcb, + 0x37, 0x74, 0xcc, 0x85, 0xed, 0xa0, 0x34, 0x6c, 0xfa, 0x37, 0x84, 0x6a, + 0x94, 0x55, 0x3b, 0x1e, 0x14, 0xab, 0x26, 0x7b, 0x3e, 0xac, 0xc3, 0x79, + 0xcd, 0x1b, 0x00, 0x02, 0xb3, 0x01, 0xc3, 0x10, 0xdd, 0x56, 0x7d, 0x0e, + 0x69, 0x39, 0x3c, 0x17, 0xa3, 0xae, 0x9c, 0x2d, 0xc7, 0x5a, 0x0b, 0x7c, + 0xd0, 0xac, 0xa1, 0x91, 0x6a, 0x6d, 0xc0, 0x3f, 0x98, 0xf1, 0x21, 0xf5, + 0xa5, 0x7c, 0xbc, 0x70, 0x0d, 0x7b, 0x2f, 0x0d, 0x5a, 0xa5, 0x4a, 0x5a, + 0xff, 0x51, 0xbf, 0x7f, 0xb5, 0x4f, 0x2c, 0xba, 0xa9, 0x46, 0x81, 0x6b, + 0xac, 0xc6, 0x62, 0x2d, 0xd7, 0xb5, 0x04, 0x5f, 0xd4, 0x5f, 0x1f, 0x6b, + 0x11, 0x7d, 0xe3, 0x58, 0x1f, 0xb5, 0xbf, 0x16, 0x43, 0x88, 0x05, 0xf5, + 0xa4, 0x7b, 0xb5, 0x0e, 0xf4, 0x01, 0xb6, 0x90, 0x69, 0x52, 0x0a, 0x5e, + 0x9b, 0x87, 0x51, 0x5e, 0xd5, 0xed, 0x2c, 0xcc, 0x58, 0xad, 0xe6, 0x77, + 0xa2, 0xc5, 0x7c, 0x1e, 0xc5, 0x92, 0xbe, 0xed, 0x3a, 0x9a, 0x97, 0xed, + 0x56, 0xc8, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, + 0x16, 0xe8, 0x24, 0xe3, 0x82, 0x36, 0x8e, 0x50, 0x45, 0xbe, 0xc6, 0x10, + 0x02, 0xb9, 0x6d, 0xf9, 0xed, 0x8f, 0x64, 0x35, 0x4d, 0x2c, 0x9f, 0x99, + 0xdc, 0xee, 0xfa, 0x63, 0x99, 0xc4, 0xb8, 0x3d, 0x77, 0xea, 0xda, 0xd5, + 0x95, 0x8b, 0x8e, 0x76, 0x02, 0x9c, 0x62, 0xa0, 0xad, 0xfe, 0x80, 0x61, + 0x72, 0x59, 0xd6, 0x9f, 0x16, 0x2e, 0x09, 0x71, 0xb8, 0xd7, 0x65, 0x25, + 0xc2, 0x5b, 0x40, 0x67, 0x8e, 0xd6, 0xf8, 0xdf, 0x67, 0x29, 0x19, 0xa2, + 0xa6, 0x07, 0xf3, 0xc8, 0x91, 0x7d, 0xf2, 0x50, 0x71, 0xba, 0x5c, 0x2d, + 0xa7, 0xae, 0xc4, 0xd5, 0xeb, 0xb9, 0x0d, 0x2d, 0x23, 0xe5, 0x8c, 0x65, + 0xf5, 0xf8, 0x97, 0x69, 0xde, 0x25, 0x6f, 0xea, 0x12, 0x72, 0x3e, 0xb9, + 0xa7, 0x8d, 0xcf, 0xa5, 0x66, 0xee, 0x4e, 0x2e, 0x66, 0x6b, 0xec, 0x77, + 0x7f, 0x53, 0xdc, 0x29, 0x73, 0x5e, 0xe9, 0x2f, 0x79, 0xac, 0x8d, 0x0f, + 0x44, 0x09, 0x5d, 0x25, 0x1d, 0x78, 0xb6, 0xe9, 0xd0, 0xfa, 0x8f, 0x5f, + 0x9c, 0xf0, 0xe0, 0xfc, 0x62, 0x9f, 0x52, 0x6b, 0x5b, 0x8e, 0x3f, 0xdf, + 0xb4, 0xf1, 0xdf, 0x35, 0xd0, 0x8f, 0x5a, 0xc9, 0x1f, 0x08, 0x86, 0xaa, + 0x5a, 0x9e, 0xe8, 0xb0, 0xaa, 0xd4, 0xcd, 0x2a, 0x5b, 0x4f, 0x7f, 0x39, + 0x9f, 0x7f, 0x21, 0xf2, 0xfd, 0x05, 0x96, 0x53, 0x09, 0xfd, 0x36, 0x4c, + 0xcd, 0x98, 0x74, 0xf5, 0xbd, 0xcd, 0x9e, 0x14, 0x15, 0x05, 0xb9, 0x3d, + 0x5f, 0x8a, 0x02, 0x86, 0x10, 0xd7, 0xd4, 0x01, 0x20, 0xd9, 0x8c, 0x65, + 0x7d, 0x9d, 0x39, 0x25, 0xbc, 0xce, 0x1a, 0xb1, 0x76, 0x92, 0xc3, 0x03, + 0xed, 0xa2, 0x41, 0x31, 0x0d, 0xc0, 0x40, 0x94, 0x01, 0xbc, 0x9b, 0xe9, + 0x5e, 0x3e, 0x8c, 0x49, 0xf6, 0x98, 0x0c, 0x39, 0x79, 0xdc, 0xd1, 0x1b, + 0xc5, 0xb2, 0x20, 0xb4, 0x6c, 0xb4, 0x4f, 0xce, 0xf4, 0x6c, 0x0b, 0xef, + 0x85, 0xf2, 0x7d, 0x9a, 0x90, 0x58, 0x1b, 0x51, 0x56, 0x52, 0xac, 0x75, + 0x9f, 0x17, 0xe6, 0x48, 0xaf, 0x18, 0x4c, 0xd8, 0x67, 0xe8, 0xd2, 0x61, + 0xbc, 0xa0, 0x95, 0xc9, 0x78, 0xd8, 0xa2, 0x1d, 0x47, 0x59, 0x30, 0xcf, + 0xf3, 0x79, 0x06, 0xd4, 0x25, 0xf8, 0x9c, 0x5c, 0x28, 0xee, 0xb0, 0xd2, + 0xb6, 0xaf, 0x34, 0x0e, 0xe5, 0xe4, 0x16, 0x2e, 0x05, 0x45, 0x23, 0xc1, + 0x88, 0x90, 0x4a, 0x8f, 0xff, 0xfb, 0xe2, 0xc0, 0xb7, 0xae, 0xb5, 0x50, + 0xc9, 0x26, 0xf0, 0xa2, 0xf5, 0x21, 0x23, 0x79, 0x23, 0xb6, 0x8f, 0x57, + 0x64, 0xd1, 0x27, 0xc2, 0x07, 0x63, 0xa6, 0x54, 0x1f, 0x2f, 0xca, 0x16, + 0xb8, 0x28, 0x51, 0x2a, 0x92, 0xe0, 0x06, 0x36, 0x55, 0x00, 0x6c, 0x99, + 0x31, 0xa7, 0x56, 0xb3, 0x7b, 0x15, 0xcd, 0xc1, 0x32, 0x3a, 0xc0, 0x37, + 0x1f, 0xea, 0x29, 0xb6, 0x75, 0xdf, 0x8a, 0x17, 0x09, 0x45, 0xc2, 0x6e, + 0xe2, 0x4c, 0xa5, 0x93, 0x9b, 0x17, 0x08, 0x27, 0x75, 0x33, 0xdb, 0x1f, + 0xab, 0x37, 0xad, 0x8e, 0xaa, 0xef, 0x0b, 0x82, 0xaa, 0xa7, 0xae, 0x2c, + 0x43, 0x4d, 0x8f, 0xa0, 0x43, 0xd7, 0xa1, 0x34, 0xeb, 0xc0, 0x4e, 0xbd, + 0x64, 0xfc, 0xc8, 0x6a, 0x56, 0xa8, 0xfc, 0x9e, 0x2d, 0x5f, 0x7a, 0xa3, + 0x72, 0x06, 0x79, 0x38, 0x33, 0x05, 0xa7, 0xf0, 0x09, 0x48, 0x55, 0xfe, + 0x3f, 0xab, 0x25, 0x8e, 0x76, 0x1d, 0x12, 0x5a, 0x20, 0x68, 0xfb, 0x51, + 0x51, 0x33, 0x40, 0x37, 0x0c, 0x90, 0x98, 0x6f, 0x66, 0x3f, 0x40, 0xa2, + 0x2e, 0x3c, 0xd1, 0x22, 0x51, 0x54, 0x25, 0x7e, 0x4c, 0x5d, 0x96, 0xb2, + 0x65, 0x0f, 0xa3, 0xdf, 0x8e, 0x97, 0xfe, 0xeb, 0xe7, 0xc6, 0x22, 0x2a, + 0x47, 0x3a, 0x78, 0x1b, 0x39, 0x2e, 0xd6, 0xbc, 0x35, 0xb4, 0xf4, 0xc3, + 0xf2, 0x6a, 0x12, 0xc9, 0xe7, 0x6c, 0x9a, 0xfc, 0xed, 0xbc, 0x11, 0xc7, + 0x71, 0x09, 0x8f, 0x56, 0xc1, 0xd8, 0xb6, 0x92, 0x35, 0x97, 0x8e, 0x71, + 0xd2, 0xbb, 0xb4, 0xed, 0xf0, 0x7e, 0xff, 0x58, 0xd9, 0x95, 0x26, 0xea, + 0xa9, 0x4d, 0x38, 0x8d, 0x4e, 0x8e, 0x53, 0xae, 0x7e, 0xe6, 0xe6, 0x82, + 0x35, 0x96, 0xab, 0x0f, 0x04, 0x0f, 0xf2, 0xac, 0x1b, 0xcd, 0x07, 0x17, + 0x1b, 0x25, 0x2f, 0x92, 0xaf, 0x19, 0xa2, 0x1b, 0xa0, 0x7a, 0xc7, 0x4f, + 0xb8, 0x1b, 0x89, 0x21, 0xb5, 0xe2, 0x24, 0xe9, 0x78, 0xae, 0x7d, 0xd7, + 0xcc, 0x8e, 0x3f, 0xa7, 0xe9, 0xbe, 0xe6, 0x79, 0x0f, 0xdf, 0x86, 0xe9, + 0xb9, 0xcd, 0x82, 0x7b, 0xf5, 0x04, 0x89, 0xa0, 0x73, 0x5d, 0xa2, 0x4e, + 0xd6, 0xa0, 0x60, 0x21, 0xe2, 0xfe, 0xd3, 0xf4, 0x19, 0x8b, 0x6a, 0x03, + 0x12, 0x9c, 0x51, 0x9a, 0x41, 0x4e, 0xf6, 0xb4, 0x6e, 0x0c, 0x43, 0xf5, + 0x00, 0x00, 0x78, 0x12, 0xdd, 0x21, 0xa8, 0xc7, 0x21, 0xa1, 0x4e, 0x44, + 0x10, 0xd0, 0xdb, 0x6f, 0x0b, 0x4c, 0xe7, 0x7a, 0x8c, 0x0c, 0xaa, 0xb6, + 0x9a, 0x7d, 0xa9, 0xff, 0x5a, 0x2e, 0x15, 0x9e, 0x6f, 0xea, 0xe1, 0x42, + 0x0c, 0x9c, 0x5a, 0x3b, 0xd5, 0xe6, 0xde, 0x23, 0x3f, 0x9c, 0x45, 0x20, + 0x67, 0x96, 0x50, 0x16, 0x80, 0x42, 0xe7, 0x67, 0x7d, 0x24, 0xdc, 0x00, + 0xaa, 0x01, 0x8a, 0xa3, 0x61, 0xfe, 0x9a, 0xce, 0xc1, 0xe5, 0x2e, 0x19, + 0x85, 0x04, 0xe6, 0x7b, 0xe8, 0x7a, 0xbc, 0x9d, 0xfe, 0x71, 0x29, 0x1d, + 0x17, 0xae, 0x6b, 0x1a, 0x64, 0xd7, 0xfe, 0x18, 0x29, 0x07, 0x9b, 0x49, + 0x43, 0xba, 0x29, 0x37, 0xa8, 0xb0, 0x26, 0x27, 0x6b, 0x7d, 0xde, 0x49, + 0x12, 0x90, 0x05, 0xe2, 0x2c, 0xd8, 0x08, 0xd0, 0x5d, 0x74, 0xa7, 0x15, + 0xbe, 0x34, 0x34, 0x6d, 0xad, 0xfb, 0xa8, 0x01, 0x4a, 0x6c, 0x98, 0xba, + 0x84, 0x38, 0xbd, 0x05, 0xe8, 0x87, 0x27, 0x91, 0x3f, 0xb8, 0xe9, 0x06, + 0x27, 0xda, 0x56, 0x07, 0xaa, 0xea, 0xf4, 0x80, 0x5c, 0x12, 0x44, 0xbe, + 0x23, 0xb3, 0x63, 0x9f, 0x5f, 0x37, 0xa7, 0x53, 0x4c, 0xfc, 0x4d, 0x87, + 0xeb, 0x91, 0xe8, 0xd7, 0x5a, 0xd6, 0xca, 0x67, 0x2d, 0x2f, 0x5a, 0x0e, + 0xc7, 0x82, 0x78, 0xa4, 0xf3, 0x56, 0x07, 0xa5, 0xab, 0x6d, 0x09, 0xd2, + 0x0d, 0x08, 0x6b, 0x6e, 0x1f, 0xc1, 0xf2, 0x91, 0x1a, 0x39, 0xfe, 0x14, + 0x56, 0x3f, 0xeb, 0x9f, 0x14, 0xc2, 0xb3, 0xb2, 0xc2, 0x8d, 0xc2, 0xee, + 0x7e, 0xf0, 0x7d, 0x92, 0xd2, 0xc3, 0x57, 0x3e, 0x2c, 0x07, 0x1b, 0x6a, + 0x9b, 0x3b, 0x79, 0x59, 0xc9, 0x22, 0x96, 0x6c, 0x3e, 0x37, 0xd3, 0x0e, + 0x5c, 0xf6, 0x8f, 0xa9, 0xaa, 0xc9, 0xa4, 0x4b, 0xaf, 0x5d, 0x1a, 0xb6, + 0xf3, 0x91, 0x32, 0x4f, 0xca, 0x72, 0xa0, 0x42, 0x01, 0x51, 0xaf, 0x19, + 0x89, 0xc4, 0xcc, 0x9b, 0xf3, 0x52, 0xe9, 0xa6, 0xf2, 0x71, 0x6f, 0x5a, + 0x38, 0x02, 0xb8, 0x75, 0x88, 0x5f, 0x8d, 0x12, 0xc5, 0x55, 0x4f, 0xd1, + 0xba, 0xf2, 0x24, 0xdc, 0x63, 0x5f, 0x93, 0xc7, 0xf3, 0xe7, 0x59, 0xac, + 0xc3, 0xed, 0xbc, 0x02, 0xe3, 0xad, 0xb2, 0x8e, 0x2c, 0x2d, 0x47, 0xb4, + 0x34, 0x8d, 0xae, 0x44, 0xc8, 0x5f, 0x14, 0xe8, 0x8e, 0x7b, 0xc3, 0x60, + 0x53, 0x9a, 0x51, 0xea, 0x7f, 0x2f, 0xb6, 0x62, 0x61, 0xf7, 0xc0, 0x18, + 0x0f, 0x20, 0x79, 0x13, 0x5c, 0xe8, 0xca, 0x04, 0x29, 0x5f, 0x70, 0x4d, + 0x88, 0xa2, 0x43, 0x20, 0x57, 0x33, 0x04, 0x74, 0x8e, 0x7c, 0x89, 0xd4, + 0x56, 0x8f, 0x93, 0x86, 0x81, 0x6c, 0x11, 0xfc, 0x32, 0x0e, 0xb0, 0x3e, + 0xe5, 0x13, 0xbf, 0x76, 0x62, 0xcc, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x04, 0x00, 0x00, 0x0e, 0xf8, 0x8f, 0xde, 0xfd, 0xfd, 0xcf, 0xd1, + 0x6f, 0x9f, 0xf2, 0xb6, 0xb6, 0x59, 0xb2, 0x73, 0x1c, 0x3c, 0x0d, 0xb0, + 0x4d, 0xb8, 0x96, 0xc6, 0xeb, 0xe5, 0xf8, 0x0d, 0x3e, 0xd7, 0x0c, 0xbd, + 0x9c, 0xaa, 0xd5, 0x1c, 0x19, 0x9a, 0x4c, 0x8e, 0xfa, 0xac, 0x68, 0x74, + 0x16, 0x06, 0xb5, 0x49, 0xe7, 0xd5, 0x6f, 0x4f, 0xcc, 0xd9, 0x02, 0x74, + 0xd6, 0x08, 0x73, 0x7c, 0xa9, 0xfa, 0x3e, 0x50, 0x87, 0xf7, 0xfb, 0xa6, + 0x94, 0xdc, 0xb1, 0x40, 0xec, 0xa7, 0xa9, 0x39, 0xff, 0x40, 0x4a, 0x97, + 0x9b, 0xcc, 0x57, 0x66, 0x68, 0xd6, 0xa8, 0x4d, 0x13, 0x06, 0x0e, 0x03, + 0xc4, 0xdf, 0x7a, 0xe4, 0x2f, 0x0e, 0xd7, 0x54, 0xe0, 0xbd, 0x93, 0xeb, + 0x82, 0xd8, 0x05, 0x2d, 0xa2, 0xf0, 0x4e, 0xd0, 0xf9, 0x3e, 0x3e, 0x6b, + 0x3d, 0x08, 0x39, 0x4e, 0x35, 0x13, 0x7b, 0x3b, 0x39, 0x2c, 0x47, 0x2c, + 0x61, 0x9f, 0xfd, 0x59, 0x88, 0x5f, 0x65, 0x08, 0xa9, 0x66, 0xec, 0xb5, + 0x21, 0xf3, 0xe9, 0xba, 0x11, 0x63, 0x24, 0x6c, 0xf4, 0x50, 0x3a, 0xe5, + 0x0c, 0x06, 0x39, 0x69, 0x2f, 0xca, 0x0f, 0x48, 0xbe, 0x95, 0x7d, 0x13, + 0x3d, 0xa5, 0x75, 0x69, 0x85, 0xc8, 0xb3, 0x72, 0x72, 0x3c, 0x4f, 0x96, + 0xe7, 0xb7, 0xbd, 0xe7, 0x76, 0xba, 0xac, 0xc0, 0x07, 0x4d, 0xc1, 0xed, + 0xb9, 0xf0, 0x91, 0x2e, 0x36, 0xb7, 0x5b, 0x1c, 0xb7, 0xd6, 0xb3, 0x45, + 0x7d, 0x0a, 0xf5, 0x43, 0xdd, 0x7a, 0x8b, 0x4e, 0x18, 0xf2, 0xf3, 0x19, + 0xcd, 0x4a, 0xda, 0x3c, 0x1b, 0x05, 0x27, 0x67, 0x43, 0xa9, 0x8e, 0xe7, + 0x4a, 0x95, 0xa9, 0xad, 0x6c, 0x8c, 0xb2, 0x2e, 0x12, 0xcb, 0xf3, 0xeb, + 0x65, 0x26, 0xf4, 0x3e, 0x86, 0xee, 0x7e, 0xd9, 0xba, 0xce, 0x8d, 0x15, + 0x3e, 0xa8, 0x40, 0x59, 0x1d, 0x27, 0x78, 0x75, 0xf0, 0xf9, 0x33, 0xb5, + 0x32, 0xa9, 0x66, 0xe6, 0x2e, 0x2e, 0x3d, 0xf5, 0x4a, 0xf0, 0x97, 0x2d, + 0xe7, 0x43, 0x85, 0x43, 0x61, 0x25, 0x15, 0x13, 0x9e, 0x8e, 0xf6, 0x78, + 0xe8, 0x67, 0xba, 0xc2, 0x6d, 0xda, 0x46, 0x25, 0x76, 0xd9, 0x9b, 0x69, + 0x95, 0x4b, 0x50, 0x8c, 0xb7, 0x36, 0x49, 0xbc, 0xd7, 0x39, 0x69, 0xb9, + 0xc1, 0x5f, 0x5f, 0xcc, 0x83, 0x4c, 0x16, 0xb8, 0x0c, 0x85, 0xf1, 0xa4, + 0x57, 0x6c, 0x22, 0x1f, 0x60, 0x0c, 0xff, 0xb6, 0xc9, 0xf7, 0x21, 0x2d, + 0x35, 0x78, 0x31, 0x79, 0xd0, 0x6d, 0x61, 0xec, 0x61, 0x04, 0x75, 0x5c, + 0x06, 0xc3, 0x53, 0x1b, 0xb5, 0xdc, 0x23, 0xb9, 0xd9, 0x07, 0xd1, 0xd0, + 0xb3, 0xa5, 0xab, 0xd9, 0xbe, 0xb7, 0xdc, 0xae, 0x3f, 0x3e, 0xd7, 0x2a, + 0x79, 0x3f, 0x9c, 0x27, 0x81, 0x8d, 0x61, 0xe8, 0x46, 0x8f, 0x05, 0xf4, + 0x9c, 0x30, 0x35, 0x9a, 0x2f, 0x62, 0x84, 0x7c, 0xa5, 0x95, 0x68, 0x34, + 0xe6, 0xf0, 0xb9, 0x42, 0xd4, 0x37, 0xc6, 0xd2, 0x35, 0x1f, 0x7b, 0xe0, + 0xa6, 0x92, 0xcf, 0xf7, 0x0f, 0x08, 0x10, 0x79, 0xbd, 0xa8, 0x7c, 0x4e, + 0xef, 0xf1, 0x01, 0x8d, 0x1b, 0x0c, 0x98, 0x46, 0x28, 0xdc, 0xd5, 0xa8, + 0xcf, 0x67, 0x7d, 0x87, 0x2a, 0x8f, 0xdd, 0x52, 0x43, 0x5a, 0x55, 0x80, + 0x88, 0xa6, 0xcd, 0x9c, 0x5d, 0x36, 0xae, 0xef, 0x61, 0x43, 0xec, 0xf0, + 0x7f, 0x92, 0x21, 0x1f, 0xa2, 0xa3, 0x76, 0x0e, 0x5d, 0xf3, 0xa7, 0xe7, + 0x7d, 0xb0, 0x2c, 0x94, 0x36, 0x95, 0x34, 0x4e, 0x04, 0xfb, 0x51, 0xf9, + 0xe6, 0x7e, 0x56, 0x7a, 0x59, 0xce, 0x0a, 0x45, 0x7e, 0xeb, 0xc4, 0xbc, + 0xfd, 0x20, 0xaa, 0x34, 0x6b, 0xee, 0x3b, 0x09, 0xe8, 0x00, 0x4b, 0xfc, + 0x68, 0x24, 0x43, 0xdb, 0x09, 0x58, 0xd0, 0xb6, 0xbf, 0xaf, 0x1d, 0x7f, + 0x8a, 0x4c, 0x9e, 0x51, 0x97, 0x97, 0xe1, 0x0c, 0x0d, 0xaf, 0xd1, 0x1e, + 0x62, 0xad, 0x70, 0xa5, 0x8a, 0x24, 0x2f, 0x4a, 0xa6, 0x55, 0xb1, 0x44, + 0x09, 0x88, 0xab, 0xa5, 0x45, 0x28, 0xa0, 0x34, 0x9e, 0x14, 0x2c, 0xf9, + 0x0f, 0xb8, 0x33, 0x8f, 0xcc, 0xba, 0x50, 0x34, 0x4c, 0x96, 0x89, 0x09, + 0xb9, 0xa8, 0xfb, 0xac, 0x59, 0x73, 0xea, 0x61, 0xbc, 0x0d, 0x24, 0x3a, + 0x20, 0xc2, 0x76, 0xfc, 0x2e, 0xce, 0xfb, 0x75, 0x00, 0xca, 0x58, 0xbd, + 0xab, 0x61, 0x9b, 0x13, 0x2b, 0xa3, 0xf6, 0x15, 0x55, 0x83, 0x23, 0xc4, + 0xf3, 0x4c, 0x89, 0xc5, 0x4a, 0x18, 0x5c, 0x8d, 0x41, 0xcc, 0x06, 0x7b, + 0xe3, 0x2a, 0x1f, 0x6a, 0x57, 0xbc, 0x54, 0x61, 0x0c, 0xf2, 0xec, 0xbf, + 0xb0, 0xf0, 0x21, 0xde, 0xfc, 0xe4, 0xef, 0xce, 0x47, 0xc8, 0xdc, 0x11, + 0xc7, 0x8a, 0x12, 0x97, 0x68, 0x1d, 0x9e, 0x9a, 0xbf, 0xad, 0x62, 0x7e, + 0x4b, 0x88, 0xd7, 0x20, 0x22, 0xce, 0x5e, 0xe3, 0x87, 0x12, 0xa3, 0x05, + 0xef, 0x1f, 0x05, 0xb1, 0xbd, 0x1b, 0x80, 0x43, 0x84, 0x33, 0x8b, 0x87, + 0xa5, 0xc2, 0xe1, 0x49, 0xa8, 0x75, 0x49, 0x9b, 0x1b, 0x64, 0x8a, 0xd0, + 0x86, 0x10, 0xa8, 0x72, 0xeb, 0x2e, 0xe7, 0x3f, 0xaa, 0x6b, 0x4a, 0x22, + 0xae, 0x17, 0x8f, 0x10, 0x22, 0x03, 0x66, 0x67, 0x35, 0x40, 0x29, 0x1e, + 0xf2, 0x05, 0x36, 0xd5, 0xed, 0xe2, 0x2a, 0xcc, 0x77, 0xe2, 0x16, 0xef, + 0xa7, 0x9b, 0xe1, 0x1b, 0xba, 0xf3, 0xf5, 0x74, 0x6c, 0x2a, 0x98, 0x8a, + 0x14, 0xaf, 0x2c, 0xab, 0xfb, 0x51, 0x53, 0x75, 0x17, 0xcb, 0x5c, 0x86, + 0xb5, 0x60, 0x70, 0x29, 0x65, 0x69, 0x49, 0x42, 0x4f, 0x42, 0x6b, 0xc7, + 0xdb, 0x98, 0x7d, 0x1e, 0xf8, 0x45, 0xb2, 0x33, 0xd6, 0x34, 0x26, 0xa6, + 0x7f, 0x76, 0x31, 0x13, 0x13, 0x9d, 0xd2, 0xb0, 0x30, 0x0b, 0x0b, 0x3e, + 0x1a, 0x84, 0xb0, 0xbd, 0x81, 0x34, 0x25, 0x73, 0x99, 0x87, 0x1a, 0xc8, + 0x44, 0x34, 0x9d, 0x1a, 0x3d, 0x76, 0x44, 0x1d, 0xe2, 0x22, 0xad, 0x3d, + 0xb2, 0xa3, 0x1c, 0xd5, 0x27, 0x8c, 0xc6, 0x84, 0xdf, 0x33, 0xbe, 0xb2, + 0xa7, 0xb9, 0xc5, 0x6e, 0x48, 0xdc, 0xe9, 0xf8, 0xef, 0xfc, 0xaa, 0x1f, + 0x5e, 0x41, 0x48, 0x1e, 0xe0, 0xb9, 0xd6, 0x6e, 0x7a, 0x9c, 0xa3, 0x98, + 0x4b, 0xfa, 0x90, 0xa4, 0x58, 0x33, 0x85, 0x3b, 0x11, 0x44, 0x83, 0x4b, + 0x1e, 0x0e, 0x5d, 0x11, 0x36, 0x15, 0xe1, 0xbf, 0x15, 0x04, 0x8e, 0x88, + 0xc6, 0x18, 0x53, 0xc3, 0x8d, 0x28, 0x86, 0x25, 0xef, 0x55, 0x7b, 0xf6, + 0x85, 0xf8, 0xed, 0x3b, 0xcf, 0x5d, 0xa6, 0xc7, 0x66, 0xb7, 0xbe, 0x14, + 0xf0, 0x62, 0x89, 0x1f, 0x32, 0x1e, 0x86, 0x2a, 0x93, 0xd5, 0xca, 0x37, + 0x03, 0x0b, 0xf8, 0x0f, 0xca, 0x50, 0x6c, 0x16, 0x2b, 0xf0, 0x77, 0xca, + 0xbb, 0x8e, 0x95, 0x11, 0xef, 0x5b, 0xbe, 0x2f, 0x62, 0x50, 0xb8, 0x3d, + 0xff, 0xfa, 0x30, 0x21, 0xb2, 0x86, 0x3f, 0x50, 0x57, 0x98, 0x79, 0x15, + 0xce, 0x3e, 0xbf, 0x49, 0x58, 0xb0, 0xb5, 0xd7, 0xbe, 0x01, 0x55, 0xee, + 0x60, 0x14, 0x9d, 0x5b, 0x57, 0x48, 0x05, 0x72, 0x6a, 0x23, 0x29, 0xeb, + 0xf3, 0x36, 0x2a, 0xc1, 0xda, 0x5e, 0x4a, 0x63, 0xc4, 0x6b, 0x04, 0xe8, + 0xe8, 0xc1, 0xb5, 0xc4, 0x2d, 0x60, 0x1f, 0xa0, 0x2b, 0x33, 0xa5, 0xb7, + 0x82, 0x59, 0x21, 0xba, 0x13, 0xda, 0x79, 0xda, 0x5a, 0xb1, 0x82, 0x5b, + 0x52, 0x7f, 0x0c, 0x70, 0x75, 0x65, 0xe0, 0x44, 0xb3, 0xca, 0xd0, 0x09, + 0x38, 0x24, 0x83, 0x8e, 0x0c, 0x4c, 0xef, 0x96, 0xe4, 0x04, 0x30, 0x46, + 0x23, 0x6a, 0x28, 0x13, 0x1d, 0x37, 0x14, 0x75, 0x6e, 0xd0, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x21, 0xa2, 0xf0, 0x7d, + 0x29, 0x8f, 0x62, 0x2e, 0xf4, 0x0e, 0x14, 0x9b, 0x60, 0x38, 0xc0, 0x95, + 0xfb, 0x3c, 0x90, 0x5a, 0xa0, 0x1f, 0x30, 0x09, 0xfc, 0x6d, 0xa9, 0xd1, + 0x7b, 0x0b, 0x7c, 0x78, 0xf9, 0xf6, 0xa8, 0x5e, 0xa6, 0x7a, 0xf6, 0x1c, + 0xab, 0x1b, 0x0e, 0xa9, 0x08, 0xfd, 0xd9, 0x97, 0x08, 0x24, 0x2b, 0xda, + 0x08, 0x8b, 0x0c, 0x07, 0x70, 0x15, 0xa8, 0x0c, 0x86, 0xfc, 0xd1, 0x84, + 0xba, 0xd0, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, + 0x35, 0x7a, 0xab, 0xaa, 0xbe, 0xd7, 0xad, 0x22, 0x99, 0x46, 0xbb, 0x78, + 0xfd, 0x47, 0x8f, 0x2a, 0x4a, 0xa6, 0x2f, 0x8d, 0x15, 0x07, 0xed, 0x26, + 0x1d, 0xb3, 0x12, 0xd3, 0x88, 0x0f, 0xf1, 0x75, 0x2a, 0x07, 0x62, 0xac, + 0xbf, 0x52, 0x4a, 0xc3, 0x12, 0xe5, 0x3c, 0xea, 0xa6, 0x1e, 0x57, 0x90, + 0x56, 0x60, 0x7d, 0xcf, 0x4b, 0x65, 0xaf, 0xee, 0x17, 0x56, 0xbe, 0xd2, + 0x38, 0x3f, 0xd6, 0xbc, 0xef, 0xa7, 0x32, 0xb7, 0x10, 0xe9, 0xbd, 0x97, + 0x45, 0x92, 0x3c, 0xd3, 0x35, 0x2e, 0x59, 0x37, 0x65, 0x5c, 0x7f, 0xd0, + 0x99, 0x9c, 0x01, 0xe9, 0x1f, 0x65, 0xe9, 0xec, 0x0f, 0x2d, 0x46, 0xbc, + 0xd4, 0x8f, 0x51, 0x1c, 0xa0, 0xa4, 0x9b, 0x4f, 0x95, 0x54, 0xb0, 0x50, + 0x74, 0xfa, 0x0f, 0xe6, 0x55, 0x81, 0xce, 0x0f, 0xd1, 0x25, 0x56, 0xc8, + 0x2f, 0x3a, 0x65, 0xd4, 0x86, 0x4a, 0x8e, 0xff, 0x5a, 0xcc, 0x67, 0x96, + 0xcc, 0x65, 0x0d, 0x20, 0xee, 0xba, 0x6b, 0xcb, 0xde, 0x10, 0x2f, 0xbf, + 0x67, 0x6d, 0xbe, 0xef, 0x72, 0xfc, 0x25, 0x62, 0xbf, 0xbb, 0xc5, 0xe0, + 0x7b, 0x4c, 0x32, 0xc5, 0xdb, 0x9f, 0xb5, 0xe2, 0x75, 0x8a, 0xba, 0xbb, + 0x69, 0x28, 0xb6, 0x41, 0x25, 0x83, 0x67, 0x35, 0x1b, 0xd7, 0xb3, 0xd7, + 0x58, 0x54, 0x8a, 0x0b, 0x7c, 0xf3, 0x05, 0xcf, 0x2c, 0x78, 0x70, 0xc6, + 0xed, 0x7e, 0x56, 0xb6, 0x4e, 0x48, 0xaa, 0x57, 0xc4, 0xb0, 0xb2, 0xa0, + 0xca, 0x50, 0xe1, 0xc7, 0x41, 0xea, 0xac, 0x5f, 0x18, 0x13, 0xe5, 0x85, + 0x78, 0x3f, 0x05, 0xf3, 0xfd, 0x74, 0x7a, 0x42, 0x61, 0x91, 0x19, 0xc6, + 0x19, 0xe9, 0xd2, 0x78, 0x2c, 0xb1, 0xa3, 0x7f, 0x62, 0xea, 0x2a, 0x35, + 0x1c, 0x55, 0xa3, 0xf7, 0xdc, 0xec, 0x48, 0x23, 0x99, 0x8d, 0xe1, 0x4d, + 0x45, 0xad, 0x92, 0xc6, 0xf4, 0xa2, 0xe5, 0xe6, 0x58, 0xe4, 0xd5, 0x37, + 0xd0, 0x47, 0x0b, 0x64, 0x68, 0x48, 0x7e, 0xeb, 0xbe, 0x5e, 0x74, 0xd1, + 0xc4, 0xa5, 0x60, 0xd0, 0x30, 0x62, 0xbc, 0x81, 0xc4, 0x01, 0x68, 0x18, + 0xf3, 0xac, 0x9d, 0xb1, 0x4d, 0xdd, 0x8b, 0xd2, 0x54, 0x5d, 0xd1, 0x1c, + 0xee, 0x75, 0x9e, 0x99, 0x42, 0x69, 0x38, 0xcc, 0x66, 0x24, 0xd9, 0x8f, + 0x70, 0x98, 0xc3, 0x5e, 0x08, 0xf0, 0xd8, 0x2d, 0xe6, 0x52, 0x48, 0xdf, + 0xd0, 0x03, 0x04, 0x92, 0xab, 0xa1, 0xa1, 0x2f, 0x7d, 0x84, 0xb2, 0x82, + 0x51, 0x56, 0x74, 0x4a, 0x94, 0xff, 0xd2, 0xe4, 0x4e, 0x1a, 0xbd, 0x18, + 0xab, 0x33, 0x68, 0x0e, 0x4f, 0x99, 0x1d, 0x7e, 0x02, 0x3f, 0x1f, 0x50, + 0x05, 0xf8, 0x59, 0x47, 0x97, 0x98, 0x60, 0xb1, 0x30, 0xb1, 0x14, 0xac, + 0x2c, 0x0a, 0xa8, 0x97, 0x83, 0xf5, 0x5a, 0x5c, 0x87, 0xe5, 0x36, 0x26, + 0xec, 0xb4, 0x94, 0x46, 0x9a, 0xad, 0x2b, 0x9a, 0xb7, 0xac, 0xc4, 0x1a, + 0x55, 0x53, 0xc0, 0x16, 0x91, 0x1c, 0xd6, 0xaa, 0x6b, 0xdd, 0x85, 0x6a, + 0x54, 0xec, 0x7c, 0xa1, 0xd5, 0x18, 0x00, 0x74, 0xd2, 0xf1, 0x7e, 0xad, + 0x7c, 0xa8, 0x85, 0x9b, 0xc0, 0x9f, 0x4f, 0x3b, 0xd9, 0x08, 0xc8, 0x9d, + 0x31, 0x22, 0x7a, 0x53, 0xa8, 0xbd, 0x00, 0xdf, 0xe8, 0x39, 0x52, 0xe9, + 0x14, 0x74, 0x7b, 0x53, 0xf9, 0xbd, 0x29, 0x8e, 0x5d, 0xf2, 0x35, 0x3b, + 0xe3, 0x48, 0xbf, 0xa0, 0xc4, 0x3d, 0x40, 0xb4, 0xf2, 0x7c, 0xd0, 0xe3, + 0x17, 0x11, 0x5b, 0xd6, 0x55, 0xd2, 0x54, 0xcf, 0x20, 0x8d, 0x74, 0x4a, + 0x6b, 0xe9, 0x5d, 0xfe, 0x72, 0x14, 0x6a, 0x11, 0x8b, 0x14, 0x19, 0xba, + 0x63, 0xe4, 0x6b, 0x39, 0xb4, 0x90, 0x67, 0x79, 0x56, 0x31, 0xd3, 0xb5, + 0xeb, 0x9e, 0x95, 0x4b, 0x1e, 0x04, 0x20, 0xd8, 0xbe, 0xe8, 0x1c, 0xd7, + 0x95, 0xcb, 0x57, 0x60, 0xe6, 0x11, 0x35, 0x42, 0x90, 0xfd, 0xb2, 0xe4, + 0x9b, 0x24, 0x70, 0xc0, 0xc3, 0xa9, 0x8a, 0xc9, 0x46, 0xd0, 0xea, 0xc9, + 0x93, 0x7d, 0x9f, 0x64, 0x12, 0x54, 0x09, 0xb7, 0xc2, 0x4d, 0x6e, 0xcc, + 0x60, 0x07, 0x36, 0x31, 0x64, 0x3d, 0x1e, 0xd3, 0x86, 0x47, 0x47, 0x42, + 0x76, 0xb6, 0xf0, 0xe5, 0xb4, 0xe7, 0xbe, 0x47, 0x91, 0x78, 0xbe, 0x06, + 0xf1, 0x6e, 0x58, 0xce, 0x32, 0x13, 0x26, 0x34, 0x92, 0xae, 0xb2, 0x29, + 0xd0, 0x30, 0x55, 0xfd, 0x89, 0x6a, 0xbf, 0x3e, 0xdf, 0x11, 0x39, 0xe4, + 0xfd, 0x56, 0xd7, 0x2f, 0x89, 0x96, 0x08, 0x54, 0xaa, 0xab, 0x8b, 0xfa, + 0x65, 0xe5, 0x64, 0xff, 0x24, 0x25, 0x8f, 0x7d, 0xf6, 0xb1, 0x7f, 0x2f, + 0xa6, 0xf6, 0x46, 0xab, 0x61, 0xfd, 0x47, 0xad, 0x6d, 0x38, 0x6d, 0xc1, + 0xe9, 0x4a, 0xf1, 0x85, 0x05, 0x0e, 0x69, 0x48, 0x7c, 0xa6, 0x76, 0x61, + 0xe3, 0x94, 0xf2, 0xd6, 0x7a, 0x9c, 0x79, 0xc0, 0x2a, 0x51, 0x23, 0xc6, + 0xaf, 0x29, 0x04, 0x0f, 0x47, 0xc2, 0x93, 0xd7, 0x64, 0xe5, 0x37, 0x2e, + 0x53, 0x3b, 0xb7, 0x7c, 0x9c, 0xb4, 0x63, 0x13, 0xc7, 0x56, 0x90, 0xe9, + 0x53, 0xd5, 0x86, 0x2b, 0x96, 0x41, 0x42, 0x56, 0xc5, 0x16, 0xd7, 0x9e, + 0x30, 0xce, 0xa1, 0x0d, 0x93, 0x5d, 0x11, 0x07, 0xb2, 0x95, 0xfd, 0xf6, + 0x0b, 0x28, 0x95, 0x1a, 0x8f, 0xfa, 0xe1, 0x57, 0x7e, 0x06, 0xff, 0x18, + 0xaf, 0xe3, 0x4f, 0x3c, 0x34, 0x5b, 0xd4, 0x46, 0x1a, 0xd1, 0xd1, 0x7e, + 0x55, 0xba, 0x5d, 0x2a, 0x1f, 0x42, 0x49, 0x95, 0x75, 0x5f, 0x80, 0x60, + 0x02, 0x01, 0xdb, 0x36, 0xad, 0x68, 0x69, 0x1e, 0x0b, 0x90, 0x3f, 0xa6, + 0xb6, 0x2f, 0x66, 0xa6, 0x7d, 0x81, 0x8c, 0xa0, 0xee, 0x05, 0x95, 0xbc, + 0xb3, 0x7c, 0x18, 0xd4, 0x1b, 0x40, 0x96, 0xf5, 0x05, 0x9d, 0x27, 0x3b, + 0x78, 0xfc, 0x19, 0x18, 0xc0, 0x61, 0xa0, 0xd6, 0xf9, 0xc0, 0x3f, 0xe5, + 0x48, 0x35, 0x0f, 0x8b, 0x0d, 0xfb, 0x31, 0xb7, 0x32, 0x40, 0x1d, 0x69, + 0x12, 0x5a, 0x23, 0xf0, 0xce, 0xe9, 0x5e, 0xa6, 0x68, 0x6b, 0xe1, 0xe2, + 0x68, 0x07, 0x02, 0x0d, 0x7a, 0xc2, 0x0a, 0x40, 0x10, 0x5e, 0x94, 0xba, + 0x77, 0x1d, 0xf7, 0xac, 0xec, 0x79, 0xa9, 0xa1, 0x8a, 0xb8, 0x49, 0x32, + 0x08, 0xe0, 0x18, 0xa8, 0x3d, 0x69, 0x41, 0x5d, 0x30, 0x3b, 0xb6, 0x91, + 0x46, 0x8d, 0x81, 0x10, 0xb0, 0xc2, 0xed, 0xa0, 0x4e, 0x59, 0x48, 0xd8, + 0x64, 0x7d, 0x2d, 0x46, 0xf2, 0x8a, 0x2e, 0x5d, 0x0c, 0x4d, 0x9f, 0xfe, + 0x7b, 0x5e, 0xbf, 0x1a, 0x78, 0xdf, 0xfc, 0x0f, 0x04, 0x37, 0x72, 0x1a, + 0x09, 0xb8, 0x6e, 0x1b, 0xf1, 0x18, 0x7d, 0x83, 0x44, 0xaa, 0x9b, 0x71, + 0xe1, 0x03, 0x04, 0x83, 0xe5, 0xaa, 0xc0, 0xd4, 0xa7, 0x80, 0x10, 0x35, + 0x09, 0xae, 0xf7, 0xe1, 0x5e, 0x7c, 0x31, 0x20, 0x43, 0x82, 0xda, 0x07, + 0x39, 0xfe, 0x8f, 0x9d, 0x70, 0x3c, 0x57, 0x43, 0x01, 0x51, 0x37, 0x2e, + 0x97, 0xef, 0xcf, 0x05, 0x44, 0x75, 0x69, 0xf7, 0xdb, 0xda, 0x80, 0x78, + 0x0c, 0xcc, 0xc1, 0x49, 0xac, 0x3b, 0x7e, 0x27, 0x6a, 0xbb, 0xdf, 0x45, + 0x5b, 0x3b, 0x29, 0xf6, 0x1b, 0xa9, 0x25, 0xf9, 0x2f, 0xcf, 0x37, 0x71, + 0x33, 0xb4, 0x90, 0xd7, 0x9b, 0x87, 0x41, 0x15, 0xd1, 0xa6, 0x39, 0xa7, + 0xa9, 0xcd, 0x66, 0x29, 0x59, 0xb4, 0x53, 0x12, 0xa1, 0x20, 0xd5, 0x04, + 0xca, 0x40, 0x31, 0xfa, 0x6f, 0xbb, 0x92, 0x04, 0xf3, 0xc2, 0x10, 0x0d, + 0xc1, 0x19, 0x78, 0x8c, 0x82, 0xed, 0x92, 0x3a, 0x6b, 0xd1, 0x3d, 0xe8, + 0xac, 0x55, 0xe4, 0x8c, 0xc6, 0xd4, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x01, 0x00, 0x00, 0xc2, 0x1d, 0x86, 0xe4, 0xf6, 0xa1, 0xbe, 0xf5, + 0xf3, 0x36, 0x9d, 0x32, 0x80, 0x17, 0x3b, 0x1f, 0x18, 0x21, 0xed, 0xa7, + 0xf5, 0xaf, 0xf1, 0x94, 0xe2, 0xa7, 0x08, 0xd5, 0xca, 0x18, 0x45, 0xf5, + 0x68, 0x94, 0x82, 0x61, 0xf7, 0xb7, 0xb2, 0xfa, 0xd4, 0x5e, 0x32, 0xd0, + 0xf0, 0x20, 0x66, 0x83, 0xd1, 0x6b, 0x3c, 0xdf, 0x73, 0xeb, 0x73, 0x82, + 0x09, 0x9b, 0xd0, 0xc5, 0xb0, 0x9f, 0x01, 0x77, 0x85, 0xcc, 0x6e, 0x23, + 0xb7, 0x00, 0x45, 0xe0, 0xa6, 0x01, 0x29, 0x1d, 0x8b, 0xc4, 0xe0, 0xc2, + 0xe0, 0x4f, 0x3b, 0x07, 0xd5, 0xac, 0x6b, 0x88, 0xb8, 0xa4, 0xe2, 0x5c, + 0x19, 0xe9, 0x98, 0x72, 0xa5, 0x6b, 0xf5, 0xa4, 0xf7, 0x15, 0xaf, 0xfb, + 0xb4, 0x80, 0x9a, 0xe3, 0xa5, 0x35, 0x2f, 0x45, 0x81, 0xf1, 0x8b, 0x2d, + 0x26, 0x5c, 0x65, 0xa9, 0x5b, 0x6e, 0x83, 0xc3, 0x62, 0x2f, 0x84, 0xef, + 0x11, 0xa5, 0x58, 0x48, 0xe9, 0x67, 0x7e, 0xd3, 0x0b, 0x5d, 0x51, 0x80, + 0x39, 0x08, 0x8e, 0xc1, 0x0d, 0x04, 0x11, 0x5f, 0x72, 0x64, 0x1f, 0x83, + 0xf8, 0xd3, 0x09, 0x38, 0xb6, 0x7f, 0x50, 0x78, 0x27, 0x20, 0xe5, 0xbd, + 0x16, 0xbf, 0x51, 0xd8, 0x4f, 0x67, 0x60, 0xf6, 0x9e, 0xff, 0x08, 0xfe, + 0xc6, 0x96, 0xd6, 0x64, 0x94, 0x28, 0xc6, 0x9a, 0x09, 0x1a, 0x34, 0x08, + 0x31, 0x4b, 0x0b, 0x97, 0x5a, 0x18, 0x72, 0x49, 0xe9, 0x1d, 0xbb, 0x9c, + 0xed, 0x7e, 0xb5, 0xc5, 0xa7, 0xf4, 0x25, 0x7a, 0x26, 0xe9, 0x15, 0x61, + 0x85, 0x32, 0xc9, 0xb3, 0xcf, 0x95, 0xbf, 0x35, 0x10, 0x2d, 0x71, 0xfe, + 0x03, 0xd6, 0x69, 0x75, 0x8d, 0xb7, 0x16, 0xa7, 0x3d, 0x0e, 0xb7, 0x55, + 0x6d, 0xa7, 0x9f, 0x10, 0x7e, 0x7e, 0xff, 0x39, 0xee, 0x8e, 0xa7, 0x81, + 0x7d, 0x11, 0xea, 0xa9, 0xd6, 0xed, 0x54, 0xf8, 0xd2, 0xd5, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0xf9, 0xde, 0x41, 0xe7, + 0xa6, 0x88, 0x53, 0x76, 0x5a, 0x26, 0xc3, 0x5c, 0xf2, 0x58, 0x68, 0x9c, + 0xc7, 0x4e, 0x53, 0x18, 0x53, 0x67, 0x39, 0x23, 0x96, 0xb0, 0xef, 0x58, + 0x29, 0xe1, 0x68, 0xd8, 0xce, 0xc0, 0x41, 0xc2, 0x35, 0x5f, 0x74, 0xfa, + 0xdf, 0xc7, 0x0f, 0x80, 0x50, 0xd1, 0xf6, 0x5a, 0x3a, 0x81, 0xe0, 0xd9, + 0x9b, 0x47, 0x96, 0xcd, 0xc5, 0x0f, 0x91, 0x12, 0x81, 0x77, 0x1e, 0xef, + 0x2e, 0xba, 0x16, 0x51, 0x70, 0x78, 0xdc, 0xa3, 0x84, 0x12, 0x7c, 0x9e, + 0x21, 0x7d, 0xa3, 0x5f, 0xce, 0xa1, 0x25, 0x84, 0x99, 0xa4, 0x2d, 0xa6, + 0x0f, 0x95, 0xef, 0xef, 0x31, 0xe6, 0xf2, 0x18, 0x08, 0x47, 0xd2, 0x5a, + 0x39, 0x01, 0x7a, 0xca, 0xd3, 0x03, 0xb1, 0xc2, 0x48, 0xf4, 0x1f, 0x6d, + 0xc2, 0x8c, 0x5c, 0xda, 0xf5, 0x10, 0xed, 0xfc, 0x2e, 0x0c, 0xb3, 0x52, + 0xaa, 0xa9, 0xed, 0xbc, 0x41, 0xcc, 0xd4, 0x4b, 0x1c, 0xd0, 0xa3, 0x1d, + 0xf4, 0xe7, 0x48, 0x34, 0x4e, 0xcf, 0x3b, 0xb3, 0x71, 0x06, 0xbe, 0x0c, + 0x35, 0xbb, 0xb4, 0x17, 0xd8, 0x8b, 0xba, 0xdd, 0x32, 0x30, 0x51, 0xb1, + 0xb1, 0xd6, 0x3a, 0xdc, 0x3b, 0x25, 0x9a, 0x57, 0xc7, 0x4d, 0xd3, 0x75, + 0x93, 0x59, 0x3e, 0x9b, 0x10, 0xcf, 0xdb, 0x38, 0x75, 0x51, 0xb2, 0x2a, + 0x48, 0x78, 0xfc, 0xaa, 0xe3, 0x91, 0xe7, 0x93, 0xe7, 0x0a, 0x07, 0x2c, + 0xf8, 0x88, 0x93, 0xde, 0x2f, 0xba, 0x7b, 0x72, 0xcd, 0x92, 0xdd, 0xb1, + 0xac, 0x1e, 0xe4, 0xe3, 0x5d, 0xa4, 0x7f, 0x86, 0xa7, 0xcb, 0xb5, 0x81, + 0x86, 0xf1, 0xf5, 0xad, 0xd6, 0x36, 0x08, 0x09, 0x9f, 0x75, 0x6f, 0x4a, + 0x5b, 0x30, 0xf8, 0xaf, 0xd2, 0xbc, 0xb5, 0xbe, 0xf2, 0xeb, 0x9b, 0xbc, + 0x11, 0xd4, 0x0c, 0x14, 0xa6, 0x6f, 0x43, 0xd3, 0xc9, 0x4e, 0xca, 0x9b, + 0x4e, 0x46, 0x60, 0x4c, 0x63, 0xcc, 0x07, 0x36, 0x8c, 0xf2, 0xd1, 0x93, + 0x7a, 0x51, 0x49, 0x15, 0xbf, 0xbf, 0x9e, 0x82, 0x21, 0x06, 0xa0, 0x39, + 0x11, 0x1d, 0x6c, 0x41, 0x72, 0xcd, 0x2a, 0x8a, 0x4a, 0xd0, 0x13, 0x6c, + 0x56, 0xf4, 0x00, 0x48, 0xaf, 0xab, 0xdf, 0xa9, 0xe9, 0xa6, 0xaa, 0x06, + 0x61, 0x79, 0xc4, 0x57, 0x42, 0xca, 0x12, 0x18, 0xcf, 0x81, 0xec, 0x79, + 0x19, 0xd2, 0xd2, 0xe3, 0x1d, 0xc6, 0x6c, 0xd0, 0xd6, 0x0a, 0xfb, 0x70, + 0x42, 0x28, 0x25, 0x23, 0xb6, 0x23, 0x15, 0x28, 0x5e, 0x9f, 0x49, 0xf2, + 0x7b, 0x69, 0x74, 0xa5, 0xb9, 0x26, 0x81, 0xfe, 0x39, 0x3e, 0x3f, 0xc8, + 0x7e, 0x9e, 0x5e, 0x8e, 0xf2, 0xdb, 0x6b, 0xfd, 0xe1, 0xc3, 0x01, 0x4a, + 0xba, 0x8f, 0x33, 0x71, 0x09, 0x80, 0x5d, 0x9c, 0x58, 0x64, 0xb7, 0x90, + 0x13, 0x2a, 0xe9, 0x1d, 0x07, 0x2c, 0x06, 0x70, 0x43, 0x0d, 0xb6, 0x57, + 0x02, 0x3c, 0xbe, 0x3c, 0x42, 0xab, 0x77, 0x15, 0x0e, 0x98, 0xfb, 0xf2, + 0x1d, 0x14, 0xd9, 0xb8, 0xd1, 0x59, 0x2a, 0x67, 0x6f, 0xfc, 0x59, 0x39, + 0x33, 0xe0, 0x49, 0x0b, 0x4e, 0x65, 0x81, 0x9f, 0x71, 0xf2, 0xa5, 0x90, + 0x4f, 0x24, 0xc7, 0x05, 0xfb, 0x77, 0x1e, 0x14, 0xca, 0x2f, 0xfc, 0xac, + 0xec, 0xbf, 0xa2, 0x69, 0x15, 0x0a, 0x6b, 0xa9, 0xa0, 0x74, 0xee, 0xad, + 0xa9, 0x50, 0x4d, 0x4d, 0xab, 0x6e, 0xc1, 0xb3, 0xda, 0xbb, 0xbd, 0xab, + 0x00, 0x05, 0x14, 0xc1, 0xc4, 0x53, 0x7b, 0x78, 0x97, 0x68, 0x3c, 0x05, + 0xf2, 0xed, 0x87, 0xca, 0x86, 0xd1, 0xdf, 0xda, 0xb3, 0x2f, 0x17, 0x87, + 0x87, 0x2f, 0xd8, 0xe9, 0xb2, 0x96, 0xdc, 0x7f, 0x22, 0xf1, 0x2a, 0x9f, + 0xfe, 0x54, 0x55, 0xa1, 0x96, 0xab, 0x9f, 0x61, 0x74, 0xcd, 0x4d, 0x77, + 0x38, 0x02, 0x23, 0x29, 0x28, 0x5b, 0xfc, 0x86, 0x17, 0x40, 0xd4, 0x42, + 0x2a, 0x9b, 0x84, 0xf7, 0x67, 0x2b, 0x3a, 0xc1, 0x31, 0x89, 0x4b, 0x67, + 0xd1, 0x7d, 0x6b, 0x36, 0xec, 0x69, 0x6b, 0x24, 0xca, 0xd6, 0x2d, 0xbb, + 0x21, 0xc8, 0x0c, 0x53, 0x41, 0x29, 0x0b, 0xc1, 0xfe, 0xd5, 0xa3, 0x4c, + 0x66, 0x2f, 0xc7, 0xf1, 0xa8, 0xc0, 0x3d, 0x9a, 0xb9, 0x09, 0x50, 0x3f, + 0x09, 0x87, 0xa4, 0x3f, 0x7a, 0x33, 0xef, 0xf0, 0xfb, 0x77, 0x02, 0x7d, + 0x92, 0xaf, 0x73, 0xaa, 0xcc, 0x3f, 0x66, 0x56, 0xd0, 0x21, 0xd1, 0xe8, + 0x0e, 0x47, 0x03, 0x5e, 0x3b, 0xe9, 0xa2, 0xe3, 0x83, 0x0b, 0x73, 0xd3, + 0xaa, 0x94, 0x80, 0xef, 0x7c, 0xdf, 0xde, 0x86, 0xc3, 0xa9, 0x62, 0x34, + 0x76, 0xee, 0x4d, 0x15, 0x73, 0x7b, 0xd7, 0x6d, 0xd4, 0x21, 0x05, 0xd4, + 0xcf, 0xf3, 0x54, 0xdc, 0x49, 0x5f, 0x5a, 0x2a, 0x37, 0x19, 0x89, 0x61, + 0x1d, 0x95, 0x17, 0x8b, 0x09, 0x95, 0x5d, 0x9f, 0xde, 0x86, 0x03, 0x93, + 0x76, 0xec, 0x54, 0xec, 0x13, 0xc3, 0xf9, 0x38, 0x8f, 0xa9, 0x11, 0xf0, + 0x9a, 0x0e, 0x5e, 0x38, 0x69, 0xeb, 0x62, 0x41, 0x9e, 0xd0, 0x1b, 0x59, + 0x8c, 0xfd, 0x16, 0xfa, 0xd8, 0x99, 0x0d, 0x83, 0x7e, 0xba, 0x5b, 0xc6, + 0x59, 0xe1, 0xae, 0xba, 0xb9, 0xb8, 0xba, 0xa5, 0x4d, 0x20, 0x00, 0xc9, + 0x0c, 0xe1, 0x77, 0xdf, 0xc4, 0x95, 0xca, 0x7c, 0xa5, 0xef, 0x0a, 0xed, + 0x9b, 0x31, 0x06, 0xe1, 0xc9, 0xa3, 0x88, 0x0a, 0xcc, 0x3d, 0xc8, 0xb6, + 0x01, 0xe2, 0xa9, 0x29, 0x03, 0x8a, 0x28, 0xf8, 0x0d, 0x70, 0x77, 0xb9, + 0xe1, 0x1b, 0x06, 0x19, 0x86, 0xc1, 0xd3, 0xcf, 0x6b, 0x9c, 0x09, 0x70, + 0x50, 0xed, 0xb5, 0xf6, 0x69, 0xcc, 0xac, 0x30, 0x6a, 0x1f, 0x1d, 0xe6, + 0x75, 0x33, 0xab, 0x55, 0x48, 0xfa, 0x81, 0xb8, 0x06, 0x3a, 0x78, 0xee, + 0xde, 0xef, 0xe2, 0x17, 0xc4, 0x3e, 0xe5, 0x22, 0xa7, 0xd1, 0x45, 0x5b, + 0x57, 0xb0, 0xde, 0x69, 0x30, 0xd1, 0x9a, 0xd7, 0x6b, 0x0e, 0x7a, 0x30, + 0x0d, 0xb5, 0xec, 0x60, 0xa7, 0x05, 0x87, 0x42, 0x4b, 0x92, 0x1f, 0x68, + 0x8e, 0x1a, 0x90, 0x84, 0x27, 0x2a, 0xc0, 0xd2, 0xff, 0xbc, 0x8e, 0x34, + 0x53, 0x9d, 0x04, 0x50, 0xcb, 0x79, 0xd9, 0x55, 0xd5, 0x4d, 0x3c, 0xe2, + 0xb4, 0x9b, 0x57, 0x07, 0x1f, 0xce, 0xd0, 0xa7, 0x84, 0xe1, 0xb7, 0x3a, + 0xaf, 0xc5, 0x67, 0x64, 0xbc, 0x02, 0xbe, 0xb0, 0x65, 0x7e, 0xb0, 0x4c, + 0xc2, 0x2d, 0xcd, 0xf8, 0x60, 0xcb, 0xfe, 0xd1, 0x8d, 0x14, 0x5a, 0xd3, + 0x38, 0xd4, 0x71, 0x5a, 0xca, 0xbb, 0xfe, 0x0e, 0x54, 0xf9, 0xb4, 0x25, + 0xa5, 0x71, 0x13, 0x95, 0x14, 0xdc, 0x86, 0xb8, 0x21, 0xa7, 0x2e, 0x13, + 0xc6, 0x2f, 0xce, 0xe7, 0x6c, 0xb8, 0x0d, 0xc9, 0xe4, 0xc4, 0x64, 0x12, + 0x78, 0x1c, 0x95, 0x92, 0xc2, 0xec, 0xaa, 0xd3, 0xc3, 0x3a, 0xd2, 0xe8, + 0x95, 0xf0, 0x6b, 0x03, 0x8c, 0xcf, 0x6b, 0xdb, 0x21, 0xa0, 0xcf, 0xf4, + 0x05, 0xc8, 0xe7, 0x77, 0x05, 0x55, 0x7b, 0x6b, 0xfa, 0x96, 0xf1, 0x7c, + 0x30, 0x62, 0x75, 0xbe, 0x6e, 0xea, 0xba, 0x9f, 0x40, 0x2e, 0x9a, 0x86, + 0x93, 0xcc, 0x38, 0xf7, 0xee, 0xd8, 0xbb, 0x24, 0xcd, 0x85, 0x3e, 0x85, + 0x16, 0x8c, 0x33, 0x23, 0x73, 0xe6, 0x43, 0xc4, 0x67, 0xbf, 0xef, 0x85, + 0xb1, 0x44, 0xf9, 0x55, 0x93, 0x4d, 0x0b, 0x8e, 0xc1, 0x42, 0x13, 0xc6, + 0xc8, 0x09, 0x63, 0xab, 0xb3, 0xc7, 0xc4, 0xa4, 0x8b, 0x72, 0xfb, 0xa5, + 0x99, 0xa1, 0x5d, 0x07, 0x02, 0x82, 0x56, 0x11, 0x3c, 0xc2, 0x5a, 0x55, + 0xf9, 0x3a, 0x93, 0x61, 0x89, 0x46, 0xb7, 0x6a, 0x42, 0x76, 0x1e, 0x70, + 0xde, 0xd9, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, + 0x32, 0xc1, 0x61, 0xaa, 0xdb, 0xe9, 0xae, 0x88, 0xcb, 0xf7, 0x28, 0xdd, + 0x82, 0x62, 0x61, 0x41, 0x4e, 0xbb, 0xf9, 0xb7, 0xe8, 0x81, 0x99, 0x18, + 0xe2, 0xa7, 0xb4, 0x7c, 0xb7, 0x08, 0x44, 0x6f, 0x24, 0xb3, 0xda, 0x57, + 0x62, 0x29, 0xc7, 0xa6, 0x84, 0xb1, 0x5d, 0xc5, 0x00, 0x4c, 0x30, 0x16, + 0xf0, 0x0a, 0x74, 0x73, 0xec, 0xaf, 0xb5, 0xde, 0xb0, 0xa7, 0x75, 0x22, + 0x8f, 0x9e, 0x43, 0x01, 0x68, 0xae, 0x91, 0xeb, 0x46, 0x52, 0x3f, 0x2c, + 0x4e, 0xc5, 0xd0, 0xc8, 0x15, 0xea, 0x99, 0xc2, 0x37, 0x5b, 0x68, 0xb5, + 0xce, 0x41, 0x92, 0xbf, 0xd6, 0xdb, 0x85, 0xad, 0x08, 0xd1, 0x11, 0x93, + 0xe8, 0xd4, 0x78, 0x43, 0x3b, 0x7d, 0xcb, 0x42, 0x84, 0xf3, 0x61, 0x88, + 0x9e, 0x6a, 0x73, 0xb9, 0x78, 0x17, 0x9a, 0x9f, 0xfb, 0x97, 0xcb, 0xd6, + 0xb5, 0x3f, 0x00, 0x41, 0xb0, 0x30, 0x2f, 0x6f, 0x89, 0xdd, 0xfa, 0x13, + 0xd1, 0x07, 0xbe, 0x2f, 0xea, 0x91, 0x62, 0xaa, 0xed, 0xcb, 0xfd, 0x07, + 0x82, 0xbb, 0x3f, 0xf4, 0xa6, 0x94, 0x66, 0x71, 0x20, 0x61, 0xac, 0x84, + 0x04, 0x70, 0xf2, 0xd3, 0xdf, 0xac, 0x44, 0xfd, 0x47, 0x26, 0x81, 0x64, + 0xb3, 0xa6, 0x90, 0x2b, 0xd2, 0x2c, 0xd0, 0x77, 0x81, 0x53, 0x45, 0x78, + 0x5f, 0x30, 0x77, 0x91, 0x83, 0x13, 0x33, 0xd1, 0x91, 0xa6, 0x35, 0x21, + 0xcb, 0x26, 0x54, 0x0a, 0xf7, 0x70, 0x5e, 0xdb, 0xd8, 0x92, 0xc7, 0xdf, + 0xf9, 0x2a, 0x46, 0x91, 0x22, 0x3b, 0xe6, 0xe1, 0x91, 0xeb, 0xa6, 0x78, + 0x81, 0x57, 0xf3, 0x04, 0xdf, 0x34, 0x55, 0x74, 0x0a, 0xfe, 0xf2, 0xbd, + 0xb3, 0xeb, 0xa3, 0x8e, 0x71, 0x15, 0xa9, 0x2f, 0x53, 0xe2, 0xa1, 0x45, + 0xdf, 0xe8, 0x29, 0x40, 0xf1, 0x4b, 0x23, 0xdb, 0x8e, 0xee, 0x19, 0xa8, + 0xd4, 0x15, 0x90, 0x8c, 0x04, 0x46, 0x81, 0x49, 0x92, 0xe5, 0xe1, 0xfe, + 0x99, 0x06, 0xfc, 0x3e, 0x43, 0x58, 0x3b, 0x19, 0x7f, 0xd2, 0x13, 0x65, + 0xc2, 0x64, 0x27, 0x6d, 0x93, 0x6a, 0xcf, 0x48, 0x2a, 0x3d, 0xdd, 0x79, + 0x9f, 0x05, 0x32, 0xeb, 0xfd, 0xb4, 0xd2, 0x1d, 0x16, 0x61, 0x3d, 0x17, + 0x4c, 0xb8, 0xad, 0x63, 0x0e, 0x6b, 0x8a, 0x4a, 0x34, 0x4c, 0xb5, 0x3c, + 0x0f, 0x05, 0x28, 0x8c, 0x8b, 0xdf, 0xf4, 0xa0, 0x49, 0xbf, 0x34, 0x6c, + 0x6a, 0x5f, 0x40, 0x95, 0x48, 0x4b, 0x93, 0x1e, 0x61, 0x6d, 0x58, 0xc3, + 0x86, 0x98, 0x70, 0x11, 0x4e, 0x44, 0x65, 0xc1, 0x0d, 0xea, 0x2f, 0xda, + 0x38, 0x16, 0xbd, 0xd4, 0x7b, 0x3e, 0x31, 0xee, 0x42, 0x4c, 0xdc, 0xe9, + 0x8b, 0x1f, 0xa9, 0xcf, 0xab, 0x60, 0xb5, 0xb1, 0xd2, 0xf2, 0x6a, 0xe9, + 0xbc, 0xcc, 0xcb, 0x60, 0x4a, 0xca, 0x70, 0x79, 0x64, 0x9d, 0x07, 0x1e, + 0xdb, 0xef, 0x34, 0xaf, 0x17, 0x93, 0x6b, 0x60, 0x73, 0x2d, 0x8c, 0x08, + 0x27, 0x1e, 0x46, 0x9f, 0xcb, 0x33, 0xdd, 0x76, 0xef, 0x17, 0x58, 0x9a, + 0x5f, 0x82, 0x78, 0x0f, 0xbf, 0xe7, 0x0f, 0x3a, 0x1e, 0xa8, 0x30, 0xbf, + 0xff, 0xc7, 0xc7, 0x82, 0x8b, 0xc3, 0x65, 0x04, 0xfd, 0x45, 0xc9, 0x88, + 0x99, 0x8e, 0x44, 0xc5, 0x23, 0x1e, 0xbf, 0xf1, 0x95, 0x70, 0x35, 0xe6, + 0x56, 0x4a, 0x53, 0xb2, 0xac, 0x0c, 0xfd, 0xf5, 0x61, 0x26, 0x5b, 0x70, + 0xd6, 0x4c, 0xfc, 0x0f, 0xcc, 0x53, 0x6e, 0x25, 0xca, 0x1d, 0x0c, 0x56, + 0xf7, 0x9c, 0x95, 0xf6, 0x3c, 0x08, 0x0c, 0x64, 0xb1, 0x1c, 0x5c, 0xe6, + 0x25, 0xa4, 0xa3, 0xb7, 0xaf, 0x8b, 0xbc, 0xe1, 0x68, 0xdf, 0x10, 0xab, + 0xbb, 0xd5, 0x30, 0x64, 0x42, 0xf6, 0xe6, 0x9a, 0xb5, 0x59, 0x12, 0x76, + 0x92, 0xac, 0x29, 0xe9, 0x45, 0xdb, 0x2e, 0x62, 0x22, 0x58, 0x24, 0x89, + 0xc8, 0x6a, 0x2a, 0xa7, 0x3f, 0x04, 0x53, 0x4e, 0x07, 0x41, 0x4e, 0x5f, + 0x95, 0x5f, 0x6e, 0x14, 0x5b, 0xa7, 0xa7, 0xd3, 0x5a, 0xa2, 0x95, 0x4a, + 0xc8, 0xe9, 0x3c, 0x5a, 0x84, 0x50, 0xbc, 0xe1, 0x9c, 0x7a, 0x16, 0xe5, + 0xc7, 0x04, 0x9d, 0x60, 0x2e, 0x7d, 0xb3, 0x77, 0x5d, 0x86, 0x2e, 0xac, + 0x57, 0x2a, 0x31, 0x26, 0x23, 0x6e, 0xcc, 0x7f, 0xb8, 0x36, 0x29, 0xa9, + 0xa8, 0xd9, 0xc6, 0x75, 0xee, 0x16, 0x23, 0x27, 0x0f, 0xe1, 0xb0, 0x3d, + 0x91, 0x3a, 0x26, 0x4a, 0x60, 0x72, 0x14, 0xf9, 0x3c, 0x66, 0x66, 0xe8, + 0x7d, 0x4a, 0x6f, 0x7e, 0x63, 0x58, 0x6a, 0x28, 0x78, 0x50, 0xef, 0x3b, + 0x9d, 0xeb, 0xb6, 0x4b, 0x5d, 0x55, 0x80, 0x84, 0x97, 0x9b, 0x74, 0x4b, + 0x5c, 0x09, 0x1d, 0xe7, 0x57, 0xfc, 0x40, 0x3f, 0xa9, 0xbd, 0xdf, 0x61, + 0x2a, 0x89, 0x62, 0x51, 0xfc, 0x24, 0xee, 0xee, 0x97, 0x10, 0xca, 0xb6, + 0x0e, 0x8e, 0x71, 0x67, 0x2a, 0x79, 0x4f, 0xc4, 0xe6, 0x3e, 0x27, 0xc2, + 0x9b, 0x85, 0xfd, 0xde, 0xfb, 0x58, 0x75, 0xf3, 0x1c, 0x31, 0xa2, 0x56, + 0x3e, 0xdc, 0x24, 0xf4, 0x4f, 0xcb, 0x5a, 0x1a, 0x77, 0x5c, 0x28, 0xd1, + 0x5a, 0x55, 0xa9, 0x8c, 0xb5, 0xdd, 0x77, 0x93, 0x58, 0xd8, 0x2f, 0x7d, + 0x5a, 0x67, 0xa1, 0x95, 0x0a, 0xd2, 0x6a, 0x93, 0xa6, 0xf0, 0x5f, 0x7f, + 0x0a, 0x29, 0xdb, 0x1d, 0x8c, 0xa7, 0x12, 0x0a, 0xf4, 0xc9, 0xcd, 0x70, + 0xd1, 0xbd, 0x48, 0xd4, 0x9a, 0xbb, 0xbb, 0x24, 0xbf, 0x52, 0x25, 0xb9, + 0x75, 0xc2, 0x17, 0x36, 0x6f, 0x4a, 0xc0, 0x53, 0x6d, 0x38, 0xfb, 0x7a, + 0x60, 0xc8, 0x5d, 0x03, 0xc1, 0x1c, 0x0c, 0x31, 0xf0, 0x59, 0xed, 0x0a, + 0x5f, 0x84, 0xf2, 0x89, 0x6c, 0xb4, 0xd5, 0x24, 0x2d, 0x2a, 0xda, 0xbe, + 0x74, 0x1d, 0x22, 0xe2, 0xc6, 0xf0, 0x9b, 0x98, 0x5a, 0x41, 0x11, 0x4c, + 0x51, 0x97, 0x16, 0xa7, 0xc9, 0xd8, 0x53, 0x12, 0x53, 0xdd, 0x22, 0xa9, + 0xf2, 0xae, 0x52, 0x49, 0x02, 0xf9, 0x5c, 0x78, 0x00, 0xa2, 0x64, 0xff, + 0x91, 0x62, 0x20, 0x6a, 0x87, 0x6a, 0x40, 0x01, 0x85, 0x30, 0xf5, 0xdd, + 0xa7, 0x64, 0x0a, 0x85, 0x8d, 0x37, 0x99, 0xcb, 0x03, 0xc8, 0x29, 0x56, + 0x7e, 0x75, 0x4f, 0xa1, 0xc3, 0x76, 0xce, 0xdb, 0xa3, 0xb4, 0x7e, 0x91, + 0x95, 0xbe, 0x53, 0x0e, 0x20, 0xc9, 0xe7, 0x71, 0x78, 0xad, 0x3d, 0x4c, + 0xbb, 0x59, 0xb9, 0x77, 0xcf, 0x7d, 0x7b, 0xff, 0x15, 0xdb, 0x1d, 0xae, + 0x1f, 0xbe, 0x33, 0x88, 0x01, 0x04, 0x95, 0xe5, 0xe9, 0x6a, 0x1c, 0xbf, + 0xc8, 0xc3, 0x33, 0x3b, 0xd8, 0x2f, 0x75, 0x4a, 0xc3, 0x6f, 0x09, 0x88, + 0x26, 0x46, 0x90, 0x89, 0x53, 0x12, 0x27, 0xc2, 0x7d, 0x23, 0x6b, 0xc4, + 0xe3, 0x0a, 0x0f, 0xc2, 0x86, 0x6d, 0x20, 0x35, 0x82, 0x33, 0xec, 0xdd, + 0xa7, 0x6a, 0xc3, 0xa8, 0x11, 0xdc, 0x02, 0xd9, 0x05, 0x1b, 0x04, 0x75, + 0x92, 0x6c, 0x08, 0x9e, 0x38, 0x72, 0xd9, 0x7d, 0x9b, 0xbc, 0xfd, 0xca, + 0xb8, 0x06, 0x0e, 0x24, 0x89, 0x90, 0xde, 0x52, 0xe4, 0xd1, 0xcc, 0x99, + 0x87, 0x0b, 0x87, 0xbb, 0x5c, 0xa9, 0xab, 0xec, 0xb5, 0xe4, 0xdd, 0x5d, + 0xfa, 0xb1, 0x97, 0x5f, 0x61, 0xf7, 0x58, 0xd6, 0x08, 0x02, 0xf2, 0x51, + 0x7c, 0x7a, 0xe6, 0xf1, 0xcb, 0x43, 0xd0, 0x21, 0x09, 0xb8, 0x82, 0xa9, + 0x52, 0xd9, 0xa8, 0x7f, 0x2b, 0xe1, 0x0f, 0x31, 0xbc, 0x16, 0xa2, 0xce, + 0x35, 0x55, 0x2e, 0xd6, 0xda, 0x38, 0xd9, 0xc2, 0x5e, 0xca, 0x27, 0xd9, + 0xa6, 0xd6, 0x4b, 0xa2, 0x73, 0xc4, 0xce, 0x66, 0x30, 0x60, 0xa2, 0x01, + 0xfa, 0xc1, 0xd6, 0xc8, 0xea, 0xdd, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x70, 0xe2, 0x62, 0x68, 0xff, 0x60, 0x67, 0x64, + 0x88, 0xdd, 0x81, 0x79, 0x82, 0xf5, 0x46, 0xf9, 0x7e, 0x0e, 0xa9, 0x26, + 0xf6, 0xcf, 0x5d, 0xef, 0x10, 0x11, 0xe1, 0x71, 0x72, 0x77, 0xcf, 0x02, + 0x7b, 0xf1, 0x6e, 0xc4, 0xb4, 0xfa, 0x2a, 0x12, 0xfe, 0x7e, 0x3c, 0x66, + 0xef, 0x41, 0x98, 0x3a, 0x1f, 0xa9, 0x14, 0x8f, 0x46, 0x22, 0xa0, 0xc2, + 0xee, 0x93, 0x25, 0x34, 0xf2, 0xb7, 0x6d, 0x0a, 0x36, 0xde, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0xd4, 0x17, 0x62, 0x25, + 0xfd, 0x5b, 0x75, 0xeb, 0xec, 0x06, 0xc9, 0x39, 0x86, 0x6d, 0xc5, 0x60, + 0x2d, 0x33, 0x3d, 0xce, 0x6a, 0x9f, 0x07, 0x3b, 0xb9, 0x70, 0x0f, 0xc7, + 0x13, 0x46, 0x35, 0x46, 0x26, 0xe4, 0xbc, 0x6e, 0x54, 0x89, 0x29, 0xd5, + 0xa4, 0x94, 0xa0, 0x3a, 0x7a, 0x61, 0xcf, 0xd1, 0x48, 0x27, 0x7a, 0x72, + 0x95, 0xde, 0x93, 0xd1, 0x19, 0x1f, 0xc9, 0xc8, 0x8f, 0x0d, 0xce, 0x34, + 0x03, 0x39, 0x0a, 0x92, 0x16, 0x09, 0xc4, 0x49, 0xf9, 0x30, 0x2e, 0x19, + 0xd1, 0x69, 0x7e, 0x78, 0x00, 0x25, 0x30, 0x6f, 0x6b, 0xe1, 0xbe, 0xad, + 0xb2, 0x05, 0xde, 0xc7, 0xc2, 0xf7, 0xd5, 0xa7, 0x4d, 0x03, 0x6f, 0x6b, + 0xcd, 0xcb, 0x42, 0xfa, 0x88, 0x16, 0xd5, 0xa6, 0x60, 0x08, 0xd4, 0xa5, + 0x5b, 0x3b, 0x7b, 0xa2, 0xca, 0xa3, 0xa2, 0x5d, 0x63, 0x7f, 0xc0, 0x37, + 0xc5, 0x7e, 0x99, 0x04, 0x5d, 0x9a, 0xb9, 0xa5, 0xac, 0xd1, 0xe2, 0x5d, + 0xb2, 0x2b, 0x7e, 0xbb, 0xb9, 0x66, 0x13, 0xa7, 0x30, 0xbf, 0x80, 0x0c, + 0x2b, 0x8d, 0x45, 0xe1, 0x8d, 0x96, 0x25, 0x27, 0x47, 0x3d, 0x21, 0x7d, + 0x1c, 0x42, 0xac, 0x31, 0x26, 0x47, 0x59, 0xb3, 0x44, 0x85, 0xf2, 0x8e, + 0x7d, 0x01, 0x96, 0x6d, 0xb2, 0x64, 0xc3, 0xfc, 0xa7, 0x82, 0x06, 0x4a, + 0x87, 0x75, 0x9b, 0x99, 0x47, 0x7e, 0xa6, 0x4d, 0x2c, 0x36, 0xff, 0xac, + 0x2b, 0x77, 0x96, 0x52, 0x14, 0x8d, 0x07, 0x0d, 0x28, 0x9d, 0x84, 0xa2, + 0xda, 0xd6, 0x45, 0x3a, 0xd4, 0xe6, 0xb7, 0x9a, 0xf3, 0x34, 0xe3, 0xda, + 0x39, 0xdf, 0x35, 0x9c, 0xe4, 0x87, 0x55, 0xc8, 0x43, 0xd0, 0x61, 0x46, + 0x52, 0x2f, 0x75, 0x63, 0xbb, 0x98, 0x97, 0xeb, 0xfb, 0x15, 0xaf, 0x8e, + 0x96, 0xdc, 0xff, 0x0a, 0x90, 0xda, 0x09, 0x63, 0x28, 0x7b, 0x92, 0x73, + 0x0b, 0xd4, 0x2b, 0x72, 0x2a, 0x86, 0x32, 0xc3, 0xc1, 0x3e, 0xe4, 0x2c, + 0x07, 0x89, 0x53, 0xb7, 0xfe, 0x78, 0x6c, 0x95, 0xb4, 0x62, 0x4d, 0x4b, + 0xfe, 0x6c, 0xfc, 0x5e, 0x4e, 0xa7, 0x8c, 0x07, 0x4f, 0x85, 0x27, 0xe0, + 0x7b, 0xd9, 0x7a, 0xe5, 0x1d, 0xbc, 0x36, 0xda, 0x8e, 0x21, 0xff, 0xb3, + 0x60, 0x2c, 0x5e, 0x23, 0x0f, 0xde, 0x3f, 0xae, 0xa5, 0x3a, 0x50, 0xa9, + 0x99, 0x39, 0x45, 0xaf, 0xd3, 0x5f, 0x4a, 0x15, 0xad, 0x9c, 0x66, 0x7f, + 0x92, 0xe0, 0x02, 0x81, 0x3e, 0x06, 0x6a, 0x5e, 0xd0, 0x0c, 0x42, 0xe7, + 0xcf, 0xe2, 0xeb, 0xa3, 0xe0, 0xf7, 0x2d, 0x8a, 0x21, 0xdb, 0x64, 0x28, + 0x2a, 0xb3, 0x2b, 0xc4, 0xc9, 0xd5, 0x60, 0xaf, 0xfc, 0x15, 0xa1, 0x44, + 0x9c, 0x96, 0x04, 0x42, 0x1c, 0x55, 0x8c, 0xa5, 0xce, 0x80, 0xce, 0x75, + 0x64, 0xa9, 0xf6, 0xa5, 0x5a, 0x0f, 0x8a, 0x4b, 0x8b, 0x72, 0xcf, 0x3e, + 0xd7, 0xeb, 0xe1, 0xd0, 0xd3, 0x2d, 0x04, 0x6c, 0x9e, 0x02, 0x75, 0x43, + 0x5c, 0xc1, 0x57, 0x66, 0xd9, 0x14, 0x5b, 0x08, 0x10, 0x44, 0x8d, 0x8e, + 0x89, 0xd1, 0x65, 0x27, 0x2a, 0x0b, 0x99, 0x6f, 0x09, 0xa6, 0x20, 0xa5, + 0x75, 0x24, 0xe4, 0xf7, 0xf5, 0xe0, 0xed, 0x79, 0x37, 0x18, 0x13, 0x1c, + 0xd9, 0xd1, 0xf5, 0x69, 0x0c, 0xa5, 0x02, 0xdf, 0x6a, 0xfd, 0x2e, 0x35, + 0x8e, 0xd0, 0x41, 0x91, 0x61, 0x0f, 0x5c, 0xdd, 0x70, 0xbf, 0x1c, 0x49, + 0xcb, 0xe9, 0xc9, 0x33, 0xc4, 0x99, 0x1e, 0x8b, 0x75, 0x48, 0xc2, 0x58, + 0xa4, 0x70, 0x1f, 0xbb, 0xcd, 0xd3, 0x0e, 0x79, 0x25, 0xbe, 0x53, 0xfa, + 0x32, 0x32, 0xf6, 0xb9, 0xf0, 0x0a, 0x52, 0x5b, 0xe0, 0x69, 0xff, 0x43, + 0xda, 0x98, 0x1f, 0xee, 0x54, 0x60, 0xf8, 0x24, 0x43, 0xc5, 0x37, 0x72, + 0xd1, 0xfc, 0x99, 0x9a, 0x3e, 0x24, 0xe0, 0xd9, 0xc2, 0x61, 0x47, 0xb3, + 0x26, 0x09, 0x85, 0x74, 0xa1, 0x2b, 0x4a, 0x70, 0xd0, 0x1b, 0x90, 0x03, + 0x25, 0xd9, 0x22, 0xc2, 0x16, 0x22, 0x3a, 0x62, 0x20, 0xd4, 0x13, 0xce, + 0xa2, 0xc7, 0x02, 0xfb, 0x9a, 0xbf, 0xf1, 0x1c, 0x80, 0x01, 0x97, 0x90, + 0x7f, 0x5a, 0x98, 0x70, 0x30, 0x61, 0x77, 0xe5, 0xd4, 0x3b, 0x03, 0x42, + 0x57, 0x31, 0x5e, 0xc6, 0x64, 0xe1, 0xf4, 0x64, 0x77, 0x21, 0x9b, 0x44, + 0x1c, 0xd9, 0x8c, 0x95, 0x8a, 0xf1, 0xcb, 0x82, 0xac, 0xc1, 0x26, 0x31, + 0xf2, 0x22, 0x41, 0xab, 0xbb, 0x23, 0xd3, 0x8d, 0xcc, 0x5c, 0x9d, 0x9b, + 0x1d, 0x9c, 0x4d, 0xf3, 0x62, 0xde, 0x15, 0x6a, 0x94, 0x8d, 0x24, 0xe7, + 0x52, 0x8d, 0x2a, 0xa4, 0x1d, 0x54, 0x5a, 0xda, 0xaf, 0xab, 0x05, 0x27, + 0x4b, 0xbb, 0xb4, 0xda, 0x0c, 0xb9, 0x20, 0xb3, 0xaf, 0x4a, 0xeb, 0x37, + 0xe5, 0x43, 0xe4, 0xc1, 0xf6, 0x9e, 0xf8, 0x6c, 0xd8, 0xa1, 0x0c, 0xf9, + 0xd1, 0x4b, 0x96, 0xa0, 0x6d, 0x38, 0x64, 0x41, 0xd3, 0x14, 0xfb, 0xad, + 0x89, 0xa9, 0xf7, 0x36, 0x01, 0x0f, 0xbe, 0x8e, 0xd7, 0x76, 0xc6, 0x70, + 0x22, 0x32, 0x8b, 0x08, 0xca, 0x95, 0xbf, 0xcf, 0x5e, 0xb8, 0xc0, 0x3f, + 0xd9, 0xaa, 0x84, 0xab, 0x30, 0x5b, 0xe3, 0x7a, 0x61, 0x32, 0xe5, 0x54, + 0x01, 0x5e, 0xb6, 0x1c, 0x9c, 0x78, 0x52, 0x2a, 0xa7, 0xf5, 0x29, 0xa6, + 0x0f, 0x14, 0xa5, 0x3a, 0x34, 0xd4, 0xf5, 0xc2, 0xb2, 0x8d, 0x12, 0x7b, + 0x8a, 0x64, 0x00, 0xfd, 0x02, 0x0e, 0x02, 0x26, 0x5a, 0xb9, 0xeb, 0xfd, + 0x30, 0xce, 0x51, 0xec, 0x5f, 0xbc, 0xee, 0x53, 0x21, 0xec, 0x0e, 0xee, + 0xc4, 0x28, 0x1a, 0xec, 0x2a, 0x39, 0x4e, 0xe1, 0x50, 0x11, 0x3f, 0x16, + 0xdd, 0xbf, 0xaf, 0x3e, 0xbe, 0xd4, 0xfe, 0x34, 0x1e, 0x62, 0x3f, 0x5a, + 0xea, 0x05, 0xfc, 0xd5, 0x45, 0x08, 0x47, 0xce, 0x38, 0x3f, 0x75, 0x7e, + 0x0c, 0x3a, 0x2a, 0x14, 0xa7, 0x61, 0xba, 0x3a, 0xa1, 0x41, 0xa2, 0x72, + 0x19, 0xfa, 0x33, 0x43, 0xa7, 0xf4, 0x4e, 0x5b, 0xf9, 0xb1, 0x45, 0x16, + 0x57, 0x8e, 0xb1, 0xad, 0x7d, 0x88, 0xd3, 0x93, 0xa2, 0x08, 0xf3, 0x96, + 0x4d, 0x84, 0x63, 0x08, 0xfa, 0x9d, 0xf3, 0x04, 0x33, 0xbd, 0x7e, 0x7a, + 0xc7, 0x63, 0xc5, 0x31, 0x5a, 0x82, 0x33, 0x90, 0x56, 0x44, 0xe9, 0xd3, + 0xc4, 0xd4, 0x76, 0x29, 0x2f, 0xdb, 0xa3, 0x9d, 0xff, 0xd4, 0xd2, 0xb1, + 0xce, 0xf1, 0xcb, 0x7f, 0x10, 0x3b, 0x90, 0xa4, 0x1b, 0xa0, 0x9b, 0xa7, + 0xfa, 0x27, 0x40, 0x11, 0x35, 0xc9, 0x7f, 0x01, 0x97, 0x76, 0x9f, 0x33, + 0xc5, 0xd6, 0x8d, 0x20, 0x07, 0x73, 0x93, 0x0b, 0x24, 0x88, 0x4e, 0x73, + 0x68, 0x79, 0x92, 0x20, 0x2a, 0x71, 0xed, 0x22, 0x0b, 0xfb, 0x42, 0xb5, + 0xd9, 0xc3, 0xaa, 0xed, 0x45, 0x03, 0x64, 0xde, 0x6f, 0x25, 0x8e, 0x3b, + 0x9a, 0xef, 0xc5, 0x63, 0xc2, 0x7f, 0x34, 0xd0, 0x1b, 0x20, 0xa3, 0xab, + 0x9d, 0x54, 0x41, 0x0e, 0x7b, 0x2e, 0x96, 0x12, 0x75, 0x58, 0xdf, 0xd5, + 0xaa, 0x3c, 0xf2, 0x26, 0xc1, 0xf1, 0x18, 0x37, 0x56, 0xf2, 0xd2, 0x86, + 0x6f, 0xd4, 0x9f, 0x57, 0x2b, 0x32, 0xe9, 0x08, 0x94, 0x53, 0x40, 0xc5, + 0x4d, 0x77, 0x39, 0xc6, 0x4c, 0x63, 0x53, 0xf9, 0xbf, 0x35, 0x08, 0xc5, + 0x0d, 0xd0, 0x89, 0x82, 0xa7, 0x2d, 0x6a, 0xb4, 0x22, 0xb1, 0x10, 0x7f, + 0xcf, 0x2e, 0x21, 0x27, 0x9c, 0x12, 0xc6, 0x0e, 0xca, 0xd2, 0x32, 0xb1, + 0x6d, 0xfd, 0x59, 0x12, 0x23, 0x60, 0x46, 0x89, 0xe0, 0x75, 0x5e, 0xc9, + 0xf4, 0x3d, 0x8a, 0x89, 0xd4, 0x23, 0xc2, 0xbe, 0x30, 0x32, 0x4a, 0x95, + 0x42, 0xe2, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, + 0xa7, 0x0b, 0x48, 0xe2, 0xeb, 0xd7, 0x12, 0x42, 0x4c, 0x71, 0xfb, 0x25, + 0x17, 0x23, 0x0e, 0x01, 0xa6, 0x21, 0xb9, 0x17, 0x6e, 0xf0, 0x24, 0x66, + 0x9e, 0x9d, 0x0f, 0x71, 0xf8, 0x5b, 0x79, 0xb0, 0x1b, 0x1f, 0xe7, 0xa2, + 0xc0, 0x17, 0x16, 0x08, 0x5e, 0x24, 0x7b, 0xf9, 0x7a, 0x1e, 0x70, 0xe2, + 0x05, 0x40, 0x16, 0x56, 0xe7, 0x79, 0xf2, 0x30, 0xa3, 0xdc, 0xe3, 0x7a, + 0x7e, 0x22, 0x88, 0xc0, 0xf7, 0xc8, 0x5c, 0x93, 0x95, 0x86, 0x02, 0x6c, + 0x73, 0x76, 0xef, 0x03, 0x2d, 0xcb, 0xa5, 0x22, 0xfe, 0x05, 0xbb, 0xe6, + 0xfd, 0x19, 0x8c, 0x8b, 0x67, 0x58, 0x81, 0x81, 0x2d, 0x36, 0xd0, 0xc1, + 0x20, 0xb2, 0x87, 0x87, 0xdb, 0xe4, 0xe5, 0xd1, 0xd1, 0xd5, 0x81, 0x34, + 0x4c, 0xd6, 0x09, 0xa2, 0x5d, 0xcc, 0x99, 0x12, 0xa5, 0x06, 0x0f, 0x06, + 0x7e, 0xbb, 0x67, 0x26, 0x69, 0x15, 0x6e, 0x5f, 0xb1, 0x8e, 0xd6, 0x34, + 0xfc, 0x4d, 0xd9, 0x03, 0xb7, 0x5a, 0xf4, 0xaa, 0x03, 0x00, 0x88, 0x6b, + 0x5a, 0xc9, 0xf2, 0xfb, 0x67, 0x72, 0xbc, 0xf7, 0xb9, 0xdc, 0x97, 0xdf, + 0x80, 0x91, 0xfa, 0x30, 0x18, 0x02, 0x89, 0xc7, 0xc9, 0x62, 0x1d, 0xc0, + 0x0b, 0xa6, 0xfe, 0x7e, 0xb9, 0xa9, 0x1f, 0x11, 0x71, 0xe1, 0xd1, 0xfe, + 0x8d, 0x90, 0x2c, 0x09, 0x82, 0x2e, 0x36, 0x79, 0xa5, 0x75, 0x54, 0xfb, + 0xd3, 0x3c, 0xb4, 0x18, 0x2f, 0x4e, 0x3f, 0x37, 0xc4, 0xf8, 0xc5, 0x59, + 0xa3, 0xfd, 0x0c, 0x62, 0x9e, 0xa8, 0x7a, 0x56, 0xc5, 0x97, 0x89, 0x35, + 0xc7, 0xb0, 0x29, 0x87, 0xbf, 0x6a, 0xdc, 0xb1, 0x2f, 0x01, 0xf4, 0x0d, + 0x7c, 0x25, 0x95, 0x39, 0x81, 0xdd, 0x1a, 0x81, 0x36, 0xc0, 0x6b, 0xbf, + 0x6b, 0x4d, 0xea, 0x23, 0xc0, 0x3e, 0x5c, 0x39, 0xe5, 0x6b, 0x59, 0xa0, + 0x50, 0x02, 0x99, 0xdf, 0x4e, 0xe3, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x04, 0x00, 0x00, 0x17, 0x88, 0xf8, 0xda, 0x3d, 0x57, 0x83, 0x63, + 0x76, 0xa0, 0x5c, 0x13, 0x1a, 0x00, 0x64, 0x30, 0x19, 0xfd, 0x2e, 0x9c, + 0x64, 0xb6, 0xda, 0x51, 0x7b, 0x55, 0xe8, 0xc4, 0x67, 0x1b, 0xda, 0xfc, + 0x4c, 0xd0, 0x27, 0x58, 0x56, 0xa1, 0x52, 0xd2, 0xb8, 0xd8, 0xd5, 0x94, + 0x69, 0xcf, 0xd0, 0xd5, 0x72, 0xeb, 0x2b, 0x05, 0xf3, 0x12, 0xa6, 0xac, + 0xa6, 0xf7, 0x90, 0x24, 0x1f, 0x22, 0x97, 0x5e, 0x8b, 0x7c, 0x2c, 0x30, + 0x61, 0x11, 0x9b, 0xdf, 0x83, 0x2b, 0x10, 0x09, 0x42, 0x77, 0x2b, 0xd9, + 0x43, 0xb3, 0x27, 0x69, 0x75, 0xf2, 0x2e, 0x72, 0xed, 0x50, 0xea, 0xbf, + 0x7f, 0x47, 0x39, 0x9c, 0xf8, 0x1e, 0xce, 0x6f, 0xdd, 0xe8, 0x40, 0xc5, + 0x14, 0x01, 0x7e, 0xbb, 0x0f, 0x43, 0x2d, 0x36, 0x70, 0x54, 0xc6, 0xbe, + 0x69, 0x24, 0xd1, 0x65, 0x49, 0x77, 0xf0, 0xd2, 0x99, 0xb4, 0x50, 0x8d, + 0x98, 0xcb, 0xbf, 0x7a, 0x7c, 0x65, 0xd3, 0x46, 0xcf, 0x90, 0x69, 0x56, + 0x15, 0xa2, 0xae, 0x11, 0x94, 0x60, 0xf9, 0x45, 0x17, 0x54, 0x6b, 0xbd, + 0xeb, 0xd8, 0x74, 0x41, 0x5c, 0xf6, 0x49, 0x0a, 0x14, 0xce, 0x43, 0x1f, + 0x67, 0xc3, 0x6c, 0xf4, 0x01, 0xce, 0x3f, 0x85, 0xed, 0x19, 0xa1, 0xf7, + 0x1b, 0xf8, 0x46, 0x45, 0xb4, 0xe9, 0xa7, 0x1f, 0x2a, 0x65, 0x00, 0x2a, + 0xd3, 0x8b, 0x6a, 0x3b, 0xac, 0x78, 0xab, 0xf4, 0xc8, 0x62, 0x76, 0xc8, + 0x24, 0xf8, 0xf8, 0x08, 0xe0, 0x64, 0x00, 0x64, 0x74, 0x9e, 0x55, 0x2e, + 0xf8, 0xc9, 0xc8, 0x58, 0x0e, 0x1f, 0x27, 0x32, 0xfd, 0x30, 0x24, 0x68, + 0xc8, 0xa4, 0x8c, 0x1c, 0xf3, 0xa7, 0x32, 0xae, 0x84, 0x0a, 0x8a, 0x1e, + 0x11, 0xce, 0xb2, 0x02, 0xf1, 0xb3, 0x5f, 0x7d, 0x5e, 0x54, 0x8c, 0xe0, + 0xeb, 0x46, 0x6e, 0x8a, 0x5f, 0x3f, 0x71, 0x47, 0x2a, 0x8a, 0xe6, 0xf0, + 0xb0, 0x04, 0x49, 0x64, 0xb3, 0x7e, 0x16, 0x09, 0x83, 0x5f, 0x12, 0xe0, + 0x85, 0xb7, 0x36, 0xc0, 0x8a, 0xa5, 0xcd, 0xae, 0xc0, 0xb4, 0xa2, 0x62, + 0x9b, 0xfa, 0x64, 0x18, 0x16, 0x8e, 0xb6, 0x50, 0xf2, 0x9b, 0xc4, 0x7d, + 0x0c, 0x4c, 0x8b, 0x58, 0xcf, 0x9b, 0x87, 0x09, 0xb1, 0x37, 0xbb, 0xaf, + 0xa7, 0x72, 0x79, 0x81, 0x09, 0x55, 0xa1, 0x6a, 0x87, 0xb0, 0x7d, 0xc8, + 0xb0, 0xc1, 0xa4, 0xa9, 0xdf, 0xcf, 0x95, 0x77, 0x36, 0x8e, 0x2b, 0xae, + 0xeb, 0x4b, 0xf9, 0x2a, 0x83, 0x6c, 0x53, 0x3c, 0x89, 0xa6, 0x08, 0xae, + 0x00, 0x4e, 0xb8, 0xf6, 0x34, 0x7c, 0xc6, 0x76, 0x87, 0x1a, 0x02, 0xb0, + 0x89, 0xa3, 0x0f, 0x00, 0xc6, 0x7b, 0xeb, 0xf7, 0x95, 0x40, 0xc5, 0x0d, + 0x6f, 0x74, 0xd8, 0x21, 0x2f, 0x9f, 0x24, 0xac, 0x43, 0xdb, 0x3a, 0x39, + 0x6c, 0x34, 0x59, 0x62, 0x66, 0xbc, 0x28, 0x7f, 0x8c, 0x64, 0x62, 0x8c, + 0x28, 0x6c, 0xf5, 0x79, 0x24, 0xb1, 0x00, 0x9c, 0x58, 0x6b, 0x09, 0xef, + 0xb0, 0x73, 0xcd, 0x47, 0xbb, 0x52, 0xfd, 0x26, 0x6a, 0xff, 0xb9, 0xf1, + 0xd5, 0x82, 0x59, 0x01, 0xfa, 0x87, 0x14, 0x24, 0x10, 0xb0, 0xf7, 0xdf, + 0xf9, 0x3f, 0x67, 0x19, 0xbd, 0xc7, 0x85, 0xb0, 0xad, 0x47, 0xa8, 0x4c, + 0x3e, 0xb6, 0x2e, 0x8a, 0xb3, 0xcc, 0x35, 0xa0, 0x48, 0xc7, 0x90, 0x81, + 0xb7, 0x53, 0x1c, 0x38, 0x63, 0xf2, 0x2f, 0xa0, 0x71, 0x82, 0xe2, 0x56, + 0xdb, 0x68, 0xe8, 0x5f, 0xf8, 0x42, 0xf2, 0xf6, 0xb8, 0x10, 0x6b, 0x54, + 0x21, 0xa0, 0xc1, 0xfe, 0xcb, 0xce, 0x12, 0xa2, 0x49, 0x51, 0x86, 0x53, + 0x56, 0xec, 0x33, 0xb3, 0x72, 0xce, 0xa4, 0x46, 0xe3, 0x37, 0xcb, 0xc0, + 0x95, 0xaa, 0xe2, 0xa3, 0xc5, 0xe9, 0x36, 0x40, 0xfe, 0xf7, 0xe2, 0x5a, + 0x6d, 0x58, 0x39, 0xb2, 0x41, 0x5d, 0xe2, 0x71, 0x72, 0xd0, 0xf0, 0x5c, + 0x16, 0x88, 0x95, 0x30, 0x0a, 0xfb, 0x8d, 0xda, 0x14, 0x80, 0xf4, 0x15, + 0xf2, 0xf6, 0xac, 0xf3, 0xd8, 0x8d, 0x13, 0x24, 0x2c, 0x74, 0x60, 0x6e, + 0x8c, 0xa1, 0x59, 0xcf, 0x74, 0x7c, 0x2d, 0x0b, 0xbb, 0x06, 0x5c, 0x9d, + 0xcd, 0xf3, 0x1e, 0x4a, 0xba, 0x3f, 0x9c, 0x4a, 0xc4, 0xd7, 0xf9, 0xf0, + 0xa5, 0x56, 0x7f, 0xb0, 0xa2, 0x57, 0xd0, 0xc3, 0xaa, 0xa7, 0xd0, 0x49, + 0xe2, 0x28, 0x9b, 0xc4, 0x64, 0x0c, 0xe0, 0x71, 0x9c, 0x05, 0x04, 0x95, + 0x00, 0x1f, 0x7b, 0xa9, 0xb9, 0xb3, 0x2b, 0x8f, 0x0b, 0x45, 0x1e, 0x23, + 0xaa, 0x27, 0x89, 0x4a, 0xb0, 0x7d, 0x03, 0xdf, 0xae, 0xdb, 0xcb, 0xc4, + 0xec, 0x3b, 0x02, 0xe2, 0x85, 0x3a, 0xb7, 0x25, 0xfb, 0xab, 0xca, 0xc1, + 0x33, 0x00, 0x5b, 0xd2, 0xcf, 0xb0, 0x11, 0x1d, 0x51, 0xb5, 0x5b, 0xea, + 0x94, 0xf7, 0xa0, 0x98, 0x33, 0xba, 0x58, 0xfc, 0x12, 0xea, 0xdd, 0x89, + 0xbd, 0x63, 0x03, 0xbe, 0x7e, 0x3b, 0x69, 0xc4, 0x9d, 0x57, 0x0f, 0xd6, + 0xbe, 0xea, 0x5b, 0xd0, 0x97, 0x63, 0x89, 0xb0, 0xa0, 0xc0, 0xd6, 0x39, + 0xc1, 0x69, 0x12, 0x6a, 0xfb, 0xac, 0x74, 0x7f, 0xfb, 0xf4, 0x7f, 0x38, + 0x44, 0x4c, 0x8a, 0xa2, 0x41, 0x15, 0xc0, 0x54, 0xc0, 0xed, 0x14, 0x83, + 0xef, 0xbc, 0x9c, 0xc7, 0xdd, 0x21, 0xd6, 0xf0, 0x9b, 0x7f, 0x09, 0xd5, + 0x96, 0xe5, 0xf7, 0xc5, 0xa9, 0xb3, 0x41, 0xb0, 0x9d, 0xeb, 0x49, 0x68, + 0x9d, 0x2b, 0xea, 0x47, 0x80, 0x3b, 0x54, 0xb8, 0xf4, 0x14, 0x5e, 0xd6, + 0x66, 0x89, 0x04, 0xb3, 0x00, 0xa3, 0xa8, 0x32, 0x62, 0x2e, 0xc3, 0x15, + 0xc6, 0x93, 0x7d, 0x40, 0x32, 0xb1, 0x6b, 0x60, 0xd3, 0x52, 0xdf, 0x09, + 0x8c, 0x80, 0x2b, 0x01, 0xe7, 0x97, 0x8d, 0xbb, 0x14, 0xd6, 0x10, 0x15, + 0x64, 0x00, 0x4a, 0x2c, 0x67, 0xca, 0xd0, 0xa1, 0x37, 0x33, 0x7b, 0xa1, + 0x2a, 0x5b, 0x5b, 0x78, 0xf8, 0x2f, 0xdd, 0x76, 0xab, 0x8a, 0xc3, 0xe3, + 0x37, 0x00, 0xd1, 0x29, 0xb0, 0x96, 0x1d, 0x18, 0xbe, 0x5d, 0x32, 0x7e, + 0xb7, 0x11, 0xa9, 0x78, 0x72, 0xa2, 0x2d, 0x29, 0x1c, 0x32, 0xa4, 0xff, + 0xc7, 0xce, 0xfe, 0xaf, 0xb7, 0x17, 0x43, 0xe5, 0x2f, 0xae, 0x45, 0xd3, + 0xaf, 0x10, 0xe3, 0xd0, 0x58, 0xb6, 0xee, 0xee, 0x7a, 0xb5, 0x06, 0x70, + 0x26, 0x7e, 0x2d, 0x5b, 0xd5, 0xe1, 0x7b, 0x9a, 0x37, 0x02, 0xfc, 0x1d, + 0x08, 0x4f, 0x1a, 0xf5, 0x44, 0x63, 0xde, 0x4b, 0x14, 0x68, 0x54, 0x0b, + 0x6a, 0x22, 0x4e, 0x02, 0x65, 0xcd, 0xf4, 0x04, 0xec, 0xcc, 0x8a, 0x0b, + 0xe0, 0x59, 0xf8, 0x65, 0x25, 0x63, 0xed, 0x0f, 0xa6, 0xc5, 0x3c, 0xcb, + 0x5d, 0xc5, 0xd8, 0x9f, 0x5a, 0xd3, 0x88, 0x3d, 0xd4, 0x2c, 0xb3, 0x04, + 0xf6, 0x97, 0xc7, 0xe2, 0xfd, 0xb6, 0xf4, 0x7d, 0x0d, 0xb9, 0x75, 0x7e, + 0x9d, 0x81, 0xdc, 0xdf, 0x8e, 0x90, 0x40, 0x0c, 0x7b, 0x45, 0xfe, 0x68, + 0xfd, 0xff, 0x1c, 0xf1, 0x16, 0x09, 0x33, 0x74, 0x27, 0x7b, 0x4d, 0xd9, + 0x9b, 0x48, 0x6d, 0x84, 0xeb, 0x96, 0x8f, 0x4b, 0x82, 0x73, 0xd5, 0x69, + 0x7d, 0x14, 0x45, 0x8c, 0xb8, 0x71, 0x87, 0x70, 0x09, 0x26, 0xfc, 0x89, + 0x6f, 0x0f, 0xb6, 0xc1, 0xd6, 0xe1, 0xbf, 0xdb, 0x85, 0x8f, 0x94, 0xad, + 0x94, 0x01, 0x01, 0xbb, 0x3f, 0xc0, 0xb5, 0xff, 0xf5, 0xbb, 0x4f, 0x50, + 0x09, 0xca, 0x7d, 0x36, 0x47, 0x66, 0x9a, 0x8c, 0xee, 0x84, 0x73, 0x9a, + 0x1f, 0x49, 0x75, 0xb4, 0xab, 0x66, 0xf7, 0x3b, 0xfe, 0x81, 0x67, 0xc9, + 0xd1, 0x16, 0xde, 0x1f, 0xc2, 0x24, 0xed, 0x6a, 0x5a, 0xe7, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0xc5, 0xd7, 0x14, 0x84, + 0xf8, 0xcf, 0x9b, 0xf4, 0xb7, 0x6f, 0x47, 0x90, 0x47, 0x30, 0x80, 0x4b, + 0x9e, 0x32, 0x25, 0xa9, 0xf1, 0x33, 0xb5, 0xde, 0xa1, 0x68, 0xf4, 0xe2, + 0x85, 0x1f, 0x07, 0x2f, 0xcc, 0x00, 0xfc, 0xaa, 0x7c, 0xa6, 0x20, 0x61, + 0x71, 0x7a, 0x48, 0xe5, 0x2e, 0x29, 0xa3, 0xfa, 0x37, 0x9a, 0x95, 0x3f, + 0xaa, 0x68, 0x93, 0xe3, 0x2e, 0xc5, 0xa2, 0x7b, 0x94, 0x5e, 0x60, 0x5f, + 0x10, 0x85, 0xf3, 0x23, 0x2d, 0x42, 0x4c, 0x13, 0x29, 0xc8, 0x8d, 0x78, + 0x6e, 0xd6, 0x8c, 0xe6, 0xfc, 0xb6, 0x2a, 0xa6, 0x3b, 0xf9, 0xab, 0x61, + 0x7c, 0x08, 0x8a, 0x3b, 0x70, 0xbe, 0x57, 0xaa, 0xda, 0x1f, 0x33, 0x4a, + 0x70, 0x17, 0x25, 0x0d, 0x3f, 0x60, 0x3d, 0xc8, 0x2e, 0xbd, 0x3b, 0x12, + 0x0b, 0x63, 0x5e, 0x3f, 0xf5, 0x6b, 0x1f, 0x0b, 0xd9, 0x33, 0x85, 0x23, + 0x71, 0x24, 0x9a, 0xb3, 0xdf, 0x5c, 0x1f, 0xef, 0x14, 0x33, 0xc8, 0x66, + 0x85, 0xb7, 0xf0, 0x56, 0x68, 0x1d, 0x51, 0x52, 0xaf, 0x80, 0x3c, 0xe2, + 0x59, 0x06, 0xf1, 0xd1, 0x9f, 0xb6, 0xc6, 0x80, 0x4e, 0x06, 0xea, 0x28, + 0xab, 0x17, 0x8f, 0x45, 0x7a, 0xf6, 0xb4, 0x93, 0xb7, 0x43, 0x9e, 0xc6, + 0xd4, 0x29, 0x00, 0x62, 0xab, 0x51, 0x7a, 0x72, 0xe5, 0xc1, 0xd4, 0x10, + 0xcd, 0xd6, 0x17, 0x54, 0xe4, 0x20, 0x84, 0x50, 0xe4, 0xf9, 0x00, 0x13, + 0xfd, 0xa6, 0x9f, 0xef, 0x19, 0xd4, 0x60, 0x2a, 0x42, 0x07, 0xcd, 0xd5, + 0xa1, 0x01, 0x6d, 0x07, 0x01, 0x32, 0x61, 0x3c, 0x65, 0x9a, 0x8f, 0x5d, + 0x33, 0xf3, 0xcb, 0x29, 0x0b, 0x8c, 0xe7, 0x3b, 0x83, 0x44, 0xb1, 0x3a, + 0x4f, 0x8e, 0x09, 0x15, 0x14, 0x69, 0x84, 0xa1, 0xbb, 0x15, 0xfd, 0xea, + 0xde, 0xbe, 0x5b, 0x6a, 0xc0, 0x95, 0x04, 0x46, 0x4d, 0x8a, 0xaa, 0xac, + 0xbc, 0x2f, 0xad, 0x12, 0x15, 0x8a, 0x53, 0x4c, 0x94, 0xb8, 0xca, 0x42, + 0x96, 0x3a, 0xf4, 0x7a, 0x18, 0x9d, 0x5b, 0x24, 0x9a, 0xce, 0xa8, 0x99, + 0xd4, 0x37, 0x32, 0xf6, 0xf2, 0xac, 0xaf, 0x3f, 0xf5, 0x3b, 0xfe, 0xda, + 0x13, 0x9a, 0xab, 0x4f, 0x55, 0xc0, 0x2c, 0x21, 0x2b, 0x65, 0x71, 0x1f, + 0xc5, 0x04, 0x32, 0xc9, 0x94, 0xe5, 0xfa, 0x6f, 0xd8, 0x2a, 0xbc, 0x70, + 0x85, 0x55, 0xdc, 0x62, 0xb7, 0x3a, 0x20, 0x0e, 0xe7, 0x67, 0x3c, 0xfe, + 0xcb, 0x83, 0x6a, 0x15, 0x6e, 0x4a, 0x35, 0x65, 0xea, 0xc1, 0xb9, 0x4d, + 0x35, 0xf9, 0x4b, 0xcf, 0xd8, 0xfd, 0xa5, 0xff, 0xff, 0x67, 0x70, 0x04, + 0xae, 0xa2, 0xa4, 0x12, 0x4b, 0x83, 0x4f, 0xc2, 0x96, 0xf0, 0x21, 0x2b, + 0x14, 0x21, 0x73, 0x42, 0x14, 0x99, 0x07, 0xe5, 0xa9, 0x52, 0x4c, 0xeb, + 0xbe, 0xc3, 0x11, 0x2e, 0x27, 0xda, 0x69, 0x94, 0xd5, 0xf6, 0xc6, 0x77, + 0x0a, 0x00, 0x5d, 0x9a, 0x82, 0xaa, 0x21, 0xfc, 0x86, 0x9b, 0xd0, 0xc4, + 0xc4, 0x1f, 0x53, 0x41, 0x7a, 0x92, 0xab, 0x1c, 0x12, 0xf6, 0xd5, 0x48, + 0xfb, 0x29, 0x4d, 0xb4, 0xd2, 0x12, 0xee, 0xc5, 0xea, 0x18, 0x33, 0xf1, + 0x4d, 0x0a, 0x10, 0x43, 0xa5, 0x35, 0xb1, 0x63, 0xc4, 0xfb, 0x38, 0x1e, + 0xef, 0xac, 0x3f, 0x97, 0x41, 0xc6, 0x96, 0x3e, 0x60, 0x13, 0xc8, 0xe3, + 0xbe, 0x61, 0xe9, 0xb6, 0x26, 0x16, 0x14, 0xf8, 0x82, 0x0d, 0x6e, 0x75, + 0x2f, 0xd7, 0x9c, 0x3a, 0x4a, 0xda, 0xd8, 0x2b, 0x35, 0xd4, 0x20, 0x32, + 0xd4, 0x4f, 0x0f, 0xe4, 0xdc, 0xd5, 0x0f, 0xfe, 0xa6, 0x81, 0x28, 0xb4, + 0x24, 0x3e, 0xb7, 0x0f, 0xb0, 0xb2, 0x5b, 0x05, 0x76, 0xbb, 0x24, 0x49, + 0x6a, 0x01, 0x68, 0x3f, 0x03, 0x96, 0xbc, 0x0c, 0x77, 0x48, 0x5f, 0xe8, + 0x39, 0xf4, 0xb0, 0x84, 0x42, 0x0e, 0x6a, 0xb9, 0xab, 0xf2, 0x95, 0x97, + 0xa7, 0x5e, 0x29, 0x34, 0x9d, 0x50, 0xc0, 0x4b, 0x40, 0x72, 0xa1, 0x7c, + 0x79, 0x5e, 0x95, 0xbe, 0xd6, 0x17, 0x43, 0x0a, 0xc9, 0x27, 0x25, 0x43, + 0xd7, 0x99, 0xd5, 0x48, 0xd8, 0x98, 0xb5, 0x2b, 0x7f, 0xe3, 0xbd, 0x1d, + 0xc0, 0xd1, 0x04, 0xd5, 0xa4, 0xe1, 0x68, 0xbe, 0x96, 0xf1, 0x2e, 0x5e, + 0x37, 0x8d, 0x39, 0x4e, 0xe4, 0xcc, 0x5e, 0xd7, 0xdd, 0x59, 0x7e, 0xe8, + 0xae, 0x48, 0xb5, 0xec, 0x2c, 0xf7, 0x68, 0x96, 0x00, 0xe5, 0xec, 0x03, + 0x6f, 0x98, 0x3a, 0x9a, 0x4f, 0xd9, 0xf1, 0x2f, 0xfe, 0x76, 0xcf, 0x8f, + 0x0b, 0x3d, 0x8a, 0x14, 0x00, 0x83, 0xcb, 0xca, 0xe3, 0x34, 0x81, 0xb5, + 0x91, 0x64, 0x2b, 0x12, 0x24, 0x86, 0x9c, 0xae, 0x3c, 0x7f, 0x53, 0x22, + 0xd4, 0x94, 0x90, 0x44, 0x6b, 0x35, 0xd2, 0xce, 0x8e, 0x95, 0xe2, 0xbe, + 0x46, 0x50, 0x3f, 0x3d, 0xc3, 0xcd, 0xef, 0x47, 0x99, 0xb5, 0xf2, 0xd4, + 0x6f, 0xf4, 0xfa, 0xa2, 0xfc, 0x1e, 0xe3, 0x99, 0x49, 0xfd, 0x1a, 0x6e, + 0x0d, 0xb5, 0xf1, 0xc8, 0x05, 0x22, 0x29, 0xca, 0x03, 0xb8, 0x15, 0x3b, + 0x01, 0x8a, 0x95, 0x74, 0x48, 0x93, 0x61, 0x35, 0xde, 0xeb, 0xa9, 0xc4, + 0x56, 0xa9, 0xd7, 0xde, 0x4b, 0xe5, 0x4b, 0xa1, 0x42, 0x6a, 0x5f, 0xe3, + 0xb2, 0xc7, 0xda, 0xfb, 0xc7, 0x70, 0x64, 0xe0, 0x68, 0x19, 0xc6, 0x11, + 0x77, 0x2b, 0x5f, 0xba, 0x1d, 0x58, 0x77, 0x98, 0x2c, 0x91, 0xb4, 0xd2, + 0xea, 0x1b, 0xdc, 0xe8, 0xfa, 0x82, 0xf3, 0x6e, 0xac, 0x88, 0x15, 0x16, + 0x1a, 0x53, 0xb3, 0x01, 0x94, 0x03, 0x47, 0x20, 0xdb, 0x71, 0xcb, 0x71, + 0xe8, 0x62, 0xad, 0x34, 0x2b, 0xa3, 0xa5, 0xe9, 0xa6, 0x82, 0x0e, 0x16, + 0x61, 0xbc, 0x29, 0x6b, 0xb1, 0x60, 0x67, 0x80, 0x9a, 0x9f, 0xc4, 0x82, + 0xf6, 0xb0, 0x7a, 0x16, 0x9c, 0x25, 0x04, 0xeb, 0xfd, 0xe0, 0x18, 0xd3, + 0xfc, 0xeb, 0xe1, 0x3c, 0x2b, 0x29, 0x7b, 0x32, 0x4e, 0xd3, 0x6d, 0xe1, + 0x27, 0xda, 0xc9, 0x14, 0x5c, 0x7f, 0xfa, 0x70, 0x41, 0x8e, 0xb4, 0xa3, + 0xde, 0x36, 0x92, 0x67, 0x97, 0xe2, 0xec, 0x85, 0x8b, 0x76, 0x08, 0x3c, + 0x32, 0x58, 0xd4, 0x7f, 0x6f, 0x91, 0x03, 0xdb, 0x19, 0x3e, 0xc4, 0x8b, + 0x3c, 0xb7, 0x75, 0x90, 0x71, 0x7a, 0x21, 0x9d, 0xa7, 0x77, 0xbf, 0xf5, + 0x92, 0x57, 0x46, 0x07, 0xa7, 0xbb, 0x0c, 0x42, 0xca, 0x4f, 0x5a, 0x27, + 0x45, 0x69, 0xfe, 0x6d, 0x78, 0x43, 0x77, 0xc4, 0xb4, 0x43, 0xff, 0x37, + 0x0d, 0xb7, 0xfa, 0xe9, 0x9e, 0x06, 0x70, 0x53, 0xfd, 0xf6, 0xa0, 0x28, + 0x84, 0x46, 0xcd, 0x61, 0xa2, 0x95, 0xc4, 0x1e, 0x6a, 0x13, 0xa1, 0x7f, + 0xaf, 0xe1, 0x73, 0x85, 0xb0, 0x53, 0x9c, 0x08, 0xb6, 0x1d, 0x4d, 0xb4, + 0x0b, 0xfb, 0x1f, 0x0c, 0x7b, 0x17, 0x06, 0x73, 0xa7, 0x22, 0x1f, 0xb0, + 0xd8, 0x45, 0x6e, 0xe5, 0xde, 0x48, 0xb7, 0x9f, 0x5a, 0xa8, 0xd1, 0xc3, + 0x04, 0xd1, 0x87, 0xec, 0x15, 0x3e, 0xd1, 0xc7, 0x57, 0x01, 0x46, 0x4b, + 0x28, 0xa8, 0x79, 0x5a, 0x7e, 0x0b, 0x56, 0x56, 0x28, 0xda, 0x35, 0xea, + 0x4c, 0x14, 0x81, 0xae, 0xc0, 0x0d, 0x12, 0xfe, 0x2d, 0xb7, 0x95, 0x4d, + 0xea, 0x78, 0xb6, 0x53, 0xcf, 0xac, 0x8a, 0xfc, 0xc9, 0x07, 0x9f, 0x93, + 0xf0, 0x11, 0x86, 0x13, 0xe9, 0xca, 0x3d, 0xce, 0xb1, 0xfd, 0x1a, 0x0a, + 0x8b, 0x11, 0x82, 0x94, 0x6a, 0xae, 0xc5, 0x80, 0x6a, 0x3b, 0xa8, 0x7c, + 0xb4, 0x53, 0x4e, 0xa9, 0x04, 0x1a, 0x4f, 0xb0, 0xb9, 0x95, 0x96, 0xa5, + 0xfd, 0xce, 0xdc, 0x57, 0x00, 0x48, 0x16, 0xe2, 0x40, 0xae, 0x04, 0xf5, + 0x83, 0x60, 0x23, 0xd9, 0x8e, 0x59, 0x56, 0x20, 0x50, 0x38, 0xc4, 0xde, + 0x88, 0x9f, 0x91, 0x06, 0xdb, 0x8f, 0x84, 0xa2, 0xaf, 0x61, 0xdd, 0x48, + 0x03, 0x4f, 0xc4, 0xb8, 0xed, 0x12, 0xd2, 0x74, 0x08, 0xb9, 0x51, 0x63, + 0xb5, 0xfe, 0x09, 0x7f, 0x7b, 0x8c, 0x5e, 0xd7, 0x27, 0xe5, 0x79, 0xe6, + 0x33, 0x60, 0x54, 0xe1, 0x21, 0xda, 0xca, 0x8b, 0x81, 0xdf, 0xb6, 0xa7, + 0x2e, 0x9d, 0x0f, 0xfc, 0x05, 0x80, 0x67, 0xcb, 0xc5, 0xdf, 0xc7, 0x13, + 0xee, 0xb5, 0x40, 0x8e, 0xa7, 0x0c, 0xcb, 0xf2, 0x45, 0x15, 0x29, 0xb1, + 0xb8, 0x02, 0x23, 0x61, 0x38, 0xf1, 0x16, 0xa1, 0x0c, 0xa1, 0xc9, 0x40, + 0x8c, 0xd0, 0x48, 0x4b, 0xce, 0x9c, 0x1e, 0x53, 0x40, 0x44, 0xf6, 0x17, + 0x16, 0xc6, 0x5c, 0xb0, 0x2a, 0x29, 0x59, 0x87, 0x67, 0x85, 0xa7, 0x81, + 0x84, 0xe9, 0x4f, 0xe5, 0x4e, 0x13, 0x5a, 0x11, 0xa1, 0x24, 0x62, 0xe9, + 0x7a, 0xea, 0x51, 0xaa, 0x45, 0xf3, 0x1d, 0x2a, 0xaf, 0x01, 0x28, 0x35, + 0xda, 0xb4, 0xe7, 0xab, 0xc1, 0xb9, 0x3c, 0x45, 0xa2, 0x0b, 0x5d, 0x40, + 0x09, 0xac, 0x62, 0x16, 0xd3, 0x1f, 0x9f, 0xc7, 0x1a, 0x56, 0xb7, 0x27, + 0xd1, 0x1b, 0xe1, 0xb5, 0x82, 0x9e, 0xe8, 0xd3, 0x5c, 0x0f, 0xe8, 0x87, + 0x61, 0xc6, 0x20, 0xb7, 0x31, 0x3f, 0x0d, 0xb3, 0x0a, 0x5a, 0xce, 0x06, + 0xa5, 0xe9, 0xfd, 0xf3, 0x29, 0x1a, 0xcd, 0x86, 0x0e, 0x31, 0x29, 0xaa, + 0xb7, 0x32, 0xf1, 0x10, 0x4e, 0x92, 0x12, 0x00, 0xc0, 0xac, 0x50, 0x4b, + 0x52, 0x59, 0x51, 0x7c, 0xa8, 0x0c, 0xf7, 0xcb, 0x16, 0x73, 0x7b, 0x90, + 0xa8, 0x57, 0x79, 0xb4, 0x73, 0x53, 0xd7, 0xed, 0xba, 0x46, 0xc5, 0x06, + 0x53, 0x02, 0xc7, 0x58, 0x4c, 0x09, 0x0c, 0xa5, 0x01, 0x13, 0x18, 0x39, + 0x4b, 0x4e, 0xc2, 0x0d, 0xd6, 0xdf, 0xaa, 0x7e, 0x46, 0xba, 0x6e, 0xcc, + 0x25, 0x42, 0xd0, 0xb3, 0x31, 0xdc, 0xdf, 0x7d, 0xf1, 0xc3, 0x73, 0xca, + 0x7a, 0xf6, 0xcb, 0x23, 0x81, 0x8d, 0xbe, 0x0b, 0xf2, 0x79, 0x8d, 0x14, + 0xa4, 0xc8, 0x36, 0x18, 0x49, 0xc8, 0x0d, 0xd7, 0xc9, 0xdd, 0x35, 0xeb, + 0xec, 0x52, 0x56, 0xae, 0xf2, 0xd2, 0x51, 0x91, 0x39, 0xbc, 0xb0, 0x49, + 0xb7, 0xf2, 0x1b, 0x64, 0x83, 0x5a, 0xa6, 0x97, 0xc2, 0x15, 0x95, 0xdc, + 0x11, 0xd2, 0x89, 0xc0, 0x6a, 0xb1, 0x44, 0x43, 0x38, 0xb6, 0x54, 0x0f, + 0xdc, 0xcb, 0xed, 0x26, 0x27, 0xd9, 0x46, 0x56, 0x4e, 0x6a, 0x54, 0x74, + 0x0f, 0x45, 0xfc, 0xb6, 0x93, 0xab, 0x3c, 0xd1, 0x86, 0x51, 0xaf, 0xa9, + 0x4a, 0xc0, 0x9c, 0x78, 0xc1, 0xb1, 0xc7, 0xf1, 0x9c, 0xd1, 0xd0, 0x32, + 0x4e, 0x4b, 0x02, 0x36, 0x68, 0x38, 0x88, 0x56, 0xc0, 0x2b, 0x12, 0x05, + 0x3b, 0xb9, 0xf6, 0xa2, 0x37, 0xe7, 0xbc, 0x81, 0xf9, 0x75, 0x51, 0x27, + 0x56, 0x0d, 0x55, 0xd1, 0x6a, 0xe0, 0xcf, 0x87, 0x0a, 0x44, 0xc6, 0x57, + 0xe1, 0x1b, 0xc0, 0x2c, 0xcf, 0xab, 0x77, 0xe9, 0x14, 0xf5, 0x34, 0x89, + 0xfb, 0xc9, 0xf2, 0x87, 0x5c, 0x75, 0xba, 0x51, 0x9a, 0x49, 0xe9, 0x23, + 0x23, 0xf4, 0xc9, 0xd1, 0x2f, 0x87, 0xf6, 0x75, 0x38, 0x97, 0x48, 0xb8, + 0x30, 0x46, 0x1d, 0x46, 0x65, 0x03, 0x10, 0xcf, 0xfb, 0x36, 0xf2, 0xb1, + 0xaf, 0x31, 0x02, 0x7b, 0x74, 0xfe, 0x9f, 0x8c, 0x73, 0x04, 0xfd, 0xb5, + 0xae, 0x2e, 0x27, 0x9c, 0xd8, 0x73, 0xbc, 0xc3, 0x4a, 0x76, 0x93, 0x66, + 0xf6, 0xb7, 0x90, 0xc4, 0x42, 0x3d, 0xcd, 0xb5, 0xf1, 0x75, 0xbf, 0xb7, + 0xdd, 0x8e, 0xb7, 0xcd, 0x90, 0x35, 0xf5, 0x95, 0x3d, 0xe4, 0x4e, 0xb0, + 0x7c, 0x5f, 0xad, 0xff, 0x75, 0x38, 0xc4, 0xc7, 0xed, 0xec, 0x70, 0xcc, + 0x9f, 0xf9, 0x77, 0xa1, 0x00, 0x2f, 0xf1, 0xa2, 0xc9, 0x74, 0xdc, 0x18, + 0x14, 0xd0, 0x2f, 0x86, 0x66, 0xa7, 0x5b, 0x39, 0x5c, 0xba, 0x0e, 0x77, + 0x16, 0x04, 0xc3, 0x02, 0x42, 0x3b, 0x66, 0x29, 0xee, 0x65, 0x00, 0xd4, + 0x22, 0x5a, 0x77, 0x74, 0xd4, 0xc3, 0xf3, 0x00, 0xdf, 0x6b, 0xc3, 0x15, + 0x89, 0x0e, 0xb1, 0xbc, 0xac, 0xe8, 0x44, 0x2f, 0x80, 0x34, 0x34, 0x8b, + 0x0c, 0x48, 0x45, 0xc2, 0x6a, 0xa3, 0x67, 0xd7, 0x3d, 0x36, 0xf3, 0x3f, + 0xe5, 0xf0, 0x5b, 0xe8, 0xad, 0x41, 0xd5, 0x82, 0xc1, 0x28, 0xab, 0x77, + 0xe8, 0x7f, 0xb3, 0xf6, 0xd2, 0x0c, 0xe4, 0x03, 0xcf, 0xe4, 0x72, 0xdb, + 0x7b, 0x81, 0xf4, 0xf3, 0x48, 0x74, 0xe1, 0x91, 0xb8, 0xf8, 0x4c, 0x2c, + 0x60, 0x99, 0x3e, 0x1e, 0x4f, 0xaf, 0x12, 0xab, 0x52, 0xef, 0xc7, 0x60, + 0xd2, 0xfe, 0x62, 0x55, 0xc8, 0x18, 0xad, 0x60, 0xa7, 0x5d, 0xde, 0x4d, + 0xfc, 0x6d, 0xe1, 0x10, 0x7c, 0xf9, 0xa2, 0x64, 0x00, 0x16, 0x1f, 0x44, + 0x7c, 0xe2, 0x72, 0x37, 0xd9, 0x92, 0xad, 0xfc, 0x62, 0x53, 0xbe, 0xb6, + 0xe0, 0xc8, 0xe0, 0xa2, 0xef, 0x22, 0x4b, 0x70, 0x3a, 0x4f, 0xc9, 0xed, + 0x6b, 0xbc, 0x17, 0x0a, 0xcf, 0x6a, 0x2c, 0xd3, 0xd2, 0x6b, 0x02, 0x45, + 0xfa, 0x9e, 0xc2, 0x21, 0x28, 0xfc, 0x07, 0x68, 0xd6, 0xb8, 0x9f, 0x2a, + 0x0b, 0x7a, 0x0e, 0xbc, 0x4e, 0xee, 0x84, 0x38, 0xe4, 0x8e, 0x70, 0xc3, + 0xc4, 0xad, 0x74, 0x87, 0x2d, 0x16, 0x4f, 0xa1, 0xf8, 0x20, 0xf5, 0xde, + 0xa3, 0xc5, 0x0c, 0x3b, 0xde, 0x44, 0x48, 0x0f, 0x3c, 0xdc, 0x7e, 0x10, + 0x8b, 0x87, 0xc4, 0x3b, 0xb0, 0x95, 0xbf, 0x61, 0x1e, 0xad, 0x07, 0x52, + 0xfd, 0x0b, 0x84, 0xa9, 0x46, 0xb0, 0x32, 0xd5, 0x22, 0x80, 0x35, 0x26, + 0x41, 0xf8, 0x11, 0x72, 0xb1, 0x31, 0x6f, 0x5a, 0x75, 0xcc, 0x67, 0xe0, + 0xb2, 0x50, 0x89, 0xb2, 0x66, 0x6e, 0xee, 0xa0, 0x41, 0x8d, 0x00, 0x2a, + 0xa7, 0x9d, 0xa5, 0x11, 0x2b, 0x07, 0x95, 0x3a, 0x55, 0x8c, 0x67, 0xb1, + 0xe5, 0x2d, 0xd4, 0xd1, 0x3e, 0x29, 0xed, 0xa5, 0x59, 0x97, 0x7b, 0xdf, + 0x92, 0x10, 0x0b, 0x04, 0x89, 0x27, 0xa0, 0xa2, 0x93, 0x18, 0x7f, 0x47, + 0x84, 0x1c, 0xc6, 0xd6, 0x8f, 0x73, 0x81, 0xa0, 0xfa, 0xe5, 0x3e, 0xd8, + 0xbf, 0x56, 0x1a, 0x76, 0xf4, 0xc4, 0x0f, 0x7a, 0x29, 0x9d, 0x32, 0x5d, + 0x41, 0xe0, 0x07, 0xb9, 0xd3, 0x3f, 0x7e, 0xff, 0x90, 0x89, 0xce, 0xdc, + 0xf1, 0x1d, 0x54, 0xb6, 0x67, 0x7f, 0x4d, 0x71, 0x9a, 0x4a, 0x5f, 0x80, + 0x0d, 0x5c, 0x77, 0xd5, 0x50, 0x7c, 0x41, 0x56, 0x7e, 0x99, 0x0a, 0xeb, + 0x66, 0x1f, 0xd2, 0x55, 0xc3, 0xc6, 0x6c, 0xc5, 0xfc, 0x34, 0x40, 0x2c, + 0x05, 0x29, 0x05, 0x7c, 0xca, 0xe6, 0x8d, 0xd3, 0xb0, 0xca, 0x84, 0x27, + 0x50, 0x7c, 0x6b, 0x17, 0x1b, 0x22, 0xe4, 0x7f, 0xe6, 0x44, 0x94, 0x06, + 0x4b, 0xb3, 0xb7, 0xbb, 0x98, 0x81, 0x44, 0x0b, 0xf5, 0x66, 0xcb, 0xad, + 0xf2, 0x9a, 0xe1, 0x47, 0xf3, 0x97, 0xa9, 0xb2, 0xc2, 0xca, 0xcd, 0x98, + 0x78, 0x60, 0xdc, 0x6e, 0x87, 0x55, 0x47, 0xf3, 0xae, 0x84, 0xdd, 0x9a, + 0xe9, 0x1a, 0x63, 0x83, 0xea, 0x23, 0x09, 0x67, 0x34, 0x83, 0x00, 0x6e, + 0x5e, 0x58, 0xb8, 0x89, 0x04, 0x08, 0x0a, 0x55, 0x9e, 0x78, 0xc9, 0xff, + 0xb9, 0xb5, 0x2c, 0xdd, 0x3b, 0x0c, 0x58, 0x07, 0x8b, 0xb4, 0x6a, 0xc4, + 0x64, 0xa3, 0x5e, 0x5b, 0xfe, 0x4d, 0xd0, 0x74, 0x01, 0x1b, 0xdf, 0x10, + 0x45, 0x2b, 0xd6, 0x9e, 0xa9, 0x60, 0x1f, 0xad, 0x46, 0xa1, 0x8c, 0xf8, + 0xf6, 0xa9, 0x8a, 0x27, 0xea, 0x51, 0x37, 0x84, 0xcf, 0xe5, 0xd7, 0x51, + 0xd6, 0x40, 0x39, 0x39, 0x5f, 0xf6, 0x96, 0x33, 0xd9, 0x86, 0x8d, 0x38, + 0xb6, 0x26, 0x04, 0x14, 0x07, 0x46, 0x3e, 0xd0, 0xc5, 0xf6, 0x0d, 0xa0, + 0x47, 0x2b, 0xc8, 0x73, 0x18, 0x6b, 0xd3, 0x0e, 0x18, 0xcc, 0x43, 0x98, + 0xd0, 0xcf, 0x1c, 0xe4, 0x4a, 0x41, 0x6a, 0x56, 0x2d, 0xf0, 0x93, 0x89, + 0x81, 0x6c, 0xce, 0x04, 0x1a, 0x23, 0x05, 0x91, 0x4f, 0x48, 0x44, 0x3a, + 0xaa, 0x03, 0xa5, 0x4a, 0xa9, 0x20, 0x2c, 0xbe, 0x6a, 0x81, 0xe6, 0xa9, + 0xf8, 0xf0, 0x2b, 0x29, 0xa1, 0xe0, 0xc4, 0xce, 0xf5, 0xda, 0x25, 0x70, + 0x49, 0xcc, 0xa0, 0x4b, 0x24, 0x49, 0x4f, 0x11, 0xc4, 0x3b, 0x22, 0x89, + 0x9a, 0xb4, 0xf4, 0xcd, 0xa3, 0xee, 0xb0, 0x76, 0x13, 0xc4, 0xbb, 0xaf, + 0x03, 0x7f, 0x27, 0xf3, 0x38, 0xbc, 0xde, 0x7c, 0x0c, 0x39, 0x14, 0xb7, + 0x14, 0xbb, 0x5c, 0xae, 0x89, 0xf8, 0xf7, 0xd6, 0x00, 0x78, 0xf4, 0xb0, + 0x52, 0x16, 0xf5, 0x54, 0xc5, 0x93, 0xf7, 0x6d, 0x0d, 0xe8, 0x58, 0xe2, + 0xa1, 0xa7, 0xdc, 0x49, 0xdb, 0xc8, 0x79, 0xbc, 0xc3, 0x97, 0x7b, 0x6c, + 0x82, 0x7b, 0xbe, 0xe9, 0x79, 0xac, 0x4a, 0xa4, 0x7c, 0x49, 0x83, 0x58, + 0x3a, 0xe4, 0xf5, 0x68, 0x5c, 0xb7, 0x7f, 0x2d, 0xfe, 0x6b, 0x96, 0xc7, + 0x8b, 0x67, 0xb5, 0xd0, 0xa1, 0x0a, 0x16, 0x62, 0x64, 0x53, 0xea, 0x29, + 0x80, 0x93, 0xf9, 0xd6, 0xa0, 0xc5, 0x1b, 0x3a, 0x1e, 0xab, 0x51, 0x88, + 0xe0, 0x9e, 0xd4, 0xf6, 0xbf, 0x70, 0x2d, 0x29, 0x2e, 0x08, 0xa9, 0x31, + 0x78, 0x0a, 0x15, 0x30, 0x9f, 0x2e, 0xc8, 0x41, 0x65, 0x8e, 0x97, 0x51, + 0x5e, 0x73, 0x46, 0x42, 0x74, 0x84, 0xfd, 0x9b, 0x4a, 0x8a, 0x68, 0x28, + 0x45, 0xd0, 0x5d, 0x65, 0x08, 0xb3, 0xf5, 0x40, 0x8a, 0x29, 0x8e, 0x70, + 0x02, 0x49, 0x6a, 0x01, 0xd6, 0x41, 0x4a, 0xf8, 0x15, 0xa3, 0x70, 0x59, + 0xe9, 0xa2, 0xe2, 0x76, 0x8c, 0x60, 0x33, 0xb3, 0xfa, 0x8b, 0xb4, 0x90, + 0x6f, 0x92, 0xc8, 0x21, 0x59, 0xc0, 0x3a, 0x30, 0x46, 0xeb, 0x49, 0xd8, + 0x85, 0x63, 0x5a, 0x23, 0x87, 0xe1, 0xa7, 0xc0, 0x1a, 0xb0, 0xc7, 0xc4, + 0x40, 0x4d, 0x11, 0x9c, 0xe3, 0xd4, 0x6b, 0xef, 0x68, 0xc8, 0x2c, 0x31, + 0xcd, 0x3e, 0xee, 0x55, 0x10, 0x67, 0x77, 0x7b, 0x30, 0xc1, 0xd0, 0x23, + 0x6c, 0x65, 0x6f, 0xfb, 0x2e, 0x62, 0x33, 0x42, 0x63, 0xdc, 0xca, 0x86, + 0xf1, 0x0e, 0xb3, 0xb0, 0x69, 0x11, 0x65, 0xe1, 0x6e, 0x6c, 0x03, 0x49, + 0x79, 0xe8, 0xf1, 0x2e, 0x8d, 0x94, 0xc8, 0xa8, 0x98, 0x2d, 0x3f, 0xfe, + 0xbd, 0x2d, 0x75, 0x45, 0xd1, 0x7a, 0x09, 0xf8, 0x90, 0x49, 0xbd, 0x4a, + 0x3b, 0xa4, 0xa3, 0x26, 0xb8, 0x62, 0x66, 0x97, 0xd9, 0xc1, 0xca, 0x12, + 0x49, 0xe1, 0x27, 0x93, 0x4f, 0x60, 0xfa, 0xb3, 0x4f, 0x4c, 0xdb, 0x87, + 0x6c, 0x3b, 0x50, 0x47, 0xe2, 0xd8, 0x5b, 0x13, 0x99, 0xf0, 0x2b, 0xbb, + 0x32, 0x33, 0xfd, 0x7d, 0x15, 0x0f, 0x2c, 0xee, 0x85, 0x83, 0xc0, 0x53, + 0x79, 0x3e, 0x51, 0xfe, 0x7c, 0x06, 0x73, 0x49, 0x49, 0x4f, 0x5a, 0x22, + 0x36, 0x8f, 0x30, 0x8a, 0xef, 0x84, 0xd6, 0x15, 0x26, 0x48, 0xe7, 0x1e, + 0xb1, 0xaa, 0x82, 0xd0, 0xc7, 0x0b, 0x97, 0x7b, 0x6c, 0x2d, 0x49, 0x7e, + 0x6d, 0xe7, 0xa3, 0x05, 0x80, 0xd7, 0x42, 0xa9, 0xc6, 0x66, 0x98, 0x30, + 0xe3, 0x8a, 0x79, 0x86, 0x9c, 0x2b, 0xbc, 0x4a, 0xe6, 0x0d, 0xc5, 0xe5, + 0x1a, 0x92, 0xd9, 0xef, 0x63, 0x52, 0x03, 0x88, 0x36, 0xc5, 0x83, 0x65, + 0xf8, 0xf1, 0x87, 0xce, 0x43, 0xfe, 0x89, 0x58, 0x07, 0x6a, 0xad, 0x85, + 0x37, 0x0f, 0xdf, 0x9e, 0xa5, 0x62, 0xa9, 0xd2, 0x41, 0x3f, 0x7f, 0xb7, + 0xf1, 0xe2, 0x58, 0xb5, 0xda, 0xdf, 0xd1, 0xba, 0x36, 0x2c, 0xe7, 0x43, + 0x31, 0x07, 0xc5, 0xf5, 0x79, 0xc9, 0x31, 0xd7, 0x1d, 0x97, 0x57, 0x9a, + 0x8e, 0x3f, 0xac, 0x00, 0x49, 0x00, 0x2f, 0xad, 0xac, 0xe7, 0x65, 0x7c, + 0xbf, 0xec, 0x85, 0x57, 0xe6, 0xcc, 0x07, 0x34, 0x02, 0x36, 0xa8, 0x6a, + 0x9f, 0x3a, 0x9a, 0x2f, 0x34, 0x93, 0x1f, 0x7d, 0x38, 0x54, 0xe3, 0x54, + 0x54, 0xee, 0x84, 0x55, 0xe1, 0x0d, 0xc1, 0x08, 0x3e, 0x33, 0x9e, 0x2a, + 0xc3, 0x6a, 0x83, 0xc4, 0x75, 0xed, 0xbc, 0x5f, 0xd9, 0x04, 0xd7, 0x77, + 0x91, 0xb1, 0xa0, 0xf2, 0xef, 0x81, 0xb0, 0x8b, 0x53, 0x5f, 0x71, 0xec, + 0xa5, 0x0b, 0xbe, 0xf2, 0x92, 0x7e, 0x0a, 0x34, 0xeb, 0x5d, 0x65, 0xc7, + 0xa9, 0x44, 0x10, 0xfb, 0xd3, 0xef, 0xe1, 0xbc, 0x06, 0x65, 0x68, 0x22, + 0xfb, 0x43, 0x2c, 0xcf, 0x8e, 0x6a, 0x28, 0xdb, 0x0b, 0xf4, 0xaf, 0x01, + 0x65, 0x97, 0xd6, 0xe5, 0x91, 0x20, 0x13, 0x2c, 0xb1, 0xc2, 0xd3, 0xc3, + 0x76, 0x90, 0xf8, 0xcd, 0x00, 0xde, 0x93, 0xf8, 0x4e, 0xcc, 0xdc, 0xca, + 0x9a, 0xf0, 0xbd, 0x9b, 0xd6, 0x57, 0xb1, 0x13, 0xd9, 0xe0, 0xe1, 0x9e, + 0x21, 0x74, 0xa9, 0x76, 0xc0, 0x0c, 0xad, 0x4f, 0x5d, 0xfe, 0x23, 0x32, + 0x5a, 0x10, 0x75, 0x5b, 0x05, 0xdf, 0xdc, 0x5b, 0x94, 0xcb, 0xe1, 0x9f, + 0x13, 0x51, 0xf5, 0x50, 0x36, 0x3b, 0xf2, 0x90, 0x9c, 0x9a, 0xc8, 0x10, + 0x88, 0xa9, 0xec, 0x22, 0x1e, 0x96, 0x70, 0xe8, 0x9e, 0x69, 0xc1, 0x22, + 0xd9, 0x14, 0x15, 0x2e, 0xbc, 0x03, 0x96, 0x9e, 0x1d, 0x00, 0x10, 0x16, + 0x4f, 0x56, 0xf0, 0x29, 0x47, 0x0a, 0x45, 0x34, 0x27, 0x21, 0x3b, 0x67, + 0x33, 0xf9, 0xdd, 0x29, 0x3a, 0xf2, 0xe4, 0x56, 0x34, 0x46, 0xbe, 0xd8, + 0x42, 0x29, 0x11, 0x7f, 0x30, 0xc1, 0xbe, 0xa5, 0xc8, 0x9d, 0x7b, 0x2e, + 0x4e, 0xcf, 0xba, 0x91, 0xb4, 0xbf, 0x0a, 0x04, 0x00, 0x49, 0x83, 0x6b, + 0x46, 0x5f, 0x3b, 0xfa, 0xf7, 0x40, 0x8d, 0x85, 0x47, 0x14, 0x58, 0xb3, + 0xa5, 0x66, 0x30, 0xfd, 0x4a, 0x80, 0xa4, 0x61, 0x3b, 0x7c, 0xb4, 0xcc, + 0x34, 0x8c, 0xc6, 0xb6, 0x10, 0xa9, 0x76, 0xc9, 0x11, 0xd7, 0x8a, 0x51, + 0x86, 0x17, 0x89, 0x28, 0xab, 0xd5, 0x03, 0x88, 0x74, 0x5b, 0x81, 0xbd, + 0x3a, 0x57, 0xfe, 0x66, 0x25, 0xd0, 0x92, 0x15, 0x84, 0x02, 0x0f, 0x51, + 0xa8, 0x58, 0xcf, 0x77, 0x65, 0x10, 0x61, 0xe8, 0xe6, 0xab, 0xb1, 0xba, + 0x3b, 0x08, 0xd6, 0xba, 0x5f, 0xf5, 0x74, 0xc5, 0x07, 0x60, 0xfd, 0xd3, + 0xc8, 0x52, 0x4e, 0xdb, 0xc3, 0xe3, 0x6d, 0x81, 0x20, 0x51, 0x01, 0x9a, + 0x5e, 0x32, 0x4e, 0x80, 0x5a, 0xcb, 0x83, 0xd7, 0xa4, 0xd9, 0xfb, 0xed, + 0x3d, 0x80, 0xa1, 0x83, 0x81, 0x91, 0xc0, 0x0b, 0xff, 0x67, 0xd8, 0x8b, + 0xd0, 0x12, 0x0b, 0xd4, 0x2b, 0x8e, 0x0d, 0x0f, 0xfc, 0xc7, 0xb3, 0xf1, + 0xe3, 0xf3, 0x5e, 0x0c, 0xb6, 0x6b, 0x9d, 0xdc, 0x22, 0x70, 0x31, 0x54, + 0xe8, 0x41, 0xfe, 0xa1, 0xe1, 0x4f, 0xfa, 0x81, 0xfb, 0xae, 0x72, 0x16, + 0xb8, 0x87, 0xc9, 0x31, 0x9d, 0x42, 0x47, 0x4a, 0x20, 0xae, 0x63, 0x16, + 0x0d, 0xfa, 0xf1, 0x27, 0x19, 0x47, 0xee, 0x45, 0x84, 0x29, 0x9a, 0xb6, + 0x42, 0xef, 0xbd, 0x15, 0xa8, 0x34, 0x33, 0x38, 0x9c, 0x9d, 0xbb, 0x5c, + 0x03, 0xf3, 0xcf, 0xcf, 0x6d, 0x2e, 0xd5, 0x88, 0xf8, 0xdd, 0xfc, 0xc0, + 0x4a, 0xdb, 0x69, 0xd9, 0x62, 0x89, 0x24, 0x46, 0xee, 0xa4, 0xb9, 0x95, + 0xe6, 0xaf, 0x7d, 0x53, 0xec, 0x41, 0xae, 0x70, 0xfe, 0x4f, 0x31, 0xe3, + 0xa2, 0x59, 0x2c, 0xa1, 0x53, 0x8b, 0xb6, 0x3b, 0x39, 0xc1, 0xa4, 0xa7, + 0x9e, 0xaa, 0x00, 0x60, 0x9a, 0x5f, 0x56, 0x51, 0xf3, 0x7b, 0x28, 0x84, + 0x36, 0x1a, 0xc1, 0x2d, 0xc8, 0xed, 0xf8, 0x48, 0x48, 0x1d, 0x39, 0x4d, + 0x3d, 0xce, 0x30, 0x90, 0x29, 0x33, 0x6f, 0x9a, 0xce, 0x58, 0xe7, 0x88, + 0xac, 0x59, 0xce, 0x85, 0x5a, 0x52, 0x2b, 0x6c, 0xb7, 0xe9, 0x2e, 0xa9, + 0xd9, 0x9a, 0xea, 0x1c, 0x47, 0xb2, 0x59, 0xff, 0x73, 0x76, 0x21, 0x40, + 0xe1, 0xde, 0x32, 0xb8, 0x73, 0x3d, 0xa5, 0x44, 0x66, 0x79, 0xa1, 0xfe, + 0xaf, 0xf6, 0x8a, 0x97, 0x09, 0x5c, 0x8b, 0x64, 0x38, 0x9f, 0xe1, 0x59, + 0x38, 0x18, 0xe9, 0xc0, 0xd6, 0xa2, 0xac, 0x74, 0xa9, 0xfd, 0x4a, 0x0d, + 0xf6, 0x47, 0x00, 0x2b, 0x09, 0x46, 0x38, 0x1c, 0xa4, 0x9f, 0x63, 0x20, + 0x18, 0x75, 0x5a, 0xb8, 0xc4, 0xbc, 0xd6, 0x6b, 0xc8, 0x14, 0x72, 0x03, + 0xe4, 0x05, 0xd4, 0x4e, 0x66, 0x20, 0x42, 0xa2, 0x8f, 0x96, 0xe7, 0xaf, + 0xd3, 0xfb, 0xa8, 0x88, 0x9b, 0xe3, 0xaa, 0xcd, 0xab, 0xce, 0x8f, 0x07, + 0x6d, 0xef, 0x98, 0xce, 0xdb, 0x42, 0x5b, 0xf4, 0x61, 0x57, 0x62, 0x27, + 0x8a, 0x53, 0x5e, 0xf8, 0x3e, 0xf6, 0x7f, 0xde, 0x5e, 0x3b, 0x1b, 0x13, + 0x2e, 0x30, 0x46, 0x4b, 0x6b, 0xb7, 0xbb, 0x33, 0x31, 0xc0, 0xfa, 0x40, + 0xab, 0x68, 0x72, 0xe3, 0x92, 0x30, 0x47, 0xd6, 0x30, 0x60, 0x42, 0x5b, + 0x88, 0x8d, 0xa6, 0x56, 0xe4, 0xac, 0x33, 0x2e, 0xca, 0x05, 0x1f, 0x60, + 0xaf, 0xde, 0x7f, 0xa9, 0xda, 0x3f, 0xa8, 0x21, 0xf6, 0xfc, 0x98, 0x7d, + 0xc4, 0x1e, 0xb0, 0xa9, 0x56, 0x2d, 0x8d, 0xea, 0x03, 0x51, 0x48, 0xac, + 0xe8, 0x22, 0xc7, 0x8b, 0xef, 0x91, 0x0e, 0xcf, 0x0c, 0xe9, 0x38, 0x43, + 0x99, 0xa8, 0x98, 0x4f, 0xfa, 0xe3, 0x03, 0xa6, 0x4f, 0xd4, 0x0d, 0x98, + 0x5b, 0x50, 0x28, 0xd7, 0xe7, 0x46, 0xd7, 0xad, 0x43, 0xb8, 0x56, 0x2a, + 0x2f, 0x7c, 0x39, 0x67, 0xf4, 0x62, 0x0e, 0xc0, 0xa8, 0x87, 0xb5, 0x81, + 0xe2, 0x13, 0x9f, 0xe4, 0xdd, 0x72, 0xf2, 0x07, 0xca, 0xac, 0x6d, 0xb2, + 0x96, 0x53, 0x5a, 0x8f, 0x66, 0x3c, 0xb4, 0xc1, 0x4f, 0x9a, 0x82, 0x55, + 0xcf, 0x0e, 0x27, 0x5f, 0xc7, 0xd2, 0x28, 0x27, 0x7f, 0x22, 0x6e, 0xa5, + 0xe7, 0x32, 0x56, 0x51, 0x18, 0xe0, 0x85, 0x6d, 0x1f, 0xfc, 0x25, 0x08, + 0x18, 0x60, 0x57, 0xfc, 0x66, 0x94, 0x2c, 0x4c, 0xbe, 0x00, 0xab, 0x9e, + 0x73, 0x9b, 0x06, 0xd3, 0xb5, 0x24, 0xa8, 0x8f, 0xb1, 0x33, 0x99, 0x4c, + 0xb4, 0x13, 0x07, 0xcd, 0x04, 0xdd, 0x77, 0xdc, 0xee, 0x96, 0x02, 0x59, + 0xe8, 0x22, 0x07, 0x16, 0x2e, 0x41, 0xc9, 0xc4, 0x59, 0x70, 0x37, 0x0f, + 0x14, 0xc9, 0xcf, 0x90, 0x57, 0xc2, 0x0d, 0xa3, 0xd7, 0x66, 0xb6, 0x7d, + 0x10, 0xd4, 0xfc, 0x18, 0x66, 0xad, 0xea, 0x5e, 0x64, 0x6c, 0x12, 0x66, + 0x3d, 0x96, 0xa5, 0xa8, 0x9c, 0x49, 0x5c, 0xd4, 0x8d, 0x1c, 0xc3, 0x38, + 0xfe, 0x53, 0xc2, 0x71, 0xd1, 0xc6, 0x41, 0xe2, 0xb9, 0x17, 0x74, 0x6e, + 0xcc, 0xf8, 0x72, 0x28, 0x38, 0x4e, 0x54, 0x9b, 0x0e, 0xa3, 0x3a, 0x43, + 0x5c, 0xd5, 0x83, 0x06, 0xbb, 0x46, 0x16, 0x6e, 0xe3, 0x8a, 0xd5, 0x1e, + 0x7f, 0x88, 0x62, 0xac, 0x35, 0x89, 0xfb, 0xbe, 0x96, 0x1d, 0x87, 0x37, + 0xb7, 0x91, 0x63, 0xae, 0x77, 0x7b, 0x66, 0x60, 0xc1, 0x3e, 0x80, 0x56, + 0xb1, 0xc8, 0x0d, 0x16, 0xde, 0x38, 0x82, 0x66, 0x99, 0x2b, 0x35, 0xd8, + 0xb4, 0x5b, 0x4b, 0x3e, 0x93, 0x96, 0x59, 0xf8, 0x96, 0x7e, 0x7b, 0x27, + 0xf4, 0x62, 0xb7, 0xda, 0x89, 0xa7, 0x34, 0x47, 0xed, 0xb3, 0x42, 0x20, + 0xeb, 0xcd, 0xf6, 0xa3, 0x9f, 0xf7, 0x48, 0x91, 0x17, 0xd2, 0x21, 0xed, + 0x5a, 0x22, 0x39, 0xc9, 0x76, 0x95, 0x36, 0xd9, 0x97, 0x0f, 0x19, 0xce, + 0xd3, 0xbc, 0x74, 0x7d, 0x53, 0x37, 0x3b, 0x4a, 0x97, 0xb7, 0xf8, 0x7e, + 0xdd, 0x4c, 0x5f, 0xae, 0x5c, 0x0b, 0xab, 0x4c, 0x34, 0xa1, 0x7e, 0x34, + 0x35, 0xf4, 0xfc, 0x92, 0xab, 0x2e, 0x6a, 0x15, 0xce, 0x84, 0xae, 0x70, + 0xae, 0x85, 0x21, 0xe6, 0x41, 0x13, 0x31, 0xe0, 0x8f, 0xab, 0x82, 0xe3, + 0x09, 0xaf, 0xa4, 0x7c, 0xb4, 0xb9, 0xb7, 0xc0, 0x67, 0x08, 0xc9, 0x9d, + 0xcd, 0x0b, 0x3c, 0xa0, 0x0c, 0xde, 0x49, 0x2f, 0x40, 0x19, 0x95, 0x64, + 0xb9, 0x7c, 0x2a, 0x72, 0xdd, 0xa2, 0x92, 0x0a, 0x21, 0xeb, 0x8c, 0xc3, + 0x6d, 0x52, 0xe7, 0x05, 0x50, 0x01, 0x55, 0x19, 0x2f, 0xbd, 0x1b, 0x72, + 0x73, 0xfe, 0x82, 0x9f, 0xbf, 0xa0, 0xfe, 0x19, 0x7c, 0x42, 0x6d, 0x76, + 0x32, 0x47, 0x36, 0x15, 0x2e, 0xde, 0xe8, 0xe6, 0xca, 0x07, 0xa3, 0x6b, + 0x40, 0x99, 0x96, 0xcd, 0x19, 0xea, 0x7e, 0xc9, 0x87, 0x9d, 0x3d, 0xa0, + 0x82, 0x88, 0xe7, 0xe4, 0x34, 0x9f, 0xa5, 0x27, 0xdf, 0xae, 0x03, 0x37, + 0xa8, 0x35, 0x64, 0x02, 0x09, 0x09, 0x9e, 0xec, 0x38, 0x0a, 0xff, 0x79, + 0x8c, 0x9a, 0x87, 0x66, 0xcd, 0xe4, 0xf4, 0x9d, 0xa9, 0x07, 0x96, 0x36, + 0xae, 0x2e, 0x4e, 0xc5, 0xe9, 0x86, 0xb2, 0x8e, 0x71, 0x5d, 0xe8, 0xee, + 0x84, 0xf3, 0x30, 0x2a, 0x58, 0x1a, 0x80, 0xb8, 0xaa, 0xb8, 0x1d, 0xc4, + 0xae, 0x59, 0x91, 0xf3, 0x16, 0x9b, 0xa3, 0x8a, 0xa3, 0x26, 0xb2, 0x0a, + 0xe5, 0x58, 0xb7, 0x96, 0x87, 0xfb, 0x00, 0xe4, 0x50, 0x7c, 0xb1, 0x77, + 0x3a, 0x18, 0xc2, 0xe3, 0xc1, 0x12, 0xa6, 0x0d, 0x06, 0xeb, 0x80, 0x6c, + 0x5a, 0xee, 0x34, 0xcc, 0x1c, 0x87, 0x35, 0x46, 0x1d, 0x05, 0x83, 0xd8, + 0x91, 0x22, 0xaa, 0xf6, 0xad, 0x87, 0xab, 0x76, 0x18, 0x79, 0xe2, 0x09, + 0xc3, 0xa3, 0x15, 0x67, 0x3a, 0x7c, 0x0f, 0xa0, 0x4c, 0x7b, 0xfc, 0xfc, + 0xdd, 0x5c, 0xe4, 0x86, 0x58, 0x13, 0xb8, 0x97, 0xae, 0x8c, 0x75, 0xc8, + 0x02, 0x1e, 0x33, 0x45, 0xa9, 0x54, 0x09, 0x15, 0x53, 0x4f, 0x28, 0x47, + 0x4d, 0x5f, 0xd0, 0xc7, 0x09, 0xbd, 0x93, 0xb0, 0x08, 0x79, 0x05, 0xbc, + 0xbc, 0xaf, 0x2c, 0xbd, 0xbb, 0x21, 0xd1, 0x60, 0xb8, 0x81, 0x4c, 0x6c, + 0x5e, 0x45, 0x39, 0xa3, 0x31, 0x54, 0xb7, 0x82, 0xef, 0x86, 0xe4, 0x5e, + 0xca, 0xd6, 0xb8, 0x31, 0xa2, 0x4c, 0x84, 0x5b, 0xac, 0xe5, 0x29, 0xbf, + 0xbf, 0x89, 0xb4, 0x4c, 0xd3, 0x69, 0x66, 0x50, 0xeb, 0xda, 0x7d, 0x00, + 0xbb, 0x45, 0x0f, 0xe1, 0xd1, 0x30, 0x1a, 0xc6, 0x94, 0x66, 0xdc, 0x01, + 0x75, 0xce, 0xf8, 0xfc, 0xd9, 0xce, 0xcf, 0x1f, 0x9e, 0x5a, 0x55, 0xa4, + 0x3e, 0xe6, 0x51, 0xc7, 0x74, 0x40, 0x82, 0x09, 0xea, 0xa0, 0xf5, 0xb2, + 0x70, 0x9f, 0x0e, 0xfb, 0x46, 0x8a, 0x69, 0xbf, 0x07, 0x92, 0xdc, 0x74, + 0x03, 0x70, 0xc6, 0x44, 0x81, 0x66, 0x40, 0xc7, 0xf5, 0xb8, 0xf0, 0x45, + 0x0f, 0xca, 0xd8, 0xb0, 0x9e, 0x48, 0x94, 0xff, 0x85, 0xcb, 0x7b, 0xec, + 0x67, 0x5d, 0xfe, 0xe9, 0x13, 0xd1, 0x67, 0x95, 0xd9, 0x35, 0x9e, 0x8a, + 0x53, 0x4d, 0x6b, 0x9d, 0x42, 0x53, 0xb1, 0x6b, 0x51, 0x1e, 0x35, 0x40, + 0x81, 0x92, 0x91, 0x5f, 0x1f, 0x8e, 0xbe, 0x37, 0xd3, 0x85, 0xab, 0x85, + 0x37, 0x1c, 0x0f, 0xae, 0xd9, 0xf7, 0xa2, 0x75, 0x3d, 0xd9, 0xd7, 0x2a, + 0x80, 0xb0, 0x4c, 0x14, 0x04, 0x40, 0xc5, 0xba, 0x0e, 0xbe, 0xab, 0xcc, + 0x38, 0x35, 0x62, 0x6c, 0xa5, 0xce, 0x49, 0x15, 0x2a, 0x10, 0xb5, 0x6a, + 0xd2, 0x3b, 0xd2, 0x6a, 0xad, 0x2e, 0x34, 0x46, 0x8b, 0x78, 0x57, 0x6e, + 0xc4, 0xde, 0x65, 0x68, 0x05, 0x8f, 0xd6, 0x6e, 0x34, 0xb9, 0xaa, 0x80, + 0x77, 0xff, 0x6c, 0x1a, 0x37, 0x87, 0xdd, 0x33, 0x13, 0x33, 0xa7, 0xa9, + 0x3a, 0x90, 0x32, 0x7b, 0x9b, 0x21, 0x31, 0xc8, 0xf5, 0x4c, 0xa6, 0x73, + 0x42, 0x79, 0x46, 0x14, 0x1b, 0xef, 0xf4, 0x78, 0xd9, 0x7e, 0x6f, 0x31, + 0xaa, 0x59, 0x97, 0x34, 0xe5, 0xe6, 0x67, 0xf3, 0x86, 0xf5, 0x61, 0xe7, + 0x51, 0x6d, 0xce, 0xb3, 0xdc, 0x86, 0xc7, 0x55, 0x43, 0xfa, 0x38, 0x78, + 0xb0, 0x8d, 0x03, 0x9c, 0xe4, 0x6c, 0xca, 0x73, 0x94, 0xa1, 0x0c, 0xb8, + 0x11, 0xda, 0x0c, 0x0b, 0x18, 0x1b, 0xd0, 0x99, 0xe7, 0xa9, 0x0d, 0xc3, + 0x36, 0xd7, 0x8c, 0x16, 0xad, 0x16, 0x1f, 0xb2, 0x3c, 0x07, 0x32, 0x11, + 0x6c, 0xd2, 0x8f, 0x33, 0x37, 0x5c, 0x3e, 0x4f, 0x7a, 0x76, 0xf7, 0x85, + 0xcc, 0x68, 0x1a, 0xf9, 0x26, 0x74, 0x42, 0xc9, 0xea, 0x21, 0x7e, 0x74, + 0x3c, 0x4f, 0xde, 0xfb, 0xd7, 0x83, 0x62, 0x12, 0xc7, 0x4f, 0xfc, 0x47, + 0x18, 0x9d, 0xc5, 0xf5, 0xe9, 0xd7, 0xaa, 0x76, 0x20, 0x99, 0x79, 0xae, + 0x9b, 0x7a, 0xde, 0x8b, 0x95, 0xc2, 0xa5, 0xa3, 0x6a, 0x30, 0x9b, 0x99, + 0x63, 0x34, 0x7c, 0xd1, 0x53, 0xa1, 0x6c, 0xd6, 0xed, 0x7d, 0x8c, 0xba, + 0xc8, 0x21, 0xf3, 0xe1, 0x31, 0x55, 0x3d, 0x88, 0x87, 0x04, 0xc7, 0xc9, + 0x65, 0x0c, 0x53, 0x1e, 0xd4, 0xd9, 0xaa, 0xda, 0xc2, 0x14, 0x88, 0xf2, + 0x07, 0x2c, 0x12, 0x4d, 0x79, 0x54, 0xaa, 0xd9, 0x47, 0x95, 0xf9, 0x7e, + 0x26, 0x89, 0x4b, 0x63, 0x7e, 0x44, 0x06, 0x0e, 0xe2, 0x8d, 0x9a, 0x0a, + 0xc3, 0xee, 0x55, 0x13, 0x55, 0x04, 0xcc, 0xb5, 0x2e, 0xa0, 0x0d, 0xec, + 0x76, 0x84, 0xc1, 0x1e, 0xdd, 0xe6, 0xfa, 0x54, 0x6e, 0x38, 0x30, 0x6f, + 0xcc, 0xa4, 0x8d, 0x76, 0x1e, 0xa3, 0x8e, 0x2c, 0x5e, 0x37, 0xeb, 0x0b, + 0xf4, 0xb5, 0x80, 0xde, 0x58, 0x13, 0x5a, 0x52, 0xdc, 0x65, 0x99, 0x1a, + 0x1b, 0x75, 0x0c, 0xbd, 0x83, 0xe8, 0x90, 0x8e, 0xa9, 0xbf, 0x42, 0x22, + 0xe1, 0x3a, 0x31, 0x4e, 0x54, 0xad, 0xd4, 0x6f, 0x80, 0xb4, 0xb5, 0x82, + 0x05, 0x20, 0xd7, 0x38, 0xd7, 0xeb, 0x25, 0x33, 0xe9, 0x4b, 0xc3, 0x5e, + 0xd1, 0x11, 0xb0, 0xd9, 0x8e, 0x90, 0x48, 0x2a, 0xe3, 0xa0, 0x60, 0x16, + 0x70, 0xe3, 0xd1, 0x45, 0x11, 0x64, 0x91, 0x69, 0x87, 0x1c, 0xbb, 0x91, + 0xc4, 0x43, 0x12, 0x62, 0x99, 0x69, 0xe5, 0x96, 0x01, 0x15, 0xdb, 0xdf, + 0x05, 0x55, 0x34, 0xbb, 0xd6, 0x76, 0x89, 0xcd, 0xb5, 0x4f, 0x2e, 0xa7, + 0x6e, 0x15, 0xc9, 0xc0, 0x8e, 0xa8, 0x63, 0x79, 0x12, 0xfb, 0x7e, 0x69, + 0x8f, 0x52, 0x5e, 0xe7, 0x76, 0x16, 0x28, 0x76, 0xca, 0xcb, 0xd8, 0x0e, + 0x4a, 0x93, 0x9d, 0x16, 0x68, 0x98, 0xf8, 0xc3, 0x39, 0xb2, 0x2d, 0xea, + 0xba, 0x72, 0x16, 0x33, 0xb7, 0xec, 0x61, 0x9e, 0x94, 0x32, 0x01, 0x22, + 0xde, 0x66, 0xfd, 0x68, 0xfa, 0xcf, 0xf2, 0x52, 0x4f, 0x02, 0xe8, 0x25, + 0xd3, 0xa3, 0x5b, 0x29, 0xae, 0xe9, 0x62, 0xfa, 0xd6, 0x1a, 0x50, 0x80, + 0x95, 0x96, 0xdf, 0x00, 0xfc, 0x23, 0xf1, 0x95, 0xef, 0xbb, 0xf5, 0x23, + 0x9d, 0x6b, 0xd6, 0xed, 0xb4, 0xe2, 0x4a, 0xf6, 0xb8, 0x20, 0x83, 0x6b, + 0x45, 0x92, 0x29, 0x5a, 0x02, 0xe9, 0xf7, 0x8e, 0x5c, 0x02, 0xde, 0xb4, + 0x9a, 0xdf, 0x18, 0x10, 0x17, 0x7f, 0xd8, 0x2e, 0x17, 0xc0, 0xf0, 0x6b, + 0x3b, 0x88, 0x09, 0x58, 0xf2, 0x18, 0x22, 0x09, 0x80, 0x4a, 0xe0, 0x51, + 0x6f, 0x7a, 0x70, 0x09, 0x1f, 0xe5, 0xfa, 0xa9, 0x4d, 0x24, 0x1f, 0x18, + 0x1c, 0x74, 0xcd, 0x87, 0x04, 0xfd, 0x85, 0x33, 0x4c, 0x28, 0xbd, 0xa3, + 0x66, 0x6c, 0x99, 0x7e, 0x50, 0x5e, 0xb5, 0x22, 0x33, 0x92, 0xd4, 0xd8, + 0x82, 0x4e, 0x38, 0xbe, 0xcb, 0x3d, 0x5f, 0x19, 0xd1, 0x0f, 0x8b, 0xa1, + 0x78, 0x08, 0x1c, 0x10, 0x0b, 0x77, 0xa7, 0x39, 0x2e, 0x91, 0x83, 0xee, + 0x1d, 0x36, 0xd8, 0x77, 0x87, 0x8a, 0x38, 0x45, 0x3c, 0xbd, 0xb9, 0x88, + 0xbb, 0x1b, 0x20, 0xd1, 0x95, 0xb9, 0x8f, 0x03, 0x46, 0xfa, 0xab, 0x70, + 0x68, 0x26, 0xd9, 0xb1, 0x25, 0x52, 0x5a, 0x77, 0x2d, 0x92, 0xc2, 0x1d, + 0xb6, 0x6e, 0xec, 0x67, 0xef, 0x34, 0xe2, 0x64, 0xb3, 0xa0, 0xae, 0x0c, + 0xd9, 0x36, 0xa1, 0xc7, 0xd8, 0xbf, 0x7a, 0x43, 0xbf, 0xc0, 0xc6, 0x90, + 0x60, 0x6a, 0x23, 0xc0, 0x6a, 0x5d, 0x62, 0x18, 0xac, 0xc1, 0x20, 0x35, + 0x17, 0xba, 0x4e, 0x54, 0xb7, 0xec, 0xd4, 0xad, 0x99, 0x94, 0xa4, 0xda, + 0x57, 0xe7, 0x46, 0xed, 0x47, 0xd1, 0xb4, 0xa2, 0x3e, 0x0f, 0x4a, 0xb6, + 0xa6, 0x68, 0x3e, 0x94, 0xb9, 0x18, 0x30, 0xe0, 0x75, 0x08, 0xe8, 0xf3, + 0x21, 0x79, 0x26, 0x68, 0x6a, 0x65, 0xb6, 0xbe, 0x03, 0x98, 0x8f, 0x04, + 0xad, 0x1e, 0xb0, 0x54, 0xd2, 0x28, 0xdd, 0x4a, 0xe9, 0xf3, 0xa0, 0x06, + 0xbf, 0x0b, 0x2a, 0xee, 0xf8, 0x03, 0x7e, 0x1d, 0x37, 0xc1, 0x32, 0xd1, + 0x41, 0xf4, 0x9b, 0xc5, 0x02, 0x10, 0x6f, 0x55, 0x5a, 0xec, 0x5b, 0xe7, + 0x61, 0x05, 0x17, 0xf0, 0xf8, 0xc6, 0x89, 0xe8, 0xad, 0x32, 0x57, 0x14, + 0xe5, 0xf8, 0xf5, 0x88, 0xd9, 0x73, 0x17, 0x10, 0xa7, 0xc3, 0xf8, 0x78, + 0x0b, 0x66, 0xab, 0x63, 0x4f, 0x96, 0x5d, 0xdf, 0x36, 0x83, 0xc4, 0x6f, + 0x20, 0xbd, 0xcb, 0x4c, 0xd2, 0xfa, 0x35, 0x87, 0xd8, 0xb6, 0xbb, 0xcc, + 0xb6, 0xd2, 0x85, 0x03, 0x6a, 0xea, 0xbb, 0x6d, 0x2f, 0xa2, 0x06, 0xc0, + 0xd6, 0x68, 0xd9, 0x7f, 0xd6, 0xa2, 0x3b, 0x08, 0x6a, 0x98, 0x26, 0x6d, + 0x9a, 0x2b, 0x68, 0x51, 0x78, 0xde, 0xa6, 0x96, 0x50, 0x7b, 0xfc, 0x03, + 0x43, 0xf8, 0x21, 0x01, 0x9d, 0xe2, 0x89, 0x65, 0x47, 0xae, 0x9c, 0x45, + 0x5e, 0xa5, 0xce, 0x97, 0xb3, 0xe6, 0xf6, 0xd4, 0x5a, 0xe8, 0x6b, 0x87, + 0xd6, 0xdf, 0xfb, 0x1f, 0xaf, 0xfb, 0xaf, 0x19, 0xa5, 0xfd, 0xba, 0xe0, + 0x22, 0x2f, 0x91, 0x97, 0xdf, 0xae, 0xe9, 0x39, 0xb1, 0xe4, 0xd3, 0x10, + 0xcb, 0xb3, 0x03, 0xb5, 0x0b, 0xf0, 0xd9, 0x70, 0x1e, 0x9c, 0x63, 0x6f, + 0x3a, 0xcf, 0x3c, 0x1b, 0x86, 0xa3, 0xad, 0x1a, 0xe7, 0x4c, 0x09, 0xd0, + 0x80, 0xf6, 0x8b, 0x72, 0x96, 0x53, 0x7e, 0x66, 0xfb, 0x7c, 0x7c, 0x8a, + 0xb0, 0x60, 0xa6, 0x4c, 0x20, 0xc4, 0x63, 0x69, 0x6a, 0xc3, 0x53, 0xf8, + 0x9a, 0x28, 0x30, 0x9d, 0x6f, 0x0e, 0x1b, 0xb2, 0x2c, 0xe6, 0x94, 0x9f, + 0xfc, 0xc0, 0x8d, 0x71, 0xbe, 0x37, 0xa6, 0xc9, 0xbd, 0x3c, 0x4a, 0xf3, + 0xc4, 0xb3, 0x88, 0x4c, 0x45, 0x26, 0x4e, 0x2f, 0x83, 0x16, 0x70, 0xb6, + 0xc7, 0xb2, 0x36, 0xf0, 0x0c, 0x67, 0xd2, 0x0a, 0xd3, 0xd9, 0x7c, 0x35, + 0x29, 0xac, 0xd4, 0x9c, 0x6d, 0xfc, 0xec, 0x58, 0x92, 0xf0, 0xba, 0x32, + 0x00, 0xae, 0xb1, 0xeb, 0x4d, 0x8c, 0x1a, 0x20, 0xe7, 0x5c, 0xfc, 0x9a, + 0x4d, 0x51, 0x24, 0x7b, 0x52, 0xeb, 0x13, 0x3d, 0xb4, 0xab, 0xda, 0xb3, + 0x74, 0x39, 0xd2, 0xf8, 0x2d, 0xef, 0x9b, 0x0f, 0xae, 0xf5, 0x3c, 0x99, + 0x34, 0xbe, 0x15, 0x5c, 0x9f, 0x5d, 0xae, 0xf4, 0x72, 0xc2, 0xac, 0x06, + 0xbe, 0xad, 0xe4, 0x68, 0xea, 0xd5, 0xa1, 0xdc, 0xdb, 0xf4, 0x61, 0x51, + 0xf5, 0x1a, 0x62, 0x15, 0xfd, 0x00, 0x51, 0x35, 0x53, 0x6c, 0x39, 0x3e, + 0xdb, 0x60, 0x0a, 0x52, 0xc1, 0x52, 0x3c, 0xd7, 0xab, 0x73, 0xea, 0x1e, + 0x38, 0x38, 0x65, 0x35, 0x35, 0x2b, 0x28, 0x04, 0x5c, 0x82, 0xea, 0x4a, + 0x9e, 0x96, 0x72, 0xa4, 0x8e, 0x42, 0xfd, 0x55, 0xa8, 0x66, 0x7a, 0x40, + 0xc9, 0xf2, 0xc2, 0x1e, 0x5d, 0x09, 0x90, 0x32, 0x18, 0xdb, 0x11, 0x4c, + 0x6c, 0x9c, 0x27, 0x62, 0x0a, 0xe6, 0xc1, 0xdf, 0xf2, 0x6a, 0x8c, 0x26, + 0xb4, 0xfb, 0xda, 0xa9, 0x08, 0x10, 0x3a, 0xf0, 0xe1, 0x64, 0xe5, 0x03, + 0x81, 0x7d, 0x15, 0x74, 0xa1, 0x8d, 0x10, 0xc8, 0xbb, 0x6a, 0x7c, 0x60, + 0xa1, 0x09, 0x35, 0x19, 0x2d, 0x70, 0xb5, 0x36, 0xc8, 0x8b, 0x66, 0x5f, + 0xe0, 0xe7, 0xea, 0x70, 0x2f, 0x5d, 0x3f, 0xae, 0x5e, 0x25, 0x84, 0xdd, + 0x9b, 0x69, 0x44, 0x37, 0x7c, 0x6b, 0x9e, 0x81, 0x18, 0x36, 0x4b, 0xff, + 0x86, 0x44, 0x2a, 0x39, 0x66, 0x7f, 0x71, 0x43, 0xe7, 0x65, 0xfe, 0xfd, + 0x34, 0xb9, 0xd9, 0x5a, 0x00, 0xd1, 0x41, 0x43, 0xc7, 0xbc, 0x65, 0x68, + 0xb7, 0x73, 0xff, 0x19, 0xd3, 0xed, 0x15, 0xa4, 0x67, 0xa1, 0x53, 0x0e, + 0xa6, 0xfb, 0x25, 0xce, 0x9d, 0x5b, 0x73, 0x08, 0xf3, 0x3b, 0x69, 0xe4, + 0x94, 0x9b, 0x94, 0x03, 0xb3, 0x8a, 0x2e, 0x07, 0x0c, 0xef, 0x18, 0x4c, + 0x2b, 0x1c, 0x83, 0x9f, 0x25, 0x20, 0x29, 0x72, 0x11, 0xa0, 0xaa, 0xed, + 0x0c, 0xf9, 0xce, 0x94, 0x0d, 0x7a, 0xb6, 0xb3, 0xa4, 0x57, 0xd6, 0x61, + 0xca, 0x1a, 0x0e, 0x89, 0x6d, 0x99, 0x4d, 0x06, 0xcd, 0x83, 0x7e, 0x09, + 0x14, 0x5b, 0xe7, 0x4c, 0x72, 0xa8, 0x98, 0xc8, 0x27, 0xf3, 0x70, 0x89, + 0x87, 0x11, 0xbb, 0x98, 0x82, 0x77, 0x9d, 0xaa, 0x95, 0x8c, 0xc1, 0xf8, + 0x39, 0x27, 0xd5, 0x64, 0x59, 0x6a, 0x8c, 0xbe, 0xe2, 0xe1, 0xd1, 0x6b, + 0xe3, 0xaf, 0x30, 0x6f, 0xf4, 0x9e, 0x35, 0x0b, 0x10, 0x24, 0x77, 0xd8, + 0xa4, 0x30, 0x2e, 0xf7, 0x97, 0xfd, 0xef, 0x1e, 0x9e, 0xf2, 0xbd, 0xf2, + 0x41, 0x73, 0x19, 0xe6, 0x7b, 0x7f, 0x74, 0x11, 0x91, 0x38, 0xc5, 0xac, + 0xd5, 0xb0, 0x48, 0xc4, 0xe9, 0x41, 0xd4, 0x50, 0x76, 0x13, 0xbf, 0xec, + 0xe8, 0x3a, 0xa8, 0x84, 0x42, 0x98, 0x12, 0x64, 0x95, 0x85, 0x79, 0x29, + 0xea, 0x3a, 0xf9, 0xa4, 0x5c, 0x9c, 0x35, 0x01, 0x68, 0x71, 0xb9, 0x5b, + 0xbe, 0xaa, 0x76, 0x9e, 0x63, 0x1c, 0xc1, 0x83, 0x94, 0xc6, 0x89, 0x2b, + 0x1d, 0x00, 0x43, 0x74, 0x00, 0x41, 0x93, 0x58, 0x52, 0xf9, 0x13, 0xfe, + 0x9f, 0x7a, 0xb7, 0x3d, 0x6b, 0x70, 0x4e, 0x4f, 0x8f, 0xf4, 0x9c, 0xe4, + 0x97, 0x62, 0xaf, 0x69, 0x45, 0xec, 0xf4, 0x53, 0x71, 0xdc, 0xc7, 0x8d, + 0x6f, 0xb2, 0x9d, 0xec, 0x43, 0xdd, 0xc0, 0xe5, 0xd1, 0x6c, 0x1a, 0x82, + 0x19, 0xf6, 0x18, 0xd3, 0x59, 0x0e, 0x07, 0x81, 0x5a, 0x23, 0x10, 0x8b, + 0xaa, 0x0b, 0x99, 0xc8, 0x34, 0xc2, 0xd0, 0xa9, 0x69, 0x7f, 0x54, 0xe3, + 0xc4, 0xa0, 0xe7, 0x4b, 0x31, 0x90, 0xe7, 0x3b, 0x45, 0x9b, 0x7f, 0xae, + 0xd2, 0xab, 0x22, 0xb9, 0xfc, 0x07, 0x39, 0x4b, 0x45, 0x83, 0x8d, 0x41, + 0x7a, 0x52, 0xb2, 0xae, 0x71, 0x78, 0x17, 0x63, 0xfa, 0xbe, 0x59, 0xca, + 0xf0, 0xfd, 0x68, 0xe5, 0xc4, 0x9a, 0x74, 0x3d, 0xec, 0xd4, 0x8b, 0xa1, + 0x2c, 0x31, 0x4d, 0x73, 0xfd, 0x5c, 0x1e, 0xeb, 0x5f, 0xf6, 0x42, 0x0d, + 0x79, 0x5f, 0x64, 0x10, 0xae, 0xb2, 0xf6, 0x9e, 0xa8, 0xab, 0xa5, 0x2b, + 0x9a, 0xcf, 0x25, 0xfa, 0xa2, 0xb3, 0xdc, 0x30, 0x3d, 0x08, 0x4e, 0xbb, + 0x7b, 0x0c, 0x28, 0x34, 0x9d, 0xda, 0xc4, 0x94, 0xa4, 0xf4, 0x1e, 0x78, + 0x8b, 0xa9, 0xd3, 0xa7, 0x1c, 0x2a, 0x27, 0x14, 0xa0, 0x44, 0x1a, 0x9a, + 0x87, 0x72, 0xa5, 0x6d, 0x69, 0x46, 0xe5, 0xc1, 0x4f, 0x29, 0x87, 0xc0, + 0xa7, 0xa8, 0x96, 0xde, 0xa9, 0x63, 0x08, 0xd8, 0x4a, 0xa1, 0x25, 0x43, + 0x76, 0x41, 0xf7, 0x9f, 0x17, 0xe3, 0xe1, 0x4b, 0xc6, 0x2b, 0x79, 0xea, + 0xd5, 0xa7, 0x72, 0x16, 0x0a, 0x8c, 0xcd, 0x49, 0x70, 0x75, 0xd4, 0x59, + 0x4a, 0x19, 0x7b, 0x31, 0x02, 0x7a, 0x3a, 0x20, 0x15, 0x62, 0x7e, 0x4e, + 0x6f, 0xac, 0xd0, 0xd1, 0x29, 0xbd, 0x2d, 0xa1, 0xc6, 0x3e, 0xa6, 0x1a, + 0x26, 0x18, 0x96, 0x98, 0x12, 0x56, 0x37, 0xbf, 0xb4, 0x91, 0x57, 0xe8, + 0xda, 0x61, 0x7c, 0x2f, 0x3e, 0xd4, 0x51, 0xfe, 0xe8, 0x5b, 0x00, 0x30, + 0x08, 0xf6, 0x4e, 0x69, 0xa8, 0x1a, 0x2b, 0x82, 0x41, 0x85, 0xa9, 0xd9, + 0x3c, 0xc8, 0x02, 0x91, 0x99, 0xd4, 0xa2, 0xfd, 0x9d, 0x1b, 0x08, 0xfc, + 0x41, 0x3e, 0x10, 0x6b, 0x80, 0x74, 0x3d, 0x72, 0x61, 0x97, 0xdd, 0x96, + 0xec, 0xf4, 0xd6, 0x6d, 0x68, 0x02, 0x6e, 0xbb, 0x55, 0x9d, 0x6f, 0x11, + 0xde, 0xd1, 0xad, 0x6d, 0x42, 0x96, 0x2c, 0x42, 0x1e, 0xa9, 0x19, 0x42, + 0x22, 0x38, 0x38, 0x18, 0x3c, 0x4b, 0xc1, 0x9c, 0x0f, 0xe1, 0x34, 0x61, + 0x06, 0x77, 0x54, 0x04, 0xe0, 0x87, 0x94, 0x5c, 0xc9, 0xa1, 0x35, 0x55, + 0x3d, 0x4a, 0xf2, 0x4f, 0x05, 0x11, 0x98, 0x6f, 0x3c, 0x85, 0x84, 0xe6, + 0xf8, 0x71, 0x8a, 0xdf, 0xe9, 0x9a, 0xe3, 0x70, 0xd6, 0x36, 0xd6, 0xc8, + 0x66, 0x3e, 0xba, 0x7c, 0x0a, 0x23, 0x0a, 0xd0, 0xb6, 0x66, 0x68, 0xa8, + 0xdf, 0x37, 0x17, 0xfb, 0xdd, 0x9c, 0x8b, 0xc7, 0x8e, 0xc4, 0x4f, 0x40, + 0x08, 0x23, 0x58, 0x15, 0xa2, 0xba, 0xef, 0xdf, 0x67, 0xcd, 0x1f, 0xb6, + 0xc4, 0xea, 0xce, 0x81, 0x38, 0x58, 0x92, 0x57, 0xcf, 0x83, 0x47, 0x29, + 0x9f, 0xde, 0x9b, 0xde, 0x01, 0xfe, 0x68, 0x91, 0x67, 0x06, 0x9d, 0x31, + 0xd0, 0xb9, 0xc3, 0xbb, 0xc3, 0x6b, 0xa0, 0x04, 0x1e, 0x34, 0xd5, 0x38, + 0xd4, 0xac, 0x70, 0xae, 0xab, 0xb2, 0xbd, 0x4b, 0xa0, 0xad, 0x2b, 0x82, + 0xaf, 0x8c, 0x90, 0x4d, 0xd3, 0xca, 0x71, 0x35, 0x75, 0x89, 0xe5, 0x42, + 0x91, 0x46, 0x8d, 0x18, 0x04, 0x7a, 0xb9, 0xaa, 0x3b, 0xe7, 0x1e, 0x8c, + 0x4e, 0xf9, 0x6e, 0x74, 0xaa, 0x2e, 0x36, 0x86, 0xfb, 0xef, 0x9c, 0xd7, + 0xba, 0x5e, 0x2e, 0x3c, 0x40, 0xce, 0x8b, 0x2b, 0x94, 0x55, 0xf2, 0xd4, + 0x7d, 0xbf, 0x8c, 0x8a, 0xa8, 0x59, 0x84, 0x6f, 0x32, 0x95, 0xc5, 0xcc, + 0xad, 0xee, 0x30, 0x23, 0x7c, 0x54, 0xea, 0x60, 0xb8, 0x88, 0x12, 0x45, + 0x03, 0xbc, 0xe3, 0x92, 0x9f, 0xa8, 0x5b, 0x07, 0x97, 0x53, 0x0d, 0xe1, + 0xe3, 0x3d, 0xdf, 0xf2, 0x2a, 0x12, 0xee, 0xdf, 0x73, 0x8d, 0x41, 0xf4, + 0xe4, 0x2c, 0xb4, 0xd4, 0x9e, 0xfe, 0xf2, 0xe6, 0xa0, 0x9e, 0x2a, 0x3a, + 0x36, 0x26, 0x7e, 0xd9, 0xe1, 0x22, 0xee, 0x0b, 0x5b, 0x48, 0xd2, 0xa9, + 0x55, 0xab, 0x50, 0x7c, 0xf6, 0xc8, 0x56, 0x31, 0xbb, 0x51, 0xe9, 0x31, + 0x4d, 0xaa, 0x13, 0x3a, 0x99, 0x9f, 0x8c, 0x59, 0x6a, 0xc9, 0xf1, 0x0a, + 0x89, 0xcc, 0x39, 0x98, 0xbd, 0xc3, 0x93, 0x97, 0x28, 0xe5, 0x73, 0x94, + 0xf2, 0x0a, 0x7a, 0x09, 0x38, 0x0b, 0xab, 0xd8, 0x49, 0x98, 0x14, 0x34, + 0x32, 0x9d, 0xef, 0x9d, 0x47, 0xdb, 0x82, 0xb9, 0x84, 0xd6, 0xd7, 0x9f, + 0xf7, 0xdf, 0x79, 0x5b, 0xe8, 0x92, 0x44, 0x31, 0x5d, 0x42, 0x80, 0x90, + 0x8d, 0x36, 0xa2, 0x39, 0x02, 0x64, 0x21, 0xa2, 0xb8, 0xfc, 0xff, 0xff, + 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xc8, 0xeb, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0xd8, 0x03, 0x00, 0x00, 0xdc, 0x03, 0x00, 0x00, + 0xe0, 0x03, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0xa8, 0x03, 0x00, 0x00, + 0x50, 0x03, 0x00, 0x00, 0x04, 0x03, 0x00, 0x00, 0xac, 0x02, 0x00, 0x00, + 0x74, 0x02, 0x00, 0x00, 0x2c, 0x02, 0x00, 0x00, 0xf4, 0x01, 0x00, 0x00, + 0xac, 0x01, 0x00, 0x00, 0x74, 0x01, 0x00, 0x00, 0x2c, 0x01, 0x00, 0x00, + 0xe4, 0x00, 0x00, 0x00, 0x9c, 0x00, 0x00, 0x00, 0x64, 0x00, 0x00, 0x00, + 0x28, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x9e, 0xfc, 0xff, 0xff, + 0x0c, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x0d, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x33, 0x00, 0x00, 0x00, 0x5e, 0xfd, 0xff, 0xff, 0x1c, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x08, 0x00, 0x04, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x01, 0x00, 0x00, 0x00, + 0x33, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x32, 0x00, 0x00, 0x00, + 0x96, 0xfd, 0xff, 0xff, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, + 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, + 0x88, 0xfd, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, 0x32, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x2f, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, + 0x31, 0x00, 0x00, 0x00, 0xca, 0xfd, 0xff, 0xff, 0x14, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x06, 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x0a, 0x00, 0x00, 0x00, 0x78, 0xfd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x01, + 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x2f, 0x00, 0x00, 0x00, + 0x05, 0x00, 0x00, 0x00, 0x2a, 0x00, 0x00, 0x00, 0x2b, 0x00, 0x00, 0x00, + 0x2c, 0x00, 0x00, 0x00, 0x2d, 0x00, 0x00, 0x00, 0x2e, 0x00, 0x00, 0x00, + 0x0e, 0xfe, 0xff, 0xff, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, + 0xbc, 0xfd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x2a, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, + 0x25, 0x00, 0x00, 0x00, 0x26, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, + 0x28, 0x00, 0x00, 0x00, 0x29, 0x00, 0x00, 0x00, 0x52, 0xfe, 0xff, 0xff, + 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0xfe, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x25, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x21, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, + 0x24, 0x00, 0x00, 0x00, 0x96, 0xfe, 0xff, 0xff, 0x14, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x08, 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x07, 0x00, 0x00, 0x00, 0x88, 0xfe, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x1d, 0x00, 0x00, 0x00, + 0x1e, 0x00, 0x00, 0x00, 0x1f, 0x00, 0x00, 0x00, 0xca, 0xfe, 0xff, 0xff, + 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x78, 0xfe, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x1d, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x19, 0x00, 0x00, 0x00, 0x1a, 0x00, 0x00, 0x00, 0x1b, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x0e, 0xff, 0xff, 0xff, 0x14, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x08, 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x05, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x15, 0x00, 0x00, 0x00, + 0x16, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, 0x42, 0xff, 0xff, 0xff, + 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xf0, 0xfe, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x15, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x11, 0x00, 0x00, 0x00, 0x12, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x86, 0xff, 0xff, 0xff, 0x14, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x08, 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x78, 0xff, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x0d, 0x00, 0x00, 0x00, + 0x0e, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0xba, 0xff, 0xff, 0xff, + 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x68, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0d, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, + 0x09, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x14, 0x00, + 0x10, 0x00, 0x0c, 0x00, 0x0b, 0x00, 0x04, 0x00, 0x0e, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x14, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x04, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x07, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x0c, 0x00, 0x0b, 0x00, 0x04, 0x00, 0x0e, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x08, 0x00, 0x07, 0x00, + 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0a, 0x00, + 0x10, 0x00, 0x0c, 0x00, 0x08, 0x00, 0x04, 0x00, 0x0a, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x0d, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x34, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x34, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, + 0x34, 0x10, 0x00, 0x00, 0xd4, 0x0f, 0x00, 0x00, 0x7c, 0x0f, 0x00, 0x00, + 0x38, 0x0f, 0x00, 0x00, 0xdc, 0x0e, 0x00, 0x00, 0x74, 0x0e, 0x00, 0x00, + 0x2c, 0x0e, 0x00, 0x00, 0xe8, 0x0d, 0x00, 0x00, 0x90, 0x0d, 0x00, 0x00, + 0x48, 0x0d, 0x00, 0x00, 0x04, 0x0d, 0x00, 0x00, 0xc0, 0x0c, 0x00, 0x00, + 0x64, 0x0c, 0x00, 0x00, 0x0c, 0x0c, 0x00, 0x00, 0xc4, 0x0b, 0x00, 0x00, + 0x80, 0x0b, 0x00, 0x00, 0x28, 0x0b, 0x00, 0x00, 0xe0, 0x0a, 0x00, 0x00, + 0x9c, 0x0a, 0x00, 0x00, 0x58, 0x0a, 0x00, 0x00, 0xfc, 0x09, 0x00, 0x00, + 0xa4, 0x09, 0x00, 0x00, 0x5c, 0x09, 0x00, 0x00, 0x18, 0x09, 0x00, 0x00, + 0xc0, 0x08, 0x00, 0x00, 0x78, 0x08, 0x00, 0x00, 0x34, 0x08, 0x00, 0x00, + 0xf0, 0x07, 0x00, 0x00, 0x94, 0x07, 0x00, 0x00, 0x3c, 0x07, 0x00, 0x00, + 0xf4, 0x06, 0x00, 0x00, 0xb0, 0x06, 0x00, 0x00, 0x58, 0x06, 0x00, 0x00, + 0x10, 0x06, 0x00, 0x00, 0xcc, 0x05, 0x00, 0x00, 0x88, 0x05, 0x00, 0x00, + 0x2c, 0x05, 0x00, 0x00, 0xd4, 0x04, 0x00, 0x00, 0x8c, 0x04, 0x00, 0x00, + 0x48, 0x04, 0x00, 0x00, 0x04, 0x04, 0x00, 0x00, 0xa8, 0x03, 0x00, 0x00, + 0x50, 0x03, 0x00, 0x00, 0x08, 0x03, 0x00, 0x00, 0xc4, 0x02, 0x00, 0x00, + 0x80, 0x02, 0x00, 0x00, 0x24, 0x02, 0x00, 0x00, 0xcc, 0x01, 0x00, 0x00, + 0x84, 0x01, 0x00, 0x00, 0x40, 0x01, 0x00, 0x00, 0xe8, 0x00, 0x00, 0x00, + 0x8c, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0xb2, 0xf0, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, + 0x28, 0x00, 0x00, 0x00, 0xfc, 0xf0, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x01, 0x00, 0x38, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0xf2, 0xf0, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x07, 0x28, 0x00, 0x00, 0x00, 0x3c, 0xf1, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x0a, 0xd7, 0x23, 0x3a, 0x02, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, 0x32, 0xf1, 0xff, 0xff, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, 0x40, 0x00, 0x00, 0x00, + 0x24, 0xf1, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x80, 0x00, 0x80, 0x37, 0x01, 0x00, 0x00, 0x00, + 0xc2, 0xff, 0x7f, 0x3f, 0x01, 0x00, 0x00, 0x00, 0xd2, 0x6f, 0x75, 0x36, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x8a, 0xf1, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x3c, 0x00, 0x00, 0x00, 0x7c, 0xf1, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x06, 0x16, 0x49, 0x3d, 0x01, 0x00, 0x00, 0x00, + 0x87, 0x19, 0xb1, 0x40, 0x01, 0x00, 0x00, 0x00, 0x58, 0x80, 0xdf, 0xc0, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x3a, 0xf2, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x1f, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, 0x2c, 0xf2, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x5d, 0xd1, 0xce, 0x39, 0x01, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x7a, 0xf2, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x1e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, + 0x6c, 0xf2, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x23, 0x20, 0xb6, 0x3b, + 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x62, 0xf2, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x3c, 0x00, 0x00, 0x00, 0x54, 0xf2, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0x01, 0x00, 0x00, 0x00, 0xa2, 0x5a, 0x91, 0x3d, 0x01, 0x00, 0x00, 0x00, + 0x47, 0xc9, 0x90, 0x41, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x00, 0xf4, 0xff, 0xff, 0x00, 0x00, 0x00, 0x01, 0x0c, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x07, 0x3c, 0x00, 0x00, 0x00, 0xac, 0xf2, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x81, 0xb7, 0xf1, 0x39, + 0x01, 0x00, 0x00, 0x00, 0x9e, 0xb5, 0x71, 0x41, 0x01, 0x00, 0x00, 0x00, + 0x33, 0x20, 0x70, 0xc1, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x04, 0x00, 0x00, 0x6a, 0xf3, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x1d, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, + 0x5c, 0xf3, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x7a, 0x08, 0x97, 0x35, + 0x01, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0xaa, 0xf3, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, + 0x24, 0x00, 0x00, 0x00, 0x9c, 0xf3, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x2f, 0xf5, 0x1f, 0x3b, + 0x02, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0xea, 0xf3, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x1b, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, 0xdc, 0xf3, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xc7, 0xea, 0x1a, 0x3c, 0x02, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0xd2, 0xf3, 0xff, 0xff, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x3c, 0x00, 0x00, 0x00, + 0xc4, 0xf3, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, + 0xb2, 0x78, 0x3f, 0x3d, 0x01, 0x00, 0x00, 0x00, 0x39, 0xb9, 0x3e, 0x41, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x70, 0xf5, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x01, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, + 0x3c, 0x00, 0x00, 0x00, 0x1c, 0xf4, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x89, 0x25, 0xf2, 0x39, 0x01, 0x00, 0x00, 0x00, + 0xde, 0xdc, 0x1d, 0x41, 0x01, 0x00, 0x00, 0x00, 0xa5, 0x23, 0x72, 0xc1, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, + 0xda, 0xf4, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x1a, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, 0xcc, 0xf4, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x42, 0xe0, 0x90, 0x35, 0x01, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x1a, 0xf5, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x19, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, 0x24, 0x00, 0x00, 0x00, + 0x0c, 0xf5, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x1a, 0x2a, 0x19, 0x3b, 0x02, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x5a, 0xf5, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x28, 0x00, 0x00, 0x00, 0x4c, 0xf5, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0xe9, 0x36, 0xdd, 0x3b, 0x02, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x42, 0xf5, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x3c, 0x00, 0x00, 0x00, 0x34, 0xf5, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x80, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, 0xdd, 0x43, 0x7e, 0x3d, + 0x01, 0x00, 0x00, 0x00, 0x99, 0x45, 0x7d, 0x41, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0xe0, 0xf6, 0xff, 0xff, 0x00, 0x00, 0x00, 0x01, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, 0x3c, 0x00, 0x00, 0x00, + 0x8c, 0xf5, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x5c, 0xfd, 0xa9, 0x39, 0x01, 0x00, 0x00, 0x00, 0x1e, 0xaa, 0x87, 0x40, + 0x01, 0x00, 0x00, 0x00, 0x08, 0xfc, 0x29, 0xc1, 0x02, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x4a, 0xf6, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, + 0x28, 0x00, 0x00, 0x00, 0x3c, 0xf6, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x55, 0xf7, 0x52, 0x35, 0x01, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x8a, 0xf6, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x07, 0x24, 0x00, 0x00, 0x00, 0x7c, 0xf6, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0xd0, 0xda, 0x1e, 0x3b, 0x02, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0xca, 0xf6, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x15, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, + 0xbc, 0xf6, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x8e, 0x0b, 0xa8, 0x3b, + 0x02, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0xb2, 0xf6, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x3c, 0x00, 0x00, 0x00, 0xa4, 0xf6, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xf5, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0x01, 0x00, 0x00, 0x00, 0x12, 0x1c, 0x6e, 0x3d, 0x01, 0x00, 0x00, 0x00, + 0xdd, 0x4a, 0x00, 0x41, 0x01, 0x00, 0x00, 0x00, 0x31, 0xc6, 0xd9, 0xc0, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x62, 0xf7, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, 0x54, 0xf7, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x80, 0x9d, 0x16, 0x39, 0x01, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0xa2, 0xf7, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x13, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, + 0x94, 0xf7, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0xa4, 0x34, 0xab, 0x3b, + 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x8a, 0xf7, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x3c, 0x00, 0x00, 0x00, 0x7c, 0xf7, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0x01, 0x00, 0x00, 0x00, 0x2e, 0x36, 0xe1, 0x3c, 0x01, 0x00, 0x00, 0x00, + 0xf8, 0x54, 0xe0, 0x40, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x28, 0xf9, 0xff, 0xff, 0x00, 0x00, 0x00, 0x01, 0x0c, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x07, 0x3c, 0x00, 0x00, 0x00, 0xd4, 0xf7, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0xe1, 0xd0, 0xa2, 0x39, + 0x01, 0x00, 0x00, 0x00, 0x9b, 0xcf, 0x22, 0x41, 0x01, 0x00, 0x00, 0x00, + 0xea, 0x23, 0x12, 0xc1, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x02, 0x00, 0x00, 0x92, 0xf8, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x12, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, + 0x84, 0xf8, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x99, 0xd3, 0xf7, 0x34, + 0x01, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0xd2, 0xf8, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, + 0x24, 0x00, 0x00, 0x00, 0xc4, 0xf8, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10, 0xd5, 0xc2, 0x3a, + 0x02, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, + 0x12, 0xf9, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, 0x04, 0xf9, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x8f, 0x84, 0xa2, 0x3b, 0x02, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0xfa, 0xf8, 0xff, 0xff, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x3c, 0x00, 0x00, 0x00, + 0xec, 0xf8, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0xf7, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, + 0x64, 0xeb, 0x8e, 0x3d, 0x01, 0x00, 0x00, 0x00, 0x3b, 0xf3, 0x17, 0x41, + 0x01, 0x00, 0x00, 0x00, 0xb7, 0xc5, 0x04, 0xc1, 0x02, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0xaa, 0xf9, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, + 0x28, 0x00, 0x00, 0x00, 0x9c, 0xf9, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x92, 0xa8, 0x98, 0x39, 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0xea, 0xf9, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, 0xdc, 0xf9, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x58, 0x76, 0xb9, 0x3b, 0x02, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0xd2, 0xf9, 0xff, 0xff, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x3c, 0x00, 0x00, 0x00, + 0xc4, 0xf9, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, + 0x43, 0xb8, 0x52, 0x3d, 0x01, 0x00, 0x00, 0x00, 0x8b, 0xe5, 0x51, 0x41, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x70, 0xfb, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x01, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, + 0x3c, 0x00, 0x00, 0x00, 0x1c, 0xfa, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xe3, 0xa1, 0xf0, 0x39, 0x01, 0x00, 0x00, 0x00, + 0x02, 0xa0, 0x70, 0x41, 0x01, 0x00, 0x00, 0x00, 0x87, 0x08, 0x65, 0xc1, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, + 0xda, 0xfa, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x0d, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, 0xcc, 0xfa, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xcc, 0x98, 0x41, 0x35, 0x01, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x1a, 0xfb, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, 0x24, 0x00, 0x00, 0x00, + 0x0c, 0xfb, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xed, 0xf5, 0xcd, 0x3a, 0x02, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x5a, 0xfb, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x28, 0x00, 0x00, 0x00, 0x4c, 0xfb, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x9d, 0xca, 0xd4, 0x3b, 0x02, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x42, 0xfb, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x3c, 0x00, 0x00, 0x00, 0x34, 0xfb, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x58, 0x58, 0xce, 0x3d, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x49, 0x41, 0x01, 0x00, 0x00, 0x00, + 0x01, 0x06, 0x52, 0xc1, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0xf2, 0xfb, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x0a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, + 0xe4, 0xfb, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x9b, 0x9c, 0xe1, 0x39, + 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x32, 0xfc, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x28, 0x00, 0x00, 0x00, 0x24, 0xfc, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0xf8, 0xb6, 0xc3, 0x3b, 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x1a, 0xfc, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x3c, 0x00, 0x00, 0x00, 0x0c, 0xfc, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x80, 0xff, 0xff, 0xff, + 0xff, 0xff, 0xff, 0xff, 0x01, 0x00, 0x00, 0x00, 0x94, 0x8d, 0x93, 0x3d, + 0x01, 0x00, 0x00, 0x00, 0x06, 0xfa, 0x92, 0x41, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0xb8, 0xfd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x01, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, 0x3c, 0x00, 0x00, 0x00, + 0x64, 0xfc, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x7a, 0xf6, 0x5f, 0x3a, 0x01, 0x00, 0x00, 0x00, 0xba, 0xf4, 0xdf, 0x41, + 0x01, 0x00, 0x00, 0x00, 0xf4, 0x7c, 0xcf, 0xc1, 0x02, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x22, 0xfd, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, + 0x28, 0x00, 0x00, 0x00, 0x14, 0xfd, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x46, 0x2f, 0xc4, 0x35, 0x01, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x62, 0xfd, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x07, 0x24, 0x00, 0x00, 0x00, 0x54, 0xfd, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x8f, 0x3f, 0xe0, 0x3a, 0x02, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x00, 0x00, 0xa2, 0xfd, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, + 0x94, 0xfd, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x25, 0xd7, 0xa9, 0x3b, + 0x02, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x8a, 0xfd, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x3c, 0x00, 0x00, 0x00, 0x7c, 0xfd, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xe3, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0x01, 0x00, 0x00, 0x00, 0xc4, 0xf4, 0x39, 0x3e, 0x01, 0x00, 0x00, 0x00, + 0xf4, 0x1f, 0xe3, 0x41, 0x01, 0x00, 0x00, 0x00, 0xaa, 0x55, 0x8f, 0xc1, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x3a, 0xfe, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x02, 0x28, 0x00, 0x00, 0x00, 0x2c, 0xfe, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x8b, 0x00, 0x4b, 0x3a, 0x01, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x7a, 0xfe, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x28, 0x00, 0x00, 0x00, + 0x6c, 0xfe, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0xd7, 0xdf, 0xc3, 0x3b, + 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x62, 0xfe, 0xff, 0xff, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x3c, 0x00, 0x00, 0x00, 0x54, 0xfe, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0x01, 0x00, 0x00, 0x00, 0x68, 0xa8, 0x04, 0x3e, 0x01, 0x00, 0x00, 0x00, + 0xc0, 0x23, 0x04, 0x42, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x14, 0x00, 0x10, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x07, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x07, 0x3c, 0x00, 0x00, 0x00, + 0xbc, 0xfe, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x1c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x3b, 0xda, 0x75, 0x3b, 0x01, 0x00, 0x00, 0x00, 0x4f, 0xd8, 0xf5, 0x42, + 0x01, 0x00, 0x00, 0x00, 0xa8, 0x2a, 0x61, 0xc2, 0x02, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x7a, 0xff, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, + 0x28, 0x00, 0x00, 0x00, 0x6c, 0xff, 0xff, 0xff, 0x08, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0xcf, 0x37, 0x69, 0x37, 0x01, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0xba, 0xff, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x07, 0x28, 0x00, 0x00, 0x00, 0xac, 0xff, 0xff, 0xff, + 0x08, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x14, 0xd8, 0x72, 0x3b, 0x02, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, + 0x14, 0x00, 0x10, 0x00, 0x0f, 0x00, 0x08, 0x00, 0x00, 0x00, 0x04, 0x00, + 0x0e, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x09, 0x30, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x0c, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x04, 0x00, 0x0c, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0xd4, 0x42, 0x16, 0x3c, 0x02, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x60, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x10, 0x00, 0x0c, 0x00, + 0x0b, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x0e, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, 0x4c, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x14, 0x00, 0x10, 0x00, 0x0c, 0x00, 0x08, 0x00, 0x04, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xa8, 0x41, 0x5b, 0x3d, 0x01, 0x00, 0x00, 0x00, + 0x66, 0x66, 0x5a, 0x41, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, + 0x0f, 0x00, 0x00, 0x00, 0xc4, 0x00, 0x00, 0x00, 0xb4, 0x00, 0x00, 0x00, + 0xa4, 0x00, 0x00, 0x00, 0x98, 0x00, 0x00, 0x00, 0x8c, 0x00, 0x00, 0x00, + 0x80, 0x00, 0x00, 0x00, 0x74, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, + 0x5c, 0x00, 0x00, 0x00, 0x50, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00, + 0x38, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x0c, 0x00, 0x0b, 0x00, + 0x00, 0x00, 0x04, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x06, 0x96, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x72, + 0x9e, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x19, 0xa6, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x09, 0xae, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x1b, + 0xb6, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x1b, 0xbe, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x1b, 0xc6, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, + 0xce, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x1b, 0xd6, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x09, 0xde, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x1b, + 0xe6, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, 0xfa, 0xff, 0xff, 0xff, + 0x00, 0x1b, 0x06, 0x00, 0x06, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x00, 0x09, 0x06, 0x00, 0x08, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x1b}; + +const unsigned int g_keyword_scrambled_model_data_length = 34520; + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.h b/src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.h new file mode 100644 index 0000000..3db2849 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.h @@ -0,0 +1,25 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_BENCHMARKS_KEYWORD_SCRAMBLED_MODEL_DATA_H_ +#define TENSORFLOW_LITE_MICRO_BENCHMARKS_KEYWORD_SCRAMBLED_MODEL_DATA_H_ + +extern const unsigned char g_keyword_scrambled_model_data[]; +extern const unsigned int g_keyword_scrambled_model_data_length; + +#endif // TENSORFLOW_LITE_MICRO_BENCHMARKS_KEYWORD_SCRAMBLED_MODEL_DATA_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/compatibility.h b/src/tensorflow_arm/tensorflow/lite/micro/compatibility.h similarity index 96% rename from src/tensorflow/lite/micro/compatibility.h rename to src/tensorflow_arm/tensorflow/lite/micro/compatibility.h index 49acb28..0fef69b 100644 --- a/src/tensorflow/lite/micro/compatibility.h +++ b/src/tensorflow_arm/tensorflow/lite/micro/compatibility.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -30,3 +31,5 @@ limitations under the License. #endif #endif // TENSORFLOW_LITE_MICRO_COMPATIBILITY_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/debug_log.h b/src/tensorflow_arm/tensorflow/lite/micro/debug_log.h similarity index 83% rename from src/tensorflow/lite/micro/debug_log.h rename to src/tensorflow_arm/tensorflow/lite/micro/debug_log.h index 1004ab9..460603f 100644 --- a/src/tensorflow/lite/micro/debug_log.h +++ b/src/tensorflow_arm/tensorflow/lite/micro/debug_log.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,19 @@ limitations under the License. #ifndef TENSORFLOW_LITE_MICRO_DEBUG_LOG_H_ #define TENSORFLOW_LITE_MICRO_DEBUG_LOG_H_ +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus + // This function should be implemented by each target platform, and provide a // way for strings to be output to some text stream. For more information, see // tensorflow/lite/micro/debug_log.cc. -extern "C" void DebugLog(const char* s); +void DebugLog(const char* s); + +#ifdef __cplusplus +} // extern "C" +#endif // __cplusplus #endif // TENSORFLOW_LITE_MICRO_DEBUG_LOG_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/activation_utils.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/activation_utils.h new file mode 100644 index 0000000..c9c562c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/activation_utils.h @@ -0,0 +1,60 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_ACTIVATION_UTILS_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_ACTIVATION_UTILS_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/cppmath.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/max.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/min.h" + +namespace tflite { +namespace ops { +namespace micro { + +// Returns the floating point value for a fused activation: +inline float ActivationValFloat(TfLiteFusedActivation act, float a) { + switch (act) { + case kTfLiteActNone: + return a; + case kTfLiteActRelu: + return TfLiteMax(0.0f, a); + case kTfLiteActReluN1To1: + return TfLiteMax(-1.0f, TfLiteMin(a, 1.0f)); + case kTfLiteActRelu6: + return TfLiteMax(0.0f, TfLiteMin(a, 6.0f)); + case kTfLiteActTanh: + return std::tanh(a); + case kTfLiteActSignBit: + return std::signbit(a); + case kTfLiteActSigmoid: + return 1.0f / (1.0f + std::exp(-a)); + } + return 0.0f; // To indicate an unsupported activation (i.e. when a new fused + // activation is added to the enum and not handled here). +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_ACTIVATION_UTILS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/activations.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/activations.cpp new file mode 100644 index 0000000..3d3d77f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/activations.cpp @@ -0,0 +1,291 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace activations { +namespace { + +struct ReluOpData { + ReluParams params; +}; + +struct Relu6OpData { + int8_t six_int8; + int8_t zero_int8; + uint8_t six_uint8; + uint8_t zero_uint8; +}; + +} // namespace + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +template +inline void ReluQuantized(const ReluOpData& data, + const RuntimeShape& input_shape, + const RuntimeShape& output_shape, const T* input_data, + T* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + const int32_t val = static_cast(input_data[i]); + int32_t clamped = + data.params.output_offset + + MultiplyByQuantizedMultiplier(val - data.params.input_offset, + data.params.output_multiplier, + data.params.output_shift); + clamped = std::max(data.params.quantized_activation_min, clamped); + clamped = std::min(data.params.quantized_activation_max, clamped); + output_data[i] = static_cast(clamped); + } +} + +template +inline void CalculateReluOpData(const TfLiteTensor* input, TfLiteTensor* output, + ReluOpData* data) { + float act_min = 0.0; + float act_max = std::numeric_limits::infinity(); + double real_multiplier = + static_cast(input->params.scale / output->params.scale); + + const RuntimeShape input_shape = GetTensorShape(input); + const RuntimeShape output_shape = GetTensorShape(output); + + QuantizeMultiplier(real_multiplier, &data->params.output_multiplier, + &data->params.output_shift); + + data->params.quantized_activation_min = std::max( + static_cast(std::numeric_limits::min()), + output->params.zero_point + + static_cast(roundf(act_min / output->params.scale))); + data->params.quantized_activation_max = + act_max == std::numeric_limits::infinity() + ? static_cast(std::numeric_limits::max()) + : std::min(static_cast(std::numeric_limits::max()), + output->params.zero_point + + static_cast( + roundf(act_max / output->params.scale))); + data->params.input_offset = input->params.zero_point; + data->params.output_offset = output->params.zero_point; +} + +inline void ReluFloat(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + const float val = input_data[i]; + const float lower = 0.0f; + const float clamped = val < lower ? lower : val; + output_data[i] = clamped; + } +} + +inline void Relu6Float(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + const float val = input_data[i]; + const float upper = 6.0f; + const float lower = 0.0f; + const float clamped = val > upper ? upper : val < lower ? lower : val; + output_data[i] = clamped; + } +} + +template +inline void Relu6Quantized(Q lower, Q upper, const RuntimeShape& input_shape, + const Q* input_data, + const RuntimeShape& output_shape, Q* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + const Q val = input_data[i]; + const Q clamped = val > upper ? upper : val < lower ? lower : val; + output_data[i] = clamped; + } +} + +void* ReluInit(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(ReluOpData)); +} + +TfLiteStatus ReluPrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + ReluOpData* data = static_cast(node->user_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + + if (input->type == kTfLiteInt8) { + CalculateReluOpData(input, output, data); + } else if (input->type == kTfLiteUInt8) { + CalculateReluOpData(input, output, data); + } + + return kTfLiteOk; +} + +TfLiteStatus ReluEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const ReluOpData& data = *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + switch (input->type) { + case kTfLiteFloat32: { + ReluFloat(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + + return kTfLiteOk; + } + case kTfLiteInt8: { + ReluQuantized(data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } + case kTfLiteUInt8: { + ReluQuantized(data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } + default: { + TF_LITE_KERNEL_LOG(context, "Only float32 is supported currently, got %s", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + } +} + +void* Relu6Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(Relu6OpData)); +} + +TfLiteStatus Relu6Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + Relu6OpData* data = static_cast(node->user_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + + if (input->type == kTfLiteInt8) { + data->six_int8 = FloatToQuantizedType(6.0f, input->params.scale, + input->params.zero_point); + data->zero_int8 = input->params.zero_point; + } else if (input->type == kTfLiteUInt8) { + data->six_uint8 = FloatToQuantizedType(6.0f, input->params.scale, + input->params.zero_point); + data->zero_uint8 = input->params.zero_point; + } + + return kTfLiteOk; +} + +TfLiteStatus Relu6Eval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const Relu6OpData& data = *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + switch (input->type) { + case kTfLiteFloat32: { + Relu6Float(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + + return kTfLiteOk; + } + case kTfLiteInt8: { + Relu6Quantized(data.zero_int8, data.six_int8, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } + case kTfLiteUInt8: { + Relu6Quantized(data.zero_uint8, data.six_uint8, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } + default: { + TF_LITE_KERNEL_LOG(context, "Only float32 is supported currently, got %s", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + } +} + +} // namespace activations + +TfLiteRegistration Register_RELU() { + return {/*init=*/activations::ReluInit, + /*free=*/nullptr, + /*prepare=*/activations::ReluPrepare, + /*invoke=*/activations::ReluEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_RELU6() { + return {/*init=*/activations::Relu6Init, + /*free=*/nullptr, + /*prepare=*/activations::Relu6Prepare, + /*invoke=*/activations::Relu6Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/arg_min_max.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/arg_min_max.cpp new file mode 100644 index 0000000..8a67877 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/arg_min_max.cpp @@ -0,0 +1,136 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/arg_min_max.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/micro_utils.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace arg_min_max { + +constexpr int kInputTensor = 0; +constexpr int kAxis = 1; +constexpr int kOutputTensor = 0; + +template +inline void ArgMinMaxHelper(const RuntimeShape& input1_shape, + const T1* input1_data, const T3* input2_data, + const RuntimeShape& output_shape, T2* output_data, + bool is_arg_max) { + if (is_arg_max) { + reference_ops::ArgMinMax(input1_shape, input1_data, input2_data, + output_shape, output_data, micro::Greater()); + } else { + reference_ops::ArgMinMax(input1_shape, input1_data, input2_data, + output_shape, output_data, micro::Less()); + } +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node, bool is_arg_max) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* axis = + tflite::micro::GetEvalInput(context, node, kAxis); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + +#define TF_LITE_ARG_MIN_MAX(data_type, axis_type, output_type) \ + ArgMinMaxHelper(tflite::micro::GetTensorShape(input), \ + tflite::micro::GetTensorData(input), \ + tflite::micro::GetTensorData(axis), \ + tflite::micro::GetTensorShape(output), \ + tflite::micro::GetTensorData(output), \ + is_arg_max) + if (axis->type == kTfLiteInt32) { + if (output->type == kTfLiteInt32) { + switch (input->type) { + case kTfLiteFloat32: + TF_LITE_ARG_MIN_MAX(float, int32_t, int32_t); + break; + case kTfLiteUInt8: + TF_LITE_ARG_MIN_MAX(uint8_t, int32_t, int32_t); + break; + case kTfLiteInt8: + TF_LITE_ARG_MIN_MAX(int8_t, int32_t, int32_t); + break; + default: + TF_LITE_KERNEL_LOG(context, + "Only float32, uint8_t and int8_t are " + "supported currently, got %s.", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + } else { + TF_LITE_KERNEL_LOG(context, + "Only int32_t are supported currently, got %s.", + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else { + TF_LITE_KERNEL_LOG(context, "Only int32_t are supported currently, got %s.", + TfLiteTypeGetName(axis->type)); + return kTfLiteError; + } + +#undef TF_LITE_ARG_MIN_MAX + + return kTfLiteOk; +} + +TfLiteStatus ArgMinEval(TfLiteContext* context, TfLiteNode* node) { + return Eval(context, node, false); +} + +TfLiteStatus ArgMaxEval(TfLiteContext* context, TfLiteNode* node) { + return Eval(context, node, true); +} + +} // namespace arg_min_max + +TfLiteRegistration Register_ARG_MAX() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/arg_min_max::ArgMaxEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_ARG_MIN() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/arg_min_max::ArgMinEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/ceil.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/ceil.cpp new file mode 100644 index 0000000..af9328e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/ceil.cpp @@ -0,0 +1,79 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/ceil.h" + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace ceil { + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteFloat32); + TF_LITE_ENSURE_TYPES_EQ(context, output->type, input->type); + TF_LITE_ENSURE_EQ(context, output->bytes, input->bytes); + TF_LITE_ENSURE_EQ(context, output->dims->size, input->dims->size); + for (int i = 0; i < output->dims->size; ++i) { + TF_LITE_ENSURE_EQ(context, output->dims->data[i], input->dims->data[i]); + } + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + reference_ops::Ceil(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + + return kTfLiteOk; +} +} // namespace ceil + +TfLiteRegistration Register_CEIL() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ceil::Prepare, + /*invoke=*/ceil::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/circular_buffer.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/circular_buffer.cpp new file mode 100644 index 0000000..186add9 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/circular_buffer.cpp @@ -0,0 +1,180 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +/* + * The circular buffer custom operator is used to implement strided streaming + * convolutions on TFLite Micro. Each time this operator is invoked, it checks + * whether or not to run, based on a predetermined stride in time. If the op + * runs, it inserts the input into the end of the output buffer and shifts the + * output values towards the start of the buffer. It discards the oldest value + * in the output buffer. + * + * Input: [, , , ] + * + * After shifting: + * Output: [, , , ] + * + * We make some assumptions in this custom operator: + * - Input shape must be [1, 1, 1, depth] + * - Output shape must be [1, num_slots, 1, depth] + * - Input and output types must match. + * - Input and output quantization params must be identical. + */ +namespace tflite { +namespace ops { +namespace micro { +namespace circular_buffer { + +namespace { + +// The CircularBuffer op has one input and one output tensor. +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +// TODO(b/149795762): Add this to TfLiteStatus enum. +constexpr int kTfLiteAbort = -9; + +// These fields control the stride period of a strided streaming model. This op +// returns kTfLiteAbort until cycles_until_run-- is zero. At this time, +// cycles_until_run is reset to cycles_max. +struct OpData { + int cycles_until_run; + int cycles_max; +}; + +} // namespace + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + OpData* op_data = static_cast(node->user_data); + + TF_LITE_ENSURE(context, input != nullptr); + TF_LITE_ENSURE(context, output != nullptr); + TF_LITE_ENSURE_EQ(context, input->dims->data[0], output->dims->data[0]); + TF_LITE_ENSURE_EQ(context, 1, input->dims->data[1]); + TF_LITE_ENSURE_EQ(context, input->dims->data[2], output->dims->data[2]); + TF_LITE_ENSURE_EQ(context, output->dims->data[3], input->dims->data[3]); + + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + + // The circular buffer custom operator currently only supports int8. + TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteInt8); + + // The last circular buffer layer simply accumulates outputs, and does not run + // periodically. + // TODO(b/150001379): Move this special case logic to the tflite flatbuffer. + static int cb_prepare_count = 0; + cb_prepare_count++; + // These checks specifically work for the only two streaming models supported + // on TFLM. They use the shape of the output tensor along with the layer + // number to determine if the circular buffer period should be 1 or 2. + + // These models are outlined int the following documents: + // https://docs.google.com/document/d/1lc_G2ZFhjiKFo02UHjBaljye1xsL0EkfybkaVELEE3Q/edit?usp=sharing + // https://docs.google.com/document/d/1pGc42PuWyrk-Jy1-9qeqtggvsmHr1ifz8Lmqfpr2rKA/edit?usp=sharing + if (output->dims->data[1] == 5 || output->dims->data[1] == 13 || + (cb_prepare_count == 5 && output->dims->data[2] == 2 && + output->dims->data[3] == 96)) { + op_data->cycles_max = 1; + cb_prepare_count = 0; + } else { + op_data->cycles_max = 2; + } + op_data->cycles_until_run = op_data->cycles_max; + node->user_data = op_data; + + return kTfLiteOk; +} + +// Shifts buffer over by the output depth, and write new input to end of buffer. +// num_slots is the number of samples stored in the output buffer. +// depth is the size of each sample. +void EvalInt8(const int8_t* input, int num_slots, int depth, int8_t* output) { + memmove(output, &output[depth], (num_slots - 1) * depth); + memcpy(&output[(num_slots - 1) * depth], input, depth); +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = reinterpret_cast(node->user_data); + + int num_slots = output->dims->data[1]; + int depth = output->dims->data[2] * output->dims->data[3]; + + if (input->type == kTfLiteInt8) { + EvalInt8(tflite::micro::GetTensorData(input), num_slots, depth, + tflite::micro::GetTensorData(output)); + } else { + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + + if (--data->cycles_until_run != 0) { + // Signal the interpreter to end current run if the delay before op invoke + // has not been reached. + // TODO(b/149795762): Add kTfLiteAbort to TfLiteStatus enum. + return static_cast(kTfLiteAbort); + } + + data->cycles_until_run = data->cycles_max; + + return kTfLiteOk; +} + +} // namespace circular_buffer + +TfLiteRegistration* Register_CIRCULAR_BUFFER() { + static TfLiteRegistration r = {/*init=*/circular_buffer::Init, + /*free=*/nullptr, + /*prepare=*/circular_buffer::Prepare, + /*invoke=*/circular_buffer::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; + return &r; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/add.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/add.cpp new file mode 100644 index 0000000..6eade35 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/add.cpp @@ -0,0 +1,249 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/add.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/add.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace add { + +constexpr int kInputTensor1 = 0; +constexpr int kInputTensor2 = 1; +constexpr int kOutputTensor = 0; + +struct OpData { + bool requires_broadcast; + + // These fields are used in both the general 8-bit -> 8bit quantized path, + // and the special 16-bit -> 16bit quantized path + int input1_shift; + int input2_shift; + int32_t output_activation_min; + int32_t output_activation_max; + + // These fields are used only in the general 8-bit -> 8bit quantized path + int32_t input1_multiplier; + int32_t input2_multiplier; + int32_t output_multiplier; + int output_shift; + int left_shift; + int32_t input1_offset; + int32_t input2_offset; + int32_t output_offset; +}; + +TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteAddParams* params, + const TfLiteTensor* input1, + const TfLiteTensor* input2, TfLiteTensor* output, + OpData* data) { + data->requires_broadcast = !HaveSameShapes(input1, input2); + + if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + // 8bit -> 8bit general quantized path, with general rescalings + data->input1_offset = -input1->params.zero_point; + data->input2_offset = -input2->params.zero_point; + data->output_offset = output->params.zero_point; + data->left_shift = 20; + const double twice_max_input_scale = + 2 * static_cast( + std::max(input1->params.scale, input2->params.scale)); + const double real_input1_multiplier = + static_cast(input1->params.scale) / twice_max_input_scale; + const double real_input2_multiplier = + static_cast(input2->params.scale) / twice_max_input_scale; + const double real_output_multiplier = + twice_max_input_scale / + ((1 << data->left_shift) * static_cast(output->params.scale)); + + QuantizeMultiplierSmallerThanOneExp( + real_input1_multiplier, &data->input1_multiplier, &data->input1_shift); + + QuantizeMultiplierSmallerThanOneExp( + real_input2_multiplier, &data->input2_multiplier, &data->input2_shift); + + QuantizeMultiplierSmallerThanOneExp( + real_output_multiplier, &data->output_multiplier, &data->output_shift); + + TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( + context, params->activation, output, &data->output_activation_min, + &data->output_activation_max)); + } + + return kTfLiteOk; +} + +void EvalAdd(TfLiteContext* context, TfLiteNode* node, TfLiteAddParams* params, + const OpData* data, const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) { + float output_activation_min, output_activation_max; + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); + tflite::ArithmeticParams op_params; + SetActivationParams(output_activation_min, output_activation_max, &op_params); +#define TF_LITE_ADD(opname) \ + reference_ops::opname(op_params, tflite::micro::GetTensorShape(input1), \ + tflite::micro::GetTensorData(input1), \ + tflite::micro::GetTensorShape(input2), \ + tflite::micro::GetTensorData(input2), \ + tflite::micro::GetTensorShape(output), \ + tflite::micro::GetTensorData(output)) + if (data->requires_broadcast) { + TF_LITE_ADD(BroadcastAdd4DSlow); + } else { + TF_LITE_ADD(Add); + } +#undef TF_LITE_ADD +} + +TfLiteStatus EvalAddQuantized(TfLiteContext* context, TfLiteNode* node, + TfLiteAddParams* params, const OpData* data, + const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2, + TfLiteEvalTensor* output) { + if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + tflite::ArithmeticParams op_params; + op_params.left_shift = data->left_shift; + op_params.input1_offset = data->input1_offset; + op_params.input1_multiplier = data->input1_multiplier; + op_params.input1_shift = data->input1_shift; + op_params.input2_offset = data->input2_offset; + op_params.input2_multiplier = data->input2_multiplier; + op_params.input2_shift = data->input2_shift; + op_params.output_offset = data->output_offset; + op_params.output_multiplier = data->output_multiplier; + op_params.output_shift = data->output_shift; + SetActivationParams(data->output_activation_min, + data->output_activation_max, &op_params); + bool need_broadcast = reference_ops::ProcessBroadcastShapes( + tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorShape(input2), &op_params); +#define TF_LITE_ADD(type, opname, dtype) \ + type::opname(op_params, tflite::micro::GetTensorShape(input1), \ + tflite::micro::GetTensorData(input1), \ + tflite::micro::GetTensorShape(input2), \ + tflite::micro::GetTensorData(input2), \ + tflite::micro::GetTensorShape(output), \ + tflite::micro::GetTensorData(output)); + if (output->type == kTfLiteInt8) { + if (need_broadcast) { + TF_LITE_ADD(reference_integer_ops, BroadcastAdd4DSlow, int8_t); + } else { + arm_elementwise_add_s8( + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorData(input2), + op_params.input1_offset, op_params.input1_multiplier, + op_params.input1_shift, op_params.input2_offset, + op_params.input2_multiplier, op_params.input2_shift, + op_params.left_shift, tflite::micro::GetTensorData(output), + op_params.output_offset, op_params.output_multiplier, + op_params.output_shift, op_params.quantized_activation_min, + op_params.quantized_activation_max, + MatchingElementsSize(tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorShape(output))); + } + } else { + if (need_broadcast) { + TF_LITE_ADD(reference_ops, BroadcastAdd4DSlow, uint8_t); + } else { + TF_LITE_ADD(reference_ops, Add, uint8_t); + } + } +#undef TF_LITE_ADD + } + + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1); + const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + OpData* data = static_cast(node->user_data); + auto* params = reinterpret_cast(node->builtin_data); + + TF_LITE_ENSURE_STATUS( + CalculateOpData(context, params, input1, input2, output, data)); + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + if (output->type == kTfLiteFloat32) { + EvalAdd(context, node, params, data, input1, input2, output); + } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + TF_LITE_ENSURE_OK(context, EvalAddQuantized(context, node, params, data, + input1, input2, output)); + } else { + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(output->type), output->type); + return kTfLiteError; + } + + return kTfLiteOk; +} + +} // namespace add + +TfLiteRegistration Register_ADD() { + return {/*init=*/add::Init, + /*free=*/nullptr, + /*prepare=*/add::Prepare, + /*invoke=*/add::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/conv.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/conv.cpp new file mode 100644 index 0000000..dcd8807 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/conv.cpp @@ -0,0 +1,463 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/conv.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/padding.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace { + +constexpr int kInputTensor = 0; +constexpr int kFilterTensor = 1; +constexpr int kBiasTensor = 2; +constexpr int kOutputTensor = 0; + +// Conv is quantized along dimension 0: +// https://www.tensorflow.org/lite/performance/quantization_spec +constexpr int kConvQuantizedDimension = 0; + +struct OpData { + TfLitePaddingValues padding; + + // Cached tensor zero point values for quantized operations. + int32_t input_zero_point; + int32_t filter_zero_point; + int32_t output_zero_point; + + // The scaling factor from input to output (aka the 'real multiplier') can + // be represented as a fixed point multiplier plus a left shift. + int32_t output_multiplier; + int output_shift; + + // Per channel output multiplier and shift. + int32_t* per_channel_output_multiplier; + int32_t* per_channel_output_shift; + + // The range of the fused activation layer. For example for kNone and + // uint8_t these would be 0 and 255. + int32_t output_activation_min; + int32_t output_activation_max; + + // Index to buffer for optimizations if applicable. + int buffer_idx; +}; + +inline PaddingType RuntimePaddingType(TfLitePadding padding) { + switch (padding) { + case TfLitePadding::kTfLitePaddingSame: + return PaddingType::kSame; + case TfLitePadding::kTfLitePaddingValid: + return PaddingType::kValid; + case TfLitePadding::kTfLitePaddingUnknown: + default: + return PaddingType::kNone; + } +} + +TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node, + const TfLiteConvParams* params, int width, + int height, int filter_width, int filter_height, + int out_width, int out_height, + const TfLiteType data_type, OpData* data) { + bool has_bias = node->inputs->size == 3; + // Check number of inputs/outputs + TF_LITE_ENSURE(context, has_bias || node->inputs->size == 2); + TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); + + // Matching GetWindowedOutputSize in TensorFlow. + auto padding = params->padding; + data->padding = ComputePaddingHeightWidth( + params->stride_height, params->stride_width, + params->dilation_height_factor, params->dilation_width_factor, height, + width, filter_height, filter_width, padding, &out_height, &out_width); + + // Note that quantized inference requires that all tensors have their + // parameters set. This is usually done during quantized training. + if (data_type != kTfLiteFloat32) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); + const TfLiteTensor* bias = + GetOptionalInputTensor(context, node, kBiasTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + int num_channels = filter->dims->data[kConvQuantizedDimension]; + + TF_LITE_ENSURE_STATUS(tflite::PopulateConvolutionQuantizationParams( + context, input, filter, bias, output, params->activation, + &data->output_multiplier, &data->output_shift, + &data->output_activation_min, &data->output_activation_max, + data->per_channel_output_multiplier, + reinterpret_cast(data->per_channel_output_shift), num_channels)); + } + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + int32_t buf_size = 0; + const auto params = static_cast(node->builtin_data); + OpData* data = static_cast(node->user_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); + const TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + RuntimeShape input_shape = GetTensorShape(input); + RuntimeShape output_shape = GetTensorShape(output); + + // Initialize cmsis-nn input dimensions + cmsis_nn_dims input_dims; + input_dims.n = MatchingDim(input_shape, 0, output_shape, 0); + input_dims.h = input->dims->data[1]; + input_dims.w = input->dims->data[2]; + input_dims.c = input_shape.Dims(3); + + // Initialize cmsis-nn filter dimensions + cmsis_nn_dims filter_dims; + filter_dims.n = output_shape.Dims(3); + filter_dims.h = filter->dims->data[1]; + filter_dims.w = filter->dims->data[2]; + filter_dims.c = input_dims.c; + + // Initialize cmsis-nn output dimensions + cmsis_nn_dims output_dims; + output_dims.n = input_dims.n; + output_dims.h = output->dims->data[1]; + output_dims.w = output->dims->data[2]; + output_dims.c = output_shape.Dims(3); + + // Dynamically allocate per-channel quantization parameters. + // TODO(#42883): This allocation is done even for non-int8 cases to get around + // a bug in kernel_util.cc which incorrectly uses per_channel_output_shift in + // non-int8 cases. Protect this section with a if (input->type == kTfLiteInt8) + // when the issue is fixed. + const int num_channels = filter->dims->data[kConvQuantizedDimension]; + data->per_channel_output_multiplier = + static_cast(context->AllocatePersistentBuffer( + context, num_channels * sizeof(int32_t))); + data->per_channel_output_shift = + static_cast(context->AllocatePersistentBuffer( + context, num_channels * sizeof(int32_t))); + + TF_LITE_ENSURE_STATUS(CalculateOpData( + context, node, params, input_dims.w, input_dims.h, filter_dims.w, + filter_dims.h, output_dims.w, output_dims.h, input->type, data)); + + data->input_zero_point = input->params.zero_point; + data->filter_zero_point = filter->params.zero_point; + data->output_zero_point = output->params.zero_point; + + if (input->type == kTfLiteInt8) { + // Initialize cmsis-nn convolution parameters + cmsis_nn_conv_params conv_params; + conv_params.input_offset = -input->params.zero_point; + conv_params.output_offset = output->params.zero_point; + conv_params.stride.h = params->stride_height; + conv_params.stride.w = params->stride_width; + conv_params.dilation.h = params->dilation_height_factor; + conv_params.dilation.w = params->dilation_width_factor; + conv_params.padding.h = data->padding.height; + conv_params.padding.w = data->padding.width; + conv_params.activation.min = data->output_activation_min; + conv_params.activation.max = data->output_activation_max; + + buf_size = arm_convolve_wrapper_s8_get_buffer_size( + &conv_params, &input_dims, &filter_dims, &output_dims); + } + + if (buf_size > 0) { + TF_LITE_ENSURE_STATUS(context->RequestScratchBufferInArena( + context, buf_size, &data->buffer_idx)); + } else { + data->buffer_idx = -1; + } + return kTfLiteOk; +} + +TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node, + TfLiteConvParams* params, const OpData& data, + const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, + const TfLiteEvalTensor* bias, + TfLiteEvalTensor* im2col, + TfLiteEvalTensor* hwcn_weights, + TfLiteEvalTensor* output) { + const int32_t input_offset = -data.input_zero_point; + const int32_t filter_offset = -data.filter_zero_point; + const int32_t output_offset = data.output_zero_point; + + ConvParams op_params; + op_params.padding_type = RuntimePaddingType(params->padding); + op_params.padding_values.width = data.padding.width; + op_params.padding_values.height = data.padding.height; + op_params.stride_width = params->stride_width; + op_params.stride_height = params->stride_height; + op_params.dilation_width_factor = params->dilation_width_factor; + op_params.dilation_height_factor = params->dilation_height_factor; + op_params.input_offset = input_offset; + op_params.weights_offset = filter_offset; + op_params.output_offset = output_offset; + op_params.output_multiplier = data.output_multiplier; + op_params.output_shift = -data.output_shift; + op_params.quantized_activation_min = data.output_activation_min; + op_params.quantized_activation_max = data.output_activation_max; + reference_ops::Conv(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), + tflite::micro::GetTensorShape(im2col), + tflite::micro::GetTensorData(im2col), nullptr); + return kTfLiteOk; +} + +TfLiteStatus EvalQuantizedPerChannel( + TfLiteContext* context, TfLiteNode* node, TfLiteConvParams* params, + const OpData& data, const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, const TfLiteEvalTensor* bias, + TfLiteEvalTensor* output, TfLiteEvalTensor* im2col) { + cmsis_nn_conv_params conv_params; + conv_params.dilation.h = params->dilation_height_factor; + conv_params.dilation.w = params->dilation_width_factor; + // TODO(#43557) Remove checks for dilation and call to reference + // implementation when dilation is supported in the optimized implementation + // by CMSIS-NN. + if (conv_params.dilation.h == 1 && conv_params.dilation.w == 1) { + // Initialize cmsis-nn convolution parameters + conv_params.input_offset = -data.input_zero_point; + conv_params.output_offset = data.output_zero_point; + conv_params.stride.h = params->stride_height; + conv_params.stride.w = params->stride_width; + conv_params.padding.h = data.padding.height; + conv_params.padding.w = data.padding.width; + conv_params.activation.min = data.output_activation_min; + conv_params.activation.max = data.output_activation_max; + + // Initialize cmsis-nn per channel quantization parameters + cmsis_nn_per_channel_quant_params quant_params; + quant_params.multiplier = + const_cast(data.per_channel_output_multiplier); + quant_params.shift = const_cast(data.per_channel_output_shift); + + RuntimeShape filter_shape = tflite::micro::GetTensorShape(filter); + RuntimeShape input_shape = tflite::micro::GetTensorShape(input); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + RuntimeShape bias_shape = tflite::micro::GetTensorShape(bias); + + // Consistency check. + TFLITE_DCHECK_LE(conv_params.activation.min, conv_params.activation.max); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + const int batch_size = MatchingDim(input_shape, 0, output_shape, 0); + const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3); + const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3); + if (tflite::micro::GetTensorData(bias)) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + + // Initialize cmsis-nn dimensions + // Input + cmsis_nn_dims input_dims; + input_dims.n = batch_size; + input_dims.h = input_shape.Dims(1); + input_dims.w = input_shape.Dims(2); + input_dims.c = input_depth; + + // Filter + cmsis_nn_dims filter_dims; + filter_dims.n = output_depth; + filter_dims.h = filter_shape.Dims(1); + filter_dims.w = filter_shape.Dims(2); + filter_dims.c = input_depth; + + // Bias + cmsis_nn_dims bias_dims; + bias_dims.n = 1; + bias_dims.h = 1; + bias_dims.w = 1; + bias_dims.c = output_depth; + + // Output + cmsis_nn_dims output_dims; + output_dims.n = batch_size; + output_dims.h = output_shape.Dims(1); + output_dims.w = output_shape.Dims(2); + output_dims.c = output_depth; + + // Initialize cmsis-nn context + cmsis_nn_context ctx; + ctx.buf = nullptr; + ctx.size = 0; + + if (data.buffer_idx > -1) { + ctx.buf = context->GetScratchBuffer(context, data.buffer_idx); + // Note: ctx.size is currently not used in cmsis-nn. + // The buffer should be allocated in the Prepare function through + // arm_convolve_wrapper_s8_get_buffer_size + } + + // arm_convolve_wrapper_s8 dispatches the optimized kernel accordingly with + // the parameters passed + TFLITE_DCHECK_EQ( + arm_convolve_wrapper_s8( + &ctx, &conv_params, &quant_params, &input_dims, + tflite::micro::GetTensorData(input), &filter_dims, + tflite::micro::GetTensorData(filter), &bias_dims, + tflite::micro::GetTensorData(bias), &output_dims, + tflite::micro::GetTensorData(output)), + ARM_MATH_SUCCESS); + } else { + // TODO(b/154032858): Investigate removing extra copies. + ConvParams op_params; + op_params.input_offset = -data.input_zero_point; + op_params.output_offset = data.output_zero_point; + op_params.stride_height = params->stride_height; + op_params.stride_width = params->stride_width; + op_params.dilation_height_factor = params->dilation_height_factor; + op_params.dilation_width_factor = params->dilation_width_factor; + op_params.padding_values.height = data.padding.height; + op_params.padding_values.width = data.padding.width; + op_params.quantized_activation_min = data.output_activation_min; + op_params.quantized_activation_max = data.output_activation_max; + + reference_integer_ops::ConvPerChannel( + op_params, data.per_channel_output_multiplier, + data.per_channel_output_shift, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } + return kTfLiteOk; +} + +TfLiteStatus EvalFloat(TfLiteContext* context, TfLiteNode* node, + TfLiteConvParams* params, const OpData& data, + const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, + const TfLiteEvalTensor* bias, TfLiteEvalTensor* im2col, + TfLiteEvalTensor* hwcn_weights, + TfLiteEvalTensor* output) { + float output_activation_min, output_activation_max; + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); + // TODO(b/154032858): Investigate removing extra copies. + ConvParams op_params; + op_params.padding_type = RuntimePaddingType(params->padding); + op_params.padding_values.width = data.padding.width; + op_params.padding_values.height = data.padding.height; + op_params.stride_width = params->stride_width; + op_params.stride_height = params->stride_height; + op_params.dilation_width_factor = params->dilation_width_factor; + op_params.dilation_height_factor = params->dilation_height_factor; + op_params.float_activation_min = output_activation_min; + op_params.float_activation_max = output_activation_max; + + reference_ops::Conv(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), + tflite::micro::GetTensorShape(im2col), + tflite::micro::GetTensorData(im2col)); + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* filter = + tflite::micro::GetEvalInput(context, node, kFilterTensor); + const TfLiteEvalTensor* bias = + (NumInputs(node) == 3) + ? tflite::micro::GetEvalInput(context, node, kBiasTensor) + : nullptr; + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + TF_LITE_ENSURE_EQ(context, input->type, output->type); + TF_LITE_ENSURE_MSG(context, input->type == filter->type, + "Hybrid models are not supported on TFLite Micro."); + + switch (input->type) { // Already know in/out types are same. + case kTfLiteFloat32: + EvalFloat(context, node, params, data, input, filter, bias, nullptr, + nullptr, output); + break; + case kTfLiteInt8: + return EvalQuantizedPerChannel(context, node, params, data, input, filter, + bias, output, nullptr); + break; + case kTfLiteUInt8: + return EvalQuantized(context, node, params, data, input, filter, bias, + nullptr, nullptr, output); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace + +TfLiteRegistration Register_CONV_2D() { + return {/*init=*/Init, + /*free=*/nullptr, + /*prepare=*/Prepare, + /*invoke=*/Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cpp new file mode 100644 index 0000000..de82af5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cpp @@ -0,0 +1,478 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/padding.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace { + +constexpr int kInputTensor = 0; +constexpr int kFilterTensor = 1; +constexpr int kBiasTensor = 2; +constexpr int kOutputTensor = 0; + +// Depthwise conv is quantized along dimension 3: +// https://www.tensorflow.org/lite/performance/quantization_spec +constexpr int kDepthwiseConvQuantizedDimension = 3; + +struct OpData { + TfLitePaddingValues padding; + + // Cached tensor zero point values for quantized operations. + int32_t input_zero_point; + int32_t filter_zero_point; + int32_t output_zero_point; + + // The scaling factor from input to output (aka the 'real multiplier') can + // be represented as a fixed point multiplier plus a left shift. + int32_t output_multiplier; + int output_shift; + + // Per channel output multiplier and shift. + int32_t* per_channel_output_multiplier; + int32_t* per_channel_output_shift; + + // The range of the fused activation layer. For example for kNone and + // uint8_t these would be 0 and 255. + int32_t output_activation_min; + int32_t output_activation_max; + // Index to buffer for optimizations if applicable. + int buffer_idx; +}; + +TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node, + TfLiteDepthwiseConvParams* params, int width, + int height, int filter_width, int filter_height, + const TfLiteType data_type, OpData* data) { + bool has_bias = node->inputs->size == 3; + // Check number of inputs/outputs + TF_LITE_ENSURE(context, has_bias || node->inputs->size == 2); + TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); + + int unused_output_height, unused_output_width; + // Set buffer index to a reset value + data->buffer_idx = -1; + data->padding = ComputePaddingHeightWidth( + params->stride_height, params->stride_width, 1, 1, height, width, + filter_height, filter_width, params->padding, &unused_output_height, + &unused_output_width); + + // Note that quantized inference requires that all tensors have their + // parameters set. This is usually done during quantized training. + if (data_type != kTfLiteFloat32) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); + const TfLiteTensor* bias = + GetOptionalInputTensor(context, node, kBiasTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + int num_channels = filter->dims->data[kDepthwiseConvQuantizedDimension]; + + return tflite::PopulateConvolutionQuantizationParams( + context, input, filter, bias, output, params->activation, + &data->output_multiplier, &data->output_shift, + &data->output_activation_min, &data->output_activation_max, + data->per_channel_output_multiplier, + reinterpret_cast(data->per_channel_output_shift), num_channels); + } + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + OpData* data = static_cast(node->user_data); + auto* params = + reinterpret_cast(node->builtin_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + const TfLiteType data_type = input->type; + int width = SizeOfDimension(input, 2); + int height = SizeOfDimension(input, 1); + int filter_width = SizeOfDimension(filter, 2); + int filter_height = SizeOfDimension(filter, 1); + + if (input->type == kTfLiteInt8) { + TF_LITE_ENSURE_EQ(context, filter->quantization.type, + kTfLiteAffineQuantization); + + // All per-channel quantized tensors need valid zero point and scale arrays. + const auto* affine_quantization = + reinterpret_cast( + filter->quantization.params); + TF_LITE_ENSURE(context, affine_quantization); + TF_LITE_ENSURE(context, affine_quantization->scale); + TF_LITE_ENSURE(context, affine_quantization->zero_point); + TF_LITE_ENSURE( + context, affine_quantization->scale->size == 1 || + affine_quantization->scale->size == + filter->dims->data[kDepthwiseConvQuantizedDimension]); + TF_LITE_ENSURE_EQ(context, affine_quantization->scale->size, + affine_quantization->zero_point->size); + } + + // Allocate memory for per-channel quantization parameters + const int num_channels = filter->dims->data[kDepthwiseConvQuantizedDimension]; + + data->per_channel_output_multiplier = + reinterpret_cast(context->AllocatePersistentBuffer( + context, num_channels * sizeof(int32_t))); + data->per_channel_output_shift = + reinterpret_cast(context->AllocatePersistentBuffer( + context, num_channels * sizeof(int32_t))); + + TF_LITE_ENSURE_STATUS(CalculateOpData(context, node, params, width, height, + filter_width, filter_height, data_type, + data)); + + data->input_zero_point = input->params.zero_point; + data->filter_zero_point = filter->params.zero_point; + data->output_zero_point = output->params.zero_point; + + if (input->type == kTfLiteInt8) { + RuntimeShape input_shape = GetTensorShape(input); + RuntimeShape output_shape = GetTensorShape(output); + RuntimeShape filter_shape = GetTensorShape(filter); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + const int batch_size = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(output_shape, 3, filter_shape, 3); + TFLITE_DCHECK_EQ(batch_size, 1); /* Only batch = 1 is supported */ + + cmsis_nn_dims input_dims; + input_dims.n = batch_size; + input_dims.h = height; + input_dims.w = width; + input_dims.c = input_shape.Dims(3); + + cmsis_nn_dims filter_dims; + filter_dims.n = 1; + filter_dims.h = filter_height; + filter_dims.w = filter_width; + filter_dims.c = output_depth; + + cmsis_nn_dims output_dims; + output_dims.n = batch_size; + output_dims.h = output_shape.Dims(1); + output_dims.w = output_shape.Dims(2); + output_dims.c = output_depth; + + cmsis_nn_dw_conv_params dw_conv_params; + dw_conv_params.padding.h = data->padding.height; + dw_conv_params.padding.w = data->padding.width; + + const int32_t buf_size = arm_depthwise_conv_wrapper_s8_get_buffer_size( + &dw_conv_params, &input_dims, &filter_dims, &output_dims); + + if (buf_size > 0) { + TF_LITE_ENSURE_STATUS(context->RequestScratchBufferInArena( + context, buf_size, &data->buffer_idx)); + } else { + data->buffer_idx = -1; + } + } + return kTfLiteOk; +} + +void EvalFloat(TfLiteContext* context, TfLiteNode* node, + TfLiteDepthwiseConvParams* params, const OpData* data, + const TfLiteEvalTensor* input, const TfLiteEvalTensor* filter, + const TfLiteEvalTensor* bias, TfLiteEvalTensor* output) { + float output_activation_min, output_activation_max; + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); + + tflite::DepthwiseParams op_params; + // Padding type is ignored, but still set. + op_params.padding_type = PaddingType::kSame; + op_params.padding_values.width = data->padding.width; + op_params.padding_values.height = data->padding.height; + op_params.stride_width = params->stride_width; + op_params.stride_height = params->stride_height; + op_params.dilation_width_factor = params->dilation_width_factor; + op_params.dilation_height_factor = params->dilation_height_factor; + op_params.depth_multiplier = params->depth_multiplier; + op_params.float_activation_min = output_activation_min; + op_params.float_activation_max = output_activation_max; + + tflite::reference_ops::DepthwiseConv( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); +} + +void EvalQuantizedPerChannel(TfLiteContext* context, TfLiteNode* node, + TfLiteDepthwiseConvParams* params, OpData* data, + const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, + const TfLiteEvalTensor* bias, + TfLiteEvalTensor* output) { + cmsis_nn_dw_conv_params dw_conv_params; + dw_conv_params.dilation.h = params->dilation_height_factor; + dw_conv_params.dilation.w = params->dilation_width_factor; + // Call to reference implementation can be removed when dilation is supported + // in the optimized implementations. + if (1 == dw_conv_params.dilation.h && 1 == dw_conv_params.dilation.w) { + dw_conv_params.input_offset = -data->input_zero_point; + dw_conv_params.output_offset = data->output_zero_point; + dw_conv_params.stride.h = params->stride_height; + dw_conv_params.stride.w = params->stride_width; + dw_conv_params.padding.h = data->padding.height; + dw_conv_params.padding.w = data->padding.width; + // TODO(b/130439627): Use calculated value for clamping. + dw_conv_params.activation.min = std::numeric_limits::min(); + dw_conv_params.activation.max = std::numeric_limits::max(); + dw_conv_params.ch_mult = params->depth_multiplier; + + cmsis_nn_per_channel_quant_params quant_params; + quant_params.multiplier = data->per_channel_output_multiplier; + quant_params.shift = data->per_channel_output_shift; + + RuntimeShape filter_shape = tflite::micro::GetTensorShape(filter); + RuntimeShape input_shape = tflite::micro::GetTensorShape(input); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + RuntimeShape bias_shape = tflite::micro::GetTensorShape(bias); + + TFLITE_DCHECK_LE(dw_conv_params.activation.min, + dw_conv_params.activation.max); + + const int batch_size = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); + + if (tflite::micro::GetTensorData(bias)) { + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + } + + cmsis_nn_dims input_dims; + input_dims.n = batch_size; + input_dims.h = input_shape.Dims(1); + input_dims.w = input_shape.Dims(2); + input_dims.c = input_shape.Dims(3); + + cmsis_nn_dims filter_dims; + filter_dims.n = filter_shape.Dims(0); + filter_dims.h = filter_shape.Dims(1); + filter_dims.w = filter_shape.Dims(2); + filter_dims.c = output_depth; + + cmsis_nn_dims bias_dims; + bias_dims.n = 1; + bias_dims.h = 1; + bias_dims.w = 1; + bias_dims.c = output_depth; + + cmsis_nn_dims output_dims; + output_dims.n = batch_size; + output_dims.h = output_shape.Dims(1); + output_dims.w = output_shape.Dims(2); + output_dims.c = output_depth; + + cmsis_nn_context ctx; + ctx.buf = nullptr; + /* 'size' is unused */ + ctx.size = 0; + + if (data->buffer_idx > -1) { + ctx.buf = context->GetScratchBuffer(context, data->buffer_idx); + } + + TFLITE_DCHECK_EQ( + arm_depthwise_conv_wrapper_s8( + &ctx, &dw_conv_params, &quant_params, &input_dims, + tflite::micro::GetTensorData(input), &filter_dims, + tflite::micro::GetTensorData(filter), &bias_dims, + tflite::micro::GetTensorData(bias), &output_dims, + tflite::micro::GetTensorData(output)), + ARM_MATH_SUCCESS); + } else { + DepthwiseParams op_params; + op_params.padding_type = PaddingType::kSame; + op_params.padding_values.width = data->padding.width; + op_params.padding_values.height = data->padding.height; + op_params.stride_width = params->stride_width; + op_params.stride_height = params->stride_height; + op_params.dilation_width_factor = params->dilation_width_factor; + op_params.dilation_height_factor = params->dilation_height_factor; + op_params.depth_multiplier = params->depth_multiplier; + op_params.input_offset = -data->input_zero_point; + op_params.weights_offset = 0; + op_params.output_offset = data->output_zero_point; + // TODO(b/130439627): Use calculated value for clamping. + op_params.quantized_activation_min = std::numeric_limits::min(); + op_params.quantized_activation_max = std::numeric_limits::max(); + + reference_integer_ops::DepthwiseConvPerChannel( + op_params, data->per_channel_output_multiplier, + data->per_channel_output_shift, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } +} + +void EvalQuantized(TfLiteContext* context, TfLiteNode* node, + TfLiteDepthwiseConvParams* params, const OpData* data, + const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, const TfLiteEvalTensor* bias, + TfLiteEvalTensor* output) { + const int32_t input_offset = -data->input_zero_point; + const int32_t filter_offset = -data->filter_zero_point; + const int32_t output_offset = data->output_zero_point; + + tflite::DepthwiseParams op_params; + // Padding type is ignored, but still set. + op_params.padding_type = PaddingType::kSame; + op_params.padding_values.width = data->padding.width; + op_params.padding_values.height = data->padding.height; + op_params.stride_width = params->stride_width; + op_params.stride_height = params->stride_height; + op_params.dilation_width_factor = params->dilation_width_factor; + op_params.dilation_height_factor = params->dilation_height_factor; + op_params.depth_multiplier = params->depth_multiplier; + op_params.quantized_activation_min = data->output_activation_min; + op_params.quantized_activation_max = data->output_activation_max; + op_params.input_offset = input_offset; + op_params.weights_offset = filter_offset; + op_params.output_offset = output_offset; + op_params.output_multiplier = data->output_multiplier; + // Legacy ops used mixed left and right shifts. Now all are +ve-means-left. + op_params.output_shift = -data->output_shift; + + if (1 == op_params.dilation_width_factor && + 1 == op_params.dilation_height_factor) { + RuntimeShape filter_shape = tflite::micro::GetTensorShape(filter); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + RuntimeShape input_shape = tflite::micro::GetTensorShape(input); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int input_depth = input_shape.Dims(3); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + arm_depthwise_conv_u8_basic_ver1( + tflite::micro::GetTensorData(input), input_width, input_height, + input_depth, tflite::micro::GetTensorData(filter), + filter_width, filter_height, op_params.depth_multiplier, + op_params.padding_values.width, op_params.padding_values.height, + op_params.stride_width, op_params.stride_height, + op_params.dilation_width_factor, op_params.dilation_height_factor, + tflite::micro::GetTensorData(bias), op_params.input_offset, + op_params.weights_offset, op_params.output_offset, + tflite::micro::GetTensorData(output), output_width, + output_height, op_params.quantized_activation_min, + op_params.quantized_activation_max, op_params.output_shift, + op_params.output_multiplier); + } else { + tflite::reference_ops::DepthwiseConv( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + auto* params = + reinterpret_cast(node->builtin_data); + OpData& data = *(static_cast(node->user_data)); + + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* filter = + tflite::micro::GetEvalInput(context, node, kFilterTensor); + const TfLiteEvalTensor* bias = + (NumInputs(node) == 3) + ? tflite::micro::GetEvalInput(context, node, kBiasTensor) + : nullptr; + + // TODO(aselle): Consider whether float conv and quantized conv should be + // separate ops to avoid dispatch overhead here. + switch (input->type) { // Already know in/out types are same. + case kTfLiteFloat32: + EvalFloat(context, node, params, &data, input, filter, bias, output); + break; + case kTfLiteInt8: + EvalQuantizedPerChannel(context, node, params, &data, input, filter, bias, + output); + break; + case kTfLiteUInt8: + EvalQuantized(context, node, params, &data, input, filter, bias, output); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace + +TfLiteRegistration Register_DEPTHWISE_CONV_2D() { + return {/*init=*/Init, + /*free=*/nullptr, + /*prepare=*/Prepare, + /*invoke=*/Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cpp new file mode 100644 index 0000000..814a1dd --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cpp @@ -0,0 +1,395 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/fully_connected.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/fully_connected.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace { + +struct OpData { + // The scaling factor from input to output (aka the 'real multiplier') can + // be represented as a fixed point multiplier plus a left shift. + int32_t output_multiplier; + int output_shift; + // The range of the fused activation layer. For example for kNone and + // uint8_t these would be 0 and 255. + int32_t output_activation_min; + int32_t output_activation_max; + // The index of the temporary tensor where the quantized inputs are cached. + int input_quantized_index; + // Index to buffer for optimizations if applicable. + int buffer_idx; + + // Cached tensor zero point values for quantized operations. + int32_t input_zero_point; + int32_t filter_zero_point; + int32_t output_zero_point; +}; + +constexpr int kInputTensor = 0; +constexpr int kWeightsTensor = 1; +constexpr int kBiasTensor = 2; +constexpr int kOutputTensor = 0; + +// TODO(b/169801227): This global struct is needed for the linker to drop unused +// code (for example, by using Register_FULLY_CONNECTED_INT8 instead of +// Register_FULLY_CONNECTED). +TfLiteRegistration fully_connected_registration; + +TfLiteStatus CalculateOpData(TfLiteContext* context, + TfLiteFusedActivation activation, + TfLiteType data_type, const TfLiteTensor* input, + const TfLiteTensor* filter, + const TfLiteTensor* bias, TfLiteTensor* output, + OpData* data) { + TfLiteStatus status = kTfLiteOk; + // Set buffer index to a reset value + data->buffer_idx = -1; + if (data_type != kTfLiteFloat32) { + double real_multiplier = 0.0; + TF_LITE_ENSURE_STATUS(GetQuantizedConvolutionMultipler( + context, input, filter, bias, output, &real_multiplier)); + int exponent; + QuantizeMultiplier(real_multiplier, &data->output_multiplier, &exponent); + data->output_shift = -exponent; + TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( + context, activation, output, &data->output_activation_min, + &data->output_activation_max)); + data->input_zero_point = input->params.zero_point; + data->filter_zero_point = filter->params.zero_point; + data->output_zero_point = output->params.zero_point; + } + return status; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + OpData* data = static_cast(node->user_data); + const auto params = + static_cast(node->builtin_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* filter = GetInput(context, node, kWeightsTensor); + const TfLiteTensor* bias = GetOptionalInputTensor(context, node, kBiasTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + TF_LITE_ENSURE_MSG(context, input->type == filter->type, + "Hybrid models are not supported on TFLite Micro."); + TF_LITE_ENSURE_STATUS(CalculateOpData(context, params->activation, + input->type, input, filter, bias, + output, data)); + + if (input->type == kTfLiteInt8 && nullptr != GetTensorData(bias)) { + RuntimeShape filter_shape = GetTensorShape(filter); + RuntimeShape output_shape = GetTensorShape(output); + + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2); + const int filter_dim_count = filter_shape.DimensionsCount(); + cmsis_nn_dims filter_dims; + filter_dims.n = filter_shape.Dims(filter_dim_count - 1); + filter_dims.h = 1; + filter_dims.w = 1; + filter_dims.c = output_shape.Dims(1); + + const int32_t buf_size = + arm_fully_connected_s8_get_buffer_size(&filter_dims); + + if (buf_size > 0) { + TF_LITE_ENSURE_STATUS(context->RequestScratchBufferInArena( + context, buf_size, &data->buffer_idx)); + } else { + data->buffer_idx = -1; + } + } + return kTfLiteOk; +} + +TfLiteStatus EvalQuantizedInt8(TfLiteContext* context, TfLiteNode* node, + const OpData& data, + const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, + const TfLiteEvalTensor* bias, + TfLiteEvalTensor* output) { + // The 'if' condition can be removed when null handling of bias is added to + // arm_fully_connected_s8 + if (nullptr != tflite::micro::GetTensorData(bias)) { + const RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2); + const int batches = output_shape.Dims(0); + const int output_depth = output_shape.Dims(1); + const RuntimeShape filter_shape = tflite::micro::GetTensorShape(filter); + const int filter_dim_count = filter_shape.DimensionsCount(); + const int accum_depth = filter_shape.Dims(filter_dim_count - 1); + const RuntimeShape input_shape = tflite::micro::GetTensorShape(input); + + cmsis_nn_fc_params fc_params; + fc_params.input_offset = -data.input_zero_point; + fc_params.output_offset = data.output_zero_point; + fc_params.filter_offset = -data.filter_zero_point; + fc_params.activation.min = data.output_activation_min; + fc_params.activation.max = data.output_activation_max; + + cmsis_nn_per_tensor_quant_params quant_params; + quant_params.multiplier = data.output_multiplier; + // TODO(b/138810107): Figure out whether output shift should be inverted + quant_params.shift = -data.output_shift; + + cmsis_nn_dims input_dims; + input_dims.n = batches; + input_dims.h = 1; + input_dims.w = 1; + input_dims.c = accum_depth; + + cmsis_nn_dims filter_dims; + filter_dims.n = accum_depth; + filter_dims.h = 1; + filter_dims.w = 1; + filter_dims.c = output_depth; + + cmsis_nn_dims bias_dims; + bias_dims.n = 1; + bias_dims.h = 1; + bias_dims.w = 1; + bias_dims.c = output_depth; + + cmsis_nn_dims output_dims; + output_dims.n = batches; + output_dims.h = 1; + output_dims.w = 1; + output_dims.c = output_depth; + + cmsis_nn_context ctx; + ctx.buf = nullptr; + ctx.size = 0; + + if (data.buffer_idx > -1) { + ctx.buf = context->GetScratchBuffer(context, data.buffer_idx); + } + + TF_LITE_ENSURE_EQ( + context, + arm_fully_connected_s8( + &ctx, &fc_params, &quant_params, &input_dims, + tflite::micro::GetTensorData(input), &filter_dims, + tflite::micro::GetTensorData(filter), &bias_dims, + tflite::micro::GetTensorData(bias), &output_dims, + tflite::micro::GetTensorData(output)), + ARM_MATH_SUCCESS); + } else { + tflite::FullyConnectedParams op_params; + op_params.input_offset = -data.input_zero_point; + op_params.weights_offset = -data.filter_zero_point; + op_params.output_offset = data.output_zero_point; + op_params.output_multiplier = data.output_multiplier; + // TODO(b/138810107): Figure out whether output shift should be inverted + op_params.output_shift = -data.output_shift; + op_params.quantized_activation_min = data.output_activation_min; + op_params.quantized_activation_max = data.output_activation_max; + + reference_integer_ops::FullyConnected( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } + return kTfLiteOk; +} + +TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node, + const OpData& data, const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, + const TfLiteEvalTensor* bias, + TfLiteEvalTensor* output) { + const int32_t input_offset = -data.input_zero_point; + const int32_t filter_offset = -data.filter_zero_point; + const int32_t output_offset = data.output_zero_point; + + tflite::FullyConnectedParams op_params; + op_params.input_offset = input_offset; + op_params.weights_offset = filter_offset; + op_params.output_offset = output_offset; + op_params.output_multiplier = data.output_multiplier; + // Legacy ops used mixed left and right shifts. Now all are +ve-means-left. + op_params.output_shift = -data.output_shift; + op_params.quantized_activation_min = data.output_activation_min; + op_params.quantized_activation_max = data.output_activation_max; + +#define TF_LITE_FULLY_CONNECTED(output_data_type) \ + reference_ops::FullyConnected( \ + op_params, tflite::micro::GetTensorShape(input), \ + tflite::micro::GetTensorData(input), \ + tflite::micro::GetTensorShape(filter), \ + tflite::micro::GetTensorData(filter), \ + tflite::micro::GetTensorShape(bias), \ + tflite::micro::GetTensorData(bias), \ + tflite::micro::GetTensorShape(output), \ + tflite::micro::GetTensorData(output)) + switch (output->type) { + case kTfLiteUInt8: + TF_LITE_FULLY_CONNECTED(uint8_t); + break; + case kTfLiteInt16: + TF_LITE_FULLY_CONNECTED(int16_t); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(output->type), output->type); + return kTfLiteError; + } + + return kTfLiteOk; +} + +TfLiteStatus EvalFloat(TfLiteContext* context, TfLiteNode* node, + TfLiteFusedActivation activation, + const TfLiteEvalTensor* input, + const TfLiteEvalTensor* filter, + const TfLiteEvalTensor* bias, TfLiteEvalTensor* output) { + float output_activation_min, output_activation_max; + CalculateActivationRange(activation, &output_activation_min, + &output_activation_max); + tflite::FullyConnectedParams op_params; + op_params.float_activation_min = output_activation_min; + op_params.float_activation_max = output_activation_max; + tflite::reference_ops::FullyConnected( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->builtin_data != nullptr); + const auto* params = + static_cast(node->builtin_data); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* filter = + tflite::micro::GetEvalInput(context, node, kWeightsTensor); + const TfLiteEvalTensor* bias = + tflite::micro::GetEvalInput(context, node, kBiasTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + // Checks in Prepare ensure input, output and filter types are all the same. + switch (input->type) { + case kTfLiteFloat32: + return EvalFloat(context, node, params->activation, input, filter, bias, + output); + case kTfLiteInt8: + return EvalQuantizedInt8(context, node, data, input, filter, bias, + output); + + case kTfLiteUInt8: + return EvalQuantized(context, node, data, input, filter, bias, output); + + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +// Note that the current function names are not ideal at all (this EvalInt8 +// function internally calls EvalQuantizedInt8, and there is similar name +// aliasing in the Eval function too). We will be attempting to have a more +// descriptive naming convention but holding off on that for now, since the +// renaming might be coupled with reducing code duplication and some additional +// refactoring. +TfLiteStatus EvalInt8(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* filter = + tflite::micro::GetEvalInput(context, node, kWeightsTensor); + const TfLiteEvalTensor* bias = + tflite::micro::GetEvalInput(context, node, kBiasTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + // Checks in Prepare ensure input, output and filter types are all the same. + if (input->type != kTfLiteInt8) { + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + + return EvalQuantizedInt8(context, node, data, input, filter, bias, output); +} + +} // namespace + +TfLiteRegistration Register_FULLY_CONNECTED() { + fully_connected_registration.init = Init; + fully_connected_registration.free = nullptr; + fully_connected_registration.prepare = Prepare; + fully_connected_registration.invoke = Eval; + fully_connected_registration.profiling_string = nullptr; + fully_connected_registration.builtin_code = 0; + fully_connected_registration.custom_name = nullptr; + fully_connected_registration.version = 0; + return fully_connected_registration; +} + +TfLiteRegistration Register_FULLY_CONNECTED_INT8() { + fully_connected_registration.init = Init; + fully_connected_registration.free = nullptr; + fully_connected_registration.prepare = Prepare; + fully_connected_registration.invoke = EvalInt8; + fully_connected_registration.profiling_string = nullptr; + fully_connected_registration.builtin_code = 0; + fully_connected_registration.custom_name = nullptr; + fully_connected_registration.version = 0; + return fully_connected_registration; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/mul.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/mul.cpp new file mode 100644 index 0000000..2c079b7 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/mul.cpp @@ -0,0 +1,229 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/mul.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace mul { + +constexpr int kInput1Tensor = 0; +constexpr int kInput2Tensor = 1; +constexpr int kOutputTensor = 0; + +struct OpData { + int32_t output_activation_min; + int32_t output_activation_max; + + int32_t output_multiplier; + int output_shift; + + // Cached tensor zero point values for quantized operations. + int32_t input1_zero_point; + int32_t input2_zero_point; + int32_t output_zero_point; +}; + +TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node, + TfLiteMulParams* params, OpData* data) { + const TfLiteTensor* input1 = GetInput(context, node, kInput1Tensor); + const TfLiteTensor* input2 = GetInput(context, node, kInput2Tensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + TF_LITE_ENSURE_EQ(context, NumInputs(node), 2); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + TF_LITE_ENSURE_TYPES_EQ(context, input1->type, input2->type); + + if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( + context, params->activation, output, &data->output_activation_min, + &data->output_activation_max)); + + double real_multiplier = static_cast(input1->params.scale) * + static_cast(input2->params.scale) / + static_cast(output->params.scale); + QuantizeMultiplier(real_multiplier, &data->output_multiplier, + &data->output_shift); + } + + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input1 = GetInput(context, node, kInput1Tensor); + const TfLiteTensor* input2 = GetInput(context, node, kInput2Tensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + if (output->dims->size == 0) { + return AllocateOutputDimensionsFromInput(context, input1, input2, output); + } + + TFLITE_DCHECK(node->builtin_data != nullptr); + auto* params = reinterpret_cast(node->builtin_data); + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + data->input1_zero_point = input1->params.zero_point; + data->input2_zero_point = input2->params.zero_point; + data->output_zero_point = output->params.zero_point; + CalculateOpData(context, node, params, data); + + return kTfLiteOk; +} + +void EvalQuantized(TfLiteContext* context, TfLiteNode* node, + TfLiteMulParams* params, const OpData& data, + const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) { + if (output->type == kTfLiteInt8 || output->type == kTfLiteUInt8) { + tflite::ArithmeticParams op_params; + SetActivationParams(data.output_activation_min, data.output_activation_max, + &op_params); + op_params.input1_offset = -data.input1_zero_point; + op_params.input2_offset = -data.input2_zero_point; + op_params.output_offset = data.output_zero_point; + op_params.output_multiplier = data.output_multiplier; + op_params.output_shift = data.output_shift; + bool need_broadcast = reference_ops::ProcessBroadcastShapes( + tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorShape(input2), &op_params); + +#define TF_LITE_MUL(type, opname, dtype) \ + type::opname(op_params, tflite::micro::GetTensorShape(input1), \ + tflite::micro::GetTensorData(input1), \ + tflite::micro::GetTensorShape(input2), \ + tflite::micro::GetTensorData(input2), \ + tflite::micro::GetTensorShape(output), \ + tflite::micro::GetTensorData(output)); + + if (output->type == kTfLiteInt8) { + if (need_broadcast) { + TF_LITE_MUL(reference_integer_ops, BroadcastMul4DSlow, int8_t); + } else { + arm_elementwise_mul_s8( + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorData(input2), + op_params.input1_offset, op_params.input2_offset, + tflite::micro::GetTensorData(output), + op_params.output_offset, op_params.output_multiplier, + op_params.output_shift, op_params.quantized_activation_min, + op_params.quantized_activation_max, + MatchingElementsSize(tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorShape(output))); + } + } else if (output->type == kTfLiteUInt8) { + if (need_broadcast) { + TF_LITE_MUL(reference_ops, BroadcastMul4DSlow, uint8_t); + } else { + TF_LITE_MUL(reference_ops, Mul, uint8_t); + } + } +#undef TF_LITE_MUL + } +} + +void EvalFloat(TfLiteContext* context, TfLiteNode* node, + TfLiteMulParams* params, const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) { + float output_activation_min, output_activation_max; + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); + tflite::ArithmeticParams op_params; + SetActivationParams(output_activation_min, output_activation_max, &op_params); + + bool need_broadcast = reference_ops::ProcessBroadcastShapes( + tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorShape(input2), &op_params); +#define TF_LITE_MUL(opname) \ + reference_ops::opname(op_params, tflite::micro::GetTensorShape(input1), \ + tflite::micro::GetTensorData(input1), \ + tflite::micro::GetTensorShape(input2), \ + tflite::micro::GetTensorData(input2), \ + tflite::micro::GetTensorShape(output), \ + tflite::micro::GetTensorData(output)); + + if (need_broadcast) { + TF_LITE_MUL(BroadcastMul4DSlow); + } else { + TF_LITE_MUL(Mul); + } +#undef TF_LITE_MUL +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInput1Tensor); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInput2Tensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + switch (input1->type) { + case kTfLiteUInt8: + case kTfLiteInt8: + EvalQuantized(context, node, params, data, input1, input2, output); + break; + case kTfLiteFloat32: + EvalFloat(context, node, params, input1, input2, output); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input1->type), input1->type); + return kTfLiteError; + } + + return kTfLiteOk; +} +} // namespace mul + +TfLiteRegistration Register_MUL() { + return {/* Init=*/mul::Init, + /* Free=*/nullptr, + /* Prepare=*/mul::Prepare, + /*invoke=*/mul::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cpp new file mode 100644 index 0000000..396ab62 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cpp @@ -0,0 +1,400 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/pooling.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/padding.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace pooling { + +namespace { + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +struct OpData { + TfLitePaddingValues padding; + // Index to buffer for optimizations if applicable. + int buffer_idx; + + int32_t activation_min; + int32_t activation_max; +}; + +TfLiteStatus CalculateOpData(TfLiteContext* context, + const TfLitePoolParams* params, + const TfLiteTensor* input, TfLiteTensor* output, + OpData* data) { + // input: batch, height, width, channel + int height = SizeOfDimension(input, 1); + int width = SizeOfDimension(input, 2); + + int out_height, out_width; + + data->padding = ComputePaddingHeightWidth( + params->stride_height, params->stride_width, + /*dilation_rate_height=*/1, + /*dilation_rate_width=*/1, height, width, params->filter_height, + params->filter_width, params->padding, &out_height, &out_width); + + if (input->type != kTfLiteFloat32) { + TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( + context, params->activation, output, &data->activation_min, + &data->activation_max)); + TFLITE_DCHECK_LE(data->activation_min, data->activation_max); + } + + // Set buffer index to a reset value + data->buffer_idx = -1; + + return kTfLiteOk; +} + +void AverageEvalFloat(const TfLiteContext* context, const TfLiteNode* node, + const TfLitePoolParams* params, const OpData& data, + const TfLiteEvalTensor* input, TfLiteEvalTensor* output) { + float activation_min, activation_max; + CalculateActivationRange(params->activation, &activation_min, + &activation_max); + + PoolParams op_params; + op_params.stride_height = params->stride_height; + op_params.stride_width = params->stride_width; + op_params.filter_height = params->filter_height; + op_params.filter_width = params->filter_width; + op_params.padding_values.height = data.padding.height; + op_params.padding_values.width = data.padding.width; + op_params.float_activation_min = activation_min; + op_params.float_activation_max = activation_max; + reference_ops::AveragePool(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); +} + +void AverageEvalQuantized(TfLiteContext* context, const TfLiteNode* node, + const TfLitePoolParams* params, const OpData& data, + const TfLiteEvalTensor* input, + TfLiteEvalTensor* output) { + TFLITE_DCHECK(input->type == kTfLiteUInt8 || input->type == kTfLiteInt8); + + if (input->type == kTfLiteUInt8) { + PoolParams op_params; + op_params.stride_height = params->stride_height; + op_params.stride_width = params->stride_width; + op_params.filter_height = params->filter_height; + op_params.filter_width = params->filter_width; + op_params.padding_values.height = data.padding.height; + op_params.padding_values.width = data.padding.width; + op_params.quantized_activation_min = data.activation_min; + op_params.quantized_activation_max = data.activation_max; + + reference_ops::AveragePool(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + RuntimeShape input_shape = tflite::micro::GetTensorShape(input); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + + cmsis_nn_dims input_dims; + input_dims.n = 1; + input_dims.h = input_shape.Dims(1); + input_dims.w = input_shape.Dims(2); + input_dims.c = depth; + + cmsis_nn_dims output_dims; + output_dims.n = 1; + output_dims.h = output_shape.Dims(1); + output_dims.w = output_shape.Dims(2); + output_dims.c = depth; + + cmsis_nn_pool_params pool_params; + pool_params.stride.h = params->stride_height; + pool_params.stride.w = params->stride_width; + pool_params.padding.h = data.padding.height; + pool_params.padding.w = data.padding.width; + pool_params.activation.min = data.activation_min; + pool_params.activation.max = data.activation_max; + + cmsis_nn_dims filter_dims; + filter_dims.n = 1; + filter_dims.h = params->filter_height; + filter_dims.w = params->filter_width; + filter_dims.c = 1; + + cmsis_nn_context ctx; + ctx.buf = nullptr; + ctx.size = 0; + if (data.buffer_idx > -1) { + ctx.buf = context->GetScratchBuffer(context, data.buffer_idx); + } + + TFLITE_DCHECK_EQ( + arm_avgpool_s8(&ctx, &pool_params, &input_dims, + tflite::micro::GetTensorData(input), + &filter_dims, &output_dims, + tflite::micro::GetTensorData(output)), + ARM_MATH_SUCCESS); + } +} + +void MaxEvalFloat(TfLiteContext* context, TfLiteNode* node, + TfLitePoolParams* params, const OpData& data, + const TfLiteEvalTensor* input, TfLiteEvalTensor* output) { + float activation_min, activation_max; + CalculateActivationRange(params->activation, &activation_min, + &activation_max); + tflite::PoolParams op_params; + op_params.stride_height = params->stride_height; + op_params.stride_width = params->stride_width; + op_params.filter_height = params->filter_height; + op_params.filter_width = params->filter_width; + op_params.padding_values.height = data.padding.height; + op_params.padding_values.width = data.padding.width; + op_params.float_activation_min = activation_min; + op_params.float_activation_max = activation_max; + reference_ops::MaxPool(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); +} + +void MaxEvalQuantizedUInt8(TfLiteContext* context, TfLiteNode* node, + TfLitePoolParams* params, const OpData& data, + const TfLiteEvalTensor* input, + TfLiteEvalTensor* output) { + tflite::PoolParams op_params; + op_params.stride_height = params->stride_height; + op_params.stride_width = params->stride_width; + op_params.filter_height = params->filter_height; + op_params.filter_width = params->filter_width; + op_params.padding_values.height = data.padding.height; + op_params.padding_values.width = data.padding.width; + op_params.quantized_activation_min = data.activation_min; + op_params.quantized_activation_max = data.activation_max; + reference_ops::MaxPool(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); +} + +TfLiteStatus MaxEvalInt8(TfLiteContext* context, const TfLiteNode* node, + const TfLitePoolParams* params, const OpData& data, + const TfLiteEvalTensor* input, + TfLiteEvalTensor* output) { + RuntimeShape input_shape = tflite::micro::GetTensorShape(input); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + + cmsis_nn_dims input_dims; + input_dims.n = 1; + input_dims.h = input_shape.Dims(1); + input_dims.w = input_shape.Dims(2); + input_dims.c = depth; + + cmsis_nn_dims output_dims; + output_dims.n = 1; + output_dims.h = output_shape.Dims(1); + output_dims.w = output_shape.Dims(2); + output_dims.c = depth; + + cmsis_nn_pool_params pool_params; + pool_params.stride.h = params->stride_height; + pool_params.stride.w = params->stride_width; + pool_params.padding.h = data.padding.height; + pool_params.padding.w = data.padding.width; + pool_params.activation.min = data.activation_min; + pool_params.activation.max = data.activation_max; + + cmsis_nn_dims filter_dims; + filter_dims.n = 1; + filter_dims.h = params->filter_height; + filter_dims.w = params->filter_width; + filter_dims.c = 1; + + cmsis_nn_context ctx; + ctx.buf = nullptr; + ctx.size = 0; + if (data.buffer_idx > -1) { + ctx.buf = context->GetScratchBuffer(context, data.buffer_idx); + } + + TFLITE_DCHECK_EQ( + arm_max_pool_s8(&ctx, &pool_params, &input_dims, + tflite::micro::GetTensorData(input), &filter_dims, + &output_dims, + tflite::micro::GetTensorData(output)), + ARM_MATH_SUCCESS); + + return kTfLiteOk; +} + +} // namespace + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus MaxPrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + OpData* data = static_cast(node->user_data); + auto* params = reinterpret_cast(node->builtin_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + TF_LITE_ENSURE_STATUS(CalculateOpData(context, params, input, output, data)); + + return kTfLiteOk; +} + +TfLiteStatus AveragePrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + OpData* data = static_cast(node->user_data); + auto* params = reinterpret_cast(node->builtin_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + TF_LITE_ENSURE_STATUS(CalculateOpData(context, params, input, output, data)); + + if (input->type == kTfLiteInt8) { + RuntimeShape input_shape = GetTensorShape(input); + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + + RuntimeShape output_shape = GetTensorShape(output); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + const int depth = MatchingDim(input_shape, 3, output_shape, 3); + const int output_width = output_shape.Dims(2); + + const int32_t buffer_size = + arm_avgpool_s8_get_buffer_size(output_width, depth); + + if (buffer_size > 0) { + TF_LITE_ENSURE_STATUS(context->RequestScratchBufferInArena( + context, buffer_size, &data->buffer_idx)); + } else { + data->buffer_idx = -1; + } + } + return kTfLiteOk; +} + +TfLiteStatus AverageEval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + + const OpData& data = *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + // Inputs and outputs share the same type, guaranteed by the converter. + switch (input->type) { + case kTfLiteFloat32: + AverageEvalFloat(context, node, params, data, input, output); + break; + case kTfLiteUInt8: + case kTfLiteInt8: + AverageEvalQuantized(context, node, params, data, input, output); + break; + default: + TF_LITE_KERNEL_LOG(context, "Input type %s is not currently supported", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus MaxEval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + + const OpData& data = *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + switch (input->type) { + case kTfLiteFloat32: + MaxEvalFloat(context, node, params, data, input, output); + break; + case kTfLiteUInt8: + MaxEvalQuantizedUInt8(context, node, params, data, input, output); + break; + case kTfLiteInt8: + MaxEvalInt8(context, node, params, data, input, output); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s not currently supported.", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace pooling + +TfLiteRegistration Register_AVERAGE_POOL_2D() { + return {/*init=*/pooling::Init, + /*free=*/nullptr, + /*prepare=*/pooling::AveragePrepare, + /*invoke=*/pooling::AverageEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_MAX_POOL_2D() { + return {/*init=*/pooling::Init, + /*free=*/nullptr, + /*prepare=*/pooling::MaxPrepare, + /*invoke=*/pooling::MaxEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cpp new file mode 100644 index 0000000..2fbcbed --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cpp @@ -0,0 +1,235 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/softmax.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace { + +// Softmax parameter data that persists in user_data +static constexpr int kInt16LUTArraySize = 513; + +TfLiteStatus CalculateSoftmaxParams(TfLiteContext* context, + const TfLiteTensor* input, + TfLiteTensor* output, + const TfLiteSoftmaxParams* params, + SoftmaxParams* op_data) { + if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8 || + input->type == kTfLiteInt16) { + if (input->type == kTfLiteUInt8) { + TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteUInt8); + TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0); + } else if (input->type == kTfLiteInt16) { + TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0); + TF_LITE_ENSURE_NEAR(context, output->params.scale, 1.f / 32768, + (0.001f * 1.f / 32768)); + } else { // input->type == kTfLiteInt8 + TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteInt8); + if (output->type == kTfLiteInt16) { + TF_LITE_ENSURE_EQ(context, output->params.zero_point, -32768); + TF_LITE_ENSURE_NEAR(context, output->params.scale, 1.f / 65536, + (0.001f * 1.f / 65536)); + } else { // output->type == kTfLiteint8 + TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteInt8); + TF_LITE_ENSURE_EQ(context, output->params.zero_point, -128); + TF_LITE_ENSURE(context, output->params.scale == 1.f / 256); + } + } + + static const int kScaledDiffIntegerBits = 5; + + // Calculate input_multiplier and input_left_shift + if (input->type == kTfLiteInt16) { + int input_left_shift; + double input_scale_beta_rescale = + static_cast(input->params.scale) * + static_cast(params->beta) / + (10.0 / 65535.0); // scale the input_diff such that [-65535, 0] + // correspond to [-10.0, 0.0] + QuantizeMultiplier(input_scale_beta_rescale, &op_data->input_multiplier, + &input_left_shift); + op_data->input_left_shift = input_left_shift; + } else { + int input_left_shift; + tflite::PreprocessSoftmaxScaling( + static_cast(params->beta), + static_cast(input->params.scale), kScaledDiffIntegerBits, + &op_data->input_multiplier, &input_left_shift); + op_data->input_left_shift = input_left_shift; + op_data->diff_min = + -1.0 * tflite::CalculateInputRadius(kScaledDiffIntegerBits, + op_data->input_left_shift); + } + } else { + TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteFloat32); + TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteFloat32); + op_data->beta = static_cast(params->beta); + } + return kTfLiteOk; +} + +void* SoftmaxInit(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(SoftmaxParams)); +} + +TfLiteStatus SoftmaxPrepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + const TfLiteTensor* input = GetInput(context, node, 0); + TF_LITE_ENSURE(context, input != nullptr); + TF_LITE_ENSURE(context, NumDimensions(input) >= 1); + TfLiteTensor* output = GetOutput(context, node, 0); + TF_LITE_ENSURE(context, output != nullptr); + + TF_LITE_ENSURE(context, node->user_data != nullptr); + SoftmaxParams* op_data = static_cast(node->user_data); + // Only allocate LUTs for KTfLiteInt16 data type + if (input->type == kTfLiteInt16) { + void* raw_exp_lut = context->AllocatePersistentBuffer( + context, sizeof(int16_t) * kInt16LUTArraySize); + TF_LITE_ENSURE(context, raw_exp_lut != nullptr); + op_data->exp_lut = reinterpret_cast(raw_exp_lut); + void* one_over_one_plus_x_lut = context->AllocatePersistentBuffer( + context, sizeof(int16_t) * kInt16LUTArraySize); + TF_LITE_ENSURE(context, one_over_one_plus_x_lut != nullptr); + op_data->one_over_one_plus_x_lut = + reinterpret_cast(one_over_one_plus_x_lut); + } + + if (output->type == kTfLiteInt16) { + TF_LITE_ENSURE(context, input->type == kTfLiteInt8 || + input->type == kTfLiteUInt8 || + input->type == kTfLiteInt16); + } else { + TF_LITE_ENSURE_EQ(context, input->type, output->type); + } + + // Populate LUT if required + if (input->type == kTfLiteInt16) { + TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0); + // exp LUT only used on negative values + // we consider exp(-10.0) is insignificant to accumulation + gen_lut([](float value) { return std::exp(value); }, -10.0f, 0.0f, + op_data->exp_lut, kInt16LUTArraySize); + gen_lut([](float value) { return 1.0f / (1.0f + value); }, 0.0f, 1.0f, + op_data->one_over_one_plus_x_lut, kInt16LUTArraySize); + op_data->zero_point = output->params.zero_point; + op_data->scale = output->params.scale; + } + + auto* params = static_cast(node->builtin_data); + return CalculateSoftmaxParams(context, input, output, params, op_data); +} + +// Takes a tensor and performs softmax along the last dimension. +void SoftmaxFloat(const TfLiteEvalTensor* input, TfLiteEvalTensor* output, + const SoftmaxParams& op_data) { + tflite::reference_ops::Softmax(op_data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); +} + +void SoftmaxQuantized(const TfLiteEvalTensor* input, TfLiteEvalTensor* output, + const SoftmaxParams& op_data) { + if (input->type == kTfLiteUInt8) { + tflite::reference_ops::Softmax( + op_data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else if (input->type == kTfLiteInt8) { + if (output->type == kTfLiteInt16) { + tflite::reference_ops::Softmax( + op_data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + const auto input_shape = tflite::micro::GetTensorShape(input); + const auto output_shape = tflite::micro::GetTensorShape(output); + const int trailing_dim = input_shape.DimensionsCount() - 1; + const int outer_size = + MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); + const int depth = + MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); + + arm_softmax_s8(tflite::micro::GetTensorData(input), outer_size, + depth, op_data.input_multiplier, op_data.input_left_shift, + op_data.diff_min, + tflite::micro::GetTensorData(output)); + } + } else { + tflite::reference_ops::SoftmaxInt16( + op_data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } +} + +TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0); + + TFLITE_DCHECK(node->user_data != nullptr); + const SoftmaxParams data = + *static_cast(node->user_data); + + switch (input->type) { + case kTfLiteFloat32: { + SoftmaxFloat(input, output, data); + return kTfLiteOk; + } + case kTfLiteInt8: + case kTfLiteUInt8: + case kTfLiteInt16: { + SoftmaxQuantized(input, output, data); + return kTfLiteOk; + } + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } +} + +} // namespace + +TfLiteRegistration Register_SOFTMAX() { + return {/*init=*/SoftmaxInit, + /*free=*/nullptr, + /*prepare=*/SoftmaxPrepare, + /*invoke=*/SoftmaxEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/svdf.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/svdf.cpp new file mode 100644 index 0000000..3599a55 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/cmsis-nn/svdf.cpp @@ -0,0 +1,479 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/activation_utils.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace { + +struct OpData { + int32_t effective_scale_1_a; + int32_t effective_scale_2_a; + // b versions of each scale are kept at int since the numbers are just the + // shift value - typically between [-32, 32]. + int effective_scale_1_b; + int effective_scale_2_b; + int scratch_tensor_index; + int scratch_output_tensor_index; + + // Cached tensor zero point values for quantized operations. + int input_zero_point; + int output_zero_point; +}; + +// Input tensors. +constexpr int kInputTensor = 0; +constexpr int kWeightsFeatureTensor = 1; +constexpr int kWeightsTimeTensor = 2; +constexpr int kBiasTensor = 3; +// This is a variable tensor, and will be modified by this op. +constexpr int kInputActivationStateTensor = 4; + +// Output tensor. +constexpr int kOutputTensor = 0; + +/** + * This version of SVDF is specific to TFLite Micro. It contains the following + * differences between the TFLite version: + * + * 1.) Scratch tensor allocation - scratch tensors must be known ahead of time + * for the Micro interpreter. + * 2.) Output dimensions - the TFLite version determines output size and runtime + * and resizes the output tensor. Micro runtime does not support tensor + * resizing. + */ +static inline void ApplyTimeWeightsBiasAndActivation( + int batch_size, int memory_size, int num_filters, int num_units, int rank, + const float* const __restrict__ weights_time_ptr, + const float* const __restrict__ bias_ptr, TfLiteFusedActivation activation, + float* const __restrict__ state_ptr, float* const __restrict__ scratch_ptr, + float* const __restrict__ output_ptr) { + // Compute matmul(activation_state, weights_time). + for (int b = 0; b < batch_size; ++b) { + // Perform batched vector dot product: + float* scratch_ptr_batch = scratch_ptr + b * num_filters; + const float* vector1_ptr = weights_time_ptr; + const float* vector2_ptr = state_ptr + b * memory_size * num_filters; + for (int i = 0; i < num_filters; ++i) { + *scratch_ptr_batch = 0.f; + for (int j = 0; j < memory_size; ++j) { + *scratch_ptr_batch += *vector1_ptr++ * *vector2_ptr++; + } + scratch_ptr_batch++; + } + } + + // Initialize output with bias if provided. + if (bias_ptr) { + // VectorBatchVectorAssign + for (int i = 0; i < batch_size; ++i) { + float* output_data = output_ptr + i * num_units; + const float* bias_data = bias_ptr; + for (int j = 0; j < num_units; ++j) { + *output_data++ = *bias_data++; + } + } + } else { + float* output_data = output_ptr; + for (int i = 0; i < batch_size * num_units; ++i) { + *output_data++ = 0.0f; + } + } + + // Reduction sum. + for (int b = 0; b < batch_size; ++b) { + float* output_ptr_batch = output_ptr + b * num_units; + float* scratch_ptr_batch = scratch_ptr + b * num_filters; + + // Reduction sum vector + for (int i = 0; i < num_units; ++i) { + for (int j = 0; j < rank; j++) { + output_ptr_batch[i] += *scratch_ptr_batch++; + } + } + } + + // Apply activation. + for (int b = 0; b < batch_size; ++b) { + float* output_ptr_batch = output_ptr + b * num_units; + for (int i = 0; i < num_units; ++i) { + *output_ptr_batch = + tflite::ops::micro::ActivationValFloat(activation, *output_ptr_batch); + ++output_ptr_batch; + } + } +} + +inline void EvalFloatSVDF( + TfLiteContext* context, TfLiteNode* node, const TfLiteEvalTensor* input, + const TfLiteEvalTensor* weights_feature, + const TfLiteEvalTensor* weights_time, const TfLiteEvalTensor* bias, + const TfLiteSVDFParams* params, int scratch_tensor_index, + TfLiteEvalTensor* activation_state, TfLiteEvalTensor* output) { + const int rank = params->rank; + const int batch_size = input->dims->data[0]; + const int input_size = input->dims->data[1]; + const int num_filters = weights_feature->dims->data[0]; + const int num_units = num_filters / rank; + const int memory_size = weights_time->dims->data[1]; + + const float* weights_feature_ptr = + tflite::micro::GetTensorData(weights_feature); + const float* weights_time_ptr = + tflite::micro::GetTensorData(weights_time); + const float* bias_ptr = tflite::micro::GetTensorData(bias); + const float* input_ptr = tflite::micro::GetTensorData(input); + + float* state_ptr = tflite::micro::GetTensorData(activation_state); + + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(context->GetScratchBuffer != nullptr); + + float* scratch_ptr = static_cast( + context->GetScratchBuffer(context, scratch_tensor_index)); + + float* output_ptr = tflite::micro::GetTensorData(output); + + // Left shift the activation_state. + { + float* new_state_start = state_ptr; + const float* old_state_start = state_ptr + 1; + const float* old_state_end = + state_ptr + batch_size * num_filters * memory_size; + while (old_state_start != old_state_end) { + *new_state_start++ = *old_state_start++; + } + } + + // Note: no need to clear the latest activation, matmul is not accumulative. + + // Compute conv1d(inputs, weights_feature). + // The activation_state's rightmost column is used to save current cycle + // activation. This is achieved by starting at state_ptr[memory_size - 1] and + // having the stride equal to memory_size. + + // Perform batched matrix vector multiply operation: + { + const float* matrix = weights_feature_ptr; + const float* vector = input_ptr; + float* result = &state_ptr[memory_size - 1]; + float* result_in_batch = result; + for (int i = 0; i < batch_size; ++i) { + const float* matrix_ptr = matrix; + for (int j = 0; j < num_filters; ++j) { + float dot_prod = 0.0f; + const float* vector_in_batch = vector + i * input_size; + for (int k = 0; k < input_size; ++k) { + dot_prod += *matrix_ptr++ * *vector_in_batch++; + } + *result_in_batch = dot_prod; + result_in_batch += memory_size; + } + } + } + + ApplyTimeWeightsBiasAndActivation( + batch_size, memory_size, num_filters, num_units, rank, weights_time_ptr, + bias_ptr, params->activation, state_ptr, scratch_ptr, output_ptr); +} + +void EvalIntegerSVDF(TfLiteContext* context, TfLiteNode* node, + const TfLiteEvalTensor* input_tensor, + const TfLiteEvalTensor* weights_feature_tensor, + const TfLiteEvalTensor* weights_time_tensor, + const TfLiteEvalTensor* bias_tensor, + const TfLiteSVDFParams* params, + TfLiteEvalTensor* activation_state_tensor, + TfLiteEvalTensor* output_tensor, const OpData& data) { + cmsis_nn_dims input_dims; + input_dims.n = input_tensor->dims->data[0]; + input_dims.h = input_tensor->dims->data[1]; + + cmsis_nn_dims weights_feature_dims; + weights_feature_dims.n = weights_feature_tensor->dims->data[0]; + weights_feature_dims.h = weights_feature_tensor->dims->data[1]; + + cmsis_nn_dims weights_time_dims; + weights_time_dims.n = weights_time_tensor->dims->data[0]; + weights_time_dims.h = weights_time_tensor->dims->data[1]; + + cmsis_nn_dims bias_dims; + bias_dims.n = bias_tensor->dims->data[0]; + + cmsis_nn_dims state_dims; + state_dims.n = bias_tensor->dims->data[0]; + state_dims.h = bias_tensor->dims->data[1]; + + cmsis_nn_dims output_dims; + output_dims.n = output_tensor->dims->data[0]; + output_dims.h = output_tensor->dims->data[1]; + + cmsis_nn_svdf_params svdf_params; + svdf_params.rank = params->rank; + svdf_params.input_offset = data.input_zero_point; + svdf_params.output_offset = data.output_zero_point; + + svdf_params.input_activation.min = INT16_MIN; + svdf_params.input_activation.max = INT16_MAX; + + svdf_params.output_activation.min = INT8_MIN; + svdf_params.output_activation.max = INT8_MAX; + + cmsis_nn_per_tensor_quant_params in_quant_params; + in_quant_params.multiplier = data.effective_scale_1_a; + in_quant_params.shift = data.effective_scale_1_b; + + cmsis_nn_per_tensor_quant_params out_quant_params; + out_quant_params.multiplier = data.effective_scale_2_a; + out_quant_params.shift = data.effective_scale_2_b; + + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(context->GetScratchBuffer != nullptr); + + cmsis_nn_context scratch_ctx; + scratch_ctx.buf = static_cast( + context->GetScratchBuffer(context, data.scratch_tensor_index)); + + cmsis_nn_context scratch_output_ctx; + scratch_output_ctx.buf = static_cast( + context->GetScratchBuffer(context, data.scratch_output_tensor_index)); + + int8_t* output_data = tflite::micro::GetTensorData(output_tensor); + arm_svdf_s8( + &scratch_ctx, &scratch_output_ctx, &svdf_params, &in_quant_params, + &out_quant_params, &input_dims, + (int8_t*)tflite::micro::GetTensorData(input_tensor), &state_dims, + (int16_t*)tflite::micro::GetTensorData(activation_state_tensor), + &weights_feature_dims, + (int8_t*)tflite::micro::GetTensorData(weights_feature_tensor), + &weights_time_dims, + (int16_t*)tflite::micro::GetTensorData(weights_time_tensor), + &bias_dims, (int32_t*)tflite::micro::GetTensorData(bias_tensor), + &output_dims, output_data); +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->builtin_data != nullptr); + + const auto* params = static_cast(node->builtin_data); + + // Validate Tensor Inputs (dtype depends on quantization): + // [0] = Input, {2, batch_size, input_size} + // [1] = Weights Feature, {2, num_filters, input_size} + // [2] = Weights Time, {2, num_filters, memory_size} + // [3] = Bias (optional), {1, num_units} + // [4] = Activation State (variable), + // {2, batch_size, memory_size * num_filters} + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* weights_feature = + GetInput(context, node, kWeightsFeatureTensor); + const TfLiteTensor* weights_time = + GetInput(context, node, kWeightsTimeTensor); + const TfLiteTensor* bias = GetOptionalInputTensor(context, node, kBiasTensor); + const TfLiteTensor* activation_state = + GetInput(context, node, kInputActivationStateTensor); + + // Define input constants based on input tensor definition above: + const int rank = params->rank; + const int input_size = input->dims->data[1]; + const int batch_size = input->dims->data[0]; + const int num_filters = weights_feature->dims->data[0]; + TF_LITE_ENSURE_EQ(context, num_filters % rank, 0); + const int num_units = num_filters / rank; + const int memory_size = weights_time->dims->data[1]; + + // Validate Input Tensor: + TF_LITE_ENSURE(context, + input->type == kTfLiteFloat32 || input->type == kTfLiteInt8); + TF_LITE_ENSURE_EQ(context, NumDimensions(input), 2); + + // Validate Tensor Output: + // [0] = float/int8, {2, batch_size, num_units} + TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE_EQ(context, NumDimensions(output), 2); + TF_LITE_ENSURE_EQ(context, output->dims->data[0], batch_size); + TF_LITE_ENSURE_EQ(context, output->dims->data[1], num_units); + + // Validate Weights Feature Input Tensor: + TF_LITE_ENSURE_EQ(context, NumDimensions(weights_feature), 2); + TF_LITE_ENSURE_EQ(context, weights_feature->dims->data[1], input_size); + + // Validate Weights Time Input Tensor: + TF_LITE_ENSURE_EQ(context, NumDimensions(weights_time), 2); + TF_LITE_ENSURE_EQ(context, weights_time->dims->data[0], num_filters); + TF_LITE_ENSURE_EQ(context, weights_time->dims->data[1], memory_size); + + // Validate Optional Bias Input Tensor: + if (bias != nullptr) { + TF_LITE_ENSURE_EQ(context, bias->dims->data[0], num_units); + } + + // Validate Activation State Input Tensor: + TF_LITE_ENSURE_EQ(context, NumDimensions(activation_state), 2); + TF_LITE_ENSURE_EQ(context, activation_state->dims->data[0], batch_size); + TF_LITE_ENSURE_EQ(context, activation_state->dims->data[1], + memory_size * num_filters); + // Since is_variable is not part of TFLiteEvalTensor, check is_variable here. + TF_LITE_ENSURE_EQ(context, activation_state->is_variable, true); + + TF_LITE_ENSURE_EQ(context, node->inputs->size, 5); + + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + if (input->type == kTfLiteInt8) { + TF_LITE_ENSURE_EQ(context, weights_feature->type, kTfLiteInt8); + TF_LITE_ENSURE_EQ(context, weights_time->type, kTfLiteInt16); + TF_LITE_ENSURE_EQ(context, activation_state->type, kTfLiteInt16); + if (bias != nullptr) { + TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteInt32); + } + + TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteInt8); + + const double effective_scale_1 = static_cast( + input->params.scale * weights_feature->params.scale / + activation_state->params.scale); + const double effective_scale_2 = + static_cast(activation_state->params.scale * + weights_time->params.scale / output->params.scale); + + // TODO(b/162018098): Use TF_LITE_ENSURE_NEAR when it is ready. + TF_LITE_ENSURE( + context, + std::abs(static_cast(bias->params.scale) - + static_cast(activation_state->params.scale * + weights_time->params.scale)) < 1e-5); + + QuantizeMultiplier(effective_scale_1, &(data->effective_scale_1_a), + &(data->effective_scale_1_b)); + QuantizeMultiplier(effective_scale_2, &(data->effective_scale_2_a), + &(data->effective_scale_2_b)); + + data->input_zero_point = input->params.zero_point; + data->output_zero_point = output->params.zero_point; + + TFLITE_DCHECK(context->RequestScratchBufferInArena != nullptr); + + const TfLiteStatus scratch_status = context->RequestScratchBufferInArena( + context, batch_size * num_filters * sizeof(int32_t), + &(data->scratch_tensor_index)); + TF_LITE_ENSURE_OK(context, scratch_status); + + const TfLiteStatus scratch_output_status = + context->RequestScratchBufferInArena( + context, batch_size * num_units * sizeof(int32_t), + &(data->scratch_output_tensor_index)); + TF_LITE_ENSURE_OK(context, scratch_output_status); + } else { + TF_LITE_ENSURE_EQ(context, weights_feature->type, kTfLiteFloat32); + TF_LITE_ENSURE_EQ(context, weights_time->type, kTfLiteFloat32); + TF_LITE_ENSURE_EQ(context, activation_state->type, kTfLiteFloat32); + if (bias != nullptr) { + TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteFloat32); + } + TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteFloat32); + + TFLITE_DCHECK(context->RequestScratchBufferInArena != nullptr); + const TfLiteStatus scratch_status = context->RequestScratchBufferInArena( + context, batch_size * num_filters * sizeof(float), + &(data->scratch_tensor_index)); + TF_LITE_ENSURE_OK(context, scratch_status); + } + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* weights_feature = + tflite::micro::GetEvalInput(context, node, kWeightsFeatureTensor); + const TfLiteEvalTensor* weights_time = + tflite::micro::GetEvalInput(context, node, kWeightsTimeTensor); + const TfLiteEvalTensor* bias = + (NumInputs(node) == 5) + ? tflite::micro::GetEvalInput(context, node, kBiasTensor) + : nullptr; + TfLiteEvalTensor* activation_state = tflite::micro::GetMutableEvalInput( + context, node, kInputActivationStateTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + switch (weights_feature->type) { + case kTfLiteFloat32: { + EvalFloatSVDF(context, node, input, weights_feature, weights_time, bias, + params, data.scratch_tensor_index, activation_state, + output); + return kTfLiteOk; + break; + } + + case kTfLiteInt8: { + EvalIntegerSVDF(context, node, input, weights_feature, weights_time, bias, + params, activation_state, output, data); + return kTfLiteOk; + break; + } + + default: + TF_LITE_KERNEL_LOG(context, "Type %s not currently supported.", + TfLiteTypeGetName(weights_feature->type)); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace + +TfLiteRegistration Register_SVDF() { + return {/*init=*/Init, + /*free=*/nullptr, + /*prepare=*/Prepare, + /*invoke=*/Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/comparisons.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/comparisons.cpp new file mode 100644 index 0000000..eb8d5c5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/comparisons.cpp @@ -0,0 +1,727 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/comparisons.h" + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace comparisons { +namespace { + +struct OpData { + ComparisonParams params; +}; + +constexpr int kInputTensor1 = 0; +constexpr int kInputTensor2 = 1; +constexpr int kOutputTensor = 0; + +TfLiteStatus EqualEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + RuntimeShape input1_shape = tflite::micro::GetTensorShape(input1); + RuntimeShape input2_shape = tflite::micro::GetTensorShape(input2); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + bool* output_data = tflite::micro::GetTensorData(output); + + bool requires_broadcast = !tflite::micro::HaveSameShapes(input1, input2); + switch (input1->type) { + case kTfLiteBool: + requires_broadcast + ? reference_ops::Broadcast4DSlowEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::EqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteFloat32: + requires_broadcast + ? reference_ops::Broadcast4DSlowEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::EqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt32: + requires_broadcast + ? reference_ops::Broadcast4DSlowEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::EqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt64: + requires_broadcast + ? reference_ops::Broadcast4DSlowEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::EqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteUInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::EqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::EqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input1->type), input1->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +// TODO(renjieliu): Refactor the logic to avoid duplications. +TfLiteStatus NotEqualEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + RuntimeShape input1_shape = tflite::micro::GetTensorShape(input1); + RuntimeShape input2_shape = tflite::micro::GetTensorShape(input2); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + bool* output_data = tflite::micro::GetTensorData(output); + + bool requires_broadcast = !tflite::micro::HaveSameShapes(input1, input2); + switch (input1->type) { + case kTfLiteBool: + requires_broadcast + ? reference_ops::Broadcast4DSlowNotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::NotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteFloat32: + requires_broadcast + ? reference_ops::Broadcast4DSlowNotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::NotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt32: + requires_broadcast + ? reference_ops::Broadcast4DSlowNotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::NotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt64: + requires_broadcast + ? reference_ops::Broadcast4DSlowNotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::NotEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteUInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowNotEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::NotEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowNotEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::NotEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input1->type), input1->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus GreaterEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + RuntimeShape input1_shape = tflite::micro::GetTensorShape(input1); + RuntimeShape input2_shape = tflite::micro::GetTensorShape(input2); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + bool* output_data = tflite::micro::GetTensorData(output); + + bool requires_broadcast = !tflite::micro::HaveSameShapes(input1, input2); + switch (input1->type) { + case kTfLiteFloat32: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt32: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt64: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteUInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input1->type), input1->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus GreaterEqualEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + RuntimeShape input1_shape = tflite::micro::GetTensorShape(input1); + RuntimeShape input2_shape = tflite::micro::GetTensorShape(input2); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + bool* output_data = tflite::micro::GetTensorData(output); + + bool requires_broadcast = !tflite::micro::HaveSameShapes(input1, input2); + switch (input1->type) { + case kTfLiteFloat32: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt32: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt64: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteUInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowGreaterEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::GreaterEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input1->type), input1->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus LessEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + RuntimeShape input1_shape = tflite::micro::GetTensorShape(input1); + RuntimeShape input2_shape = tflite::micro::GetTensorShape(input2); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + bool* output_data = tflite::micro::GetTensorData(output); + + bool requires_broadcast = !tflite::micro::HaveSameShapes(input1, input2); + switch (input1->type) { + case kTfLiteFloat32: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt32: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt64: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteUInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input1->type), input1->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus LessEqualEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + RuntimeShape input1_shape = tflite::micro::GetTensorShape(input1); + RuntimeShape input2_shape = tflite::micro::GetTensorShape(input2); + RuntimeShape output_shape = tflite::micro::GetTensorShape(output); + bool* output_data = tflite::micro::GetTensorData(output); + + bool requires_broadcast = !tflite::micro::HaveSameShapes(input1, input2); + switch (input1->type) { + case kTfLiteFloat32: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt32: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt64: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessEqualNoScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteUInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + case kTfLiteInt8: + requires_broadcast + ? reference_ops::Broadcast4DSlowLessEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data) + : reference_ops::LessEqualWithScaling( + data->params, input1_shape, + tflite::micro::GetTensorData(input1), input2_shape, + tflite::micro::GetTensorData(input2), output_shape, + output_data); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input1->type), input1->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1); + TF_LITE_ENSURE(context, input1 != nullptr); + const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2); + TF_LITE_ENSURE(context, input2 != nullptr); + + if (input1->type == kTfLiteUInt8 || input1->type == kTfLiteInt8) { + auto input1_offset = -input1->params.zero_point; + auto input2_offset = -input2->params.zero_point; + const int kLeftShift = 8; + + int32_t input1_multiplier; + int input1_shift; + QuantizeMultiplierSmallerThanOneExp( + static_cast(input1->params.scale), &input1_multiplier, + &input1_shift); + int32_t input2_multiplier; + int input2_shift; + QuantizeMultiplierSmallerThanOneExp( + static_cast(input2->params.scale), &input2_multiplier, + &input2_shift); + + data->params.left_shift = kLeftShift; + data->params.input1_offset = input1_offset; + data->params.input1_multiplier = input1_multiplier; + data->params.input1_shift = input1_shift; + data->params.input2_offset = input2_offset; + data->params.input2_multiplier = input2_multiplier; + data->params.input2_shift = input2_shift; + } + + return kTfLiteOk; +} + +} // namespace comparisons + +TfLiteRegistration Register_EQUAL() { + return {/*init=*/comparisons::Init, + /*free=*/nullptr, + /*prepare=*/comparisons::Prepare, + /*invoke=*/comparisons::EqualEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_NOT_EQUAL() { + return {/*init=*/comparisons::Init, + /*free=*/nullptr, + /*prepare=*/comparisons::Prepare, + /*invoke=*/comparisons::NotEqualEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_GREATER() { + return {/*init=*/comparisons::Init, + /*free=*/nullptr, + /*prepare=*/comparisons::Prepare, + /*invoke=*/comparisons::GreaterEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_GREATER_EQUAL() { + return {/*init=*/comparisons::Init, + /*free=*/nullptr, + /*prepare=*/comparisons::Prepare, + /*invoke=*/comparisons::GreaterEqualEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_LESS() { + return {/*init=*/comparisons::Init, + /*free=*/nullptr, + /*prepare=*/comparisons::Prepare, + /*invoke=*/comparisons::LessEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_LESS_EQUAL() { + return {/*init=*/comparisons::Init, + /*free=*/nullptr, + /*prepare=*/comparisons::Prepare, + /*invoke=*/comparisons::LessEqualEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/concatenation.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/concatenation.cpp new file mode 100644 index 0000000..7090ce5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/concatenation.cpp @@ -0,0 +1,279 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/concatenation.h" + +#include + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace concatenation { + +constexpr int kMaxInputNum = 10; // Maximum number of input tensors +constexpr int kOutputTensor = 0; + +struct OpData { + ConcatenationParams params; +}; + +// Handles negative axis index, coerces to positive index value. +inline int CalculatePositiveAxis(int axis, const TfLiteTensor* output_tensor) { + if (axis >= 0) { + return axis; + } else { + return NumDimensions(output_tensor) + axis; + } +} + +// The following functions are helpers to get tensor data in the format that the +// reference op implementation expects. They provide the same functionality as +// class VectorOfTensors and class VectorOfQuantizedTensors in TFLite. + +// Gets shapes from a list of tensors. +inline void GetAllInputTensorShapes(const TfLiteContext* context, + const TfLiteNode* node, + RuntimeShape all_shapes[kMaxInputNum]) { + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(node != nullptr); + for (int i = 0; i < node->inputs->size; ++i) { + const TfLiteEvalTensor* t = tflite::micro::GetEvalInput(context, node, i); + RuntimeShape shape = tflite::micro::GetTensorShape(t); + all_shapes[i].ReplaceWith(shape.DimensionsCount(), shape.DimsData()); + } +} + +// Get shape pointers from a list of shapes. +inline void GetShapesPointers(const RuntimeShape* shapes, size_t num, + const RuntimeShape* pointers[]) { + for (size_t i = 0; i < num; ++i) { + pointers[i] = &shapes[i]; + } +} + +// Gets data pointers from a list of tensors. +template +inline void GetAllInputTensorData(const TfLiteContext* context, + const TfLiteNode* node, + T* all_data[kMaxInputNum]) { + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(node != nullptr); + for (int i = 0; i < node->inputs->size; ++i) { + const TfLiteEvalTensor* t = tflite::micro::GetEvalInput(context, node, i); + all_data[i] = tflite::micro::GetTensorData(t); + } +} + +template +void EvalUnquantized(TfLiteContext* context, TfLiteNode* node) { + // Collect the shapes and data pointer of input tensors + RuntimeShape inputs_shape[kMaxInputNum]; + const RuntimeShape* inputs_shape_ptr[kMaxInputNum]; + const data_type* inputs_data[kMaxInputNum]; + GetAllInputTensorShapes(context, node, inputs_shape); + GetShapesPointers(inputs_shape, node->inputs->size, inputs_shape_ptr); + GetAllInputTensorData(context, node, inputs_data); + + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + reference_ops::Concatenation(data->params, inputs_shape_ptr, inputs_data, + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); +} + +void EvalQuantizedUInt8(TfLiteContext* context, TfLiteNode* node) { + // Collect the shapes and data pointer of input tensors + RuntimeShape inputs_shape[kMaxInputNum]; + const RuntimeShape* inputs_shape_ptr[kMaxInputNum]; + const uint8_t* inputs_data[kMaxInputNum]; + GetAllInputTensorShapes(context, node, inputs_shape); + GetShapesPointers(inputs_shape, node->inputs->size, inputs_shape_ptr); + GetAllInputTensorData(context, node, inputs_data); + + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + reference_ops::ConcatenationWithScaling( + data->params, inputs_shape_ptr, inputs_data, + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + // This function only checks the types. Additional shape validations are + // performed in the reference implementation called during Eval(). + const TfLiteConcatenationParams* params = + reinterpret_cast(node->builtin_data); + + const TfLiteTensor* input_tensor = GetInput(context, node, 0); + TF_LITE_ENSURE(context, input_tensor != nullptr); + TfLiteType input_type = input_tensor->type; + const TfLiteTensor* output_tensor = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output_tensor != nullptr); + TfLiteType output_type = output_tensor->type; + + // Check activation and input type + TF_LITE_ENSURE_EQ(context, params->activation, kTfLiteActNone); + TF_LITE_ENSURE(context, + input_type == kTfLiteFloat32 || input_type == kTfLiteUInt8 || + input_type == kTfLiteInt8 || input_type == kTfLiteInt32 || + input_type == kTfLiteInt64); + + // Output type must match input type + TF_LITE_ENSURE_EQ(context, output_type, input_type); + + // This implementation does not support large number of input tensors + const int num_inputs = NumInputs(node); + TF_LITE_ENSURE(context, num_inputs <= kMaxInputNum); + + // Shapes with dimensions >4 are not yet supported with static allocation. + for (int i = 0; i < num_inputs; ++i) { + const TfLiteTensor* input = GetInput(context, node, i); + TF_LITE_ENSURE(context, input != nullptr); + int num_dimensions = NumDimensions(input); + + if (num_dimensions > 4) { + TF_LITE_KERNEL_LOG( + context, + "Op Concatenation does not currently support num dimensions >4 " + "Tensor has %d dimensions.", + num_dimensions); + return kTfLiteError; + } + } + + // Calculate OpData. + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + + switch (output_type) { // Already know in/outtypes are same. + case kTfLiteFloat32: + case kTfLiteInt32: + case kTfLiteInt64: { + data->params.axis = CalculatePositiveAxis(params->axis, output); + data->params.inputs_count = node->inputs->size; + break; + } + case kTfLiteUInt8: + case kTfLiteInt8: { + data->params.axis = CalculatePositiveAxis(params->axis, output); + data->params.inputs_count = node->inputs->size; + + float* input_scales = + reinterpret_cast(context->AllocatePersistentBuffer( + context, node->inputs->size * sizeof(float))); + + int32_t* input_zero_points = + reinterpret_cast(context->AllocatePersistentBuffer( + context, node->inputs->size * sizeof(int32_t))); + + // Allocate persistent scale and zeropoint buffers. + // Store input scale and zero point values in OpParams: + for (int i = 0; i < node->inputs->size; ++i) { + const TfLiteTensor* t = GetInput(context, node, i); + TF_LITE_ENSURE(context, t != nullptr); + input_scales[i] = t->params.scale; + input_zero_points[i] = t->params.zero_point; + } + + data->params.input_scale = input_scales; + data->params.input_zeropoint = input_zero_points; + data->params.output_zeropoint = output->params.zero_point; + data->params.output_scale = output->params.scale; + break; + } + default: + TF_LITE_KERNEL_LOG( + context, "Op Concatenation does not currently support Type '%s'.", + TfLiteTypeGetName(output_type)); + return kTfLiteError; + } + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* output_tensor = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output_tensor != nullptr); + TfLiteType output_type = output_tensor->type; + + switch (output_type) { // Already know in/outtypes are same. + case kTfLiteFloat32: + EvalUnquantized(context, node); + break; + case kTfLiteInt32: + EvalUnquantized(context, node); + break; + case kTfLiteUInt8: + EvalQuantizedUInt8(context, node); + break; + case kTfLiteInt8: + EvalUnquantized(context, node); + break; + case kTfLiteInt64: + EvalUnquantized(context, node); + break; + + default: + TF_LITE_KERNEL_LOG( + context, "Op Concatenation does not currently support Type '%s'.", + TfLiteTypeGetName(output_type)); + return kTfLiteError; + } + + return kTfLiteOk; +} + +} // namespace concatenation + +TfLiteRegistration Register_CONCATENATION() { + return {/*init=*/concatenation::Init, + /*free=*/nullptr, + /*prepare=*/concatenation::Prepare, + /*invoke=*/concatenation::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test.h new file mode 100644 index 0000000..0fa0e95 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test.h @@ -0,0 +1,97 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_CONV_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_CONV_H_ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/micro_ops.h" +#include "tensorflow_arm/tensorflow/lite/micro/test_helpers.h" +#include "tensorflow_arm/tensorflow/lite/micro/testing/micro_test.h" + +namespace tflite { +namespace testing { + +TfLiteStatus InvokeConv(TfLiteTensor* tensors, int tensors_size, + int output_length, TfLiteConvParams* conv_params, + TfLiteRegistration registration, float* output_data); + +TfLiteStatus InvokeConv(TfLiteTensor* tensors, int tensors_size, + int output_length, TfLiteConvParams* conv_params, + TfLiteRegistration registration, int8_t* output_data); + +TfLiteStatus InvokeConv(TfLiteTensor* tensors, int tensors_size, + int output_length, TfLiteConvParams* conv_params, + TfLiteRegistration registration, uint8_t* output_data); + +TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size, + const float* expected_output_data, + int output_length, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, + float* output_data, float tolerance = 1e-5); + +TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size, + const int8_t* expected_output_data, + int output_length, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, + int8_t* output_data, float tolerance = 1e-5); + +TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size, + const uint8_t* expected_output_data, + int output_length, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, + uint8_t* output_data, float tolerance = 1e-5); + +void TestConvFloat(const int* input_dims_data, const float* input_data, + const int* filter_dims_data, const float* filter_data, + const int* bias_dims_data, const float* bias_data, + const int* output_dims_data, + const float* expected_output_data, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, float* output_data); + +void TestConvQuantizedPerLayer( + const int* input_dims_data, const float* input_data, + uint8_t* input_quantized, float input_scale, const int* filter_dims_data, + const float* filter_data, uint8_t* filter_quantized, float filter_scale, + const int* bias_dims_data, const float* bias_data, int32_t* bias_quantized, + const int* output_dims_data, const float* expected_output_data, + uint8_t* expected_output_quantized, float output_scale, + TfLiteConvParams* conv_params, TfLiteRegistration registration, + uint8_t* output_data); + +void TestConvQuantizedPerChannel( + const int* input_dims_data, const float* input_data, + int8_t* input_quantized, float input_scale, int input_zero_point, + const int* filter_dims_data, const float* filter_data, + int8_t* filter_data_quantized, const int* bias_dims_data, + const float* bias_data, int32_t* bias_data_quantized, float* bias_scales, + int* bias_zero_points, const int* output_dims_data, + const float* expected_output_data, int8_t* expected_output_data_quantized, + float output_scale, int output_zero_point, TfLiteConvParams* conv_params, + TfLiteRegistration registration, int8_t* output_data); + +} // namespace testing +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_CONV_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test_common.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test_common.cpp new file mode 100644 index 0000000..8d30931 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/conv_test_common.cpp @@ -0,0 +1,255 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/kernels/conv_test.h" + +namespace tflite { +namespace testing { + +template +TfLiteStatus InvokeConv(TfLiteTensor* tensors, int tensors_size, + int output_length, TfLiteConvParams* conv_params, + TfLiteRegistration registration, T* output_data) { + int inputs_array_data[] = {3, 0, 1, 2}; + TfLiteIntArray* inputs_array = IntArrayFromInts(inputs_array_data); + int outputs_array_data[] = {1, 3}; + TfLiteIntArray* outputs_array = IntArrayFromInts(outputs_array_data); + + micro::KernelRunner runner( + registration, tensors, tensors_size, inputs_array, outputs_array, + reinterpret_cast(conv_params), micro_test::reporter); + + const char* init_data = reinterpret_cast(conv_params); + TfLiteStatus status = runner.InitAndPrepare(init_data); + if (status != kTfLiteOk) { + return status; + } + return runner.Invoke(); +} + +template +TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size, + const T* expected_output_data, + int output_length, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, + T* output_data, float tolerance) { + TfLiteStatus status = InvokeConv(tensors, tensors_size, output_length, + conv_params, registration, output_data); + if (status != kTfLiteOk) { + return status; + } + for (int i = 0; i < output_length; ++i) { + TF_LITE_MICRO_EXPECT_NEAR(expected_output_data[i], output_data[i], + tolerance); + } + return kTfLiteOk; +} + +TfLiteStatus InvokeConv(TfLiteTensor* tensors, int tensors_size, + int output_length, TfLiteConvParams* conv_params, + TfLiteRegistration registration, float* output_data) { + return InvokeConv(tensors, tensors_size, output_length, conv_params, + registration, output_data); +} + +TfLiteStatus InvokeConv(TfLiteTensor* tensors, int tensors_size, + int output_length, TfLiteConvParams* conv_params, + TfLiteRegistration registration, int8_t* output_data) { + return InvokeConv(tensors, tensors_size, output_length, conv_params, + registration, output_data); +} + +TfLiteStatus InvokeConv(TfLiteTensor* tensors, int tensors_size, + int output_length, TfLiteConvParams* conv_params, + TfLiteRegistration registration, uint8_t* output_data) { + return InvokeConv(tensors, tensors_size, output_length, conv_params, + registration, output_data); +} + +TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size, + const float* expected_output_data, + int output_length, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, + float* output_data, float tolerance) { + return ValidateConvGoldens(tensors, tensors_size, expected_output_data, + output_length, conv_params, registration, + output_data, tolerance); +} + +TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size, + const int8_t* expected_output_data, + int output_length, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, + int8_t* output_data, float tolerance) { + return ValidateConvGoldens( + tensors, tensors_size, expected_output_data, output_length, conv_params, + registration, output_data, tolerance); +} + +TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size, + const uint8_t* expected_output_data, + int output_length, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, + uint8_t* output_data, float tolerance) { + return ValidateConvGoldens( + tensors, tensors_size, expected_output_data, output_length, conv_params, + registration, output_data, tolerance); +} + +void TestConvFloat(const int* input_dims_data, const float* input_data, + const int* filter_dims_data, const float* filter_data, + const int* bias_dims_data, const float* bias_data, + const int* output_dims_data, + const float* expected_output_data, + TfLiteConvParams* conv_params, + TfLiteRegistration registration, float* output_data) { + TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data); + TfLiteIntArray* filter_dims = IntArrayFromInts(filter_dims_data); + TfLiteIntArray* bias_dims = IntArrayFromInts(bias_dims_data); + TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data); + const int output_dims_count = ElementCount(*output_dims); + constexpr int inputs_size = 3; + constexpr int outputs_size = 1; + constexpr int tensors_size = inputs_size + outputs_size; + TfLiteTensor tensors[tensors_size] = { + CreateTensor(input_data, input_dims), + CreateTensor(filter_data, filter_dims), + CreateTensor(bias_data, bias_dims), + CreateTensor(output_data, output_dims), + }; + + TF_LITE_MICRO_EXPECT_EQ( + kTfLiteOk, ValidateConvGoldens(tensors, tensors_size, + expected_output_data, output_dims_count, + conv_params, registration, output_data)); +} + +void TestConvQuantizedPerLayer( + const int* input_dims_data, const float* input_data, + uint8_t* input_quantized, float input_scale, const int* filter_dims_data, + const float* filter_data, uint8_t* filter_quantized, float filter_scale, + const int* bias_dims_data, const float* bias_data, int32_t* bias_quantized, + const int* output_dims_data, const float* expected_output_data, + uint8_t* expected_output_quantized, float output_scale, + TfLiteConvParams* conv_params, TfLiteRegistration registration, + uint8_t* output_data) { + TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data); + TfLiteIntArray* filter_dims = IntArrayFromInts(filter_dims_data); + TfLiteIntArray* bias_dims = IntArrayFromInts(bias_dims_data); + TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data); + const int output_dims_count = ElementCount(*output_dims); + + tflite::Quantize(expected_output_data, expected_output_quantized, + output_dims_count, output_scale, 128); + + constexpr int inputs_size = 3; + constexpr int outputs_size = 1; + constexpr int tensors_size = inputs_size + outputs_size; + TfLiteTensor tensors[tensors_size] = { + CreateQuantizedTensor(input_data, input_quantized, input_dims, + input_scale, 128), + CreateQuantizedTensor(filter_data, filter_quantized, filter_dims, + filter_scale, 128), + CreateQuantizedBiasTensor(bias_data, bias_quantized, bias_dims, + input_scale, filter_scale), + CreateQuantizedTensor(output_data, output_dims, output_scale, 128)}; + + float filter_scales[] = {1, filter_scale}; + int filter_zero_points[] = {1, 128}; + TfLiteAffineQuantization filter_quant = {FloatArrayFromFloats(filter_scales), + IntArrayFromInts(filter_zero_points), + 0}; + tensors[1].quantization = {kTfLiteAffineQuantization, &filter_quant}; + + TF_LITE_MICRO_EXPECT_EQ( + kTfLiteOk, + ValidateConvGoldens(tensors, tensors_size, expected_output_quantized, + output_dims_count, conv_params, registration, + output_data)); +} + +void TestConvQuantizedPerChannel( + const int* input_dims_data, const float* input_data, + int8_t* input_quantized, float input_scale, int input_zero_point, + const int* filter_dims_data, const float* filter_data, + int8_t* filter_data_quantized, const int* bias_dims_data, + const float* bias_data, int32_t* bias_data_quantized, float* bias_scales, + int* bias_zero_points, const int* output_dims_data, + const float* expected_output_data, int8_t* expected_output_data_quantized, + float output_scale, int output_zero_point, TfLiteConvParams* conv_params, + TfLiteRegistration registration, int8_t* output_data) { + TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data); + TfLiteIntArray* filter_dims = IntArrayFromInts(filter_dims_data); + TfLiteIntArray* bias_dims = IntArrayFromInts(bias_dims_data); + TfLiteIntArray* output_dims = IntArrayFromInts(output_dims_data); + const int output_dims_count = ElementCount(*output_dims); + + int filter_zero_points[5]; + float filter_scales[5]; + TfLiteAffineQuantization filter_quant; + TfLiteAffineQuantization bias_quant; + TfLiteTensor input_tensor = CreateQuantizedTensor( + input_data, input_quantized, input_dims, input_scale, input_zero_point); + TfLiteTensor filter_tensor = CreateSymmetricPerChannelQuantizedTensor( + filter_data, filter_data_quantized, filter_dims, filter_scales, + filter_zero_points, &filter_quant, 0 /* quantized dimension */); + TfLiteTensor bias_tensor = CreatePerChannelQuantizedBiasTensor( + bias_data, bias_data_quantized, bias_dims, input_scale, &filter_scales[1], + bias_scales, bias_zero_points, &bias_quant, 0 /* quantized dimension */); + TfLiteTensor output_tensor = CreateQuantizedTensor( + output_data, output_dims, output_scale, output_zero_point); + + float input_scales[] = {1, input_scale}; + int input_zero_points[] = {1, input_zero_point}; + TfLiteAffineQuantization input_quant = {FloatArrayFromFloats(input_scales), + IntArrayFromInts(input_zero_points), + 0}; + input_tensor.quantization = {kTfLiteAffineQuantization, &input_quant}; + + float output_scales[] = {1, output_scale}; + int output_zero_points[] = {1, output_zero_point}; + TfLiteAffineQuantization output_quant = {FloatArrayFromFloats(output_scales), + IntArrayFromInts(output_zero_points), + 0}; + output_tensor.quantization = {kTfLiteAffineQuantization, &output_quant}; + + constexpr int inputs_size = 3; + constexpr int outputs_size = 1; + constexpr int tensors_size = inputs_size + outputs_size; + TfLiteTensor tensors[tensors_size] = { + input_tensor, + filter_tensor, + bias_tensor, + output_tensor, + }; + + tflite::Quantize(expected_output_data, expected_output_data_quantized, + output_dims_count, output_scale, output_zero_point); + TF_LITE_MICRO_EXPECT_EQ( + kTfLiteOk, + ValidateConvGoldens(tensors, tensors_size, expected_output_data_quantized, + output_dims_count, conv_params, registration, + output_data, 1.0 /* tolerance */)); +} + +} // namespace testing +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/dequantize.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/dequantize.cpp new file mode 100644 index 0000000..f13fecd --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/dequantize.cpp @@ -0,0 +1,169 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/dequantize.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/quantize.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/requantize.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace dequantize { + +struct OpData { + tflite::DequantizationParams quantization_params; + // The scaling factor from input to output (aka the 'real multiplier') can + // be represented as a fixed point multiplier plus a left shift. + int32_t output_multiplier; + int output_shift; + int32_t output_zero_point; +}; + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + // TODO(b/140515557): Add cached dequant to improve hybrid model performance. + const TfLiteTensor* input = GetInput(context, node, 0); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, 0); + TF_LITE_ENSURE(context, output != nullptr); + + TF_LITE_ENSURE(context, input->type == kTfLiteUInt8 || + input->type == kTfLiteInt8 || + input->type == kTfLiteInt16); + TF_LITE_ENSURE( + context, output->type == kTfLiteFloat32 || output->type == kTfLiteInt32); + + if (output->type == kTfLiteInt32) { + const double effective_output_scale = + static_cast(input->params.scale) / + static_cast(output->params.scale); + QuantizeMultiplier(effective_output_scale, &data->output_multiplier, + &data->output_shift); + } + + data->quantization_params.zero_point = input->params.zero_point; + data->quantization_params.scale = static_cast(input->params.scale); + data->output_zero_point = output->params.zero_point; + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0); + + if (output->type == kTfLiteFloat32) { + switch (input->type) { + case kTfLiteUInt8: + reference_ops::Dequantize(data->quantization_params, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + case kTfLiteInt8: + reference_ops::Dequantize(data->quantization_params, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + case kTfLiteInt16: + reference_ops::Dequantize(data->quantization_params, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else if (output->type == kTfLiteInt32) { + int flat_size = MatchingFlatSize(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorShape(output)); + switch (input->type) { + case kTfLiteInt16: { + reference_ops::Requantize( + tflite::micro::GetTensorData(input), flat_size, + data->output_multiplier, data->output_shift, + data->quantization_params.zero_point, data->output_zero_point, + tflite::micro::GetTensorData(output)); + break; + } + case kTfLiteInt8: { + reference_ops::Requantize( + tflite::micro::GetTensorData(input), flat_size, + data->output_multiplier, data->output_shift, + data->quantization_params.zero_point, data->output_zero_point, + tflite::micro::GetTensorData(output)); + break; + } + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else { + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + + return kTfLiteOk; +} + +} // namespace dequantize + +TfLiteRegistration Register_DEQUANTIZE() { + return {/*init=*/dequantize::Init, + /*free=*/nullptr, + /*prepare=*/dequantize::Prepare, + /*invoke=*/dequantize::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/detection_postprocess.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/detection_postprocess.cpp new file mode 100644 index 0000000..71af711 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/detection_postprocess.cpp @@ -0,0 +1,808 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include + +#define FLATBUFFERS_LOCALE_INDEPENDENT 0 +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flexbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace { + +/** + * This version of detection_postprocess is specific to TFLite Micro. It + * contains the following differences between the TFLite version: + * + * 1.) Temporaries (temporary tensors) - Micro use instead scratch buffer API. + * 2.) Output dimensions - the TFLite version does not support undefined out + * dimensions. So model must have static out dimensions. + */ + +// Input tensors +constexpr int kInputTensorBoxEncodings = 0; +constexpr int kInputTensorClassPredictions = 1; +constexpr int kInputTensorAnchors = 2; + +// Output tensors +constexpr int kOutputTensorDetectionBoxes = 0; +constexpr int kOutputTensorDetectionClasses = 1; +constexpr int kOutputTensorDetectionScores = 2; +constexpr int kOutputTensorNumDetections = 3; + +constexpr int kNumCoordBox = 4; +constexpr int kBatchSize = 1; + +constexpr int kNumDetectionsPerClass = 100; + +// Object Detection model produces axis-aligned boxes in two formats: +// BoxCorner represents the lower left corner (xmin, ymin) and +// the upper right corner (xmax, ymax). +// CenterSize represents the center (xcenter, ycenter), height and width. +// BoxCornerEncoding and CenterSizeEncoding are related as follows: +// ycenter = y / y_scale * anchor.h + anchor.y; +// xcenter = x / x_scale * anchor.w + anchor.x; +// half_h = 0.5*exp(h/ h_scale)) * anchor.h; +// half_w = 0.5*exp(w / w_scale)) * anchor.w; +// ymin = ycenter - half_h +// ymax = ycenter + half_h +// xmin = xcenter - half_w +// xmax = xcenter + half_w +struct BoxCornerEncoding { + float ymin; + float xmin; + float ymax; + float xmax; +}; + +struct CenterSizeEncoding { + float y; + float x; + float h; + float w; +}; +// We make sure that the memory allocations are contiguous with static assert. +static_assert(sizeof(BoxCornerEncoding) == sizeof(float) * kNumCoordBox, + "Size of BoxCornerEncoding is 4 float values"); +static_assert(sizeof(CenterSizeEncoding) == sizeof(float) * kNumCoordBox, + "Size of CenterSizeEncoding is 4 float values"); + +struct OpData { + int max_detections; + int max_classes_per_detection; // Fast Non-Max-Suppression + int detections_per_class; // Regular Non-Max-Suppression + float non_max_suppression_score_threshold; + float intersection_over_union_threshold; + int num_classes; + bool use_regular_non_max_suppression; + CenterSizeEncoding scale_values; + + // Scratch buffers indexes + int active_candidate_idx; + int decoded_boxes_idx; + int scores_idx; + int score_buffer_idx; + int keep_scores_idx; + int scores_after_regular_non_max_suppression_idx; + int sorted_values_idx; + int keep_indices_idx; + int sorted_indices_idx; + int buffer_idx; + int selected_idx; + + // Cached tensor scale and zero point values for quantized operations + TfLiteQuantizationParams input_box_encodings; + TfLiteQuantizationParams input_class_predictions; + TfLiteQuantizationParams input_anchors; +}; + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + OpData* op_data = nullptr; + + const uint8_t* buffer_t = reinterpret_cast(buffer); + const flexbuffers::Map& m = flexbuffers::GetRoot(buffer_t, length).AsMap(); + + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + op_data = reinterpret_cast( + context->AllocatePersistentBuffer(context, sizeof(OpData))); + + op_data->max_detections = m["max_detections"].AsInt32(); + op_data->max_classes_per_detection = m["max_classes_per_detection"].AsInt32(); + if (m["detections_per_class"].IsNull()) + op_data->detections_per_class = kNumDetectionsPerClass; + else + op_data->detections_per_class = m["detections_per_class"].AsInt32(); + if (m["use_regular_nms"].IsNull()) + op_data->use_regular_non_max_suppression = false; + else + op_data->use_regular_non_max_suppression = m["use_regular_nms"].AsBool(); + + op_data->non_max_suppression_score_threshold = + m["nms_score_threshold"].AsFloat(); + op_data->intersection_over_union_threshold = m["nms_iou_threshold"].AsFloat(); + op_data->num_classes = m["num_classes"].AsInt32(); + op_data->scale_values.y = m["y_scale"].AsFloat(); + op_data->scale_values.x = m["x_scale"].AsFloat(); + op_data->scale_values.h = m["h_scale"].AsFloat(); + op_data->scale_values.w = m["w_scale"].AsFloat(); + + return op_data; +} + +void Free(TfLiteContext* context, void* buffer) {} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + auto* op_data = static_cast(node->user_data); + + // Inputs: box_encodings, scores, anchors + TF_LITE_ENSURE_EQ(context, NumInputs(node), 3); + const TfLiteTensor* input_box_encodings = + GetInput(context, node, kInputTensorBoxEncodings); + const TfLiteTensor* input_class_predictions = + GetInput(context, node, kInputTensorClassPredictions); + const TfLiteTensor* input_anchors = + GetInput(context, node, kInputTensorAnchors); + TF_LITE_ENSURE_EQ(context, NumDimensions(input_box_encodings), 3); + TF_LITE_ENSURE_EQ(context, NumDimensions(input_class_predictions), 3); + TF_LITE_ENSURE_EQ(context, NumDimensions(input_anchors), 2); + + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 4); + const int num_boxes = input_box_encodings->dims->data[1]; + const int num_classes = op_data->num_classes; + + op_data->input_box_encodings.scale = input_box_encodings->params.scale; + op_data->input_box_encodings.zero_point = + input_box_encodings->params.zero_point; + op_data->input_class_predictions.scale = + input_class_predictions->params.scale; + op_data->input_class_predictions.zero_point = + input_class_predictions->params.zero_point; + op_data->input_anchors.scale = input_anchors->params.scale; + op_data->input_anchors.zero_point = input_anchors->params.zero_point; + + // Scratch tensors + context->RequestScratchBufferInArena(context, num_boxes, + &op_data->active_candidate_idx); + context->RequestScratchBufferInArena(context, + num_boxes * kNumCoordBox * sizeof(float), + &op_data->decoded_boxes_idx); + context->RequestScratchBufferInArena( + context, + input_class_predictions->dims->data[1] * + input_class_predictions->dims->data[2] * sizeof(float), + &op_data->scores_idx); + + // Additional buffers + context->RequestScratchBufferInArena(context, num_boxes * sizeof(float), + &op_data->score_buffer_idx); + context->RequestScratchBufferInArena(context, num_boxes * sizeof(float), + &op_data->keep_scores_idx); + context->RequestScratchBufferInArena( + context, op_data->max_detections * num_boxes * sizeof(float), + &op_data->scores_after_regular_non_max_suppression_idx); + context->RequestScratchBufferInArena( + context, op_data->max_detections * num_boxes * sizeof(float), + &op_data->sorted_values_idx); + context->RequestScratchBufferInArena(context, num_boxes * sizeof(int), + &op_data->keep_indices_idx); + context->RequestScratchBufferInArena( + context, op_data->max_detections * num_boxes * sizeof(int), + &op_data->sorted_indices_idx); + int buffer_size = std::max(num_classes, op_data->max_detections); + context->RequestScratchBufferInArena( + context, buffer_size * num_boxes * sizeof(int), &op_data->buffer_idx); + buffer_size = std::min(num_boxes, op_data->max_detections); + context->RequestScratchBufferInArena( + context, buffer_size * num_boxes * sizeof(int), &op_data->selected_idx); + + // Outputs: detection_boxes, detection_scores, detection_classes, + // num_detections + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 4); + + return kTfLiteOk; +} + +class Dequantizer { + public: + Dequantizer(int zero_point, float scale) + : zero_point_(zero_point), scale_(scale) {} + float operator()(uint8_t x) { + return (static_cast(x) - zero_point_) * scale_; + } + + private: + int zero_point_; + float scale_; +}; + +void DequantizeBoxEncodings(const TfLiteEvalTensor* input_box_encodings, + int idx, float quant_zero_point, float quant_scale, + int length_box_encoding, + CenterSizeEncoding* box_centersize) { + const uint8_t* boxes = + tflite::micro::GetTensorData(input_box_encodings) + + length_box_encoding * idx; + Dequantizer dequantize(quant_zero_point, quant_scale); + // See definition of the KeyPointBoxCoder at + // https://github.com/tensorflow/models/blob/master/research/object_detection/box_coders/keypoint_box_coder.py + // The first four elements are the box coordinates, which is the same as the + // FastRnnBoxCoder at + // https://github.com/tensorflow/models/blob/master/research/object_detection/box_coders/faster_rcnn_box_coder.py + box_centersize->y = dequantize(boxes[0]); + box_centersize->x = dequantize(boxes[1]); + box_centersize->h = dequantize(boxes[2]); + box_centersize->w = dequantize(boxes[3]); +} + +template +T ReInterpretTensor(const TfLiteEvalTensor* tensor) { + const float* tensor_base = tflite::micro::GetTensorData(tensor); + return reinterpret_cast(tensor_base); +} + +template +T ReInterpretTensor(TfLiteEvalTensor* tensor) { + float* tensor_base = tflite::micro::GetTensorData(tensor); + return reinterpret_cast(tensor_base); +} + +TfLiteStatus DecodeCenterSizeBoxes(TfLiteContext* context, TfLiteNode* node, + OpData* op_data) { + // Parse input tensor boxencodings + const TfLiteEvalTensor* input_box_encodings = + tflite::micro::GetEvalInput(context, node, kInputTensorBoxEncodings); + TF_LITE_ENSURE_EQ(context, input_box_encodings->dims->data[0], kBatchSize); + const int num_boxes = input_box_encodings->dims->data[1]; + TF_LITE_ENSURE(context, input_box_encodings->dims->data[2] >= kNumCoordBox); + const TfLiteEvalTensor* input_anchors = + tflite::micro::GetEvalInput(context, node, kInputTensorAnchors); + + // Decode the boxes to get (ymin, xmin, ymax, xmax) based on the anchors + CenterSizeEncoding box_centersize; + CenterSizeEncoding scale_values = op_data->scale_values; + CenterSizeEncoding anchor; + for (int idx = 0; idx < num_boxes; ++idx) { + switch (input_box_encodings->type) { + // Quantized + case kTfLiteUInt8: + DequantizeBoxEncodings( + input_box_encodings, idx, + static_cast(op_data->input_box_encodings.zero_point), + static_cast(op_data->input_box_encodings.scale), + input_box_encodings->dims->data[2], &box_centersize); + DequantizeBoxEncodings( + input_anchors, idx, + static_cast(op_data->input_anchors.zero_point), + static_cast(op_data->input_anchors.scale), kNumCoordBox, + &anchor); + break; + // Float + case kTfLiteFloat32: { + // Please see DequantizeBoxEncodings function for the support detail. + const int box_encoding_idx = idx * input_box_encodings->dims->data[2]; + const float* boxes = &(tflite::micro::GetTensorData( + input_box_encodings)[box_encoding_idx]); + box_centersize = *reinterpret_cast(boxes); + anchor = + ReInterpretTensor(input_anchors)[idx]; + break; + } + default: + // Unsupported type. + return kTfLiteError; + } + + float ycenter = static_cast(static_cast(box_centersize.y) / + static_cast(scale_values.y) * + static_cast(anchor.h) + + static_cast(anchor.y)); + + float xcenter = static_cast(static_cast(box_centersize.x) / + static_cast(scale_values.x) * + static_cast(anchor.w) + + static_cast(anchor.x)); + + float half_h = + static_cast(0.5 * + (std::exp(static_cast(box_centersize.h) / + static_cast(scale_values.h))) * + static_cast(anchor.h)); + float half_w = + static_cast(0.5 * + (std::exp(static_cast(box_centersize.w) / + static_cast(scale_values.w))) * + static_cast(anchor.w)); + + float* decoded_boxes = reinterpret_cast( + context->GetScratchBuffer(context, op_data->decoded_boxes_idx)); + auto& box = reinterpret_cast(decoded_boxes)[idx]; + box.ymin = ycenter - half_h; + box.xmin = xcenter - half_w; + box.ymax = ycenter + half_h; + box.xmax = xcenter + half_w; + } + return kTfLiteOk; +} + +void DecreasingPartialArgSort(const float* values, int num_values, + int num_to_sort, int* indices) { + std::iota(indices, indices + num_values, 0); + std::partial_sort( + indices, indices + num_to_sort, indices + num_values, + [&values](const int i, const int j) { return values[i] > values[j]; }); +} + +int SelectDetectionsAboveScoreThreshold(const float* values, int size, + const float threshold, + float* keep_values, int* keep_indices) { + int counter = 0; + for (int i = 0; i < size; i++) { + if (values[i] >= threshold) { + keep_values[counter] = values[i]; + keep_indices[counter] = i; + counter++; + } + } + return counter; +} + +bool ValidateBoxes(const float* decoded_boxes, const int num_boxes) { + for (int i = 0; i < num_boxes; ++i) { + // ymax>=ymin, xmax>=xmin + auto& box = reinterpret_cast(decoded_boxes)[i]; + if (box.ymin >= box.ymax || box.xmin >= box.xmax) { + return false; + } + } + return true; +} + +float ComputeIntersectionOverUnion(const float* decoded_boxes, const int i, + const int j) { + auto& box_i = reinterpret_cast(decoded_boxes)[i]; + auto& box_j = reinterpret_cast(decoded_boxes)[j]; + const float area_i = (box_i.ymax - box_i.ymin) * (box_i.xmax - box_i.xmin); + const float area_j = (box_j.ymax - box_j.ymin) * (box_j.xmax - box_j.xmin); + if (area_i <= 0 || area_j <= 0) return 0.0; + const float intersection_ymin = std::max(box_i.ymin, box_j.ymin); + const float intersection_xmin = std::max(box_i.xmin, box_j.xmin); + const float intersection_ymax = std::min(box_i.ymax, box_j.ymax); + const float intersection_xmax = std::min(box_i.xmax, box_j.xmax); + const float intersection_area = + std::max(intersection_ymax - intersection_ymin, 0.0) * + std::max(intersection_xmax - intersection_xmin, 0.0); + return intersection_area / (area_i + area_j - intersection_area); +} + +// NonMaxSuppressionSingleClass() prunes out the box locations with high overlap +// before selecting the highest scoring boxes (max_detections in number) +// It assumes all boxes are good in beginning and sorts based on the scores. +// If lower-scoring box has too much overlap with a higher-scoring box, +// we get rid of the lower-scoring box. +// Complexity is O(N^2) pairwise comparison between boxes +TfLiteStatus NonMaxSuppressionSingleClassHelper( + TfLiteContext* context, TfLiteNode* node, OpData* op_data, + const float* scores, int* selected, int* selected_size, + int max_detections) { + const TfLiteEvalTensor* input_box_encodings = + tflite::micro::GetEvalInput(context, node, kInputTensorBoxEncodings); + const int num_boxes = input_box_encodings->dims->data[1]; + const float non_max_suppression_score_threshold = + op_data->non_max_suppression_score_threshold; + const float intersection_over_union_threshold = + op_data->intersection_over_union_threshold; + // Maximum detections should be positive. + TF_LITE_ENSURE(context, (max_detections >= 0)); + // intersection_over_union_threshold should be positive + // and should be less than 1. + TF_LITE_ENSURE(context, (intersection_over_union_threshold > 0.0f) && + (intersection_over_union_threshold <= 1.0f)); + // Validate boxes + float* decoded_boxes = reinterpret_cast( + context->GetScratchBuffer(context, op_data->decoded_boxes_idx)); + + TF_LITE_ENSURE(context, ValidateBoxes(decoded_boxes, num_boxes)); + + // threshold scores + int* keep_indices = reinterpret_cast( + context->GetScratchBuffer(context, op_data->keep_indices_idx)); + float* keep_scores = reinterpret_cast( + context->GetScratchBuffer(context, op_data->keep_scores_idx)); + int num_scores_kept = SelectDetectionsAboveScoreThreshold( + scores, num_boxes, non_max_suppression_score_threshold, keep_scores, + keep_indices); + int* sorted_indices = reinterpret_cast( + context->GetScratchBuffer(context, op_data->sorted_indices_idx)); + + DecreasingPartialArgSort(keep_scores, num_scores_kept, num_scores_kept, + sorted_indices); + + const int num_boxes_kept = num_scores_kept; + const int output_size = std::min(num_boxes_kept, max_detections); + *selected_size = 0; + + int num_active_candidate = num_boxes_kept; + uint8_t* active_box_candidate = reinterpret_cast( + context->GetScratchBuffer(context, op_data->active_candidate_idx)); + + for (int row = 0; row < num_boxes_kept; row++) { + active_box_candidate[row] = 1; + } + for (int i = 0; i < num_boxes_kept; ++i) { + if (num_active_candidate == 0 || *selected_size >= output_size) break; + if (active_box_candidate[i] == 1) { + selected[(*selected_size)++] = keep_indices[sorted_indices[i]]; + active_box_candidate[i] = 0; + num_active_candidate--; + } else { + continue; + } + for (int j = i + 1; j < num_boxes_kept; ++j) { + if (active_box_candidate[j] == 1) { + float intersection_over_union = ComputeIntersectionOverUnion( + decoded_boxes, keep_indices[sorted_indices[i]], + keep_indices[sorted_indices[j]]); + + if (intersection_over_union > intersection_over_union_threshold) { + active_box_candidate[j] = 0; + num_active_candidate--; + } + } + } + } + + return kTfLiteOk; +} + +// This function implements a regular version of Non Maximal Suppression (NMS) +// for multiple classes where +// 1) we do NMS separately for each class across all anchors and +// 2) keep only the highest anchor scores across all classes +// 3) The worst runtime of the regular NMS is O(K*N^2) +// where N is the number of anchors and K the number of +// classes. +TfLiteStatus NonMaxSuppressionMultiClassRegularHelper(TfLiteContext* context, + TfLiteNode* node, + OpData* op_data, + const float* scores) { + const TfLiteEvalTensor* input_box_encodings = + tflite::micro::GetEvalInput(context, node, kInputTensorBoxEncodings); + const TfLiteEvalTensor* input_class_predictions = + tflite::micro::GetEvalInput(context, node, kInputTensorClassPredictions); + TfLiteEvalTensor* detection_boxes = + tflite::micro::GetEvalOutput(context, node, kOutputTensorDetectionBoxes); + TfLiteEvalTensor* detection_classes = tflite::micro::GetEvalOutput( + context, node, kOutputTensorDetectionClasses); + TfLiteEvalTensor* detection_scores = + tflite::micro::GetEvalOutput(context, node, kOutputTensorDetectionScores); + TfLiteEvalTensor* num_detections = + tflite::micro::GetEvalOutput(context, node, kOutputTensorNumDetections); + + const int num_boxes = input_box_encodings->dims->data[1]; + const int num_classes = op_data->num_classes; + const int num_detections_per_class = op_data->detections_per_class; + const int max_detections = op_data->max_detections; + const int num_classes_with_background = + input_class_predictions->dims->data[2]; + // The row index offset is 1 if background class is included and 0 otherwise. + int label_offset = num_classes_with_background - num_classes; + TF_LITE_ENSURE(context, num_detections_per_class > 0); + + // For each class, perform non-max suppression. + float* class_scores = reinterpret_cast( + context->GetScratchBuffer(context, op_data->score_buffer_idx)); + int* box_indices_after_regular_non_max_suppression = reinterpret_cast( + context->GetScratchBuffer(context, op_data->buffer_idx)); + float* scores_after_regular_non_max_suppression = + reinterpret_cast(context->GetScratchBuffer( + context, op_data->scores_after_regular_non_max_suppression_idx)); + + int size_of_sorted_indices = 0; + int* sorted_indices = reinterpret_cast( + context->GetScratchBuffer(context, op_data->sorted_indices_idx)); + float* sorted_values = reinterpret_cast( + context->GetScratchBuffer(context, op_data->sorted_values_idx)); + + for (int col = 0; col < num_classes; col++) { + for (int row = 0; row < num_boxes; row++) { + // Get scores of boxes corresponding to all anchors for single class + class_scores[row] = + *(scores + row * num_classes_with_background + col + label_offset); + } + // Perform non-maximal suppression on single class + int selected_size = 0; + int* selected = reinterpret_cast( + context->GetScratchBuffer(context, op_data->selected_idx)); + TF_LITE_ENSURE_STATUS(NonMaxSuppressionSingleClassHelper( + context, node, op_data, class_scores, selected, &selected_size, + num_detections_per_class)); + // Add selected indices from non-max suppression of boxes in this class + int output_index = size_of_sorted_indices; + for (int i = 0; i < selected_size; i++) { + int selected_index = selected[i]; + + box_indices_after_regular_non_max_suppression[output_index] = + (selected_index * num_classes_with_background + col + label_offset); + scores_after_regular_non_max_suppression[output_index] = + class_scores[selected_index]; + output_index++; + } + // Sort the max scores among the selected indices + // Get the indices for top scores + int num_indices_to_sort = std::min(output_index, max_detections); + DecreasingPartialArgSort(scores_after_regular_non_max_suppression, + output_index, num_indices_to_sort, sorted_indices); + + // Copy values to temporary vectors + for (int row = 0; row < num_indices_to_sort; row++) { + int temp = sorted_indices[row]; + sorted_indices[row] = box_indices_after_regular_non_max_suppression[temp]; + sorted_values[row] = scores_after_regular_non_max_suppression[temp]; + } + // Copy scores and indices from temporary vectors + for (int row = 0; row < num_indices_to_sort; row++) { + box_indices_after_regular_non_max_suppression[row] = sorted_indices[row]; + scores_after_regular_non_max_suppression[row] = sorted_values[row]; + } + size_of_sorted_indices = num_indices_to_sort; + } + + // Allocate output tensors + for (int output_box_index = 0; output_box_index < max_detections; + output_box_index++) { + if (output_box_index < size_of_sorted_indices) { + const int anchor_index = floor( + box_indices_after_regular_non_max_suppression[output_box_index] / + num_classes_with_background); + const int class_index = + box_indices_after_regular_non_max_suppression[output_box_index] - + anchor_index * num_classes_with_background - label_offset; + const float selected_score = + scores_after_regular_non_max_suppression[output_box_index]; + // detection_boxes + float* decoded_boxes = reinterpret_cast( + context->GetScratchBuffer(context, op_data->decoded_boxes_idx)); + ReInterpretTensor(detection_boxes)[output_box_index] = + reinterpret_cast(decoded_boxes)[anchor_index]; + // detection_classes + tflite::micro::GetTensorData(detection_classes)[output_box_index] = + class_index; + // detection_scores + tflite::micro::GetTensorData(detection_scores)[output_box_index] = + selected_score; + } else { + ReInterpretTensor( + detection_boxes)[output_box_index] = {0.0f, 0.0f, 0.0f, 0.0f}; + // detection_classes + tflite::micro::GetTensorData(detection_classes)[output_box_index] = + 0.0f; + // detection_scores + tflite::micro::GetTensorData(detection_scores)[output_box_index] = + 0.0f; + } + } + tflite::micro::GetTensorData(num_detections)[0] = + size_of_sorted_indices; + + return kTfLiteOk; +} + +// This function implements a fast version of Non Maximal Suppression for +// multiple classes where +// 1) we keep the top-k scores for each anchor and +// 2) during NMS, each anchor only uses the highest class score for sorting. +// 3) Compared to standard NMS, the worst runtime of this version is O(N^2) +// instead of O(KN^2) where N is the number of anchors and K the number of +// classes. +TfLiteStatus NonMaxSuppressionMultiClassFastHelper(TfLiteContext* context, + TfLiteNode* node, + OpData* op_data, + const float* scores) { + const TfLiteEvalTensor* input_box_encodings = + tflite::micro::GetEvalInput(context, node, kInputTensorBoxEncodings); + const TfLiteEvalTensor* input_class_predictions = + tflite::micro::GetEvalInput(context, node, kInputTensorClassPredictions); + TfLiteEvalTensor* detection_boxes = + tflite::micro::GetEvalOutput(context, node, kOutputTensorDetectionBoxes); + + TfLiteEvalTensor* detection_classes = tflite::micro::GetEvalOutput( + context, node, kOutputTensorDetectionClasses); + TfLiteEvalTensor* detection_scores = + tflite::micro::GetEvalOutput(context, node, kOutputTensorDetectionScores); + TfLiteEvalTensor* num_detections = + tflite::micro::GetEvalOutput(context, node, kOutputTensorNumDetections); + + const int num_boxes = input_box_encodings->dims->data[1]; + const int num_classes = op_data->num_classes; + const int max_categories_per_anchor = op_data->max_classes_per_detection; + const int num_classes_with_background = + input_class_predictions->dims->data[2]; + + // The row index offset is 1 if background class is included and 0 otherwise. + int label_offset = num_classes_with_background - num_classes; + TF_LITE_ENSURE(context, (max_categories_per_anchor > 0)); + const int num_categories_per_anchor = + std::min(max_categories_per_anchor, num_classes); + float* max_scores = reinterpret_cast( + context->GetScratchBuffer(context, op_data->score_buffer_idx)); + int* sorted_class_indices = reinterpret_cast( + context->GetScratchBuffer(context, op_data->buffer_idx)); + + for (int row = 0; row < num_boxes; row++) { + const float* box_scores = + scores + row * num_classes_with_background + label_offset; + int* class_indices = sorted_class_indices + row * num_classes; + DecreasingPartialArgSort(box_scores, num_classes, num_categories_per_anchor, + class_indices); + max_scores[row] = box_scores[class_indices[0]]; + } + + // Perform non-maximal suppression on max scores + int selected_size = 0; + int* selected = reinterpret_cast( + context->GetScratchBuffer(context, op_data->selected_idx)); + TF_LITE_ENSURE_STATUS(NonMaxSuppressionSingleClassHelper( + context, node, op_data, max_scores, selected, &selected_size, + op_data->max_detections)); + + // Allocate output tensors + int output_box_index = 0; + + for (int i = 0; i < selected_size; i++) { + int selected_index = selected[i]; + + const float* box_scores = + scores + selected_index * num_classes_with_background + label_offset; + const int* class_indices = + sorted_class_indices + selected_index * num_classes; + + for (int col = 0; col < num_categories_per_anchor; ++col) { + int box_offset = num_categories_per_anchor * output_box_index + col; + + // detection_boxes + float* decoded_boxes = reinterpret_cast( + context->GetScratchBuffer(context, op_data->decoded_boxes_idx)); + ReInterpretTensor(detection_boxes)[box_offset] = + reinterpret_cast(decoded_boxes)[selected_index]; + + // detection_classes + tflite::micro::GetTensorData(detection_classes)[box_offset] = + class_indices[col]; + + // detection_scores + tflite::micro::GetTensorData(detection_scores)[box_offset] = + box_scores[class_indices[col]]; + + output_box_index++; + } + } + + tflite::micro::GetTensorData(num_detections)[0] = output_box_index; + return kTfLiteOk; +} + +void DequantizeClassPredictions(const TfLiteEvalTensor* input_class_predictions, + const int num_boxes, + const int num_classes_with_background, + float* scores, OpData* op_data) { + float quant_zero_point = + static_cast(op_data->input_class_predictions.zero_point); + float quant_scale = + static_cast(op_data->input_class_predictions.scale); + Dequantizer dequantize(quant_zero_point, quant_scale); + const uint8_t* scores_quant = + tflite::micro::GetTensorData(input_class_predictions); + for (int idx = 0; idx < num_boxes * num_classes_with_background; ++idx) { + scores[idx] = dequantize(scores_quant[idx]); + } +} + +TfLiteStatus NonMaxSuppressionMultiClass(TfLiteContext* context, + TfLiteNode* node, OpData* op_data) { + // Get the input tensors + const TfLiteEvalTensor* input_box_encodings = + tflite::micro::GetEvalInput(context, node, kInputTensorBoxEncodings); + const TfLiteEvalTensor* input_class_predictions = + tflite::micro::GetEvalInput(context, node, kInputTensorClassPredictions); + const int num_boxes = input_box_encodings->dims->data[1]; + const int num_classes = op_data->num_classes; + + TF_LITE_ENSURE_EQ(context, input_class_predictions->dims->data[0], + kBatchSize); + TF_LITE_ENSURE_EQ(context, input_class_predictions->dims->data[1], num_boxes); + const int num_classes_with_background = + input_class_predictions->dims->data[2]; + + TF_LITE_ENSURE(context, (num_classes_with_background - num_classes <= 1)); + TF_LITE_ENSURE(context, (num_classes_with_background >= num_classes)); + + const float* scores; + switch (input_class_predictions->type) { + case kTfLiteUInt8: { + float* temporary_scores = reinterpret_cast( + context->GetScratchBuffer(context, op_data->scores_idx)); + DequantizeClassPredictions(input_class_predictions, num_boxes, + num_classes_with_background, temporary_scores, + op_data); + scores = temporary_scores; + } break; + case kTfLiteFloat32: + scores = tflite::micro::GetTensorData(input_class_predictions); + break; + default: + // Unsupported type. + return kTfLiteError; + } + + if (op_data->use_regular_non_max_suppression) { + TF_LITE_ENSURE_STATUS(NonMaxSuppressionMultiClassRegularHelper( + context, node, op_data, scores)); + } else { + TF_LITE_ENSURE_STATUS( + NonMaxSuppressionMultiClassFastHelper(context, node, op_data, scores)); + } + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE(context, (kBatchSize == 1)); + auto* op_data = static_cast(node->user_data); + + // These two functions correspond to two blocks in the Object Detection model. + // In future, we would like to break the custom op in two blocks, which is + // currently not feasible because we would like to input quantized inputs + // and do all calculations in float. Mixed quantized/float calculations are + // currently not supported in TFLite. + + // This fills in temporary decoded_boxes + // by transforming input_box_encodings and input_anchors from + // CenterSizeEncodings to BoxCornerEncoding + TF_LITE_ENSURE_STATUS(DecodeCenterSizeBoxes(context, node, op_data)); + + // This fills in the output tensors + // by choosing effective set of decoded boxes + // based on Non Maximal Suppression, i.e. selecting + // highest scoring non-overlapping boxes. + TF_LITE_ENSURE_STATUS(NonMaxSuppressionMultiClass(context, node, op_data)); + + return kTfLiteOk; +} +} // namespace + +TfLiteRegistration* Register_DETECTION_POSTPROCESS() { + static TfLiteRegistration r = {/*init=*/Init, + /*free=*/Free, + /*prepare=*/Prepare, + /*invoke=*/Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; + return &r; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/elementwise.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/elementwise.cpp new file mode 100644 index 0000000..83c668b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/elementwise.cpp @@ -0,0 +1,217 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace elementwise { +namespace { + +bool IsNumericSupportedType(const TfLiteType type) { + return type == kTfLiteFloat32; +} + +bool IsLogicalSupportedType(const TfLiteType type) { + return type == kTfLiteBool; +} + +typedef bool (*IsSupportedType)(TfLiteType); +template +TfLiteStatus GenericPrepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + const TfLiteTensor* input = GetInput(context, node, 0); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, 0); + TF_LITE_ENSURE(context, output != nullptr); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + if (!IsSupportedType(input->type)) { + TF_LITE_KERNEL_LOG(context, "Input data type %s (%d) is not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +template +inline TfLiteStatus EvalImpl(TfLiteContext* context, TfLiteNode* node, + T func(T), TfLiteType expected_type) { + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, expected_type); + const size_t num_elements = ElementCount(*input->dims); + const T* in_data = tflite::micro::GetTensorData(input); + T* out_data = tflite::micro::GetTensorData(output); + for (size_t i = 0; i < num_elements; ++i) { + out_data[i] = func(in_data[i]); + } + return kTfLiteOk; +} + +inline TfLiteStatus EvalNumeric(TfLiteContext* context, TfLiteNode* node, + float float_func(float)) { + return EvalImpl(context, node, float_func, kTfLiteFloat32); +} + +inline TfLiteStatus EvalLogical(TfLiteContext* context, TfLiteNode* node, + bool bool_func(bool)) { + return EvalImpl(context, node, bool_func, kTfLiteBool); +} + +TfLiteStatus AbsEval(TfLiteContext* context, TfLiteNode* node) { + return EvalNumeric(context, node, std::abs); +} + +TfLiteStatus SinEval(TfLiteContext* context, TfLiteNode* node) { + return EvalNumeric(context, node, std::sin); +} + +TfLiteStatus CosEval(TfLiteContext* context, TfLiteNode* node) { + return EvalNumeric(context, node, std::cos); +} + +TfLiteStatus LogEval(TfLiteContext* context, TfLiteNode* node) { + return EvalNumeric(context, node, std::log); +} + +TfLiteStatus SqrtEval(TfLiteContext* context, TfLiteNode* node) { + return EvalNumeric(context, node, std::sqrt); +} + +TfLiteStatus RsqrtEval(TfLiteContext* context, TfLiteNode* node) { + return EvalNumeric(context, node, [](float f) { return 1.f / std::sqrt(f); }); +} + +TfLiteStatus SquareEval(TfLiteContext* context, TfLiteNode* node) { + return EvalNumeric(context, node, [](float f) { return f * f; }); +} + +TfLiteStatus LogicalNotEval(TfLiteContext* context, TfLiteNode* node) { + return EvalLogical(context, node, [](bool v) { return !v; }); +} + +} // namespace +} // namespace elementwise + +TfLiteRegistration Register_ABS() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::AbsEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_SIN() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::SinEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_COS() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::CosEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_LOG() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::LogEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_SQRT() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::SqrtEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_RSQRT() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::RsqrtEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_SQUARE() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::SquareEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_LOGICAL_NOT() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/ + elementwise::GenericPrepare, + /*invoke=*/elementwise::LogicalNotEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.cpp new file mode 100644 index 0000000..5c25773 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.cpp @@ -0,0 +1,30 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// +// This is a stub file for non-Ethos platforms +// +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +TfLiteRegistration* Register_ETHOSU() { return nullptr; } + +const char* GetString_ETHOSU() { return ""; } + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.h new file mode 100644 index 0000000..b2207cd --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.h @@ -0,0 +1,31 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_ETHOSU_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_ETHOSU_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +TfLiteRegistration* Register_ETHOSU(); + +const char* GetString_ETHOSU(); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_ETHOSU_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.cpp new file mode 100644 index 0000000..588a323 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.cpp @@ -0,0 +1,71 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// This file is generated. See: +// tensorflow/lite/micro/kernels/detection_postprocess_test/README.md + +#include "tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.h" + +const int g_gen_data_size_none_regular_nms = 242; +const unsigned char g_gen_data_none_regular_nms[] = { + 0x6d, 0x61, 0x78, 0x5f, 0x64, 0x65, 0x74, 0x65, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x73, 0x00, 0x6d, 0x61, 0x78, 0x5f, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x65, 0x73, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x64, 0x65, 0x74, 0x65, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x00, 0x64, 0x65, 0x74, 0x65, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x73, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x63, 0x6c, 0x61, 0x73, + 0x73, 0x00, 0x75, 0x73, 0x65, 0x5f, 0x72, 0x65, 0x67, 0x75, 0x6c, 0x61, + 0x72, 0x5f, 0x6e, 0x6d, 0x73, 0x00, 0x6e, 0x6d, 0x73, 0x5f, 0x73, 0x63, + 0x6f, 0x72, 0x65, 0x5f, 0x74, 0x68, 0x72, 0x65, 0x73, 0x68, 0x6f, 0x6c, + 0x64, 0x00, 0x6e, 0x6d, 0x73, 0x5f, 0x69, 0x6f, 0x75, 0x5f, 0x74, 0x68, + 0x72, 0x65, 0x73, 0x68, 0x6f, 0x6c, 0x64, 0x00, 0x6e, 0x75, 0x6d, 0x5f, + 0x63, 0x6c, 0x61, 0x73, 0x73, 0x65, 0x73, 0x00, 0x79, 0x5f, 0x73, 0x63, + 0x61, 0x6c, 0x65, 0x00, 0x78, 0x5f, 0x73, 0x63, 0x61, 0x6c, 0x65, 0x00, + 0x68, 0x5f, 0x73, 0x63, 0x61, 0x6c, 0x65, 0x00, 0x77, 0x5f, 0x73, 0x63, + 0x61, 0x6c, 0x65, 0x00, 0x0b, 0x78, 0x12, 0x94, 0xa4, 0x43, 0x58, 0x33, + 0x6a, 0x11, 0x22, 0x2b, 0x0b, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0b, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0xa0, 0x40, + 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x3f, + 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0xa0, 0x40, 0x00, 0x00, 0x20, 0x41, 0x00, 0x00, 0x20, 0x41, + 0x06, 0x0e, 0x06, 0x06, 0x0e, 0x0e, 0x06, 0x6a, 0x0e, 0x0e, 0x0e, 0x37, + 0x26, 0x01, +}; +const int g_gen_data_size_regular_nms = 242; +const unsigned char g_gen_data_regular_nms[] = { + 0x6d, 0x61, 0x78, 0x5f, 0x64, 0x65, 0x74, 0x65, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x73, 0x00, 0x6d, 0x61, 0x78, 0x5f, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x65, 0x73, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x64, 0x65, 0x74, 0x65, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x00, 0x64, 0x65, 0x74, 0x65, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x73, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x63, 0x6c, 0x61, 0x73, + 0x73, 0x00, 0x75, 0x73, 0x65, 0x5f, 0x72, 0x65, 0x67, 0x75, 0x6c, 0x61, + 0x72, 0x5f, 0x6e, 0x6d, 0x73, 0x00, 0x6e, 0x6d, 0x73, 0x5f, 0x73, 0x63, + 0x6f, 0x72, 0x65, 0x5f, 0x74, 0x68, 0x72, 0x65, 0x73, 0x68, 0x6f, 0x6c, + 0x64, 0x00, 0x6e, 0x6d, 0x73, 0x5f, 0x69, 0x6f, 0x75, 0x5f, 0x74, 0x68, + 0x72, 0x65, 0x73, 0x68, 0x6f, 0x6c, 0x64, 0x00, 0x6e, 0x75, 0x6d, 0x5f, + 0x63, 0x6c, 0x61, 0x73, 0x73, 0x65, 0x73, 0x00, 0x79, 0x5f, 0x73, 0x63, + 0x61, 0x6c, 0x65, 0x00, 0x78, 0x5f, 0x73, 0x63, 0x61, 0x6c, 0x65, 0x00, + 0x68, 0x5f, 0x73, 0x63, 0x61, 0x6c, 0x65, 0x00, 0x77, 0x5f, 0x73, 0x63, + 0x61, 0x6c, 0x65, 0x00, 0x0b, 0x78, 0x12, 0x94, 0xa4, 0x43, 0x58, 0x33, + 0x6a, 0x11, 0x22, 0x2b, 0x0b, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0b, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0xa0, 0x40, + 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x3f, + 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0xa0, 0x40, 0x00, 0x00, 0x20, 0x41, 0x00, 0x00, 0x20, 0x41, + 0x06, 0x0e, 0x06, 0x06, 0x0e, 0x0e, 0x06, 0x6a, 0x0e, 0x0e, 0x0e, 0x37, + 0x26, 0x01, +}; + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.h new file mode 100644 index 0000000..e4df15f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/flexbuffers_generated_data.h @@ -0,0 +1,28 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_FLEXBUFFERS_GENERATED_DATA_H +#define TENSORFLOW_LITE_MICRO_KERNELS_FLEXBUFFERS_GENERATED_DATA_H + +extern const int g_gen_data_size_none_regular_nms; +extern const unsigned char g_gen_data_none_regular_nms[]; + +extern const int g_gen_data_size_regular_nms; +extern const unsigned char g_gen_data_regular_nms[]; + +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/floor.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/floor.cpp new file mode 100644 index 0000000..8f7ae18 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/floor.cpp @@ -0,0 +1,60 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/floor.h" + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace floor { + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteFloat32); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + reference_ops::Floor(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; +} +} // namespace floor + +TfLiteRegistration Register_FLOOR() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/floor::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/fully_connected.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/fully_connected.h new file mode 100644 index 0000000..3f0e936 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/fully_connected.h @@ -0,0 +1,53 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_FULLY_CONNECTED_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_FULLY_CONNECTED_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +// This is the most generic TfLiteRegistration. The actual supported types may +// still be target dependent. The only requirement is that every implementation +// (reference or optimized) must define this function. +TfLiteRegistration Register_FULLY_CONNECTED(); + +#if defined(CMSIS_NN) || defined(ARDUINO) +// The Arduino is a special case where we use the CMSIS kernels, but because of +// the current approach to building for Arduino, we do not support -DCMSIS_NN as +// part of the build. As a result, we use defined(ARDUINO) as proxy for the +// CMSIS kernels for this one special case. + +// Returns a TfLiteRegistration struct for cmsis-nn kernel variant that only +// supports int8. +TfLiteRegistration Register_FULLY_CONNECTED_INT8(); + +#else +// Note that while this block gets used for both reference and optimized kernels +// that do not have any specialized implementations, the only goal here is to +// define fallback implementation that allow reference kernels to still be used +// from applications that call a more specific kernel variant. + +inline TfLiteRegistration Register_FULLY_CONNECTED_INT8() { + return Register_FULLY_CONNECTED(); +} + +#endif +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_FULLY_CONNECTED_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/hard_swish.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/hard_swish.cpp new file mode 100644 index 0000000..ac9642e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/hard_swish.cpp @@ -0,0 +1,145 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/hard_swish.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace hard_swish { + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +void* HardSwishInit(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(HardSwishParams)); +} + +TfLiteStatus HardSwishPrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + + if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8) { + HardSwishParams* params = static_cast(node->user_data); + + params->input_zero_point = input->params.zero_point; + params->output_zero_point = output->params.zero_point; + + const float input_scale = input->params.scale; + const float hires_input_scale = (1.0f / 128.0f) * input_scale; + const float reluish_scale = 3.0f / 32768.0f; + const float output_scale = output->params.scale; + + const double output_multiplier = + static_cast(hires_input_scale / output_scale); + int32_t output_multiplier_fixedpoint_int32; + QuantizeMultiplier(output_multiplier, &output_multiplier_fixedpoint_int32, + ¶ms->output_multiplier_exponent); + DownScaleInt32ToInt16Multiplier( + output_multiplier_fixedpoint_int32, + ¶ms->output_multiplier_fixedpoint_int16); + + TF_LITE_ENSURE(context, params->output_multiplier_exponent <= 0); + + const double reluish_multiplier = + static_cast(hires_input_scale / reluish_scale); + int32_t reluish_multiplier_fixedpoint_int32; + QuantizeMultiplier(reluish_multiplier, &reluish_multiplier_fixedpoint_int32, + ¶ms->reluish_multiplier_exponent); + DownScaleInt32ToInt16Multiplier( + reluish_multiplier_fixedpoint_int32, + ¶ms->reluish_multiplier_fixedpoint_int16); + } + + return kTfLiteOk; +} + +TfLiteStatus HardSwishEval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + HardSwishParams* params = static_cast(node->user_data); + + switch (input->type) { + case kTfLiteFloat32: { + tflite::reference_ops::HardSwish( + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } break; + case kTfLiteUInt8: { + tflite::reference_ops::HardSwish( + *params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } break; + case kTfLiteInt8: { + tflite::reference_ops::HardSwish( + *params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } break; + default: { + TF_LITE_KERNEL_LOG( + context, + "Only float32/int8_t/uint8_t are supported currently, got %s", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + } + return kTfLiteOk; +} + +} // namespace hard_swish + +TfLiteRegistration Register_HARD_SWISH() { + return {/*init=*/hard_swish::HardSwishInit, + /*free=*/nullptr, + /*prepare=*/hard_swish::HardSwishPrepare, + /*invoke=*/hard_swish::HardSwishEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.cpp new file mode 100644 index 0000000..595800c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.cpp @@ -0,0 +1,169 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.h" + +namespace tflite { +namespace micro { + +namespace { +constexpr size_t kBufferAlignment = 16; +} // namespace + +// TODO(b/161841696): Consider moving away from global arena buffers: +constexpr int KernelRunner::kNumScratchBuffers_; +constexpr int KernelRunner::kKernelRunnerBufferSize_; +uint8_t KernelRunner::kKernelRunnerBuffer_[]; + +KernelRunner::KernelRunner(const TfLiteRegistration& registration, + TfLiteTensor* tensors, int tensors_size, + TfLiteIntArray* inputs, TfLiteIntArray* outputs, + void* builtin_data, ErrorReporter* error_reporter) + : allocator_(SimpleMemoryAllocator::Create( + error_reporter, kKernelRunnerBuffer_, kKernelRunnerBufferSize_)), + registration_(registration), + tensors_(tensors), + error_reporter_(error_reporter) { + // Prepare TfLiteContext: + context_.impl_ = static_cast(this); + context_.ReportError = ReportOpError; + context_.recommended_num_threads = 1; + context_.GetTensor = GetTensor; + context_.GetEvalTensor = GetEvalTensor; + context_.AllocatePersistentBuffer = AllocatePersistentBuffer; + context_.RequestScratchBufferInArena = RequestScratchBufferInArena; + context_.GetScratchBuffer = GetScratchBuffer; + + // Prepare TfLiteNode: + node_.inputs = inputs; + node_.outputs = outputs; + node_.builtin_data = builtin_data; +} + +TfLiteStatus KernelRunner::InitAndPrepare(const char* init_data, + size_t length) { + if (registration_.init) { + node_.user_data = registration_.init(&context_, init_data, length); + } + if (registration_.prepare) { + TF_LITE_ENSURE_STATUS(registration_.prepare(&context_, &node_)); + } + return kTfLiteOk; +} + +TfLiteStatus KernelRunner::Invoke() { + if (registration_.invoke == nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, + "TfLiteRegistration missing invoke function pointer!"); + return kTfLiteError; + } + return registration_.invoke(&context_, &node_); +} + +TfLiteTensor* KernelRunner::GetTensor(const struct TfLiteContext* context, + int tensor_index) { + TFLITE_DCHECK(context != nullptr); + KernelRunner* runner = reinterpret_cast(context->impl_); + TFLITE_DCHECK(runner != nullptr); + + return &runner->tensors_[tensor_index]; +} + +TfLiteEvalTensor* KernelRunner::GetEvalTensor( + const struct TfLiteContext* context, int tensor_index) { + TFLITE_DCHECK(context != nullptr); + KernelRunner* runner = reinterpret_cast(context->impl_); + TFLITE_DCHECK(runner != nullptr); + + TfLiteEvalTensor* eval_tensor = + reinterpret_cast(runner->allocator_->AllocateTemp( + sizeof(TfLiteEvalTensor), alignof(TfLiteEvalTensor))); + TFLITE_DCHECK(eval_tensor != nullptr); + + // In unit tests, the TfLiteTensor pointer contains the source of truth for + // buffers and values: + eval_tensor->data = runner->tensors_[tensor_index].data; + eval_tensor->dims = runner->tensors_[tensor_index].dims; + eval_tensor->type = runner->tensors_[tensor_index].type; + return eval_tensor; +} + +void* KernelRunner::AllocatePersistentBuffer(TfLiteContext* context, + size_t bytes) { + TFLITE_DCHECK(context != nullptr); + KernelRunner* runner = reinterpret_cast(context->impl_); + TFLITE_DCHECK(runner != nullptr); + + return runner->allocator_->AllocateFromTail(bytes, kBufferAlignment); +} + +TfLiteStatus KernelRunner::RequestScratchBufferInArena(TfLiteContext* context, + size_t bytes, + int* buffer_index) { + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(buffer_index != nullptr); + + KernelRunner* runner = reinterpret_cast(context->impl_); + TFLITE_DCHECK(runner != nullptr); + + if (runner->scratch_buffer_count_ == kNumScratchBuffers_) { + TF_LITE_REPORT_ERROR( + runner->error_reporter_, + "Exceeded the maximum number of scratch tensors allowed (%d).", + kNumScratchBuffers_); + return kTfLiteError; + } + + // For tests, we allocate scratch buffers from the tail and keep them around + // for the lifetime of model. This means that the arena size in the tests will + // be more than what we would have if the scratch buffers could share memory. + runner->scratch_buffers_[runner->scratch_buffer_count_] = + runner->allocator_->AllocateFromTail(bytes, kBufferAlignment); + TFLITE_DCHECK(runner->scratch_buffers_[runner->scratch_buffer_count_] != + nullptr); + + *buffer_index = runner->scratch_buffer_count_++; + return kTfLiteOk; +} + +void* KernelRunner::GetScratchBuffer(TfLiteContext* context, int buffer_index) { + TFLITE_DCHECK(context != nullptr); + KernelRunner* runner = reinterpret_cast(context->impl_); + TFLITE_DCHECK(runner != nullptr); + + TFLITE_DCHECK(runner->scratch_buffer_count_ <= kNumScratchBuffers_); + if (buffer_index >= runner->scratch_buffer_count_) { + return nullptr; + } + return runner->scratch_buffers_[buffer_index]; +} + +void KernelRunner::ReportOpError(struct TfLiteContext* context, + const char* format, ...) { + TFLITE_DCHECK(context != nullptr); + KernelRunner* runner = reinterpret_cast(context->impl_); + TFLITE_DCHECK(runner != nullptr); + + va_list args; + va_start(args, format); + TF_LITE_REPORT_ERROR(runner->error_reporter_, format, args); + va_end(args); +} + +} // namespace micro +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.h new file mode 100644 index 0000000..c3a5f7f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_runner.h @@ -0,0 +1,87 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_KERNEL_RUNNER_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_KERNEL_RUNNER_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h" + +namespace tflite { +namespace micro { + +// Helper class to perform a simulated kernel (i.e. TfLiteRegistration) +// lifecycle (init, prepare, invoke). All internal allocations are handled by +// this class. Simply pass in the registration, list of required tensors, inputs +// array, outputs array, and any pre-builtin data. Calling Invoke() will +// automatically walk the kernel and outputs will be ready on the TfLiteTensor +// output provided during construction. +class KernelRunner { + public: + KernelRunner(const TfLiteRegistration& registration, TfLiteTensor* tensors, + int tensors_size, TfLiteIntArray* inputs, + TfLiteIntArray* outputs, void* builtin_data, + ErrorReporter* error_reporter); + + // Calls init and prepare on the kernel (i.e. TfLiteRegistration) struct. Any + // exceptions will be reported through the error_reporter and returned as a + // status code here. + TfLiteStatus InitAndPrepare(const char* init_data = nullptr, + size_t length = 0); + + // Calls init, prepare, and invoke on a given TfLiteRegistration pointer. + // After successful invoke, results will be available in the output tensor as + // passed into the constructor of this class. + TfLiteStatus Invoke(); + + protected: + static TfLiteTensor* GetTensor(const struct TfLiteContext* context, + int tensor_index); + static TfLiteEvalTensor* GetEvalTensor(const struct TfLiteContext* context, + int tensor_index); + static void* AllocatePersistentBuffer(TfLiteContext* context, size_t bytes); + static TfLiteStatus RequestScratchBufferInArena(TfLiteContext* context, + size_t bytes, + int* buffer_index); + static void* GetScratchBuffer(TfLiteContext* context, int buffer_index); + static void ReportOpError(struct TfLiteContext* context, const char* format, + ...); + + private: + static constexpr int kNumScratchBuffers_ = 12; + + static constexpr int kKernelRunnerBufferSize_ = 10000; + static uint8_t kKernelRunnerBuffer_[kKernelRunnerBufferSize_]; + + SimpleMemoryAllocator* allocator_ = nullptr; + const TfLiteRegistration& registration_; + TfLiteTensor* tensors_ = nullptr; + ErrorReporter* error_reporter_ = nullptr; + + TfLiteContext context_ = {}; + TfLiteNode node_ = {}; + + int scratch_buffer_count_ = 0; + uint8_t* scratch_buffers_[kNumScratchBuffers_]; +}; + +} // namespace micro +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_KERNEL_RUNNER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h new file mode 100644 index 0000000..e643c83 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h @@ -0,0 +1,78 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_KERNEL_UTIL_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_KERNEL_UTIL_H_ + +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace micro { + +// Returns a mutable tensor for a given input index. is_variable must be checked +// during prepare when the full TfLiteTensor is available. +inline TfLiteEvalTensor* GetMutableEvalInput(const TfLiteContext* context, + const TfLiteNode* node, + int index) { + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(node != nullptr); + return context->GetEvalTensor(context, node->inputs->data[index]); +} + +// Returns the TfLiteEvalTensor struct for a given input index in a node. +inline const TfLiteEvalTensor* GetEvalInput(const TfLiteContext* context, + const TfLiteNode* node, int index) { + return GetMutableEvalInput(context, node, index); +} + +// Returns the TfLiteEvalTensor struct for a given output index in a node. +inline TfLiteEvalTensor* GetEvalOutput(const TfLiteContext* context, + const TfLiteNode* node, int index) { + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(node != nullptr); + return context->GetEvalTensor(context, node->outputs->data[index]); +} + +// Returns data for a TfLiteEvalTensor struct. +template +T* GetTensorData(TfLiteEvalTensor* tensor) { + return tensor != nullptr ? reinterpret_cast(tensor->data.raw) : nullptr; +} + +// Returns const data for a TfLiteEvalTensor struct. +template +const T* GetTensorData(const TfLiteEvalTensor* tensor) { + TFLITE_DCHECK(tensor != nullptr); + return reinterpret_cast(tensor->data.raw); +} + +// Returns the shape of a TfLiteEvalTensor struct. +const RuntimeShape GetTensorShape(const TfLiteEvalTensor* tensor); + +// Return true if the given tensors have the same shape. +bool HaveSameShapes(const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2); + +} // namespace micro +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_KERNEL_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util_full.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util_full.cpp new file mode 100644 index 0000000..20fd0f6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util_full.cpp @@ -0,0 +1,44 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { +namespace micro { + +bool HaveSameShapes(const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2) { + TFLITE_DCHECK(input1 != nullptr); + TFLITE_DCHECK(input2 != nullptr); + return TfLiteIntArrayEqual(input1->dims, input2->dims); +} + +const RuntimeShape GetTensorShape(const TfLiteEvalTensor* tensor) { + if (tensor == nullptr || tensor->dims == nullptr) { + return RuntimeShape(); + } + TfLiteIntArray* dims = tensor->dims; + const int dims_size = dims->size; + const int32_t* dims_data = reinterpret_cast(dims->data); + return RuntimeShape(dims_size, dims_data); +} + +} // namespace micro +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/l2norm.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/l2norm.cpp new file mode 100644 index 0000000..ab5b5a8 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/l2norm.cpp @@ -0,0 +1,160 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/l2normalization.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace l2norm { + +namespace { + +// This file has two implementation of L2Norm. +enum KernelType { + kReference, + kGenericOptimized, +}; + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +} // namespace + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + auto* params = reinterpret_cast(node->builtin_data); + L2NormalizationParams* data = + static_cast(node->user_data); + + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + + TF_LITE_ENSURE(context, NumDimensions(input) <= 4); + + TF_LITE_ENSURE(context, output->type == kTfLiteFloat32 || + output->type == kTfLiteUInt8 || + output->type == kTfLiteInt8); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + + if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + data->input_zero_point = input->params.zero_point; + } else if (output->type == kTfLiteFloat32) { + data->input_zero_point = 0; + } + + // TODO(ahentz): For some reason our implementations don't support + // activations. + TF_LITE_ENSURE_EQ(context, params->activation, kTfLiteActNone); + + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, + sizeof(L2NormalizationParams)); +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const L2NormalizationParams& data = + *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + // TODO(b/143912164): instead of hardcode the epsilon here, we should read it + // from tensorflow, i.e., adding a params. + // We don't compute epsilon for quantized kernel: + // + // epsilon_float = (epsilon_quant - zp) * scale + // so + // espsilon_quant = epsilon_float / scale + zp + // We know epsilon_float is just a very small number to avoid division by + // zero error, and scale is > 1, so the integer value of epsilon for quant + // is just dominated by the zero point. + // Also, GetInvSqrtQuantizedMultiplierExp handles the scenario where the sum + // of input value squared is zero case well. + // So we don't even need to do handle the epsilon for quantized kernel case. + const float epsilon = 1e-6f; + if (output->type == kTfLiteFloat32) { + reference_ops::L2Normalization(data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), + epsilon); + } else if (output->type == kTfLiteUInt8) { + reference_ops::L2Normalization( + data, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else if (output->type == kTfLiteInt8) { + const auto input_shape = tflite::micro::GetTensorShape(input); + const auto output_shape = tflite::micro::GetTensorShape(output); + const int trailing_dim = input_shape.DimensionsCount() - 1; + const int depth = + MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); + const int outer_size = + MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); + reference_integer_ops::L2Normalization( + data.input_zero_point, outer_size, depth, + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorData(output)); + } else { + TF_LITE_KERNEL_LOG(context, "Output type is %s, requires float.", + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + + return kTfLiteOk; +} + +} // namespace l2norm + +TfLiteRegistration Register_L2NORM_REF() { + return {/*init=*/l2norm::Init, + /*free=*/nullptr, + /*prepare=*/l2norm::Prepare, + /*invoke=*/l2norm::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_L2_NORMALIZATION() { return Register_L2NORM_REF(); } + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/logical.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/logical.cpp new file mode 100644 index 0000000..ddbb30c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/logical.cpp @@ -0,0 +1,108 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/binary_function.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace logical { +namespace { + +// Input/output tensor index. +constexpr int kInputTensor1 = 0; +constexpr int kInputTensor2 = 1; +constexpr int kOutputTensor = 0; + +TfLiteStatus LogicalImpl(TfLiteContext* context, TfLiteNode* node, + bool (*func)(bool, bool)) { + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + if (tflite::micro::HaveSameShapes(input1, input2)) { + reference_ops::BinaryFunction( + tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), func); + } else { + reference_ops::BroadcastBinaryFunction4DSlow( + tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), func); + } + + return kTfLiteOk; +} + +bool LogicalOr(bool x, bool y) { return x || y; } + +TfLiteStatus LogicalOrEval(TfLiteContext* context, TfLiteNode* node) { + return LogicalImpl(context, node, LogicalOr); +} + +bool LogicalAnd(bool x, bool y) { return x && y; } + +TfLiteStatus LogicalAndEval(TfLiteContext* context, TfLiteNode* node) { + return LogicalImpl(context, node, LogicalAnd); +} + +} // namespace +} // namespace logical + +TfLiteRegistration Register_LOGICAL_OR() { + // Init, Free, Prepare, Eval are satisfying the Interface required by + // TfLiteRegistration. + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/logical::LogicalOrEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_LOGICAL_AND() { + // Init, Free, Prepare, Eval are satisfying the Interface required by + // TfLiteRegistration. + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/logical::LogicalAndEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/logistic.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/logistic.cpp new file mode 100644 index 0000000..16ff55c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/logistic.cpp @@ -0,0 +1,153 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/logistic.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace activations { +namespace { +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +struct OpData { + int32_t input_zero_point; + int32_t input_range_radius; + int32_t input_multiplier; + int input_left_shift; +}; + +TfLiteStatus CalculateArithmeticOpData(TfLiteContext* context, TfLiteNode* node, + OpData* data) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + if (input->type == kTfLiteInt8) { + TF_LITE_ENSURE_EQ(context, output->params.zero_point, + std::numeric_limits::min()); + + static constexpr int kInputIntegerBits = 4; + const double input_real_multiplier = + static_cast(input->params.scale) * + static_cast(1 << (31 - kInputIntegerBits)); + + data->input_zero_point = input->params.zero_point; + + const double q = std::frexp(input_real_multiplier, &data->input_left_shift); + data->input_multiplier = static_cast(TfLiteRound(q * (1ll << 31))); + + data->input_range_radius = + CalculateInputRadius(kInputIntegerBits, data->input_left_shift, 31); + } + return kTfLiteOk; +} +} // namespace + +void* LogisticInit(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus LogisticPrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + return CalculateArithmeticOpData(context, node, data); +} + +TfLiteStatus LogisticEval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + if (input->type == kTfLiteFloat32) { + switch (output->type) { + case kTfLiteFloat32: { + reference_ops::Logistic(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else if (input->type == kTfLiteInt8) { + switch (output->type) { + case kTfLiteInt8: { + reference_integer_ops::Logistic( + data->input_zero_point, data->input_range_radius, + data->input_multiplier, data->input_left_shift, + NumElements(input->dims), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else { + // TODO(b/141211002): Also support other data types once we have supported + // temporary tensors in TFLM. + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace activations + +TfLiteRegistration Register_LOGISTIC() { + return {/*init=*/activations::LogisticInit, + /*free=*/nullptr, + /*prepare=*/activations::LogisticPrepare, + /*invoke=*/activations::LogisticEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/maximum_minimum.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/maximum_minimum.cpp new file mode 100644 index 0000000..a7a355b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/maximum_minimum.cpp @@ -0,0 +1,151 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/maximum_minimum.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace maximum_minimum { +namespace { + +// This file has a reference implementation of TFMaximum/TFMinimum. +enum KernelType { + kReference, +}; + +constexpr int kInputTensor1 = 0; +constexpr int kInputTensor2 = 1; +constexpr int kOutputTensor = 0; + +struct OpContext { + OpContext(TfLiteContext* context, TfLiteNode* node) { + input1 = tflite::micro::GetEvalInput(context, node, kInputTensor1); + input2 = tflite::micro::GetEvalInput(context, node, kInputTensor2); + output = tflite::micro::GetEvalOutput(context, node, kOutputTensor); + } + const TfLiteEvalTensor* input1; + const TfLiteEvalTensor* input2; + TfLiteEvalTensor* output; +}; + +struct MaximumOp { + template + static data_type op(data_type el1, data_type el2) { + return el1 > el2 ? el1 : el2; + } +}; + +struct MinimumOp { + template + static data_type op(data_type el1, data_type el2) { + return el1 < el2 ? el1 : el2; + } +}; + +} // namespace + +template +void TFLiteOperation(TfLiteContext* context, TfLiteNode* node, + const OpContext& op_context) { + reference_ops::MaximumMinimumBroadcastSlow( + tflite::micro::GetTensorShape(op_context.input1), + tflite::micro::GetTensorData(op_context.input1), + tflite::micro::GetTensorShape(op_context.input2), + tflite::micro::GetTensorData(op_context.input2), + tflite::micro::GetTensorShape(op_context.output), + tflite::micro::GetTensorData(op_context.output), + op_type::template op); +} + +template +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + OpContext op_context(context, node); + + if (kernel_type == kReference) { + switch (op_context.output->type) { + case kTfLiteFloat32: + TFLiteOperation(context, node, op_context); + break; + case kTfLiteUInt8: + TFLiteOperation(context, node, op_context); + break; + case kTfLiteInt8: + TFLiteOperation(context, node, op_context); + break; + case kTfLiteInt32: + TFLiteOperation(context, node, op_context); + break; + case kTfLiteInt64: + TFLiteOperation(context, node, op_context); + break; + default: + TF_LITE_KERNEL_LOG(context, + "Type %s (%d) is not supported by Maximum/Minimum.", + TfLiteTypeGetName(op_context.output->type), + op_context.output->type); + return kTfLiteError; + } + } else { + TF_LITE_KERNEL_LOG(context, + "Kernel type not supported by Maximum/Minimum."); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace maximum_minimum + +TfLiteRegistration Register_MAXIMUM() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/ + maximum_minimum::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_MINIMUM() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/ + maximum_minimum::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/micro_ops.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/micro_ops.h new file mode 100644 index 0000000..6653e8f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/micro_ops.h @@ -0,0 +1,104 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_MICRO_OPS_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_MICRO_OPS_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +// Forward declaration of all micro op kernel registration methods. These +// registrations are included with the standard `BuiltinOpResolver`. +// +// This header is particularly useful in cases where only a subset of ops are +// needed. In such cases, the client can selectively add only the registrations +// their model requires, using a custom `(Micro)MutableOpResolver`. Selective +// registration in turn allows the linker to strip unused kernels. + +namespace tflite { + +// TFLM is incrementally moving towards a flat tflite namespace +// (https://abseil.io/tips/130). Any new ops (or cleanup of existing ops should +// have their Register function declarations in the tflite namespace. + +TfLiteRegistration Register_CONV_2D(); +TfLiteRegistration Register_DEPTHWISE_CONV_2D(); +TfLiteRegistration Register_QUANTIZE(); +TfLiteRegistration Register_SHAPE(); +TfLiteRegistration Register_SOFTMAX(); +TfLiteRegistration Register_SVDF(); +TfLiteRegistration Register_TRANSPOSE_CONV_2D(); + +namespace ops { +namespace micro { + +TfLiteRegistration Register_ABS(); +TfLiteRegistration Register_ADD(); +TfLiteRegistration Register_ARG_MAX(); +TfLiteRegistration Register_ARG_MIN(); +TfLiteRegistration Register_AVERAGE_POOL_2D(); +TfLiteRegistration Register_CEIL(); +// TODO(b/160234179): Change custom OPs to also return by value. +TfLiteRegistration* Register_CIRCULAR_BUFFER(); +TfLiteRegistration Register_CONCATENATION(); +TfLiteRegistration Register_COS(); +TfLiteRegistration Register_DEQUANTIZE(); +TfLiteRegistration Register_EQUAL(); +TfLiteRegistration Register_FLOOR(); +TfLiteRegistration Register_GREATER(); +TfLiteRegistration Register_GREATER_EQUAL(); +TfLiteRegistration Register_HARD_SWISH(); +TfLiteRegistration Register_LESS(); +TfLiteRegistration Register_LESS_EQUAL(); +TfLiteRegistration Register_LOG(); +TfLiteRegistration Register_LOGICAL_AND(); +TfLiteRegistration Register_LOGICAL_NOT(); +TfLiteRegistration Register_LOGICAL_OR(); +TfLiteRegistration Register_LOGISTIC(); +TfLiteRegistration Register_MAXIMUM(); +TfLiteRegistration Register_MAX_POOL_2D(); +TfLiteRegistration Register_MEAN(); +TfLiteRegistration Register_MINIMUM(); +TfLiteRegistration Register_MUL(); +TfLiteRegistration Register_NEG(); +TfLiteRegistration Register_NOT_EQUAL(); +TfLiteRegistration Register_PACK(); +TfLiteRegistration Register_PAD(); +TfLiteRegistration Register_PADV2(); +TfLiteRegistration Register_PRELU(); +TfLiteRegistration Register_REDUCE_MAX(); +TfLiteRegistration Register_RELU(); +TfLiteRegistration Register_RELU6(); +TfLiteRegistration Register_RESHAPE(); +TfLiteRegistration Register_RESIZE_NEAREST_NEIGHBOR(); +TfLiteRegistration Register_ROUND(); +TfLiteRegistration Register_RSQRT(); +TfLiteRegistration Register_SIN(); +TfLiteRegistration Register_SPLIT(); +TfLiteRegistration Register_SPLIT_V(); +TfLiteRegistration Register_SQRT(); +TfLiteRegistration Register_SQUARE(); +TfLiteRegistration Register_STRIDED_SLICE(); +TfLiteRegistration Register_SUB(); +TfLiteRegistration Register_UNPACK(); +TfLiteRegistration Register_L2_NORMALIZATION(); +TfLiteRegistration Register_TANH(); + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_MICRO_OPS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/micro_utils.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/micro_utils.h similarity index 95% rename from src/tensorflow/lite/micro/kernels/micro_utils.h rename to src/tensorflow_arm/tensorflow/lite/micro/kernels/micro_utils.h index 85db263..76a4e21 100644 --- a/src/tensorflow/lite/micro/kernels/micro_utils.h +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/micro_utils.h @@ -1,8 +1,12 @@ +#if !defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -35,3 +39,5 @@ struct Less { } // namespace ops } // namespace tflite #endif // TENSORFLOW_LITE_MICRO_KERNELS_MICRO_UTILS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/neg.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/neg.cpp new file mode 100644 index 0000000..c25d289 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/neg.cpp @@ -0,0 +1,69 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/neg.h" + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace neg { + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + switch (input->type) { + // TODO(wangtz): handle for kTfLiteInt8 + case kTfLiteFloat32: + reference_ops::Negate(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace neg + +TfLiteRegistration Register_NEG() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/neg::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/pack.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/pack.cpp new file mode 100644 index 0000000..9d19a96 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/pack.cpp @@ -0,0 +1,130 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace pack { +namespace { + +constexpr int kOutputTensor = 0; + +template +TfLiteStatus PackImpl(TfLiteContext* context, TfLiteNode* node, + TfLiteEvalTensor* output, int values_count, int axis) { + const TfLiteEvalTensor* input0 = + tflite::micro::GetEvalInput(context, node, 0); + + const int dimensions = output->dims->size; + const TfLiteIntArray* input_dims = input0->dims; + const TfLiteIntArray* output_dims = output->dims; + + if (axis < 0) { + axis += dimensions; + } + + int outer_size = 1; + for (int i = 0; i < axis; ++i) { + outer_size *= output_dims->data[i]; + } + int copy_size = 1; + for (int i = axis + 1; i < dimensions; ++i) { + copy_size *= output_dims->data[i]; + } + int input_size = 1; + for (int i = 0; i < input_dims->size; ++i) { + input_size *= input_dims->data[i]; + } + TFLITE_DCHECK_EQ(input_size, copy_size * outer_size); + + T* output_data = tflite::micro::GetTensorData(output); + + for (int i = 0; i < values_count; ++i) { + const TfLiteEvalTensor* t = tflite::micro::GetEvalInput(context, node, i); + const T* input_data = tflite::micro::GetTensorData(t); + for (int k = 0; k < outer_size; ++k) { + const T* input_ptr = input_data + copy_size * k; + int loc = k * values_count * copy_size + i * copy_size; + T* output_ptr = output_data + loc; + for (int j = 0; j < copy_size; ++j) output_ptr[j] = input_ptr[j]; + } + } + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLitePackParams* data = + reinterpret_cast(node->builtin_data); + + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + switch (output->type) { + case kTfLiteFloat32: { + return PackImpl(context, node, output, data->values_count, + data->axis); + } + case kTfLiteUInt8: { + return PackImpl(context, node, output, data->values_count, + data->axis); + } + case kTfLiteInt8: { + return PackImpl(context, node, output, data->values_count, + data->axis); + } + case kTfLiteInt32: { + return PackImpl(context, node, output, data->values_count, + data->axis); + } + case kTfLiteInt64: { + return PackImpl(context, node, output, data->values_count, + data->axis); + } + default: { + TF_LITE_KERNEL_LOG(context, "Type '%s' is not supported by pack.", + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } + + return kTfLiteOk; +} + +} // namespace +} // namespace pack + +TfLiteRegistration Register_PACK() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/pack::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/pad.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/pad.cpp new file mode 100644 index 0000000..4d4f635 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/pad.cpp @@ -0,0 +1,257 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/pad.h" + +#include + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/portable_tensor.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace pad { +namespace { + +struct OpData { + PadParams params; + int32_t output_zero_point; +}; + +} // namespace + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + OpData* data = static_cast(node->user_data); + + TF_LITE_ENSURE(context, NumInputs(node) == 2 || NumInputs(node) == 3); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + const TfLiteTensor* input = GetInput(context, node, /*index=*/0); + TF_LITE_ENSURE(context, input != nullptr); + const TfLiteTensor* paddings = GetInput(context, node, /*index=*/1); + TF_LITE_ENSURE(context, paddings != nullptr); + const TfLiteTensor* constant_values = + NumInputs(node) == 3 ? GetInput(context, node, /*index=*/2) : nullptr; + TfLiteTensor* output = GetOutput(context, node, /*index=*/0); + TF_LITE_ENSURE(context, output != nullptr); + + TF_LITE_ENSURE_EQ(context, input->type, output->type); + + // Current implementations rely on the inputs being <= 4D. + TF_LITE_ENSURE(context, NumDimensions(input) <= + reference_ops::PadKernelMaxDimensionCount()); + + if (constant_values != nullptr) { + TF_LITE_ENSURE_EQ(context, input->type, constant_values->type); + // Ensure that constant_values is a scalar. + TF_LITE_ENSURE_EQ(context, NumElements(constant_values), 1); + } + + // There must be a pair of paddings for each output dimension. + TF_LITE_ENSURE_EQ(context, GetTensorShape(paddings).FlatSize(), + output->dims->size * 2); + + // On Micro, outputs must be properly sized by the converter. + // NOTE: This data is only available because the paddings buffer is stored in + // the flatbuffer: + TF_LITE_ENSURE(context, IsConstantTensor(paddings)); + const int32_t* paddings_data = GetTensorData(paddings); + for (int i = 0; i < output->dims->size; i++) { + int output_dim = output->dims->data[i]; + int expected_dim = + input->dims->data[i] + paddings_data[i * 2] + paddings_data[i * 2 + 1]; + TF_LITE_ENSURE_EQ(context, output_dim, expected_dim); + } + + // Calculate OpData: + data->params.resizing_category = ResizingCategory::kGenericResize; + const int paddings_total = GetTensorShape(paddings).FlatSize(); + if (paddings_total == 8 && (paddings_data[0] == 0 && paddings_data[1] == 0) && + (paddings_data[6] == 0 && paddings_data[7] == 0)) { + data->params.resizing_category = ResizingCategory::kImageStyle; + } + + const int num_input_dimensions = NumDimensions(input); + data->params.left_padding_count = num_input_dimensions; + data->params.right_padding_count = num_input_dimensions; + + for (int idx = num_input_dimensions - 1; idx >= 0; --idx) { + data->params.left_padding[idx] = paddings_data[idx * 2]; + data->params.right_padding[idx] = paddings_data[idx * 2 + 1]; + } + + if (input->type == kTfLiteInt8 || input->type == kTfLiteUInt8) { + if (constant_values == nullptr) { + // Quantized Pad requires that 0 is represented in the quantized + // range. + if (input->type == kTfLiteUInt8) { + TF_LITE_ENSURE(context, output->params.zero_point >= + std::numeric_limits::min()); + TF_LITE_ENSURE(context, output->params.zero_point <= + std::numeric_limits::max()); + } else { + TF_LITE_ENSURE(context, output->params.zero_point >= + std::numeric_limits::min()); + TF_LITE_ENSURE(context, output->params.zero_point <= + std::numeric_limits::max()); + } + } else { + // Quantized Pad requires that 'constant_values' is represented in the + // same quantized range as the input and output tensors. + TF_LITE_ENSURE_EQ(context, output->params.zero_point, + constant_values->params.zero_point); + TF_LITE_ENSURE_EQ(context, static_cast(output->params.scale), + static_cast(constant_values->params.scale)); + } + data->output_zero_point = output->params.zero_point; + } + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const OpData* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, /*index=*/0); + const TfLiteEvalTensor* constant_values = + NumInputs(node) == 3 + ? tflite::micro::GetEvalInput(context, node, /*index=*/2) + : nullptr; + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, /*index=*/0); + + switch (input->type) { + case kTfLiteFloat32: { + float pad_value = + constant_values == nullptr + ? 0.f + : *tflite::micro::GetTensorData(constant_values); + if (data->params.resizing_category == ResizingCategory::kImageStyle) { + reference_ops::PadImageStyle( + data->params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), &pad_value, + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + reference_ops::Pad(data->params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + &pad_value, tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } + } break; + case kTfLiteUInt8: { + uint8_t pad_value; + if (constant_values == nullptr) { + pad_value = static_cast(data->output_zero_point); + } else { + pad_value = *tflite::micro::GetTensorData(constant_values); + } + if (data->params.resizing_category == ResizingCategory::kImageStyle) { + reference_ops::PadImageStyle( + data->params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), &pad_value, + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + reference_ops::Pad(data->params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + &pad_value, tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } + } break; + case kTfLiteInt8: { + int8_t pad_value; + if (constant_values == nullptr) { + pad_value = static_cast(data->output_zero_point); + } else { + pad_value = *tflite::micro::GetTensorData(constant_values); + } + if (data->params.resizing_category == ResizingCategory::kImageStyle) { + reference_ops::PadImageStyle( + data->params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), &pad_value, + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + reference_ops::Pad(data->params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + &pad_value, tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } + } break; + case kTfLiteInt32: { + int32_t pad_value = + constant_values == nullptr + ? 0 + : *tflite::micro::GetTensorData(constant_values); + reference_ops::Pad(data->params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + &pad_value, tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } break; + default: + + TF_LITE_KERNEL_LOG(context, "Type %s not currently supported by Pad.", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } +#undef TF_LITE_PAD + return kTfLiteOk; +} + +} // namespace pad + +TfLiteRegistration Register_PAD() { + return {/*init=*/pad::Init, + /*free=*/nullptr, + /*prepare=*/pad::Prepare, + /*invoke=*/pad::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +// Also register Pad as PadV2. +TfLiteRegistration Register_PADV2() { + return {/*init=*/pad::Init, + /*free=*/nullptr, + /*prepare=*/pad::Prepare, + /*invoke=*/pad::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/prelu.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/prelu.cpp new file mode 100644 index 0000000..9dff4fa --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/prelu.cpp @@ -0,0 +1,172 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/prelu.h" + +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace activations { +namespace { + +TfLiteStatus CalculatePreluParams(const TfLiteTensor* input, + const TfLiteTensor* alpha, + TfLiteTensor* output, PreluParams* params) { + if (output->type == kTfLiteInt8 || output->type == kTfLiteUInt8 || + output->type == kTfLiteInt16) { + double real_multiplier_1 = static_cast(input->params.scale) / + static_cast(output->params.scale); + double real_multiplier_2 = static_cast(input->params.scale) * + static_cast(alpha->params.scale) / + static_cast(output->params.scale); + QuantizeMultiplier(real_multiplier_1, ¶ms->output_multiplier_1, + ¶ms->output_shift_1); + QuantizeMultiplier(real_multiplier_2, ¶ms->output_multiplier_2, + ¶ms->output_shift_2); + + params->input_offset = -input->params.zero_point; + params->alpha_offset = -alpha->params.zero_point; + params->output_offset = output->params.zero_point; + } + + return kTfLiteOk; +} + +} // namespace + +inline void BroadcastPrelu4DSlowFloat( + const RuntimeShape& unextended_input1_shape, const float* input1_data, + const RuntimeShape& unextended_input2_shape, const float* input2_data, + const RuntimeShape& unextended_output_shape, float* output_data) { + TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4); + TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4); + const RuntimeShape output_shape = + RuntimeShape::ExtendedShape(4, unextended_output_shape); + + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(unextended_input1_shape, + unextended_input2_shape, &desc1, &desc2); + + for (int b = 0; b < output_shape.Dims(0); ++b) { + for (int y = 0; y < output_shape.Dims(1); ++y) { + for (int x = 0; x < output_shape.Dims(2); ++x) { + for (int c = 0; c < output_shape.Dims(3); ++c) { + auto out_idx = Offset(output_shape, b, y, x, c); + auto in1_idx = SubscriptToIndex(desc1, b, y, x, c); + auto in2_idx = SubscriptToIndex(desc2, b, y, x, c); + auto in1_val = input1_data[in1_idx]; + auto in2_val = input2_data[in2_idx]; + output_data[out_idx] = in1_val >= 0.0f ? in1_val : in1_val * in2_val; + } + } + } + } +} + +void* PreluInit(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(PreluParams)); +} + +TfLiteStatus PreluPrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + PreluParams* params = static_cast(node->user_data); + + const TfLiteTensor* input = GetInput(context, node, 0); + TF_LITE_ENSURE(context, input != nullptr); + const TfLiteTensor* alpha = GetInput(context, node, 1); + TF_LITE_ENSURE(context, alpha != nullptr); + TfLiteTensor* output = GetOutput(context, node, 0); + TF_LITE_ENSURE(context, output != nullptr); + + return CalculatePreluParams(input, alpha, output, params); +} + +TfLiteStatus PreluEval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const PreluParams& params = + *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + const TfLiteEvalTensor* alpha = tflite::micro::GetEvalInput(context, node, 1); + TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0); + + switch (input->type) { + case kTfLiteFloat32: { + BroadcastPrelu4DSlowFloat(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(alpha), + tflite::micro::GetTensorData(alpha), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } break; + case kTfLiteUInt8: { + reference_ops::BroadcastPrelu4DSlow( + params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(alpha), + tflite::micro::GetTensorData(alpha), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } break; + case kTfLiteInt8: { + reference_ops::BroadcastPrelu4DSlow( + params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(alpha), + tflite::micro::GetTensorData(alpha), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } break; + default: + TF_LITE_KERNEL_LOG( + context, "Only float32 and uint8_t are supported currently, got %d.", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } +} + +} // namespace activations + +TfLiteRegistration Register_PRELU() { + return {/*init=*/activations::PreluInit, + /*free=*/nullptr, + /*prepare=*/activations::PreluPrepare, + /*invoke=*/activations::PreluEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.cpp new file mode 100644 index 0000000..c358959 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.cpp @@ -0,0 +1,98 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/kernels/quantize.h" + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace { + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, + sizeof(OpDataQuantizeReference)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + auto* data = static_cast(node->user_data); + + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + const TfLiteTensor* input = GetInput(context, node, 0); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, 0); + TF_LITE_ENSURE(context, output != nullptr); + + // TODO(b/128934713): Add support for fixed-point per-channel quantization. + // Currently this only support affine per-layer quantization. + TF_LITE_ENSURE_EQ(context, output->quantization.type, + kTfLiteAffineQuantization); + const auto* affine_quantization = + reinterpret_cast(output->quantization.params); + TF_LITE_ENSURE(context, affine_quantization); + TF_LITE_ENSURE(context, affine_quantization->scale); + TF_LITE_ENSURE(context, affine_quantization->scale->size == 1); + + TF_LITE_ENSURE(context, input->type == kTfLiteFloat32 || + input->type == kTfLiteInt16 || + input->type == kTfLiteInt8); + TF_LITE_ENSURE(context, output->type == kTfLiteUInt8 || + output->type == kTfLiteInt8 || + output->type == kTfLiteInt16 || + output->type == kTfLiteInt32); + + if ((input->type == kTfLiteInt16 && output->type == kTfLiteInt8) || + (input->type == kTfLiteInt8 && output->type == kTfLiteInt8) || + (input->type == kTfLiteInt16 && output->type == kTfLiteInt16) || + (input->type == kTfLiteInt16 && output->type == kTfLiteInt32)) { + double effective_scale = static_cast(input->params.scale) / + static_cast(output->params.scale); + + QuantizeMultiplier(effective_scale, &data->requantize_output_multiplier, + &data->requantize_output_shift); + } + + data->quantization_params.zero_point = output->params.zero_point; + data->quantization_params.scale = static_cast(output->params.scale); + + data->input_zero_point = input->params.zero_point; + return kTfLiteOk; +} + +} // namespace + +TfLiteRegistration Register_QUANTIZE() { + return {/*init=*/Init, + /*free=*/nullptr, + /*prepare=*/Prepare, + /*invoke=*/EvalQuantizeReference, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.h new file mode 100644 index 0000000..46eea34 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize.h @@ -0,0 +1,40 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_QUANTIZE_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_QUANTIZE_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +struct OpDataQuantizeReference { + tflite::QuantizationParams quantization_params; + // The scaling factor from input to output (aka the 'real multiplier') can + // be represented as a fixed point multiplier plus a left shift. + int32_t requantize_output_multiplier; + int requantize_output_shift; + + int32_t input_zero_point; +}; + +TfLiteStatus EvalQuantizeReference(TfLiteContext* context, TfLiteNode* node); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_QUANTIZE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize_common.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize_common.cpp new file mode 100644 index 0000000..9899495 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/quantize_common.cpp @@ -0,0 +1,125 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/quantize.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/requantize.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/quantize.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { + +TfLiteStatus EvalQuantizeReference(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + auto* data = static_cast(node->user_data); + + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0); + + if (input->type == kTfLiteFloat32) { + switch (output->type) { + case kTfLiteInt8: + reference_ops::AffineQuantize( + data->quantization_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + case kTfLiteUInt8: + reference_ops::AffineQuantize( + data->quantization_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + case kTfLiteInt16: + reference_ops::AffineQuantize( + data->quantization_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else if (input->type == kTfLiteInt16) { + size_t size = ElementCount(*input->dims); + switch (output->type) { + case kTfLiteInt8: + reference_ops::Requantize( + tflite::micro::GetTensorData(input), size, + data->requantize_output_multiplier, data->requantize_output_shift, + data->input_zero_point, data->quantization_params.zero_point, + tflite::micro::GetTensorData(output)); + break; + case kTfLiteInt16: + reference_ops::Requantize( + tflite::micro::GetTensorData(input), size, + data->requantize_output_multiplier, data->requantize_output_shift, + data->input_zero_point, data->quantization_params.zero_point, + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + case kTfLiteInt32: + reference_ops::Requantize( + tflite::micro::GetTensorData(input), size, + data->requantize_output_multiplier, data->requantize_output_shift, + data->input_zero_point, data->quantization_params.zero_point, + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else if (input->type == kTfLiteInt8) { + // Int8 to Int8 requantization, required if the input and output tensors + // have different scales and/or zero points. + size_t size = ElementCount(*input->dims); + switch (output->type) { + case kTfLiteInt8: + reference_ops::Requantize( + tflite::micro::GetTensorData(input), size, + data->requantize_output_multiplier, data->requantize_output_shift, + data->input_zero_point, data->quantization_params.zero_point, + tflite::micro::GetTensorData(output)); + break; + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + } else { + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } + + return kTfLiteOk; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/reduce.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/reduce.cpp new file mode 100644 index 0000000..550ab64 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/reduce.cpp @@ -0,0 +1,345 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/reduce.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/mean.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace reduce { + +constexpr int kMaxNumberOfAxis = 4; +constexpr int kMaxNumberOfReducedAxis = 2; + +struct OpData { + int32_t multiplier; + int shift; + int temp_buffer_idx; + int resolved_axis_idx; + int input_zp; + float input_scale; + int output_zp; + float output_scale; + int num_output_elements; +}; + +void* InitReduce(TfLiteContext* context, const char* buffer, size_t length) { + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus PrepareSimple(TfLiteContext* context, TfLiteNode* node) { + // Inputs Tensor (dtype depends on quantization): + // [0] = Input + // [1] = Axis + const TfLiteTensor* input = GetInput(context, node, 0); + + // Outputs Tensor (dtype depends on quantization): + // [0] = Output + + // Validate number of inputs and outputs + TF_LITE_ENSURE_EQ(context, node->inputs->size, 2); + TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); + + // Validate axis type + const TfLiteTensor* axis = GetInput(context, node, 1); + TF_LITE_ENSURE(context, axis != nullptr); + TF_LITE_ENSURE_TYPES_EQ(context, axis->type, kTfLiteInt32); + + if (input->type == kTfLiteInt8) { + OpData* data = static_cast(node->user_data); + const TfLiteTensor* output = GetOutput(context, node, 0); + const double real_multiplier = static_cast(input->params.scale) / + static_cast(output->params.scale); + QuantizeMultiplier(real_multiplier, &data->multiplier, &data->shift); + } + + return kTfLiteOk; +} + +TfLiteStatus PrepareMax(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_OK(context, PrepareSimple(context, node)); + + OpData* op_data = static_cast(node->user_data); + const TfLiteTensor* input = GetInput(context, node, 0); + const TfLiteTensor* output = GetOutput(context, node, 0); + const TfLiteTensor* axis = GetInput(context, node, 1); + + op_data->input_scale = input->params.scale; + op_data->output_scale = output->params.scale; + op_data->num_output_elements = NumElements(output); + + context->RequestScratchBufferInArena(context, sizeof(int) * input->dims->size, + &op_data->temp_buffer_idx); + context->RequestScratchBufferInArena( + context, sizeof(int) * static_cast(ElementCount(*axis->dims)), + &op_data->resolved_axis_idx); + + return kTfLiteOk; +} + +TfLiteStatus PrepareMeanOrSum(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input = GetInput(context, node, 0); + OpData* op_data = reinterpret_cast(node->user_data); + const TfLiteTensor* output = GetOutput(context, node, 0); + if (input->type == kTfLiteInt8) { + const double real_multiplier = static_cast(input->params.scale) / + static_cast(output->params.scale); + QuantizeMultiplier(real_multiplier, &op_data->multiplier, &op_data->shift); + } + + int output_size = NumElements(output); + if (input->type == kTfLiteInt8 || input->type == kTfLiteUInt8) { + context->RequestScratchBufferInArena(context, output_size * sizeof(int32_t), + &op_data->temp_buffer_idx); + op_data->input_zp = input->params.zero_point; + op_data->input_scale = input->params.scale; + op_data->output_zp = output->params.zero_point; + op_data->output_scale = output->params.scale; + } + + TF_LITE_ENSURE_OK(context, PrepareSimple(context, node)); + // TODO(b/144955155): Support uint8_t(b/144955155) and int8_t(b/144955018) + return kTfLiteOk; +} + +void ResolveAxis(const int* axis_data, int axis_count, + tflite::MeanParams* op_params) { + int i = 0; + for (; i < axis_count; ++i) { + op_params->axis[i] = static_cast(axis_data[i]); + } + for (; i < 4; ++i) { + op_params->axis[i] = 1; + } + op_params->axis_count = axis_count; +} + +TfLiteStatus EvalMean(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + const TfLiteEvalTensor* axis = tflite::micro::GetEvalInput(context, node, 1); + TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0); + TfLiteReducerParams* params = + reinterpret_cast(node->builtin_data); + OpData* op_data = reinterpret_cast(node->user_data); + + int num_axis = static_cast(ElementCount(*axis->dims)); + int temp_index[kMaxNumberOfAxis]; + int resolved_axis[kMaxNumberOfReducedAxis]; + + tflite::MeanParams op_params; + ResolveAxis(tflite::micro::GetTensorData(axis), num_axis, &op_params); + + // Special case mean implementation exists for 4D mean across axes 1 and 2. + bool special_case_4d_axes_1_and_2 = + input->dims->size == 4 && op_params.axis_count == 2 && + ((op_params.axis[0] == 1 && op_params.axis[1] == 2) || + (op_params.axis[0] == 2 && op_params.axis[1] == 1)); + + switch (input->type) { + case kTfLiteFloat32: { + // Defer to specialized implementation for 4D Mean across axes 1 & 2. + if (params->keep_dims && special_case_4d_axes_1_and_2) { + reference_ops::Mean(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + TF_LITE_ENSURE( + context, + reference_ops::Mean( + tflite::micro::GetTensorData(input), input->dims->data, + input->dims->size, tflite::micro::GetTensorData(output), + output->dims->data, output->dims->size, + tflite::micro::GetTensorData(axis), num_axis, + params->keep_dims, temp_index, resolved_axis, + tflite::micro::GetTensorData(output))); + } + } break; + case kTfLiteInt8: { + // Defer to specialized implementation for 4D Mean across axes 1 & 2. + if (params->keep_dims && special_case_4d_axes_1_and_2) { + reference_integer_ops::Mean( + op_params, op_data->multiplier, op_data->shift, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), op_data->input_zp, + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), op_data->output_zp); + } else if (op_data->input_zp == op_data->output_zp && + op_data->input_scale == op_data->output_scale) { + int32_t* temp_buffer = static_cast( + context->GetScratchBuffer(context, op_data->temp_buffer_idx)); + TF_LITE_ENSURE( + context, + reference_ops::Mean( + tflite::micro::GetTensorData(input), input->dims->data, + input->dims->size, tflite::micro::GetTensorData(output), + output->dims->data, output->dims->size, + tflite::micro::GetTensorData(axis), num_axis, + params->keep_dims, temp_index, resolved_axis, temp_buffer)); + } else { + int32_t* temp_buffer = static_cast( + context->GetScratchBuffer(context, op_data->temp_buffer_idx)); + TF_LITE_ENSURE( + context, + reference_ops::QuantizedMeanOrSum( + tflite::micro::GetTensorData(input), op_data->input_zp, + op_data->input_scale, input->dims->data, input->dims->size, + tflite::micro::GetTensorData(output), + op_data->output_zp, op_data->output_scale, output->dims->data, + output->dims->size, tflite::micro::GetTensorData(axis), + num_axis, params->keep_dims, temp_index, resolved_axis, + temp_buffer, false)); + } + } break; + case kTfLiteUInt8: { + // Defer to specialized implementation for 4D Mean across axes 1 & 2. + if (params->keep_dims && special_case_4d_axes_1_and_2) { + reference_ops::Mean(op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + op_data->input_zp, op_data->input_scale, + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), + op_data->output_zp, op_data->output_scale); + } else if (op_data->input_zp == op_data->output_zp && + op_data->input_scale == op_data->output_scale) { + uint32_t* temp_buffer = static_cast( + context->GetScratchBuffer(context, op_data->temp_buffer_idx)); + TF_LITE_ENSURE( + context, + reference_ops::Mean(tflite::micro::GetTensorData(input), + input->dims->data, input->dims->size, + tflite::micro::GetTensorData(output), + output->dims->data, output->dims->size, + tflite::micro::GetTensorData(axis), + num_axis, params->keep_dims, temp_index, + resolved_axis, temp_buffer)); + } else { + uint32_t* temp_buffer = static_cast( + context->GetScratchBuffer(context, op_data->temp_buffer_idx)); + TF_LITE_ENSURE( + context, + reference_ops::QuantizedMeanOrSum( + tflite::micro::GetTensorData(input), op_data->input_zp, + op_data->input_scale, input->dims->data, input->dims->size, + tflite::micro::GetTensorData(output), + op_data->output_zp, op_data->output_scale, output->dims->data, + output->dims->size, tflite::micro::GetTensorData(axis), + num_axis, params->keep_dims, temp_index, resolved_axis, + temp_buffer, false)); + } + } break; + default: + TF_LITE_ENSURE_MSG(context, false, + "Currently, only float32, int8 or uint8 input type " + "is supported."); + } + return kTfLiteOk; +} + +TfLiteStatus EvalMax(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + const TfLiteEvalTensor* axis = tflite::micro::GetEvalInput(context, node, 1); + TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + TfLiteReducerParams* params = + static_cast(node->builtin_data); + OpData* op_data = static_cast(node->user_data); + + // Interpret an axis tensor with null dimensions as a scalar + int num_axis = static_cast(ElementCount(*axis->dims)); + int* temp_buffer = static_cast( + context->GetScratchBuffer(context, op_data->temp_buffer_idx)); + int* resolved_axis = static_cast( + context->GetScratchBuffer(context, op_data->resolved_axis_idx)); + switch (input->type) { + case kTfLiteFloat32: + TF_LITE_ENSURE( + context, + reference_ops::ReduceGeneric( + tflite::micro::GetTensorData(input), input->dims->data, + input->dims->size, tflite::micro::GetTensorData(output), + output->dims->data, output->dims->size, + tflite::micro::GetTensorData(axis), num_axis, + params->keep_dims, temp_buffer, resolved_axis, + std::numeric_limits::lowest(), + [](const float current, const float in) -> float { + return (in > current) ? in : current; + })); + break; + case kTfLiteInt8: + TF_LITE_ENSURE_EQ(context, static_cast(op_data->input_scale), + static_cast(op_data->output_scale)); + TF_LITE_ENSURE_EQ(context, op_data->input_zp, op_data->output_zp); + TF_LITE_ENSURE( + context, + reference_ops::ReduceGeneric( + tflite::micro::GetTensorData(input), input->dims->data, + input->dims->size, tflite::micro::GetTensorData(output), + output->dims->data, output->dims->size, + tflite::micro::GetTensorData(axis), num_axis, + params->keep_dims, temp_buffer, resolved_axis, + std::numeric_limits::lowest(), + [](const int8_t current, const int8_t in) -> int8_t { + return (in > current) ? in : current; + })); + break; + default: + TF_LITE_KERNEL_LOG(context, + "Only float32 and int8 types are supported.\n"); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace reduce + +TfLiteRegistration Register_MEAN() { + return {/*init=*/reduce::InitReduce, + /*free=*/nullptr, + /*prepare=*/reduce::PrepareMeanOrSum, + /*invoke=*/reduce::EvalMean, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +TfLiteRegistration Register_REDUCE_MAX() { + return {/*init=*/reduce::InitReduce, + /*free=*/nullptr, + /*prepare=*/reduce::PrepareMax, + /*invoke=*/reduce::EvalMax, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/reshape.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/reshape.cpp new file mode 100644 index 0000000..e61eb92 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/reshape.cpp @@ -0,0 +1,121 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace reshape { + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +TfLiteStatus ReshapeOutput(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + // Tensorflow's Reshape allows one of the shape components to have the + // special -1 value, meaning it will be calculated automatically based on the + // input. Here we calculate what that dimension should be so that the number + // of output elements in the same as the number of input elements. + int num_input_elements = NumElements(input); + TfLiteIntArray* output_shape = output->dims; + + if (NumInputs(node) == 1 && // Legacy scalar supported with params. + output_shape->size == 1 && output_shape->data[0] == 0) { + // Legacy tflite models use a shape parameter of [0] to indicate scalars, + // so adjust accordingly. TODO(b/111614235): Allow zero-sized buffers during + // toco conversion. + output_shape->size = 0; + } + + int num_output_elements = 1; + int stretch_dim = -1; + for (int i = 0; i < output_shape->size; ++i) { + int value = output_shape->data[i]; + if (value == -1) { + TF_LITE_ENSURE_EQ(context, stretch_dim, -1); + stretch_dim = i; + } else { + num_output_elements *= value; + } + } + if (stretch_dim != -1) { + output_shape->data[stretch_dim] = num_input_elements / num_output_elements; + num_output_elements *= output_shape->data[stretch_dim]; + } + + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + TF_LITE_ENSURE_EQ(context, num_input_elements, num_output_elements); + return kTfLiteOk; +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE(context, NumInputs(node) == 1 || NumInputs(node) == 2); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + TF_LITE_ENSURE_EQ(context, ReshapeOutput(context, node), kTfLiteOk); + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + // TODO(b/162522304): storing input bytes in OpData increases some models + // significantly, possibly due to alignment issues. + size_t input_bytes; + TF_LITE_ENSURE_STATUS(TfLiteTypeSizeOf(input->type, &input_bytes)); + input_bytes *= ElementCount(*input->dims); + + // Do nothing for in-place reshape. + if (input->data.raw != output->data.raw) { + // Otherwise perform reshape with copy. + for (size_t i = 0; i < input_bytes; ++i) { + output->data.raw[i] = input->data.raw[i]; + } + } + return kTfLiteOk; +} + +} // namespace reshape + +TfLiteRegistration Register_RESHAPE() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/reshape::Prepare, + /*invoke=*/reshape::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/resize_nearest_neighbor.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/resize_nearest_neighbor.cpp new file mode 100644 index 0000000..31fb3cf --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/resize_nearest_neighbor.cpp @@ -0,0 +1,124 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace resize_nearest_neighbor { + +constexpr int kInputTensor = 0; +constexpr int kSizeTensor = 1; +constexpr int kOutputTensor = 0; + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 2); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* size = GetInput(context, node, kSizeTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + // Our current implementations rely on the input being 4D, + // and the size being 1D tensor with exactly 2 elements. + TF_LITE_ENSURE_EQ(context, NumDimensions(input), 4); + TF_LITE_ENSURE_EQ(context, NumDimensions(size), 1); + TF_LITE_ENSURE_EQ(context, size->type, kTfLiteInt32); + TF_LITE_ENSURE_EQ(context, size->dims->data[0], 2); + + output->type = input->type; + + if (!IsConstantTensor(size)) { + TF_LITE_KERNEL_LOG(context, "Dynamic tensors are unsupported in tfmicro."); + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = + reinterpret_cast(node->builtin_data); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* size = + tflite::micro::GetEvalInput(context, node, kSizeTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + tflite::ResizeNearestNeighborParams op_params; + op_params.align_corners = params->align_corners; + op_params.half_pixel_centers = false; + + if (output->type == kTfLiteFloat32) { + reference_ops::ResizeNearestNeighbor( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(size), + tflite::micro::GetTensorData(size), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else if (output->type == kTfLiteUInt8) { + reference_ops::ResizeNearestNeighbor( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(size), + tflite::micro::GetTensorData(size), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else if (output->type == kTfLiteInt8) { + reference_ops::ResizeNearestNeighbor( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(size), + tflite::micro::GetTensorData(size), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + TF_LITE_KERNEL_LOG(context, + "Output type is %d, requires float, uint8_t or int8_t.", + output->type); + return kTfLiteError; + } + + return kTfLiteOk; +} +} // namespace resize_nearest_neighbor + +TfLiteRegistration Register_RESIZE_NEAREST_NEIGHBOR() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/resize_nearest_neighbor::Prepare, + /*invoke=*/resize_nearest_neighbor::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/round.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/round.cpp new file mode 100644 index 0000000..8de89eb --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/round.cpp @@ -0,0 +1,79 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/round.h" + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace round { + +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteFloat32); + TF_LITE_ENSURE_TYPES_EQ(context, output->type, input->type); + TF_LITE_ENSURE_EQ(context, output->bytes, input->bytes); + TF_LITE_ENSURE_EQ(context, output->dims->size, input->dims->size); + for (int i = 0; i < output->dims->size; ++i) { + TF_LITE_ENSURE_EQ(context, output->dims->data[i], input->dims->data[i]); + } + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + reference_ops::Round(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + + return kTfLiteOk; +} +} // namespace round + +TfLiteRegistration Register_ROUND() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/round::Prepare, + /*invoke=*/round::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/shape.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/shape.cpp new file mode 100644 index 0000000..9ce7b8c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/shape.cpp @@ -0,0 +1,76 @@ +#if !defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { + +namespace { +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +void ExtractShape(const TfLiteEvalTensor* input, int32_t* output_data) { + for (int i = 0; i < input->dims->size; ++i) { + output_data[i] = input->dims->data[i]; + } +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + if (output->type != kTfLiteInt32) { + TF_LITE_KERNEL_LOG(context, "Output type %s (%d) not supported.", + TfLiteTypeGetName(output->type), output->type); + return kTfLiteError; + } else { + ExtractShape(input, tflite::micro::GetTensorData(output)); + } + + return kTfLiteOk; +} + +} // namespace + +TfLiteRegistration Register_SHAPE() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/Prepare, + /*invoke=*/Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/split.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/split.cpp new file mode 100644 index 0000000..07f1cb5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/split.cpp @@ -0,0 +1,138 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace split { + +template +TfLiteStatus SplitImpl(TfLiteContext* context, TfLiteNode* node, + const TfLiteEvalTensor* input, int axis_value) { + const int output_count = NumOutputs(node); + const TfLiteIntArray* input_dims = input->dims; + const TfLiteEvalTensor* output0 = + tflite::micro::GetEvalOutput(context, node, 0); + const TfLiteIntArray* output_dims = output0->dims; + + const int split_dimensions = input_dims->size; + int axis = axis_value < 0 ? axis_value + split_dimensions : axis_value; + + TFLITE_DCHECK_LT(axis, split_dimensions); + TFLITE_DCHECK_EQ(output_dims->size, split_dimensions); + + int64_t split_size = output_dims->data[axis] * output_count; + + TFLITE_DCHECK_EQ(split_size, input_dims->data[axis]); + int64_t outer_size = 1; + for (int i = 0; i < axis; ++i) { + outer_size *= input_dims->data[i]; + } + + int64_t base_inner_size = 1; + for (int i = axis + 1; i < split_dimensions; ++i) { + base_inner_size *= input_dims->data[i]; + } + + const T* input_ptr = tflite::micro::GetTensorData(input); + for (int k = 0; k < outer_size; ++k) { + for (int i = 0; i < output_count; ++i) { + TfLiteEvalTensor* t = tflite::micro::GetEvalOutput(context, node, i); + T* output_data = tflite::micro::GetTensorData(t); + const int copy_size = output_dims->data[axis] * base_inner_size; + T* output_ptr = output_data + k * copy_size; + for (int j = 0; j < copy_size; ++j) output_ptr[j] = input_ptr[j]; + input_ptr += copy_size; + } + } + + return kTfLiteOk; +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* axis = GetInput(context, node, 0); + TF_LITE_ENSURE(context, axis != nullptr); + + // Dynamic output tensors are needed if axis tensor is not constant. + // But Micro doesn't support dynamic memory allocation, so we only support + // constant axis tensor for now. + TF_LITE_ENSURE_MSG(context, IsConstantTensor(axis), + "Non constant axis tensor not supported"); + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* axis = tflite::micro::GetEvalInput(context, node, 0); + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 1); + + int axis_value = tflite::micro::GetTensorData(axis)[0]; + if (axis_value < 0) { + axis_value += input->dims->size; + } + + TF_LITE_ENSURE(context, axis_value >= 0); + TF_LITE_ENSURE(context, axis_value < input->dims->size); + + switch (input->type) { + case kTfLiteFloat32: { + return SplitImpl(context, node, input, axis_value); + } + case kTfLiteUInt8: { + return SplitImpl(context, node, input, axis_value); + } + case kTfLiteInt8: { + return SplitImpl(context, node, input, axis_value); + } + case kTfLiteInt16: { + return SplitImpl(context, node, input, axis_value); + } + case kTfLiteInt32: { + return SplitImpl(context, node, input, axis_value); + } + default: + TF_LITE_KERNEL_LOG(context, "Type %s currently not supported.", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } +#undef TF_LITE_SPLIT + + return kTfLiteOk; +} + +} // namespace split + +TfLiteRegistration Register_SPLIT() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/split::Prepare, + /*invoke=*/split::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/split_v.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/split_v.cpp new file mode 100644 index 0000000..762b175 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/split_v.cpp @@ -0,0 +1,138 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace split_v { + +template +TfLiteStatus SplitImpl(TfLiteContext* context, TfLiteNode* node, + const TfLiteEvalTensor* input, int axis_value) { + const TfLiteIntArray* input_dims = input->dims; + const TfLiteEvalTensor* output0 = + tflite::micro::GetEvalOutput(context, node, 0); + + const int split_dimensions = input_dims->size; + + TFLITE_DCHECK_LT(axis_value, split_dimensions); + TFLITE_DCHECK_EQ(output0->dims->size, split_dimensions); + + int64_t split_size = 0; + const int output_count = NumOutputs(node); + for (int i = 0; i < output_count; i++) { + split_size += + tflite::micro::GetEvalOutput(context, node, i)->dims->data[axis_value]; + } + TFLITE_DCHECK_EQ(split_size, input_dims->data[axis_value]); + int64_t outer_size = 1; + for (int i = 0; i < axis_value; ++i) { + outer_size *= input_dims->data[i]; + } + + int64_t base_inner_size = 1; + for (int i = axis_value + 1; i < split_dimensions; ++i) { + base_inner_size *= input_dims->data[i]; + } + + const T* input_ptr = tflite::micro::GetTensorData(input); + for (int k = 0; k < outer_size; ++k) { + for (int i = 0; i < output_count; ++i) { + TfLiteEvalTensor* output_tensor = + tflite::micro::GetEvalOutput(context, node, i); + T* output_data = tflite::micro::GetTensorData(output_tensor); + const int copy_size = + output_tensor->dims->data[axis_value] * base_inner_size; + T* output_ptr = output_data + k * copy_size; + for (int j = 0; j < copy_size; ++j) output_ptr[j] = input_ptr[j]; + input_ptr += copy_size; + } + } + + return kTfLiteOk; +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 3); + + // Dynamic output tensors are needed if axis tensor is not constant. + // But Micro doesn't support dynamic memory allocation, so we only support + // constant axis tensor for now. + const TfLiteTensor* axis = GetInput(context, node, 2); + TF_LITE_ENSURE_MSG(context, IsConstantTensor(axis), + "Non constant axis tensor not supported"); + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0); + const TfLiteEvalTensor* axis = tflite::micro::GetEvalInput(context, node, 2); + + int axis_value = tflite::micro::GetTensorData(axis)[0]; + if (axis_value < 0) { + axis_value += input->dims->size; + } + + TF_LITE_ENSURE(context, axis_value >= 0); + TF_LITE_ENSURE(context, axis_value < input->dims->size); + + switch (input->type) { + case kTfLiteFloat32: { + return SplitImpl(context, node, input, axis_value); + } + case kTfLiteInt8: { + return SplitImpl(context, node, input, axis_value); + } + case kTfLiteInt16: { + return SplitImpl(context, node, input, axis_value); + } + case kTfLiteInt32: { + return SplitImpl(context, node, input, axis_value); + } + default: + TF_LITE_KERNEL_LOG(context, "Type %s currently not supported.", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace split_v + +TfLiteRegistration Register_SPLIT_V() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/split_v::Prepare, + /*invoke=*/split_v::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/strided_slice.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/strided_slice.cpp new file mode 100644 index 0000000..7e01aa0 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/strided_slice.cpp @@ -0,0 +1,195 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/strided_slice.h" + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace strided_slice { + +constexpr int kInputTensor = 0; +constexpr int kBeginTensor = 1; +constexpr int kEndTensor = 2; +constexpr int kStridesTensor = 3; +constexpr int kOutputTensor = 0; + +struct StridedSliceContext { + StridedSliceContext(TfLiteContext* context, TfLiteNode* node) { + params = reinterpret_cast(node->builtin_data); + input = GetInput(context, node, kInputTensor); + begin = GetInput(context, node, kBeginTensor); + end = GetInput(context, node, kEndTensor); + strides = GetInput(context, node, kStridesTensor); + output = GetOutput(context, node, kOutputTensor); + dims = NumDimensions(input); + } + const TfLiteStridedSliceParams* params; + const TfLiteTensor* input; + const TfLiteTensor* begin; + const TfLiteTensor* end; + const TfLiteTensor* strides; + TfLiteTensor* output; + int dims; +}; + +// This Op only supports 1-4D cases and since we use the reference 4D +// implementation, the 1-3D tensors are mapped to 4D. +const int kMaxDim = 4; + +tflite::StridedSliceParams BuildStridedSliceParams( + StridedSliceContext* op_context) { + tflite::StridedSliceParams op_params; + op_params.start_indices_count = op_context->dims; + op_params.stop_indices_count = op_context->dims; + op_params.strides_count = op_context->dims; + + for (int i = 0; i < op_context->dims; ++i) { + op_params.start_indices[i] = GetTensorData(op_context->begin)[i]; + op_params.stop_indices[i] = GetTensorData(op_context->end)[i]; + op_params.strides[i] = GetTensorData(op_context->strides)[i]; + } + + op_params.begin_mask = op_context->params->begin_mask; + op_params.ellipsis_mask = 0; + op_params.end_mask = op_context->params->end_mask; + op_params.new_axis_mask = 0; + op_params.shrink_axis_mask = op_context->params->shrink_axis_mask; + return op_params; +} + +// Processes the indexing tensors (begin, end and strides) to resize the +// output tensor. This function is callable from both Prepare() and Eval() as +// long as the caller ensures the indexing tensors are present. +TfLiteStatus CheckOutputSize(TfLiteContext* context, + StridedSliceContext* op_context) { + using ::tflite::strided_slice::StartForAxis; + using ::tflite::strided_slice::StopForAxis; + TfLiteIntArray* output_shape = op_context->output->dims; + int shape_size = 0; + auto op_params = BuildStridedSliceParams(op_context); + auto input_shape = GetTensorShape(op_context->input); + for (int idx = 0; idx < op_context->dims; ++idx) { + int32_t stride = GetTensorData(op_context->strides)[idx]; + TF_LITE_ENSURE_MSG(context, stride != 0, "stride value has to be non-zero"); + int32_t begin = StartForAxis(op_params, input_shape, idx); + int32_t end = StopForAxis(op_params, input_shape, idx, begin); + + // When shrinking an axis, the end position does not matter (and can be + // incorrect when negative indexing is used, see Issue #19260). Always use + // begin + 1 to generate a length 1 slice, since begin has + // already been adjusted for negative indices by StartForAxis. + const bool shrink_axis = op_context->params->shrink_axis_mask & (1 << idx); + if (shrink_axis) { + end = begin + 1; + } + + // This is valid for both positive and negative strides + int32_t dim_shape = std::ceil((end - begin) / static_cast(stride)); + dim_shape = dim_shape < 0 ? 0 : dim_shape; + if (!shrink_axis) { + TF_LITE_ENSURE_EQ(context, output_shape->data[shape_size], dim_shape); + shape_size++; + } + } + TF_LITE_ENSURE_EQ(context, output_shape->size, shape_size); + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(StridedSliceParams)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + StridedSliceParams* op_params = + static_cast(node->user_data); + TF_LITE_ENSURE_EQ(context, NumInputs(node), 4); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + StridedSliceContext op_context(context, node); + TF_LITE_ENSURE_MSG(context, op_context.dims <= kMaxDim, + "input dim should not exceed 4"); + auto params = BuildStridedSliceParams(&op_context); + memcpy(op_params, ¶ms, sizeof(StridedSliceParams)); + return CheckOutputSize(context, &op_context); +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + const StridedSliceParams& op_params = + *(static_cast(node->user_data)); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + switch (output->type) { + case kTfLiteFloat32: + reference_ops::StridedSlice(op_params, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + case kTfLiteUInt8: + reference_ops::StridedSlice( + op_params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + case kTfLiteInt8: + reference_ops::StridedSlice(op_params, + tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + break; + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + return kTfLiteOk; +} +} // namespace strided_slice + +TfLiteRegistration Register_STRIDED_SLICE() { + return {/*init=*/strided_slice::Init, + /*free=*/nullptr, + /*prepare=*/strided_slice::Prepare, + /*invoke=*/strided_slice::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/sub.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/sub.cpp new file mode 100644 index 0000000..dd1b653 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/sub.cpp @@ -0,0 +1,259 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/sub.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace sub { + +constexpr int kInputTensor1 = 0; +constexpr int kInputTensor2 = 1; +constexpr int kOutputTensor = 0; + +struct OpData { + bool requires_broadcast; + + // These fields are used in both the general 8-bit -> 8bit quantized path, + // and the special 16-bit -> 16bit quantized path + int input1_shift; + int input2_shift; + int32_t output_activation_min; + int32_t output_activation_max; + + // These fields are used only in the general 8-bit -> 8bit quantized path + int32_t input1_multiplier; + int32_t input2_multiplier; + int32_t output_multiplier; + int output_shift; + int left_shift; + int32_t input1_offset; + int32_t input2_offset; + int32_t output_offset; +}; + +TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteSubParams* params, + const TfLiteTensor* input1, + const TfLiteTensor* input2, TfLiteTensor* output, + OpData* data) { + data->requires_broadcast = !HaveSameShapes(input1, input2); + + if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + // 8bit -> 8bit general quantized path, with general rescalings + data->input1_offset = -input1->params.zero_point; + data->input2_offset = -input2->params.zero_point; + data->output_offset = output->params.zero_point; + data->left_shift = 20; + const float twice_max_input_scale = + 2 * std::max(input1->params.scale, input2->params.scale); + const double real_input1_multiplier = + static_cast(input1->params.scale / twice_max_input_scale); + const double real_input2_multiplier = + static_cast(input2->params.scale / twice_max_input_scale); + const double real_output_multiplier = + static_cast(twice_max_input_scale / + ((1 << data->left_shift) * output->params.scale)); + + QuantizeMultiplierSmallerThanOneExp( + real_input1_multiplier, &data->input1_multiplier, &data->input1_shift); + + QuantizeMultiplierSmallerThanOneExp( + real_input2_multiplier, &data->input2_multiplier, &data->input2_shift); + + QuantizeMultiplierSmallerThanOneExp( + real_output_multiplier, &data->output_multiplier, &data->output_shift); + + TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( + context, params->activation, output, &data->output_activation_min, + &data->output_activation_max)); + } + + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + OpData* data = static_cast(node->user_data); + auto* params = reinterpret_cast(node->builtin_data); + + const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1); + TF_LITE_ENSURE(context, input1 != nullptr); + const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2); + TF_LITE_ENSURE(context, input2 != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + + TF_LITE_ENSURE_STATUS( + CalculateOpData(context, params, input1, input2, output, data)); + return kTfLiteOk; +} + +void EvalSub(TfLiteContext* context, TfLiteNode* node, TfLiteSubParams* params, + const OpData* data, const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) { + float output_activation_min, output_activation_max; + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); + tflite::ArithmeticParams op_params; + SetActivationParams(output_activation_min, output_activation_max, &op_params); + if (data->requires_broadcast) { + tflite::reference_ops::BroadcastSubSlow( + op_params, tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + tflite::reference_ops::SubWithActivation( + op_params, tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } +} + +TfLiteStatus EvalSubQuantized(TfLiteContext* context, TfLiteNode* node, + TfLiteSubParams* params, const OpData* data, + const TfLiteEvalTensor* input1, + const TfLiteEvalTensor* input2, + TfLiteEvalTensor* output) { + if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + tflite::ArithmeticParams op_params; + op_params.left_shift = data->left_shift; + op_params.input1_offset = data->input1_offset; + op_params.input1_multiplier = data->input1_multiplier; + op_params.input1_shift = data->input1_shift; + op_params.input2_offset = data->input2_offset; + op_params.input2_multiplier = data->input2_multiplier; + op_params.input2_shift = data->input2_shift; + op_params.output_offset = data->output_offset; + op_params.output_multiplier = data->output_multiplier; + op_params.output_shift = data->output_shift; + SetActivationParams(data->output_activation_min, + data->output_activation_max, &op_params); + bool need_broadcast = reference_ops::ProcessBroadcastShapes( + tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorShape(input2), &op_params); + + if (output->type == kTfLiteInt8) { + if (need_broadcast) { + tflite::reference_ops::BroadcastSubSlow( + op_params, tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + tflite::reference_ops::Sub( + op_params, tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } + } else { + if (need_broadcast) { + tflite::reference_ops::BroadcastSubSlow( + op_params, tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } else { + tflite::reference_ops::Sub( + op_params, tflite::micro::GetTensorShape(input1), + tflite::micro::GetTensorData(input1), + tflite::micro::GetTensorShape(input2), + tflite::micro::GetTensorData(input2), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + } + } + } + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + + const TfLiteEvalTensor* input1 = + tflite::micro::GetEvalInput(context, node, kInputTensor1); + const TfLiteEvalTensor* input2 = + tflite::micro::GetEvalInput(context, node, kInputTensor2); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + if (output->type == kTfLiteFloat32) { + EvalSub(context, node, params, &data, input1, input2, output); + } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { + TF_LITE_ENSURE_OK(context, EvalSubQuantized(context, node, params, &data, + input1, input2, output)); + } else { + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(output->type), output->type); + return kTfLiteError; + } + + return kTfLiteOk; +} + +} // namespace sub + +TfLiteRegistration Register_SUB() { + return {/*init=*/sub::Init, + /*free=*/nullptr, + /*prepare=*/sub::Prepare, + /*invoke=*/sub::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf.h b/src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf.h new file mode 100644 index 0000000..713c9ae --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf.h @@ -0,0 +1,54 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_KERNELS_SVDF_H_ +#define TENSORFLOW_LITE_MICRO_KERNELS_SVDF_H_ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +struct OpData { + int32_t effective_scale_1_a; + int32_t effective_scale_2_a; + // b versions of each scale are kept at int since the numbers are just the + // shift value - typically between [-32, 32]. + int effective_scale_1_b; + int effective_scale_2_b; + int scratch_tensor_index; + int scratch_output_tensor_index; + + // Cached tensor zero point values for quantized operations. + int input_zero_point; + int output_zero_point; +}; + +// TensorflowLite Micro-specific reference implementation for Integer SVDF. +void EvalIntegerSvdfReference(TfLiteContext* context, TfLiteNode* node, + const TfLiteEvalTensor* input_tensor, + const TfLiteEvalTensor* weights_feature_tensor, + const TfLiteEvalTensor* weights_time_tensor, + const TfLiteEvalTensor* bias_tensor, + const TfLiteSVDFParams* params, + TfLiteEvalTensor* activation_state_tensor, + TfLiteEvalTensor* output_tensor, + const OpData& data); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_KERNELS_SVDF_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf_common.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf_common.cpp new file mode 100644 index 0000000..650e730 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/svdf_common.cpp @@ -0,0 +1,180 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/activation_utils.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/svdf.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { + +void EvalIntegerSvdfReference(TfLiteContext* context, TfLiteNode* node, + const TfLiteEvalTensor* input_tensor, + const TfLiteEvalTensor* weights_feature_tensor, + const TfLiteEvalTensor* weights_time_tensor, + const TfLiteEvalTensor* bias_tensor, + const TfLiteSVDFParams* params, + TfLiteEvalTensor* activation_state_tensor, + TfLiteEvalTensor* output_tensor, + const OpData& data) { + const int n_rank = params->rank; + const int n_batch = input_tensor->dims->data[0]; + const int n_input = input_tensor->dims->data[1]; + const int n_filter = weights_feature_tensor->dims->data[0]; + const int n_unit = n_filter / n_rank; + const int n_memory = weights_time_tensor->dims->data[1]; + + TFLITE_DCHECK(context != nullptr); + TFLITE_DCHECK(context->GetScratchBuffer != nullptr); + + int32_t* scratch_tensor = static_cast( + context->GetScratchBuffer(context, data.scratch_tensor_index)); + int32_t* scratch_output_tensor = static_cast( + context->GetScratchBuffer(context, data.scratch_output_tensor_index)); + + // Shift states. + int16_t* const state_ptr = + tflite::micro::GetTensorData(activation_state_tensor); + + // Left shift the activation_state. + { + int16_t* new_state_start = state_ptr; + const int16_t* old_state_start = state_ptr + 1; + const int16_t* old_state_end = state_ptr + n_batch * n_filter * n_memory; + while (old_state_start != old_state_end) { + *new_state_start++ = *old_state_start++; + } + } + + // Note: no need to clear the latest activation, matmul is not accumulative. + + // Feature matmul. + { + int16_t* state = + tflite::micro::GetTensorData(activation_state_tensor); + const int8_t* input = tflite::micro::GetTensorData(input_tensor); + const int8_t* weight_feature = + tflite::micro::GetTensorData(weights_feature_tensor); + const int32_t output_max = std::numeric_limits::max(); + const int32_t output_min = std::numeric_limits::min(); + int16_t* result_in_batch = state + (n_memory - 1); + for (int b = 0; b < n_batch; b++) { + const int8_t* matrix_ptr = weight_feature; + for (int r = 0; r < n_filter; r++) { + int32_t dot_prod = 0; + const int8_t* vector_in_batch = input + b * n_input; + for (int c = 0; c < n_input; c++) { + dot_prod += + *matrix_ptr++ * (*vector_in_batch++ - data.input_zero_point); + } + dot_prod = MultiplyByQuantizedMultiplier( + dot_prod, data.effective_scale_1_a, data.effective_scale_1_b); + dot_prod = std::min(std::max(output_min, dot_prod), output_max); + // This assumes state is symmetrically quantized. Otherwise last bit of + // state should be initialized to its zero point and accumulate the + // dot_prod. + // Equivalent as the following: + // result_in_batch = zero point, which happens to be zero. + // result_in_batch += dot_prod_56. + *result_in_batch = dot_prod; + result_in_batch += n_memory; + } + } + } + + // Time. + { + for (int b = 0; b < n_batch; ++b) { + int32_t* scratch_ptr_batch = scratch_tensor + b * n_filter; + + // Perform batched vector dot product: + const int16_t* vector1_ptr = + tflite::micro::GetTensorData(weights_time_tensor); + const int16_t* vector2_ptr = + tflite::micro::GetTensorData(activation_state_tensor) + + b * n_memory * n_filter; + + for (int i = 0; i < n_filter; i++) { + *scratch_ptr_batch = 0; + for (int j = 0; j < n_memory; j++) { + *scratch_ptr_batch += *vector1_ptr++ * *vector2_ptr++; + } + scratch_ptr_batch++; + } + } + } + + // Reduce, add bias, rescale, activation. + { + // Add bias. + if (bias_tensor) { + // Vector batch assign: + const int32_t* bias_data = + tflite::micro::GetTensorData(bias_tensor); + for (int i = 0; i < n_batch; ++i) { + int32_t* output_ptr = scratch_output_tensor + i * n_unit; + const int32_t* bias_ptr = bias_data; + for (int j = 0; j < n_unit; ++j) { + *output_ptr++ = *bias_ptr++; + } + } + } else { + int32_t* output_ptr = scratch_output_tensor; + for (int i = 0; i < n_batch * n_unit; ++i) { + *output_ptr++ = 0; + } + } + + // Reduce. + for (int b = 0; b < n_batch; ++b) { + int32_t* output_temp_ptr = scratch_output_tensor + b * n_unit; + int32_t* scratch_ptr_batch = scratch_tensor + b * n_filter; + + // Reduction sum vector + for (int i = 0; i < n_unit; ++i) { + for (int j = 0; j < n_rank; ++j) { + output_temp_ptr[i] += *scratch_ptr_batch++; + } + } + } + + // Rescale. + const int32_t output_max = std::numeric_limits::max(); + const int32_t output_min = std::numeric_limits::min(); + for (int i = 0; i < n_batch * n_unit; ++i) { + int32_t x1 = scratch_output_tensor[i]; + int32_t x2 = MultiplyByQuantizedMultiplier(x1, data.effective_scale_2_a, + data.effective_scale_2_b); + int32_t x3 = x2 + data.output_zero_point; + int32_t x4 = std::min(std::max(output_min, x3), output_max); + tflite::micro::GetTensorData(output_tensor)[i] = + static_cast(x4); + } + } +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/tanh.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/tanh.cpp new file mode 100644 index 0000000..c08adc6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/tanh.cpp @@ -0,0 +1,161 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/tanh.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace activations { +namespace { +constexpr int kInputTensor = 0; +constexpr int kOutputTensor = 0; + +struct OpData { + int32_t input_zero_point; + int32_t input_range_radius; + int32_t input_multiplier; + int input_left_shift; +}; + +void* TanhInit(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus CalculateArithmeticOpData(TfLiteContext* context, TfLiteNode* node, + OpData* data) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + + TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); + + if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8) { + static constexpr int kInputIntegerBits = 4; + const double input_real_multiplier = + static_cast(input->params.scale) * + static_cast(1 << (31 - kInputIntegerBits)); + + const double q = std::frexp(input_real_multiplier, &data->input_left_shift); + data->input_multiplier = static_cast(TfLiteRound(q * (1ll << 31))); + + data->input_range_radius = + CalculateInputRadius(kInputIntegerBits, data->input_left_shift, 31); + } + return kTfLiteOk; +} + +TfLiteStatus TanhPrepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + + OpData* data = static_cast(node->user_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + data->input_zero_point = input->params.zero_point; + return CalculateArithmeticOpData(context, node, data); +} + +} // namespace + +TfLiteStatus TanhEval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + switch (input->type) { + case kTfLiteFloat32: { + reference_ops::Tanh(tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } break; + case kTfLiteInt16: { + TanhParams params; + params.input_left_shift = data.input_left_shift; + reference_ops::Tanh(params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } break; + case kTfLiteUInt8: { + TanhParams params; + params.input_zero_point = data.input_zero_point; + params.input_range_radius = data.input_range_radius; + params.input_multiplier = data.input_multiplier; + params.input_left_shift = data.input_left_shift; + reference_ops::Tanh(params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + + return kTfLiteOk; + } break; + case kTfLiteInt8: { + reference_integer_ops::Tanh( + data.input_zero_point, data.input_range_radius, data.input_multiplier, + data.input_left_shift, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output)); + return kTfLiteOk; + } break; + default: + TF_LITE_KERNEL_LOG(context, "Input %s, output %s not supported.", + TfLiteTypeGetName(input->type), + TfLiteTypeGetName(output->type)); + return kTfLiteError; + } +} + +} // namespace activations + +TfLiteRegistration Register_TANH() { + return {/*init=*/activations::TanhInit, + /*free=*/nullptr, + /*prepare=*/activations::TanhPrepare, + /*invoke=*/activations::TanhEval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/transpose_conv.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/transpose_conv.cpp new file mode 100644 index 0000000..550d293 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/transpose_conv.cpp @@ -0,0 +1,268 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/transpose_conv.h" + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/kernels/padding.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace { + +constexpr int kInputTensor = 0; +constexpr int kFilterTensor = 1; +constexpr int kBiasTensor = 2; +constexpr int kOutputTensor = 0; + +// Conv is quantized along dimension 0: +// https://www.tensorflow.org/lite/performance/quantization_spec +constexpr int kConvQuantizedDimension = 0; + +struct OpData { + ConvParams params; + + // A scratch buffer is required for quantized implementations. + int scratch_buffer_index; + + // Multiplier and shift arrays are required for the int8 implementation. + int32_t* per_channel_output_multiplier; + int32_t* per_channel_output_shift; +}; + +inline PaddingType RuntimePaddingType(TfLitePadding padding) { + switch (padding) { + case TfLitePadding::kTfLitePaddingSame: + return PaddingType::kSame; + case TfLitePadding::kTfLitePaddingValid: + return PaddingType::kValid; + case TfLitePadding::kTfLitePaddingUnknown: + default: + return PaddingType::kNone; + } +} + +TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node, + const TfLiteConvParams* params, int width, + int height, int filter_width, int filter_height, + int out_width, int out_height, + const TfLiteType data_type, OpData* data) { + bool has_bias = node->inputs->size == 3; + // Check number of inputs/outputs + TF_LITE_ENSURE(context, has_bias || node->inputs->size == 2); + TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); + + // Matching GetWindowedOutputSize in TensorFlow. + auto padding = params->padding; + TfLitePaddingValues padding_values = ComputePaddingHeightWidth( + params->stride_height, params->stride_width, + params->dilation_height_factor, params->dilation_width_factor, height, + width, filter_height, filter_width, padding, &out_height, &out_width); + + data->params.padding_type = RuntimePaddingType(padding); + data->params.padding_values.width = padding_values.width; + data->params.padding_values.height = padding_values.height; + + // Note that quantized inference requires that all tensors have their + // parameters set. This is usually done during quantized training. + if (data_type != kTfLiteFloat32) { + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); + TF_LITE_ENSURE(context, filter != nullptr); + const TfLiteTensor* bias = + GetOptionalInputTensor(context, node, kBiasTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + int output_channels = filter->dims->data[kConvQuantizedDimension]; + + TF_LITE_ENSURE_STATUS(tflite::PopulateConvolutionQuantizationParams( + context, input, filter, bias, output, params->activation, + &data->params.output_multiplier, &data->params.output_shift, + &data->params.quantized_activation_min, + &data->params.quantized_activation_max, + data->per_channel_output_multiplier, + reinterpret_cast(data->per_channel_output_shift), + output_channels)); + } + return kTfLiteOk; +} + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr); + return context->AllocatePersistentBuffer(context, sizeof(OpData)); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TFLITE_DCHECK(node->user_data != nullptr); + TFLITE_DCHECK(node->builtin_data != nullptr); + + OpData* data = static_cast(node->user_data); + const auto params = static_cast(node->builtin_data); + + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE(context, output != nullptr); + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE(context, input != nullptr); + const TfLiteTensor* filter = GetInput(context, node, kFilterTensor); + TF_LITE_ENSURE(context, filter != nullptr); + + int input_width = input->dims->data[2]; + int input_height = input->dims->data[1]; + int filter_width = filter->dims->data[2]; + int filter_height = filter->dims->data[1]; + int output_width = output->dims->data[2]; + int output_height = output->dims->data[1]; + + // Dynamically allocate per-channel quantization parameters. + const int num_channels = filter->dims->data[kConvQuantizedDimension]; + data->per_channel_output_multiplier = + static_cast(context->AllocatePersistentBuffer( + context, num_channels * sizeof(int32_t))); + data->per_channel_output_shift = + static_cast(context->AllocatePersistentBuffer( + context, num_channels * sizeof(int32_t))); + + // Quantized kernels use an int32 scratch buffer. + if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8) { + TFLITE_DCHECK(context->RequestScratchBufferInArena != nullptr); + TFLITE_DCHECK(context->RequestScratchBufferInArena( + context, + GetTensorShape(output).FlatSize() * sizeof(int32_t), + &(data->scratch_buffer_index)) == kTfLiteOk); + } + + // All per-channel quantized tensors need valid zero point and scale arrays. + if (input->type == kTfLiteInt8) { + TF_LITE_ENSURE_EQ(context, filter->quantization.type, + kTfLiteAffineQuantization); + + const auto* affine_quantization = + static_cast(filter->quantization.params); + TF_LITE_ENSURE(context, affine_quantization); + TF_LITE_ENSURE(context, affine_quantization->scale); + TF_LITE_ENSURE(context, affine_quantization->zero_point); + + TF_LITE_ENSURE(context, + affine_quantization->scale->size == 1 || + affine_quantization->scale->size == + filter->dims->data[kConvQuantizedDimension]); + TF_LITE_ENSURE_EQ(context, affine_quantization->scale->size, + affine_quantization->zero_point->size); + } + + TF_LITE_ENSURE_STATUS(CalculateOpData( + context, node, params, input_width, input_height, filter_width, + filter_height, output_width, output_height, input->type, data)); + + // Offsets (zero points) + data->params.input_offset = -input->params.zero_point; + data->params.weights_offset = -filter->params.zero_point; + data->params.output_offset = output->params.zero_point; + + // Stride + dilation + data->params.stride_width = params->stride_width; + data->params.stride_height = params->stride_height; + data->params.dilation_width_factor = params->dilation_width_factor; + data->params.dilation_height_factor = params->dilation_height_factor; + + float output_activation_min, output_activation_max; + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); + data->params.float_activation_min = output_activation_min; + data->params.float_activation_max = output_activation_max; + return kTfLiteOk; +} // namespace conv + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + const TfLiteEvalTensor* filter = + tflite::micro::GetEvalInput(context, node, kFilterTensor); + const TfLiteEvalTensor* bias = + (NumInputs(node) == 3) + ? tflite::micro::GetEvalInput(context, node, kBiasTensor) + : nullptr; + TfLiteEvalTensor* output = + tflite::micro::GetEvalOutput(context, node, kOutputTensor); + + TFLITE_DCHECK(node->user_data != nullptr); + const OpData& data = *(static_cast(node->user_data)); + + TF_LITE_ENSURE_EQ(context, input->type, output->type); + TF_LITE_ENSURE_MSG(context, input->type == filter->type, + "Hybrid models are not supported on TFLite Micro."); + + switch (input->type) { // Already know in/out types are same. + case kTfLiteFloat32: { + reference_ops::TransposeConv( + data.params, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), + tflite::micro::GetTensorShape(nullptr), nullptr); + break; + } + case kTfLiteInt8: { + int32_t* scratch_buffer = static_cast( + context->GetScratchBuffer(context, data.scratch_buffer_index)); + reference_integer_ops::TransposeConv( + data.params, data.per_channel_output_multiplier, + data.per_channel_output_shift, tflite::micro::GetTensorShape(input), + tflite::micro::GetTensorData(input), + tflite::micro::GetTensorShape(filter), + tflite::micro::GetTensorData(filter), + tflite::micro::GetTensorShape(bias), + tflite::micro::GetTensorData(bias), + tflite::micro::GetTensorShape(output), + tflite::micro::GetTensorData(output), + tflite::micro::GetTensorShape(nullptr), nullptr, scratch_buffer); + break; + } + default: + TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.", + TfLiteTypeGetName(input->type), input->type); + return kTfLiteError; + } + return kTfLiteOk; +} + +} // namespace + +TfLiteRegistration Register_TRANSPOSE_CONV_2D() { + return {/*init=*/Init, + /*free=*/nullptr, + /*prepare=*/Prepare, + /*invoke=*/Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/kernels/unpack.cpp b/src/tensorflow_arm/tensorflow/lite/micro/kernels/unpack.cpp new file mode 100644 index 0000000..516ae9b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/kernels/unpack.cpp @@ -0,0 +1,124 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/kernel_util.h" + +namespace tflite { +namespace ops { +namespace micro { +namespace unpack { +namespace { + +constexpr int kInputTensor = 0; + +template +TfLiteStatus UnpackImpl(TfLiteContext* context, TfLiteNode* node, + const TfLiteEvalTensor* input, int output_count, + int axis) { + const TfLiteEvalTensor* output0 = + tflite::micro::GetEvalOutput(context, node, 0); + const TfLiteIntArray* input_dims = input->dims; + const TfLiteIntArray* output_dims = output0->dims; + const int dimensions = input_dims->size; + + if (axis < 0) { + axis += input->dims->size; + } + + TFLITE_DCHECK_LT(axis, dimensions); + + int outer_size = 1; + for (int i = 0; i < axis; ++i) { + outer_size *= input_dims->data[i]; + } + int copy_size = 1; + for (int i = axis + 1; i < dimensions; ++i) { + copy_size *= input_dims->data[i]; + } + int output_size = 1; + for (int i = 0; i < output_dims->size; ++i) { + output_size *= output_dims->data[i]; + } + TFLITE_DCHECK_EQ(output_size, copy_size * outer_size); + + const T* input_data = tflite::micro::GetTensorData(input); + + for (int i = 0; i < output_count; ++i) { + TfLiteEvalTensor* t = tflite::micro::GetEvalOutput(context, node, i); + T* output_data = tflite::micro::GetTensorData(t); + for (int k = 0; k < outer_size; ++k) { + T* output_ptr = output_data + copy_size * k; + int loc = k * output_count * copy_size + i * copy_size; + const T* input_ptr = input_data + loc; + for (int j = 0; j < copy_size; ++j) output_ptr[j] = input_ptr[j]; + } + } + + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + TfLiteUnpackParams* data = + reinterpret_cast(node->builtin_data); + + const TfLiteEvalTensor* input = + tflite::micro::GetEvalInput(context, node, kInputTensor); + + switch (input->type) { + case kTfLiteFloat32: { + return UnpackImpl(context, node, input, data->num, data->axis); + } + case kTfLiteInt32: { + return UnpackImpl(context, node, input, data->num, data->axis); + } + case kTfLiteUInt8: { + return UnpackImpl(context, node, input, data->num, data->axis); + } + case kTfLiteInt8: { + return UnpackImpl(context, node, input, data->num, data->axis); + } + default: { + TF_LITE_KERNEL_LOG(context, "Type '%s' is not supported by unpack.", + TfLiteTypeGetName(input->type)); + return kTfLiteError; + } + } + + return kTfLiteOk; +} +} // namespace +} // namespace unpack + +TfLiteRegistration Register_UNPACK() { + return {/*init=*/nullptr, + /*free=*/nullptr, + /*prepare=*/nullptr, + /*invoke=*/unpack::Eval, + /*profiling_string=*/nullptr, + /*builtin_code=*/0, + /*custom_name=*/nullptr, + /*version=*/0}; +} + +} // namespace micro +} // namespace ops +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.cpp b/src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.cpp new file mode 100644 index 0000000..06e7728 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.cpp @@ -0,0 +1,161 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" + +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +uint8_t* AlignPointerUp(uint8_t* data, size_t alignment) { + std::uintptr_t data_as_uintptr_t = reinterpret_cast(data); + uint8_t* aligned_result = reinterpret_cast( + ((data_as_uintptr_t + (alignment - 1)) / alignment) * alignment); + return aligned_result; +} + +uint8_t* AlignPointerDown(uint8_t* data, size_t alignment) { + std::uintptr_t data_as_uintptr_t = reinterpret_cast(data); + uint8_t* aligned_result = + reinterpret_cast((data_as_uintptr_t / alignment) * alignment); + return aligned_result; +} + +size_t AlignSizeUp(size_t size, size_t alignment) { + size_t aligned_size = (((size + (alignment - 1)) / alignment) * alignment); + return aligned_size; +} + +TfLiteStatus TfLiteTypeSizeOf(TfLiteType type, size_t* size) { + switch (type) { + case kTfLiteFloat32: + *size = sizeof(float); + break; + case kTfLiteInt16: + *size = sizeof(int16_t); + break; + case kTfLiteInt32: + *size = sizeof(int32_t); + break; + case kTfLiteUInt8: + *size = sizeof(uint8_t); + break; + case kTfLiteInt8: + *size = sizeof(int8_t); + break; + case kTfLiteInt64: + *size = sizeof(int64_t); + break; + case kTfLiteUInt64: + *size = sizeof(uint64_t); + break; + case kTfLiteBool: + *size = sizeof(bool); + break; + case kTfLiteComplex64: + *size = sizeof(float) * 2; + break; + case kTfLiteComplex128: + *size = sizeof(double) * 2; + break; + default: + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus BytesRequiredForTensor(const tflite::Tensor& flatbuffer_tensor, + size_t* bytes, size_t* type_size, + ErrorReporter* error_reporter) { + int element_count = 1; + // If flatbuffer_tensor.shape == nullptr, then flatbuffer_tensor is a scalar + // so has 1 element. + if (flatbuffer_tensor.shape() != nullptr) { + for (size_t n = 0; n < flatbuffer_tensor.shape()->Length(); ++n) { + element_count *= flatbuffer_tensor.shape()->Get(n); + } + } + + TfLiteType tf_lite_type; + TF_LITE_ENSURE_STATUS(ConvertTensorType(flatbuffer_tensor.type(), + &tf_lite_type, error_reporter)); + TF_LITE_ENSURE_STATUS(TfLiteTypeSizeOf(tf_lite_type, type_size)); + *bytes = element_count * (*type_size); + return kTfLiteOk; +} + +TfLiteStatus TfLiteEvalTensorByteLength(const TfLiteEvalTensor* eval_tensor, + size_t* out_bytes) { + TFLITE_DCHECK(out_bytes != nullptr); + + int element_count = 1; + // If eval_tensor->dims == nullptr, then tensor is a scalar so has 1 element. + if (eval_tensor->dims != nullptr) { + for (int n = 0; n < eval_tensor->dims->size; ++n) { + element_count *= eval_tensor->dims->data[n]; + } + } + size_t type_size; + TF_LITE_ENSURE_STATUS(TfLiteTypeSizeOf(eval_tensor->type, &type_size)); + *out_bytes = element_count * type_size; + return kTfLiteOk; +} + +TfLiteStatus AllocateOutputDimensionsFromInput(TfLiteContext* context, + const TfLiteTensor* input1, + const TfLiteTensor* input2, + TfLiteTensor* output) { + const TfLiteTensor* input = nullptr; + + TF_LITE_ENSURE(context, input1->dims != nullptr); + TF_LITE_ENSURE(context, input2->dims != nullptr); + TF_LITE_ENSURE(context, output->dims->size == 0); + + input = input1->dims->size > input2->dims->size ? input1 : input2; + TF_LITE_ENSURE(context, output->type == input->type); + + size_t size = 0; + TfLiteTypeSizeOf(input->type, &size); + const int dimensions_count = tflite::GetTensorShape(input).DimensionsCount(); + for (int i = 0; i < dimensions_count; i++) { + size *= input->dims->data[i]; + } + + output->bytes = size; + + output->dims = + reinterpret_cast(context->AllocatePersistentBuffer( + context, TfLiteIntArrayGetSizeInBytes(size))); + + output->dims->size = input->dims->size; + for (int i = 0; i < dimensions_count; i++) { + output->dims->data[i] = input->dims->data[i]; + } + + return kTfLiteOk; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.h b/src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.h new file mode 100644 index 0000000..1f3257b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/memory_helpers.h @@ -0,0 +1,62 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_MEMORY_HELPERS_H_ +#define TENSORFLOW_LITE_MICRO_MEMORY_HELPERS_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +// Returns the next pointer address aligned to the given alignment. +uint8_t* AlignPointerUp(uint8_t* data, size_t alignment); + +// Returns the previous pointer address aligned to the given alignment. +uint8_t* AlignPointerDown(uint8_t* data, size_t alignment); + +// Returns an increased size that's a multiple of alignment. +size_t AlignSizeUp(size_t size, size_t alignment); + +// Returns size in bytes for a given TfLiteType. +TfLiteStatus TfLiteTypeSizeOf(TfLiteType type, size_t* size); + +// How many bytes are needed to hold a tensor's contents. +TfLiteStatus BytesRequiredForTensor(const tflite::Tensor& flatbuffer_tensor, + size_t* bytes, size_t* type_size, + ErrorReporter* error_reporter); + +// How many bytes are used in a TfLiteEvalTensor instance. The byte length is +// returned in out_bytes. +TfLiteStatus TfLiteEvalTensorByteLength(const TfLiteEvalTensor* eval_tensor, + size_t* out_bytes); + +// Deduce output dimensions from input and allocate given size. +// Useful for operators with two inputs where the largest input should equal the +// output dimension. +TfLiteStatus AllocateOutputDimensionsFromInput(TfLiteContext* context, + const TfLiteTensor* input1, + const TfLiteTensor* input2, + TfLiteTensor* output); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MEMORY_HELPERS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cpp b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cpp new file mode 100644 index 0000000..c5a2528 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cpp @@ -0,0 +1,440 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h" + +namespace tflite { + +// Simple stable in-place sort function. Not time-efficient for large arrays. +// Would normally be in an anonymous namespace to keep it private, but we want +// to be able to test it externally. +void ReverseSortInPlace(int* values, int* ids, int size) { + bool any_swapped; + do { + any_swapped = false; + for (int i = 1; i < size; ++i) { + if (values[i - 1] < values[i]) { + const int value_temp = values[i - 1]; + values[i - 1] = values[i]; + values[i] = value_temp; + const int id_temp = ids[i - 1]; + ids[i - 1] = ids[i]; + ids[i] = id_temp; + any_swapped = true; + } + } + } while (any_swapped); +} + +GreedyMemoryPlanner::GreedyMemoryPlanner(unsigned char* scratch_buffer, + int scratch_buffer_size) + : buffer_count_(0), need_to_calculate_offsets_(true) { + // Allocate the arrays we need within the scratch buffer arena. + max_buffer_count_ = scratch_buffer_size / per_buffer_size(); + + unsigned char* next_free = scratch_buffer; + requirements_ = reinterpret_cast(next_free); + next_free += sizeof(BufferRequirements) * max_buffer_count_; + + buffer_sizes_sorted_ = reinterpret_cast(next_free); + next_free += sizeof(int) * max_buffer_count_; + + buffer_ids_sorted_ = reinterpret_cast(next_free); + next_free += sizeof(int) * max_buffer_count_; + + buffers_sorted_by_offset_ = reinterpret_cast(next_free); + next_free += sizeof(ListEntry) * max_buffer_count_; + + buffer_offsets_ = reinterpret_cast(next_free); +} + +GreedyMemoryPlanner::~GreedyMemoryPlanner() { + // We don't own the scratch buffer, so don't deallocate anything. +} + +TfLiteStatus GreedyMemoryPlanner::AddBuffer( + tflite::ErrorReporter* error_reporter, int size, int first_time_used, + int last_time_used) { + if (buffer_count_ >= max_buffer_count_) { + TF_LITE_REPORT_ERROR(error_reporter, "Too many buffers (max is %d)", + max_buffer_count_); + return kTfLiteError; + } + BufferRequirements* current = &requirements_[buffer_count_]; + current->size = size; + current->first_time_used = first_time_used; + current->last_time_used = last_time_used; + current->offline_offset = kOnlinePlannedBuffer; + ++buffer_count_; + need_to_calculate_offsets_ = true; + return kTfLiteOk; +} + +TfLiteStatus GreedyMemoryPlanner::AddBuffer( + tflite::ErrorReporter* error_reporter, int size, int first_time_used, + int last_time_used, int offline_offset) { + BufferRequirements* current = &requirements_[buffer_count_]; + if (AddBuffer(error_reporter, size, first_time_used, last_time_used) != + kTfLiteOk) { + return kTfLiteError; + } + current->offline_offset = offline_offset; + return kTfLiteOk; +} + +bool GreedyMemoryPlanner::DoesEntryOverlapInTime( + const GreedyMemoryPlanner::ListEntry* entry, const int first_time_used, + const int last_time_used) const { + const BufferRequirements* entry_requirements = + &requirements_[entry->requirements_index]; + if (entry_requirements->first_time_used > last_time_used) { + return false; + } + if (first_time_used > entry_requirements->last_time_used) { + return false; + } + return true; +} + +GreedyMemoryPlanner::ListEntry* +GreedyMemoryPlanner::NextSimultaneouslyActiveBuffer( + const GreedyMemoryPlanner::ListEntry* start, const int first_time_used, + const int last_time_used) { + ListEntry* result = nullptr; + ListEntry* candidate_next_entry; + if (start == nullptr) { + candidate_next_entry = &buffers_sorted_by_offset_[first_entry_index_]; + } else { + if (start->next_entry_index == -1) { + return nullptr; + } + candidate_next_entry = &buffers_sorted_by_offset_[start->next_entry_index]; + } + do { + if (DoesEntryOverlapInTime(candidate_next_entry, first_time_used, + last_time_used)) { + result = candidate_next_entry; + break; + } + if (candidate_next_entry->next_entry_index == -1) { + break; + } + candidate_next_entry = + &buffers_sorted_by_offset_[candidate_next_entry->next_entry_index]; + } while (true); + return result; +} + +void GreedyMemoryPlanner::CalculateOffsetsIfNeeded() { + if (!need_to_calculate_offsets_ || (buffer_count_ == 0)) { + return; + } + need_to_calculate_offsets_ = false; + + // Start off by ordering the buffers in descending order of size. + // This helps find a more compact layout. Intuitively, you can think + // about putting the large buffers in place first, and then the + // smaller buffers can fit in the gaps, rather than fragmenting the + // gaps with small buffers at the beginning. Add offline planned offsets + // first in the list, since they have a predetermined offset. + int idx_from_tail = buffer_count_; + int idx_from_head = 0; + for (int i = 0; i < buffer_count_; ++i) { + if (requirements_[i].offline_offset == kOnlinePlannedBuffer) { + idx_from_tail--; + buffer_sizes_sorted_[idx_from_tail] = requirements_[i].size; + buffer_ids_sorted_[idx_from_tail] = i; + buffer_offsets_[i] = -1; + } else { + buffer_sizes_sorted_[idx_from_head] = requirements_[i].size; + buffer_ids_sorted_[idx_from_head] = i; + buffer_offsets_[i] = requirements_[i].offline_offset; + idx_from_head++; + } + } + + // This sorting algorithm is naive, and may end up taking a very long time + // with hundreds of buffers. Do not sort the offline planned offsets. + ReverseSortInPlace(&buffer_sizes_sorted_[idx_from_head], + &buffer_ids_sorted_[idx_from_head], + buffer_count_ - idx_from_head); + + // Initialize the first entry to the first buffer in + // buffer_ids_sorted_. + // - If there are no offline planned offsets, the largest buffer will be + // first, and the buffers will be handled in size order. + // - If offline offsets are present, these will be handled first in order + // for the greedy algorithm to utilized gaps in the offline plan. + first_entry_index_ = 0; + next_free_entry_ = 1; + ListEntry* first_entry = &buffers_sorted_by_offset_[first_entry_index_]; + first_entry->next_entry_index = -1; // to mark the entry as end of list + int buffer_id = buffer_ids_sorted_[0]; + first_entry->requirements_index = buffer_id; + if (requirements_[buffer_id].offline_offset == kOnlinePlannedBuffer) { + buffer_offsets_[buffer_id] = 0; + } + first_entry->offset = buffer_offsets_[buffer_id]; + + // Work through the rest of the buffers to find a good gap to place each one. + for (int i = 1; i < buffer_count_; ++i) { + // The id is the order the buffer was originally added by the client. + buffer_id = buffer_ids_sorted_[i]; + // Look at what size and time range the buffer needs to be active. + BufferRequirements* wanted_requirements = &requirements_[buffer_id]; + const int wanted_size = wanted_requirements->size; + const int wanted_first_time_used = wanted_requirements->first_time_used; + const int wanted_last_time_used = wanted_requirements->last_time_used; + + // Find the first buffer that's active in our time range. All placed + // buffers are stored in the order of their starting position in the arena + // so that it's easy to find the next buffer in memory, and so the gap. + // The candidate_entry variable holds the buffer that we're considering + // placing the current buffer after. + + int candidate_offset = 0; + // Loop through the offset-ordered list of buffers, looking for gaps. + if (wanted_requirements->offline_offset == kOnlinePlannedBuffer) { + ListEntry* prior_entry = nullptr; + while (true) { + // Find out what the next active buffer is. + ListEntry* next_entry = NextSimultaneouslyActiveBuffer( + prior_entry, wanted_first_time_used, wanted_last_time_used); + + if (prior_entry) { + BufferRequirements* candidate_requirements = + &requirements_[prior_entry->requirements_index]; + const int prior_entry_offset = + prior_entry->offset + candidate_requirements->size; + if (prior_entry_offset > candidate_offset) { + candidate_offset = prior_entry_offset; + } + } + if (next_entry == nullptr) { + // We're at the end of the list, so we can always append the buffer + // here. + break; + } + // Find out how much space there is between us and the next buffer. + const int gap = next_entry->offset - candidate_offset; + if (gap >= wanted_size) { + // This entry has a big enough gap between it and the next, so + // use it! + break; + } + // The gap wasn't big enough, so move on to another candidate. + prior_entry = next_entry; + } + } else { + // Offline planned offset are to be considered constant + candidate_offset = wanted_requirements->offline_offset; + } + // At this point, we've either found a gap (possibly at the end of the + // list) and want to place the buffer there, or there are no other active + // buffers in this time range and so we can put it at offset zero. + // Record the buffer's offset in our plan. + buffer_offsets_[buffer_id] = candidate_offset; + // Add the newly-placed buffer to our offset-ordered list, so that + // subsequent passes can fit in their buffers around it. + ListEntry* new_entry = &buffers_sorted_by_offset_[next_free_entry_]; + new_entry->offset = candidate_offset; + new_entry->requirements_index = buffer_id; + const int new_entry_index = next_free_entry_; + ++next_free_entry_; + + if (first_entry->offset > candidate_offset) { + // The new entry offset is smaller than the first entry offset => + // replace the first entry + first_entry = new_entry; + first_entry->next_entry_index = first_entry_index_; + first_entry_index_ = new_entry_index; + } else { + ListEntry* current_entry = first_entry; + // Make sure that we insert the buffer at the correct place in the + // buffer-offset-ordered list + while (true) { + const int next_entry_index = current_entry->next_entry_index; + if (next_entry_index == -1) { + // We're at the end of the list, so just add the new entry here. + current_entry->next_entry_index = new_entry_index; + new_entry->next_entry_index = -1; + break; + } + // not at the end of the list -> take a look at next entry + ListEntry* next_entry = &buffers_sorted_by_offset_[next_entry_index]; + if (next_entry->offset > candidate_offset) { + // We're at the right spot to do an insertion and retain the sorting + // order, so place the new entry here. + new_entry->next_entry_index = current_entry->next_entry_index; + current_entry->next_entry_index = new_entry_index; + break; + } + current_entry = next_entry; + } + } + } +} + +size_t GreedyMemoryPlanner::GetMaximumMemorySize() { + CalculateOffsetsIfNeeded(); + if (buffer_count_ == 0) { + return 0; + } + ListEntry* entry = &buffers_sorted_by_offset_[first_entry_index_]; + size_t max_size = 0; + while (entry) { + BufferRequirements* requirements = + &requirements_[entry->requirements_index]; + // TODO(b/148246793): Update all size and offset variables types from + // int to size_t + const size_t current_size = entry->offset + requirements->size; + if (current_size > max_size) { + max_size = current_size; + } + if (entry->next_entry_index == -1) { + break; + } + entry = &buffers_sorted_by_offset_[entry->next_entry_index]; + } + return max_size; +} + +void GreedyMemoryPlanner::PrintMemoryPlan(ErrorReporter* error_reporter) { + CalculateOffsetsIfNeeded(); + + for (int i = 0; i < buffer_count_; ++i) { + TF_LITE_REPORT_ERROR( + error_reporter, + "Planner buffer ID: %d, calculated offset: %d, size required: %d, " + "first_time_created: %d, " + "last_time_used: %d", + i, buffer_offsets_[i], requirements_[i].size, + requirements_[i].first_time_used, requirements_[i].last_time_used); + } + + constexpr int kLineWidth = 80; + int max_size = kLineWidth; + int max_time = 0; + for (int i = 0; i < buffer_count_; ++i) { + BufferRequirements* requirements = &requirements_[i]; + const int offset = buffer_offsets_[i]; + const int last_time_used = requirements->last_time_used; + const int size = offset + requirements->size; + if (size > max_size) { + max_size = size; + } + if (last_time_used > max_time) { + max_time = last_time_used; + } + } + + char line[kLineWidth + 1]; + for (int t = 0; t <= max_time; ++t) { + for (int c = 0; c < kLineWidth; ++c) { + line[c] = '.'; + } + for (int i = 0; i < buffer_count_; ++i) { + BufferRequirements* requirements = &requirements_[i]; + if ((t < requirements->first_time_used) || + (t > requirements->last_time_used)) { + continue; + } + const int offset = buffer_offsets_[i]; + if (offset == -1) { + continue; + } + const int size = requirements->size; + const int line_start = (offset * kLineWidth) / max_size; + const int line_end = ((offset + size) * kLineWidth) / max_size; + for (int n = line_start; n < line_end; ++n) { + if (line[n] == '.') { + char display; + if (i < 10) { + display = '0' + i; + } else if (i < 36) { + display = 'a' + (i - 10); + } else if (i < 62) { + display = 'A' + (i - 36); + } else { + display = '*'; + } + line[n] = display; + } else { + line[n] = '!'; + } + } + } + line[kLineWidth] = 0; + TF_LITE_REPORT_ERROR(error_reporter, "%s", (const char*)line); + } +} + +int GreedyMemoryPlanner::GetBufferCount() { return buffer_count_; } + +TfLiteStatus GreedyMemoryPlanner::GetOffsetForBuffer( + tflite::ErrorReporter* error_reporter, int buffer_index, int* offset) { + CalculateOffsetsIfNeeded(); + if ((buffer_index < 0) || (buffer_index >= buffer_count_)) { + TF_LITE_REPORT_ERROR(error_reporter, + "buffer index %d is outside range 0 to %d", + buffer_index, buffer_count_); + return kTfLiteError; + } + *offset = buffer_offsets_[buffer_index]; + return kTfLiteOk; +} + +bool GreedyMemoryPlanner::DoAnyBuffersOverlap(ErrorReporter* error_reporter) { + CalculateOffsetsIfNeeded(); + bool were_overlaps_found = false; + for (int i = 0; i < buffer_count_; ++i) { + BufferRequirements* a_requirements = &requirements_[i]; + const int a_start_offset = buffer_offsets_[i]; + const int a_first_time_used = a_requirements->first_time_used; + const int a_last_time_used = a_requirements->last_time_used; + const int a_end_offset = a_start_offset + a_requirements->size; + for (int j = 0; j < buffer_count_; ++j) { + if (i == j) { + continue; + } + BufferRequirements* b_requirements = &requirements_[j]; + const int b_start_offset = buffer_offsets_[j]; + const int b_first_time_used = b_requirements->first_time_used; + const int b_last_time_used = b_requirements->last_time_used; + const int b_end_offset = b_start_offset + b_requirements->size; + if ((a_first_time_used > b_last_time_used) || + (b_first_time_used > a_last_time_used)) { + // Buffers don't overlap in time. + continue; + } + if ((a_start_offset >= b_end_offset) || + (b_start_offset >= a_end_offset)) { + // No overlap in memory. + continue; + } + were_overlaps_found = true; + TF_LITE_REPORT_ERROR( + error_reporter, "Overlap: %d (%d=>%d, %d->%d) vs %d (%d=>%d, %d->%d)", + i, a_first_time_used, a_last_time_used, a_start_offset, a_end_offset, + j, b_first_time_used, b_last_time_used, b_start_offset, b_end_offset); + } + } + return were_overlaps_found; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h new file mode 100644 index 0000000..02794af --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h @@ -0,0 +1,166 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ +#define TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_planner/memory_planner.h" + +namespace tflite { + +constexpr int kOnlinePlannedBuffer = -1; + +// A memory planner that uses a greedy algorithm to arrange buffers in memory +// to minimize the overall arena size needed. +// +// The algorithm works like this: +// - The client enters the buffer information through AddBuffer(). +// - When a function like GetOffsetForBuffer() is called, the +// CalculateOffsetsIfNeeded() method is invoked. +// - If an up to date plan is not already present, one will be calculated. +// - The buffers are sorted in descending order of size. +// - The largest buffer is placed at offset zero. +// - The rest of the buffers are looped through in descending size order. +// - The other buffers that need to be in memory at the same time are found. +// - The first gap between simultaneously active buffers that the current +// buffer fits into will be used. +// - If no large-enough gap is found, the current buffer is placed after the +// last buffer that's simultaneously active. +// - This continues until all buffers are placed, and the offsets stored. +// +// This is not guaranteed to produce the best placement, since that's an +// NP-Complete problem, but in practice it should produce one that's decent. +class GreedyMemoryPlanner : public MemoryPlanner { + public: + // You need to pass in an area of memory to be used for planning. This memory + // needs to have a lifetime as long as the planner, but isn't owned by this + // object, so management should be handled by the client. This is so it can be + // stack or globally allocated if necessary on devices without dynamic memory + // allocation. How many buffers can be planned for will depend on the size of + // this scratch memory, so you should enlarge it if you see an error when + // calling AddBuffer(). The memory can be reused once you're done with the + // planner, as long as you copy the calculated offsets to another location. + // Each buffer requires about 36 bytes of scratch. + GreedyMemoryPlanner(unsigned char* scratch_buffer, int scratch_buffer_size); + ~GreedyMemoryPlanner() override; + + // Record details of a buffer we want to place. + TfLiteStatus AddBuffer(ErrorReporter* error_reporter, int size, + int first_time_used, int last_time_used) override; + + // Record details of an offline planned buffer offset we want to place. + // offline_offset is the buffer offset from the start of the arena. + TfLiteStatus AddBuffer(ErrorReporter* error_reporter, int size, + int first_time_used, int last_time_used, + int offline_offset); + + // Returns the high-water mark of used memory. This is the minimum size of a + // memory arena you'd need to allocate to hold these buffers. + size_t GetMaximumMemorySize() override; + + // How many buffers have been recorded. + int GetBufferCount() override; + + // Where a given buffer should be placed in the memory arena. + // This information is stored in the memory arena itself, so once the arena + // is used for inference, it will be overwritten. + TfLiteStatus GetOffsetForBuffer(ErrorReporter* error_reporter, + int buffer_index, int* offset) override; + + // Prints an ascii-art diagram of the buffer layout plan. + void PrintMemoryPlan(ErrorReporter* error_reporter); + + // Debug method to check whether any buffer allocations are overlapping. This + // is an O(N^2) complexity operation, so only use for testing. + bool DoAnyBuffersOverlap(ErrorReporter* error_reporter); + + // Used to store a list of buffers ordered by their offset. + struct ListEntry { + int offset; + int requirements_index; + int next_entry_index; + }; + + // Number of bytes required in order to plan a buffer. + static size_t per_buffer_size() { + const int per_buffer_size = + sizeof(BufferRequirements) + // requirements_ + sizeof(int) + // buffer_sizes_sorted_ + sizeof(int) + // buffer_ids_sorted_ + sizeof(ListEntry) + // buffers_sorted_by_offset_ + sizeof(int); // buffer_offsets_; + return per_buffer_size; + } + + private: + // Whether a buffer is active in a given time range. + bool DoesEntryOverlapInTime(const ListEntry* entry, const int first_time_used, + const int last_time_used) const; + + // Walks the list to return the next buffer that is active in a given time + // range, or a null pointer if there are none. + ListEntry* NextSimultaneouslyActiveBuffer(const ListEntry* start, + const int first_time_used, + const int last_time_used); + + // If there isn't an up to date plan, calculate a new one. + void CalculateOffsetsIfNeeded(); + + // How many buffers we can plan for, based on the arena size we're given in + // the constructor. + int max_buffer_count_; + + // The number of buffers added so far. + int buffer_count_; + + // Records the client-provided information about each buffer. + struct BufferRequirements { + int size; + int offline_offset; + int first_time_used; + int last_time_used; + }; + + // Working arrays used during the layout algorithm. + BufferRequirements* requirements_; + // buffer_sizes_sorted_ and buffer_ids_sorted_ are sorted according to: + // { + // offline planned buffers, + // online planned buffers sorted by size + // } + int* buffer_sizes_sorted_; + int* buffer_ids_sorted_; + ListEntry* buffers_sorted_by_offset_; + int next_free_entry_; // Index of the next free entry of + // buffers_sorted_by_offset_ + int first_entry_index_; // Index of the first entry (smallest offset) of + // buffers_sorted_by_offset_ + + // Stores the outcome of the plan, the location of each buffer in the arena. + int* buffer_offsets_; + + // Whether buffers have been added since the last plan was calculated. + bool need_to_calculate_offsets_; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.cpp b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.cpp new file mode 100644 index 0000000..3f10224 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.cpp @@ -0,0 +1,57 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.h" + +namespace tflite { + +LinearMemoryPlanner::LinearMemoryPlanner() + : current_buffer_count_(0), next_free_offset_(0) {} +LinearMemoryPlanner::~LinearMemoryPlanner() {} + +TfLiteStatus LinearMemoryPlanner::AddBuffer( + tflite::ErrorReporter* error_reporter, int size, int first_time_used, + int last_time_used) { + if (current_buffer_count_ >= kMaxBufferCount) { + TF_LITE_REPORT_ERROR(error_reporter, "Too many buffers (max is %d)", + kMaxBufferCount); + return kTfLiteError; + } + buffer_offsets_[current_buffer_count_] = next_free_offset_; + next_free_offset_ += size; + ++current_buffer_count_; + return kTfLiteOk; +} + +size_t LinearMemoryPlanner::GetMaximumMemorySize() { return next_free_offset_; } + +int LinearMemoryPlanner::GetBufferCount() { return current_buffer_count_; } + +TfLiteStatus LinearMemoryPlanner::GetOffsetForBuffer( + tflite::ErrorReporter* error_reporter, int buffer_index, int* offset) { + if ((buffer_index < 0) || (buffer_index >= current_buffer_count_)) { + TF_LITE_REPORT_ERROR(error_reporter, + "buffer index %d is outside range 0 to %d", + buffer_index, current_buffer_count_); + return kTfLiteError; + } + *offset = buffer_offsets_[buffer_index]; + return kTfLiteOk; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/memory_planner/linear_memory_planner.h b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.h similarity index 89% rename from src/tensorflow/lite/micro/memory_planner/linear_memory_planner.h rename to src/tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.h index 4d77e77..f4596bd 100644 --- a/src/tensorflow/lite/micro/memory_planner/linear_memory_planner.h +++ b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/linear_memory_planner.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -16,8 +17,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_LINEAR_MEMORY_PLANNER_H_ #define TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_LINEAR_MEMORY_PLANNER_H_ -#include "tensorflow/lite/micro/compatibility.h" -#include "tensorflow/lite/micro/memory_planner/memory_planner.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_planner/memory_planner.h" namespace tflite { @@ -48,3 +49,5 @@ class LinearMemoryPlanner : public MemoryPlanner { } // namespace tflite #endif // TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_LINEAR_MEMORY_PLANNER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/memory_planner/memory_planner.h b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/memory_planner.h similarity index 91% rename from src/tensorflow/lite/micro/memory_planner/memory_planner.h rename to src/tensorflow_arm/tensorflow/lite/micro/memory_planner/memory_planner.h index 4670d59..8beb5a6 100644 --- a/src/tensorflow/lite/micro/memory_planner/memory_planner.h +++ b/src/tensorflow_arm/tensorflow/lite/micro/memory_planner/memory_planner.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -16,8 +17,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_MEMORY_PLANNER_H_ #define TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_MEMORY_PLANNER_H_ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" namespace tflite { @@ -57,7 +58,7 @@ class MemoryPlanner { int size, int first_time_used, int last_time_used) = 0; - // The largest contguous block of memory that's needed to hold the layout. + // The largest contiguous block of memory that's needed to hold the layout. virtual size_t GetMaximumMemorySize() = 0; // How many buffers have been added to the planner. virtual int GetBufferCount() = 0; @@ -69,3 +70,5 @@ class MemoryPlanner { } // namespace tflite #endif // TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_MEMORY_PLANNER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.cpp b/src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.cpp new file mode 100644 index 0000000..7e76a4b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.cpp @@ -0,0 +1,1161 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/micro_allocator.h" + +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_arm/tensorflow/lite/core/api/op_resolver.h" +#include "tensorflow_arm/tensorflow/lite/core/api/tensor_utils.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_planner/memory_planner.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h" +#include "tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_utils.h" + +namespace tflite { + +namespace { + +// Maximum number of scratch buffer requests per operator. Operator kernels that +// request more than this value will receive an exception. +constexpr size_t kMaxScratchBuffersPerOp = 12; + +// Sentinel value used as a placeholder to mark a ScratchBufferRequest request +// needs a node id assignment. +constexpr int kUnassignedScratchBufferRequestIndex = -1; + +// Used to hold information used during allocation calculations. +struct AllocationInfo { + size_t bytes; + void** output_ptr; + int first_created; + int last_used; + int32_t offline_offset; + bool needs_allocating; +}; + +// We align tensor buffers to 16-byte boundaries, since this is a common +// requirement for SIMD extensions. +constexpr int kBufferAlignment = 16; +constexpr char kOfflineMemAllocMetadata[] = "OfflineMemoryAllocation"; +const TfLiteIntArray kZeroLengthIntArray = {}; + +class MicroBuiltinDataAllocator : public BuiltinDataAllocator { + public: + explicit MicroBuiltinDataAllocator(SimpleMemoryAllocator* memory_allocator) + : memory_allocator_(memory_allocator) {} + + void* Allocate(size_t size, size_t alignment_hint) override { + return memory_allocator_->AllocateFromTail(size, alignment_hint); + } + void Deallocate(void* data) override { + // Do not deallocate, builtin data needs to be available for the life time + // of the model. + } + + private: + SimpleMemoryAllocator* memory_allocator_; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +#if !defined(__clang__) +// Helper function to check flatbuffer metadata correctness. This function is +// not called by default. Hence it's not linked in to the final binary code. +TfLiteStatus CheckOfflinePlannedOffsets(const Model* model, + ErrorReporter* error_reporter) { + // Suppress compile warning for unused function + (void)CheckOfflinePlannedOffsets; + + if (model->metadata()) { + for (size_t i = 0; i < model->metadata()->size(); ++i) { + auto metadata = model->metadata()->Get(i); + if (strncmp(metadata->name()->c_str(), kOfflineMemAllocMetadata, + strlen(kOfflineMemAllocMetadata)) == 0) { + auto* subgraphs = model->subgraphs(); + const SubGraph* subgraph = (*subgraphs)[0]; + const flatbuffers::Vector>* tensors = + subgraph->tensors(); + const flatbuffers::Vector>* buffers = + model->buffers(); + int nbr_tflite_tensors = tensors->size(); + auto* buffer = (*buffers)[metadata->buffer()]; + auto* array = buffer->data(); + const uint32_t* metadata_buffer = (uint32_t*)array->data(); + int version = metadata_buffer[0]; + int subgraph_idx = metadata_buffer[1]; + const int nbr_offline_offsets = metadata_buffer[2]; +#ifndef TF_LITE_STRIP_ERROR_STRINGS + int* offline_planner_offsets = (int*)&metadata_buffer[3]; +#endif + + TF_LITE_REPORT_ERROR(error_reporter, "==== Model metadata info: ====="); + TF_LITE_REPORT_ERROR(error_reporter, + "Offline planner metadata found, version %d, " + "subgraph %d, nbr offline offsets %d", + version, subgraph_idx, nbr_offline_offsets); + for (int j = 0; j < nbr_offline_offsets; ++j) { + TF_LITE_REPORT_ERROR( + error_reporter, + "Offline planner tensor index %d, offline offset: %d", j, + offline_planner_offsets[j]); + } + + if (version != 1) { + TF_LITE_REPORT_ERROR(error_reporter, "Version not supported! (%d)\n", + version); + return kTfLiteError; + } + if (subgraph_idx != 0) { + TF_LITE_REPORT_ERROR(error_reporter, + "Only 1 subgraph supported! Subgraph idx (%d)\n", + subgraph_idx); + return kTfLiteError; + } + if (nbr_tflite_tensors != nbr_offline_offsets) { + TF_LITE_REPORT_ERROR(error_reporter, + "Nbr of offline buffer offsets (%d) in metadata " + "not equal nbr tensors (%d)\n", + nbr_offline_offsets, nbr_tflite_tensors); + return kTfLiteError; + } + } + } + } + return kTfLiteOk; +} +#endif + +// A helper class to construct AllocationInfo array. This array contains the +// lifetime of tensors / scratch_buffer and will be used to calculate the memory +// plan. Methods need to be called in order from `Init`, `Add*`, to `Finish`. +class AllocationInfoBuilder { + public: + AllocationInfoBuilder(AllocationInfo* info, size_t tensor_count, + size_t scratch_buffer_count, ErrorReporter* reporter) + : info_(info), + tensor_count_(tensor_count), + buffer_count_(scratch_buffer_count), + reporter_(reporter) {} + + // Check if model contains offline planned buffer offsets. + // - If there's no metadata available, offline_planner_offsets is not set + // - If there's metadata available, offline_planner_offsets will point to the + // first offset in the metadata buffer list. + TfLiteStatus GetOfflinePlannedOffsets( + const Model* model, const int32_t** offline_planner_offsets); + + // Add allocaiton information for the tensors. + TfLiteStatus AddTensors(const SubGraph* subgraph, + const int32_t* offline_offsets, + TfLiteEvalTensor* eval_tensors); + + // Add allocation information for the scratch buffers. + TfLiteStatus AddScratchBuffers( + internal::ScratchBufferRequest* scratch_buffer_requests, + ScratchBufferHandle* scratch_buffer_handles); + + // Returns a pointer to the built AllocationInfo array. + const AllocationInfo* Finish() const { return info_; } + + private: + AllocationInfo* info_ = nullptr; + size_t tensor_count_ = 0; + size_t buffer_count_ = 0; + ErrorReporter* reporter_ = nullptr; +}; + +TfLiteStatus AllocationInfoBuilder::AddTensors(const SubGraph* subgraph, + const int32_t* offline_offsets, + TfLiteEvalTensor* eval_tensors) { + TFLITE_DCHECK(eval_tensors != nullptr); + + // Set up allocation info for all tensors. + for (size_t i = 0; i < tensor_count_; ++i) { + AllocationInfo* current = &info_[i]; + current->output_ptr = &(eval_tensors[i].data.data); + + TF_LITE_ENSURE_STATUS( + TfLiteEvalTensorByteLength(&eval_tensors[i], ¤t->bytes)); + + current->first_created = -1; + current->last_used = -1; + current->needs_allocating = (eval_tensors[i].data.data == nullptr) && + (!subgraph->tensors()->Get(i)->is_variable()); + if (offline_offsets) { + current->offline_offset = offline_offsets[i]; + } else { + current->offline_offset = kOnlinePlannedBuffer; + } + } + + for (size_t i = 0; i < subgraph->inputs()->size(); ++i) { + const int tensor_index = subgraph->inputs()->Get(i); + AllocationInfo* current = &info_[tensor_index]; + current->first_created = 0; + } + + // Mark all outputs as persistent to the end of the invocation. + for (size_t i = 0; i < subgraph->outputs()->size(); ++i) { + const int tensor_index = subgraph->outputs()->Get(i); + AllocationInfo* current = &info_[tensor_index]; + current->last_used = subgraph->operators()->size() - 1; + } + + // Figure out when the first and last use of each tensor is. + for (int i = (subgraph->operators()->size() - 1); i >= 0; --i) { + const auto* op = subgraph->operators()->Get(i); + for (size_t n = 0; n < op->inputs()->size(); ++n) { + const int tensor_index = op->inputs()->Get(n); + AllocationInfo* current = &info_[tensor_index]; + if (((current->last_used == -1) || (current->last_used < i))) { + current->last_used = i; + } + } + for (size_t n = 0; n < op->outputs()->size(); ++n) { + const int tensor_index = op->outputs()->Get(n); + AllocationInfo* current = &info_[tensor_index]; + if ((current->first_created == -1) || (current->first_created > i)) { + current->first_created = i; + } + } + } + + // Sanity check for valid tensor lifetime. + for (size_t i = 0; i < tensor_count_; ++i) { + AllocationInfo* current = &info_[i]; + // Even though tensor appears to be read only it may still need to be + // allocated. + const bool appears_read_only = + (current->first_created == -1) && (current->last_used != -1); + const bool has_partial_lifetime = + !appears_read_only && + ((current->first_created == -1) || (current->last_used == -1)); + if (has_partial_lifetime && current->needs_allocating) { + TF_LITE_REPORT_ERROR( + reporter_, + "Logic error in memory planner, tensor %d has an invalid lifetime: " + "first_created: %d, last_used: %d", + i, current->first_created, current->last_used); + return kTfLiteError; + } + } + return kTfLiteOk; +} + +// The tensor offsets will be encoded in the metadata:[Metadata] field of the +// Model. The following encoding applies: +// +// | Metadata component | Value | +// | name:string | “OfflineMemoryAllocation” | +// | buffer:unit | Index of buffer containing memory allocation data | +// +// The buffer contents for the memory allocation is a list of 32-bit integers. +// The number of tensors, n, must be equal to the number of tensors defined in +// the model. The following encoding applies: +// +// | Offset | Value | +// | 0 | Offline allocation format version – set to 0 | +// | 1 | Subgraph index to which this allocation applies | +// | 2 | Number offsets following: n | +// | 3 | Arena byte offset of tensor #0 or -1 to allocate at runtime | +// | 4 | Arena byte offset of tensor #1 or -1 to allocate at runtime | +// | 3+(n-1) | Arena byte offset of tensor #(n-1) or -1 to allocate at runtime | +TfLiteStatus AllocationInfoBuilder::GetOfflinePlannedOffsets( + const Model* model, const int32_t** offline_planner_offsets) { + if (model->metadata()) { + for (size_t i = 0; i < model->metadata()->size(); ++i) { + auto metadata = model->metadata()->Get(i); + if (strncmp(metadata->name()->c_str(), kOfflineMemAllocMetadata, + strlen(kOfflineMemAllocMetadata)) == 0) { + const flatbuffers::Vector>* buffers = + model->buffers(); + auto* buffer = (*buffers)[metadata->buffer()]; + auto* array = buffer->data(); + const uint32_t* metadata_buffer = + reinterpret_cast(array->data()); + const size_t nbr_tensors = static_cast(metadata_buffer[2]); + *offline_planner_offsets = + reinterpret_cast(&metadata_buffer[3]); + + if (tensor_count_ != nbr_tensors) { + TF_LITE_REPORT_ERROR(reporter_, + "Nbr of offline buffer offsets (%d) in metadata " + "not equal nbr tensors (%d)\n", + nbr_tensors, tensor_count_); + return kTfLiteError; + } + } + } + } + return kTfLiteOk; +} + +TfLiteStatus AllocationInfoBuilder::AddScratchBuffers( + internal::ScratchBufferRequest* scratch_buffer_requests, + ScratchBufferHandle* scratch_buffer_handles) { + // Set up allocation info for buffers. + for (size_t i = tensor_count_; i < tensor_count_ + buffer_count_; ++i) { + internal::ScratchBufferRequest* current_request = + &(scratch_buffer_requests[i - tensor_count_]); + ScratchBufferHandle* current_handle = + &(scratch_buffer_handles[i - tensor_count_]); + + AllocationInfo* current = &info_[i]; + current->output_ptr = reinterpret_cast(¤t_handle->data); + current->bytes = current_request->bytes; + current->first_created = current_request->node_idx; + current->last_used = current_request->node_idx; + current->offline_offset = kOnlinePlannedBuffer; + current->needs_allocating = true; + } + return kTfLiteOk; +} + +TfLiteStatus CreatePlan(ErrorReporter* error_reporter, + GreedyMemoryPlanner* planner, + const AllocationInfo* allocation_info, + size_t allocation_info_size) { + // Add the tensors to our allocation plan. + for (size_t i = 0; i < allocation_info_size; ++i) { + const AllocationInfo* current = &allocation_info[i]; + if (current->needs_allocating) { + size_t aligned_bytes_required = + AlignSizeUp(current->bytes, kBufferAlignment); + if (current->offline_offset == kOnlinePlannedBuffer) { + TF_LITE_ENSURE_STATUS( + planner->AddBuffer(error_reporter, aligned_bytes_required, + current->first_created, current->last_used)); + } else { + TF_LITE_ENSURE_STATUS(planner->AddBuffer( + error_reporter, aligned_bytes_required, current->first_created, + current->last_used, current->offline_offset)); + } + } + } + return kTfLiteOk; +} + +TfLiteStatus CommitPlan(ErrorReporter* error_reporter, MemoryPlanner* planner, + uint8_t* starting_point, + const AllocationInfo* allocation_info, + size_t allocation_info_size) { + // Figure out the actual memory addresses for each buffer, based on the plan. + int planner_index = 0; + for (size_t i = 0; i < allocation_info_size; ++i) { + const AllocationInfo* current = &allocation_info[i]; + if (current->needs_allocating) { + int offset = -1; + TF_LITE_ENSURE_STATUS( + planner->GetOffsetForBuffer(error_reporter, planner_index, &offset)); + *current->output_ptr = reinterpret_cast(starting_point + offset); + ++planner_index; + } + } + return kTfLiteOk; +} +} // namespace + +namespace internal { + +// Handles architecture safe mapping of flatbuffer vectors to a TfLite*Array +// struct. Matching types are required (e.g. float and TfLiteFloatArray). +// Big-endian systems will always allocate dimension array data in the tail +// (persistent) section. +template +TfLiteStatus FlatBufferVectorToTfLiteTypeArray( + SimpleMemoryAllocator* allocator, ErrorReporter* error_reporter, + const flatbuffers::Vector* flatbuffer_array, + kTfLiteArrayType** result) { + TFLITE_DCHECK(error_reporter != nullptr); + TFLITE_DCHECK(flatbuffer_array != nullptr); + // TODO(b/159668691): Consider adding type assertion or breaking this function + // into multiple functions for each type. std::is_same is c++11 and has a + // special updated constructor in c++17 that requires a string argument. + if (FLATBUFFERS_LITTLEENDIAN) { + // On little-endian machines, TfLite*Array happens to have the same memory + // layout as flatbuffers:Vector, so we can + // reinterpret_cast the flatbuffer vector and avoid a copy and malloc. + *result = const_cast( + reinterpret_cast(flatbuffer_array)); + } else { + // Big-endian architecture can not use the same memory layout as + // flatbuffers::Vector. Allocate from the tail and + // copy values from the flatbuffer into the newly allocated chunk. + kTfLiteArrayType* array = + reinterpret_cast(allocator->AllocateFromTail( + TfLiteIntArrayGetSizeInBytes(flatbuffer_array->Length()), + alignof(kTfLiteArrayType))); + if (array == nullptr) { + TF_LITE_REPORT_ERROR( + error_reporter, + "Failed to allocate %d bytes of memory to copy an array.", + TfLiteIntArrayGetSizeInBytes(flatbuffer_array->Length())); + return kTfLiteError; + } + array->size = flatbuffer_array->Length(); + for (int i = 0; i < array->size; ++i) { + array->data[i] = flatbuffer_array->Get(i); + } + *result = array; + } + return kTfLiteOk; +} + +// Returns a pointer to any buffer associated with the flatbuffer tensor. Can +// return nullptr if no buffer is found. +void* GetFlatbufferTensorBuffer( + const tflite::Tensor& flatbuffer_tensor, + const flatbuffers::Vector>* buffers) { + // We need to figure out where the actual contents of this tensor are stored + // in memory. We'll check to see if there's a serialized buffer (pretty much + // the same as a constant op in TensorFlow) associated with this tensor first, + // and if there is update the runtime structure to point to its location in + // memory. + // First see if there's any buffer information in the serialized tensor. + // TODO(b/170379532): Add better unit tests to validate flatbuffer values. + void* out_buffer = nullptr; + if (auto* buffer = (*buffers)[flatbuffer_tensor.buffer()]) { + // If we've found a buffer, does it have any data? + if (auto* array = buffer->data()) { + // If it has any data, is the data size larger than zero? + if (array->size()) { + // We've found a buffer with valid data, so update the runtime tensor + // data structure to point to it. + out_buffer = const_cast(static_cast(array->data())); + } + } + // TODO(petewarden): It's not clear in what circumstances we could have a + // buffer in the serialized tensor, but it doesn't have any data in it. Is + // that a validly-generated file, and if so what does it mean, or is it an + // error condition? It would be good to tighten up the specification to make + // it less ambiguous. + } + return out_buffer; +} + +TfLiteStatus InitializeTfLiteTensorFromFlatbuffer( + SimpleMemoryAllocator* allocator, bool allocate_temp, + const tflite::Tensor& flatbuffer_tensor, + const flatbuffers::Vector>* buffers, + ErrorReporter* error_reporter, TfLiteTensor* result) { + TFLITE_DCHECK(result != nullptr); + + *result = {}; + // Make sure the serialized type is one we know how to deal with, and convert + // it from a flatbuffer enum into a constant used by the kernel C API. + TF_LITE_ENSURE_STATUS(ConvertTensorType(flatbuffer_tensor.type(), + &result->type, error_reporter)); + // Make sure we remember if the serialized tensor is designated as a variable. + result->is_variable = flatbuffer_tensor.is_variable(); + + result->data.data = GetFlatbufferTensorBuffer(flatbuffer_tensor, buffers); + + // TODO(petewarden): Some of these paths aren't getting enough testing + // coverage, so we should figure out some tests that exercise them. + if (result->data.data == nullptr) { + // The tensor contents haven't been set from a serialized buffer, so + // make a note that they will be allocated from memory. The actual + // allocation won't happen until later. + result->allocation_type = kTfLiteArenaRw; + } else { + // We set the data from a serialized buffer, so record tha. + result->allocation_type = kTfLiteMmapRo; + } + + // Figure out what the size in bytes of the buffer is and store it. + size_t type_size; + TF_LITE_ENSURE_STATUS(BytesRequiredForTensor( + flatbuffer_tensor, &result->bytes, &type_size, error_reporter)); + + if (flatbuffer_tensor.shape() == nullptr) { + // flatbuffer_tensor.shape() can return a nullptr in the case of a scalar + // tensor. + result->dims = const_cast(&kZeroLengthIntArray); + } else { + // TFLM doesn't allow reshaping the tensor which requires dynamic memory + // allocation so it is safe to drop the const qualifier. In the future, if + // we really want to update the tensor shape, we can always pass in a new + // TfLiteIntArray - especially we have to do so if the dimension is + TF_LITE_ENSURE_STATUS(FlatBufferVectorToTfLiteTypeArray( + allocator, error_reporter, flatbuffer_tensor.shape(), &(result->dims))); + } + + // Copy the quantization information from the serialized data. + const auto* src_quantization = flatbuffer_tensor.quantization(); + if (src_quantization && src_quantization->scale() && + (src_quantization->scale()->size() > 0) && + src_quantization->zero_point() && + (src_quantization->zero_point()->size() > 0)) { + // Always populate the TfLiteTensor.params field, even if there are + // per-channel quantization parameters. + result->params.scale = src_quantization->scale()->Get(0); + // Note that the zero_point field in the FlatBuffers schema is a 64-bit + // integer, but the zero_point field in the TfLiteQuantizationParams struct + // is a 32-bit integer. + result->params.zero_point = + static_cast(src_quantization->zero_point()->Get(0)); + + // Populate per-channel quantization params. + int channels = src_quantization->scale()->size(); + TfLiteAffineQuantization* quantization = + allocate_temp + ? reinterpret_cast( + allocator->AllocateTemp(sizeof(TfLiteAffineQuantization), + alignof(TfLiteAffineQuantization))) + : reinterpret_cast( + allocator->AllocateFromTail( + sizeof(TfLiteAffineQuantization), + alignof(TfLiteAffineQuantization))); + if (quantization == nullptr) { + TF_LITE_REPORT_ERROR(error_reporter, + "Unable to allocate TfLiteAffineQuantization.\n"); + return kTfLiteError; + } + + // TODO(b/153688719): Reduce tail allocation by using a global zero-point + // buffer. This value can not be reused from the flatbuffer since the + // zero_point is stored as a int64_t. + quantization->zero_point = + allocate_temp + ? reinterpret_cast(allocator->AllocateTemp( + TfLiteIntArrayGetSizeInBytes(channels), + alignof(TfLiteIntArray))) + : reinterpret_cast(allocator->AllocateFromTail( + TfLiteIntArrayGetSizeInBytes(channels), + alignof(TfLiteIntArray))); + if (quantization->zero_point == nullptr) { + TF_LITE_REPORT_ERROR(error_reporter, + "Unable to allocate quantization->zero_point.\n"); + return kTfLiteError; + } + + TF_LITE_ENSURE_STATUS(FlatBufferVectorToTfLiteTypeArray( + allocator, error_reporter, src_quantization->scale(), + &quantization->scale)); + + quantization->zero_point->size = channels; + int* zero_point_data = quantization->zero_point->data; + for (int i = 0; i < channels; i++) { + zero_point_data[i] = src_quantization->zero_point()->Get(i); + } + // TODO(rocky): Need to add a micro_allocator test case that fails when + // this is not copied: + quantization->quantized_dimension = src_quantization->quantized_dimension(); + + result->quantization = {kTfLiteAffineQuantization, quantization}; + } + return kTfLiteOk; +} + +TfLiteStatus InitializeTfLiteEvalTensorFromFlatbuffer( + SimpleMemoryAllocator* allocator, const tflite::Tensor& flatbuffer_tensor, + const flatbuffers::Vector>* buffers, + ErrorReporter* error_reporter, TfLiteEvalTensor* result) { + *result = {}; + // Make sure the serialized type is one we know how to deal with, and convert + // it from a flatbuffer enum into a constant used by the kernel C API. + TF_LITE_ENSURE_STATUS(ConvertTensorType(flatbuffer_tensor.type(), + &result->type, error_reporter)); + + result->data.data = GetFlatbufferTensorBuffer(flatbuffer_tensor, buffers); + + if (flatbuffer_tensor.shape() == nullptr) { + // flatbuffer_tensor.shape() can return a nullptr in the case of a scalar + // tensor. + result->dims = const_cast(&kZeroLengthIntArray); + } else { + TF_LITE_ENSURE_STATUS(FlatBufferVectorToTfLiteTypeArray( + allocator, error_reporter, flatbuffer_tensor.shape(), &(result->dims))); + } + return kTfLiteOk; +} + +} // namespace internal + +MicroAllocator::MicroAllocator(SimpleMemoryAllocator* memory_allocator, + ErrorReporter* error_reporter) + : memory_allocator_(memory_allocator), + error_reporter_(error_reporter), + model_is_allocating_(false) {} + +MicroAllocator::~MicroAllocator() {} + +MicroAllocator* MicroAllocator::Create(uint8_t* tensor_arena, size_t arena_size, + ErrorReporter* error_reporter) { + uint8_t* aligned_arena = AlignPointerUp(tensor_arena, kBufferAlignment); + size_t aligned_arena_size = tensor_arena + arena_size - aligned_arena; + return Create(SimpleMemoryAllocator::Create(error_reporter, aligned_arena, + aligned_arena_size), + error_reporter); +} + +MicroAllocator* MicroAllocator::Create(SimpleMemoryAllocator* memory_allocator, + ErrorReporter* error_reporter) { + TFLITE_DCHECK(memory_allocator != nullptr); + TFLITE_DCHECK(error_reporter != nullptr); + + uint8_t* allocator_buffer = memory_allocator->AllocateFromTail( + sizeof(MicroAllocator), alignof(MicroAllocator)); + MicroAllocator* allocator = + new (allocator_buffer) MicroAllocator(memory_allocator, error_reporter); + return allocator; +} + +TfLiteStatus MicroAllocator::StartModelAllocation( + const Model* model, const MicroOpResolver& op_resolver, + NodeAndRegistration** node_and_registrations, + TfLiteEvalTensor** eval_tensors) { + TFLITE_DCHECK(model != nullptr); + + if (model_is_allocating_) { + TF_LITE_REPORT_ERROR(error_reporter_, + "MicroAllocator: Model allocation started before " + "finishing previously allocated model"); + return kTfLiteError; + } + + model_is_allocating_ = true; + + TF_LITE_ENSURE_STATUS(InitScratchBufferData()); + TF_LITE_ENSURE_STATUS(AllocateTfLiteEvalTensors(model, eval_tensors)); + TF_LITE_ENSURE_STATUS( + AllocateNodeAndRegistrations(model, node_and_registrations)); + TF_LITE_ENSURE_STATUS(PrepareNodeAndRegistrationDataFromFlatbuffer( + model, op_resolver, *node_and_registrations)); + + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::FinishModelAllocation( + const Model* model, TfLiteEvalTensor* eval_tensors, + ScratchBufferHandle** scratch_buffer_handles) { + if (!model_is_allocating_) { + TF_LITE_REPORT_ERROR(error_reporter_, + "MicroAllocator: Model allocation finished before " + "starting allocating model"); + return kTfLiteError; + } + + const SubGraph* subgraph = GetSubGraphFromModel(model); + TFLITE_DCHECK(subgraph != nullptr); + + TF_LITE_ENSURE_STATUS(AllocateScratchBufferHandles( + scratch_buffer_handles, scratch_buffer_request_count_)); + TF_LITE_ENSURE_STATUS(CommitStaticMemoryPlan(model, subgraph, eval_tensors, + *scratch_buffer_handles)); + TF_LITE_ENSURE_STATUS(AllocateVariables(subgraph, eval_tensors)); + + model_is_allocating_ = false; + return kTfLiteOk; +} + +void* MicroAllocator::AllocatePersistentBuffer(size_t bytes) { + return memory_allocator_->AllocateFromTail(bytes, kBufferAlignment); +} + +TfLiteStatus MicroAllocator::RequestScratchBufferInArena(size_t bytes, + int* buffer_idx) { + // All scratch buffer requests are stored in the head section of the arena + // when a model is in the prepare phase. First align a scratch buffer request + // pointer to the start of the head: + internal::ScratchBufferRequest* requests = GetScratchBufferRequests(); + + // Count the number of requested scratch buffers for the current node: + size_t current_node_request_count = 0; + for (size_t i = 0; i < scratch_buffer_request_count_; ++i) { + if (requests[i].node_idx == kUnassignedScratchBufferRequestIndex) { + ++current_node_request_count; + } + } + + // First, ensure that the per-kernel request has not exceeded the limit: + if (current_node_request_count >= kMaxScratchBuffersPerOp) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Scratch buffer request exeeds limit per operator (%d)", + kMaxScratchBuffersPerOp); + return kTfLiteError; + } + + // Initialize and assign values for the request at the current index: + internal::ScratchBufferRequest* current_request = + &requests[scratch_buffer_request_count_]; + *current_request = {}; + // Assign -1 as a sentinel value that will be updated when the node finishes + // allocating: + current_request->bytes = bytes; + current_request->node_idx = kUnassignedScratchBufferRequestIndex; + + // Assign the current request index to the out-param: + *buffer_idx = scratch_buffer_request_count_; + + // Bump the request count to prepare for the next request: + ++scratch_buffer_request_count_; + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::FinishPrepareNodeAllocations(int node_id) { + // When a node has finished preparing, all temp allocations performed by the + // kernel should be cleaned up: + ResetTempAllocations(); + + // Find and update any new scratch buffer requests for the current node: + internal::ScratchBufferRequest* requests = GetScratchBufferRequests(); + + for (size_t i = 0; i < scratch_buffer_request_count_; ++i) { + // A request with a node_idx of -1 is a sentinel value used to indicate this + // was a new request for the current node. The allocator finally knows the + // node index at this point. Assign the value and update the list of new + // requests so the head section can be adjusted to allow for the next kernel + // to allocate at most kMaxScratchBuffersPerOp requests: + if (requests[i].node_idx == kUnassignedScratchBufferRequestIndex) { + requests[i].node_idx = node_id; + } + } + + // Ensure that the head is re-adjusted to allow for another at-most + // kMaxScratchBuffersPerOp scratch buffer requests in the next operator: + TF_LITE_ENSURE_STATUS(memory_allocator_->SetHeadBufferSize( + sizeof(internal::ScratchBufferRequest) * + (scratch_buffer_request_count_ + kMaxScratchBuffersPerOp), + alignof(internal::ScratchBufferRequest))); + + return kTfLiteOk; +} + +size_t MicroAllocator::used_bytes() const { + return memory_allocator_->GetUsedBytes(); +} + +TfLiteStatus MicroAllocator::AllocateNodeAndRegistrations( + const Model* model, NodeAndRegistration** node_and_registrations) { + TFLITE_DCHECK(node_and_registrations); + + const SubGraph* subgraph = GetSubGraphFromModel(model); + TFLITE_DCHECK(subgraph != nullptr); + + NodeAndRegistration* output = reinterpret_cast( + memory_allocator_->AllocateFromTail( + sizeof(NodeAndRegistration) * subgraph->operators()->size(), + alignof(NodeAndRegistration))); + if (output == nullptr) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Failed to allocate memory for node_and_registrations."); + return kTfLiteError; + } + *node_and_registrations = output; + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::PrepareNodeAndRegistrationDataFromFlatbuffer( + const Model* model, const MicroOpResolver& op_resolver, + NodeAndRegistration* node_and_registrations) { + TFLITE_DCHECK(model != nullptr); + TFLITE_DCHECK(node_and_registrations != nullptr); + + const SubGraph* subgraph = GetSubGraphFromModel(model); + TFLITE_DCHECK(subgraph != nullptr); + + TfLiteStatus status = kTfLiteOk; + auto* opcodes = model->operator_codes(); + MicroBuiltinDataAllocator builtin_data_allocator(memory_allocator_); + for (size_t i = 0; i < subgraph->operators()->size(); ++i) { + const auto* op = subgraph->operators()->Get(i); + const size_t index = op->opcode_index(); + if (index >= opcodes->size()) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Missing registration for opcode_index %d\n", index); + return kTfLiteError; + } + auto* opcode = (*opcodes)[index]; + status = + GetRegistrationFromOpCode(opcode, op_resolver, error_reporter_, + &(node_and_registrations[i].registration)); + if (status != kTfLiteOk) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Failed to get registration from op code %s\n ", + EnumNameBuiltinOperator(GetBuiltinCode(opcode))); + return status; + } + const auto* registration = node_and_registrations[i].registration; + if (registration == nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, "Skipping op for opcode_index %d\n", + index); + return kTfLiteError; + } + BuiltinOperator op_type = + static_cast(registration->builtin_code); + + const char* custom_data = nullptr; + size_t custom_data_size = 0; + unsigned char* builtin_data = nullptr; + + if (op_type == BuiltinOperator_CUSTOM) { + // Custom Ops may or may not have a non-null custom_options field. + if (op->custom_options() != nullptr) { + custom_data = + reinterpret_cast(op->custom_options()->data()); + custom_data_size = op->custom_options()->size(); + } + } else { + if (op->custom_options() != nullptr) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Unsupported behavior: found builtin operator %s with custom " + "options.\n", + EnumNameBuiltinOperator(op_type)); + return kTfLiteError; + } + + MicroOpResolver::BuiltinParseFunction parser = + op_resolver.GetOpDataParser(op_type); + if (parser == nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, "Did not find a parser for %s", + EnumNameBuiltinOperator(op_type)); + + return kTfLiteError; + } + TF_LITE_ENSURE_STATUS(parser(op, error_reporter_, &builtin_data_allocator, + (void**)(&builtin_data))); + } + + TfLiteIntArray* inputs_array; + TF_LITE_ENSURE_STATUS(internal::FlatBufferVectorToTfLiteTypeArray( + memory_allocator_, error_reporter_, op->inputs(), &inputs_array)); + + TfLiteIntArray* outputs_array; + TF_LITE_ENSURE_STATUS(internal::FlatBufferVectorToTfLiteTypeArray( + memory_allocator_, error_reporter_, op->outputs(), &outputs_array)); + + TfLiteNode* node = &(node_and_registrations[i].node); + *node = {}; + node->inputs = inputs_array; + node->outputs = outputs_array; + node->builtin_data = reinterpret_cast(builtin_data); + node->custom_initial_data = custom_data; + node->custom_initial_data_size = custom_data_size; + } + + return kTfLiteOk; +} + +TfLiteTensor* MicroAllocator::AllocatePersistentTfLiteTensor( + const Model* model, TfLiteEvalTensor* eval_tensors, int tensor_index) { + const SubGraph* subgraph = GetSubGraphFromModel(model); + TFLITE_DCHECK(subgraph != nullptr); + + // This value is allocated from persistent arena space. It is guaranteed to be + // around for the lifetime of the application. + TfLiteTensor* tensor = + AllocatePersistentTfLiteTensorInternal(model, eval_tensors, tensor_index); + + // Populate any fields from the flatbuffer, since this TfLiteTensor struct is + // allocated in the persistent section of the arena, ensure that additional + // allocations also take place in that section of the arena. + if (PopulateTfLiteTensorFromFlatbuffer(model, subgraph, tensor, tensor_index, + /*allocate_temp=*/false) != + kTfLiteOk) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Failed to populate a persistent TfLiteTensor struct " + "from flatbuffer data!"); + return nullptr; + } + + if (eval_tensors != nullptr) { + // Tensor buffers that are allocated at runtime (e.g. non-weight buffers) + // and not located in the flatbuffer are stored on the pre-allocated list of + // TfLiteEvalTensors structs. These structs are the source of truth, simply + // point the corresponding buffer to the new TfLiteTensor data value. + tensor->data.data = eval_tensors[tensor_index].data.data; + } + return tensor; +} + +TfLiteTensor* MicroAllocator::AllocateTempTfLiteTensor( + const Model* model, TfLiteEvalTensor* eval_tensors, int tensor_index) { + const SubGraph* subgraph = GetSubGraphFromModel(model); + TFLITE_DCHECK(subgraph != nullptr); + + // This value is allocated from temporary arena space. It is guaranteed to be + // around for at least the scope of the calling function. Since this struct + // allocation takes place in temp space, no need to own or cleanup. + TfLiteTensor* tensor = + reinterpret_cast(memory_allocator_->AllocateTemp( + sizeof(TfLiteTensor), alignof(TfLiteTensor))); + + // Populate any fields from the flatbuffer, since this TfLiteTensor struct is + // allocated in the temp section of the arena, ensure that additional + // allocations also take place in that section of the arena. + if (PopulateTfLiteTensorFromFlatbuffer(model, subgraph, tensor, tensor_index, + /*allocate_temp=*/true) != kTfLiteOk) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Failed to populate a temp TfLiteTensor struct from flatbuffer data!"); + return nullptr; + } + + if (eval_tensors != nullptr) { + // Tensor buffers that are allocated at runtime (e.g. non-weight buffers) + // and not located in the flatbuffer are stored on the pre-allocated list of + // TfLiteEvalTensors structs. These structs are the source of truth, simply + // point the corresponding buffer to the new TfLiteTensor data value. + tensor->data.data = eval_tensors[tensor_index].data.data; + } + return tensor; +} + +void MicroAllocator::ResetTempAllocations() { + memory_allocator_->ResetTempAllocations(); +} + +TfLiteStatus MicroAllocator::AllocateTfLiteEvalTensors( + const Model* model, TfLiteEvalTensor** eval_tensors) { + TFLITE_DCHECK(eval_tensors != nullptr); + + const SubGraph* subgraph = GetSubGraphFromModel(model); + TFLITE_DCHECK(subgraph != nullptr); + + size_t alloc_count = subgraph->tensors()->size(); + TfLiteEvalTensor* tensors = + reinterpret_cast(memory_allocator_->AllocateFromTail( + sizeof(TfLiteEvalTensor) * alloc_count, alignof(TfLiteEvalTensor))); + if (tensors == nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Failed to allocate memory for context->eval_tensors, " + "%d bytes required", + sizeof(TfLiteEvalTensor) * alloc_count); + return kTfLiteError; + } + + for (size_t i = 0; i < alloc_count; ++i) { + TfLiteStatus status = internal::InitializeTfLiteEvalTensorFromFlatbuffer( + memory_allocator_, *subgraph->tensors()->Get(i), model->buffers(), + error_reporter_, &tensors[i]); + if (status != kTfLiteOk) { + TF_LITE_REPORT_ERROR(error_reporter_, "Failed to initialize tensor %d", + i); + return kTfLiteError; + } + } + *eval_tensors = tensors; + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::AllocateVariables(const SubGraph* subgraph, + TfLiteEvalTensor* eval_tensors) { + for (size_t i = 0; i < subgraph->tensors()->size(); ++i) { + auto* tensor = subgraph->tensors()->Get(i); + if (tensor->is_variable()) { + size_t buffer_size; + TF_LITE_ENSURE_STATUS( + TfLiteEvalTensorByteLength(&eval_tensors[i], &buffer_size)); + + eval_tensors[i].data.data = + memory_allocator_->AllocateFromTail(buffer_size, kBufferAlignment); + + if (eval_tensors[i].data.data == nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Failed to allocate variable tensor of size %d", + buffer_size); + return kTfLiteError; + } + } + } + return kTfLiteOk; +} + +TfLiteTensor* MicroAllocator::AllocatePersistentTfLiteTensorInternal( + const Model* model, TfLiteEvalTensor* eval_tensors, int tensor_index) { + return reinterpret_cast(memory_allocator_->AllocateFromTail( + sizeof(TfLiteTensor), alignof(TfLiteTensor))); +} + +TfLiteStatus MicroAllocator::PopulateTfLiteTensorFromFlatbuffer( + const Model* model, const SubGraph* subgraph, TfLiteTensor* tensor, + int tensor_index, bool allocate_temp) { + // TODO(b/162311891): This method serves as a stub to ensure quantized + // allocations in the tail can be recorded. Once the interpreter has APIs for + // accessing buffers on TfLiteEvalTensor this method can be dropped. + return internal::InitializeTfLiteTensorFromFlatbuffer( + memory_allocator_, allocate_temp, *subgraph->tensors()->Get(tensor_index), + model->buffers(), error_reporter_, tensor); +} + +ErrorReporter* MicroAllocator::error_reporter() const { + return error_reporter_; +} + +const SubGraph* MicroAllocator::GetSubGraphFromModel(const Model* model) { + auto* subgraphs = model->subgraphs(); + if (subgraphs->size() != 1) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Only 1 subgraph is currently supported.\n"); + return nullptr; + } + return (*subgraphs)[0]; +} + +TfLiteStatus MicroAllocator::CommitStaticMemoryPlan( + const Model* model, const SubGraph* subgraph, + TfLiteEvalTensor* eval_tensors, + ScratchBufferHandle* scratch_buffer_handles) { + size_t head_usage = 0; + // Create static memory plan + // 1. Calculate AllocationInfo to know the lifetime of each tensor/buffer. + // 2. Add them into the planner (such as the GreedyMemoryPlanner). + // 3. Static memory planning using the planner. + // 4. Set tensor/buffer pointers based on the offsets from the previous step. + // + // Note that AllocationInfo is only needed for creating the plan. It will be + // allocated from the temp section and cleaned up at the bottom of this + // function. + + size_t allocation_info_count = + subgraph->tensors()->size() + scratch_buffer_request_count_; + size_t bytes = sizeof(AllocationInfo) * allocation_info_count; + + // Allocate an array of AllocationInfo structs from the temp section. This + // struct will be used by AllocationInfoBuilder to find buffer usage. + AllocationInfo* allocation_info = reinterpret_cast( + memory_allocator_->AllocateTemp(bytes, alignof(AllocationInfo))); + if (allocation_info == nullptr) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Failed to allocate memory for allocation_info, %d bytes required", + bytes); + return kTfLiteError; + } + + // Use the AllocationInfoBuilder class to help determine where buffers are + // used in the subgraph. + AllocationInfoBuilder builder(allocation_info, subgraph->tensors()->size(), + scratch_buffer_request_count_, error_reporter_); + + const int32_t* offline_planner_offsets = nullptr; + TF_LITE_ENSURE_STATUS( + builder.GetOfflinePlannedOffsets(model, &offline_planner_offsets)); + TF_LITE_ENSURE_STATUS( + builder.AddTensors(subgraph, offline_planner_offsets, eval_tensors)); + + internal::ScratchBufferRequest* scratch_buffer_requests = + GetScratchBufferRequests(); + + TF_LITE_ENSURE_STATUS(builder.AddScratchBuffers(scratch_buffer_requests, + scratch_buffer_handles)); + + // Remaining arena size that memory planner can use for calculating offsets. + size_t remaining_arena_size = + memory_allocator_->GetAvailableMemory(kBufferAlignment); + uint8_t* planner_arena = + memory_allocator_->AllocateTemp(remaining_arena_size, kBufferAlignment); + TF_LITE_ENSURE(error_reporter_, planner_arena != nullptr); + GreedyMemoryPlanner planner(planner_arena, remaining_arena_size); + TF_LITE_ENSURE_STATUS(CreatePlan(error_reporter_, &planner, allocation_info, + allocation_info_count)); + + // Reset all temp allocations used above: + memory_allocator_->ResetTempAllocations(); + + size_t actual_available_arena_size = + memory_allocator_->GetAvailableMemory(kBufferAlignment); + + // Make sure we have enough arena size. + if (planner.GetMaximumMemorySize() > actual_available_arena_size) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Arena size is too small for all buffers. Needed %u but only " + "%u was available.", + planner.GetMaximumMemorySize(), actual_available_arena_size); + return kTfLiteError; + } + // Commit the plan. + TF_LITE_ENSURE_STATUS(CommitPlan(error_reporter_, &planner, + memory_allocator_->GetHeadBuffer(), + allocation_info, allocation_info_count)); + head_usage = planner.GetMaximumMemorySize(); + + // The head is used to store memory plans for one model at a time during the + // model preparation stage, and is re-purposed to store scratch buffer handles + // during model invocation. The head must be as large as the greater of the + // largest model memory plan's size and the total space required for all + // scratch buffer handles. + if (max_head_buffer_usage_ < head_usage) { + max_head_buffer_usage_ = head_usage; + } + + // The head is used for storing scratch buffer allocations before finalizing a + // memory plan in this function. Ensure that the head is set to the largest + // memory plan sent through the allocator: + TF_LITE_ENSURE_STATUS(memory_allocator_->SetHeadBufferSize( + max_head_buffer_usage_, kBufferAlignment)); + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::AllocateScratchBufferHandles( + ScratchBufferHandle** scratch_buffer_handles, size_t handle_count) { + TFLITE_DCHECK(scratch_buffer_handles != nullptr); + + if (scratch_buffer_request_count_ == 0) { + // No scratch buffer requests were requested during model allocation. + return kTfLiteOk; + } + + // Allocate a consecutive block of memory store the scratch buffer handles. + // This alignment ensures quick lookup during inference time for the model: + *scratch_buffer_handles = reinterpret_cast( + memory_allocator_->AllocateFromTail( + sizeof(ScratchBufferHandle) * handle_count, + alignof(ScratchBufferHandle))); + + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::InitScratchBufferData() { + // A model is preparing to allocate resources, ensure that scratch buffer + // request counter is cleared: + scratch_buffer_request_count_ = 0; + + // All requests will be stored in the head section. Each kernel is allowed at + // most kMaxScratchBuffersPerOp requests. Adjust the head to reserve at most + // that many requests to begin: + TF_LITE_ENSURE_STATUS(memory_allocator_->SetHeadBufferSize( + sizeof(internal::ScratchBufferRequest) * kMaxScratchBuffersPerOp, + alignof(internal::ScratchBufferRequest))); + + return kTfLiteOk; +} + +internal::ScratchBufferRequest* MicroAllocator::GetScratchBufferRequests() { + return reinterpret_cast( + AlignPointerUp(memory_allocator_->GetHeadBuffer(), + alignof(internal::ScratchBufferRequest))); +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.h new file mode 100644 index 0000000..44910f0 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_allocator.h @@ -0,0 +1,282 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_MICRO_ALLOCATOR_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_ALLOCATOR_H_ + +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h" +#include "tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +namespace internal { + +// Sets up all of the data structure members for a TfLiteTensor based on the +// contents of a serialized tensor in the flatbuffer. +// TODO(b/162311891): Drop this method when the interpreter has an API for +// returning buffers on TfLiteEvalTensor. +TfLiteStatus InitializeTfLiteTensorFromFlatbuffer( + SimpleMemoryAllocator* allocator, bool allocate_temp, + const tflite::Tensor& flatbuffer_tensor, + const flatbuffers::Vector>* buffers, + ErrorReporter* error_reporter, TfLiteTensor* result); + +// Holds placeholder information for a scratch buffer request from a kernel. +// This struct is only used during the model prepare stage. Each request from a +// kernel is stored in the head section. During the prepare stage, the head +// section will at least hold kMaxScratchBuffersPerOp number of requests plus +// any requests from previous kernel requests. +// +// When the memory plan is finalized, these structs are no longer used in favor +// of a sequential, array of ScratchBufferHandle allocations in the tail +// section. These allocations are indexed by the request API defined in the +// TfLiteContext struct. +typedef struct { + // Number of bytes required by the buffer. The actual allocated size might be + // greater than `bytes` due to buffer alignment. + size_t bytes; + // Node where the buffer is allocated for. This provides useful information to + // determine the lifetime of the buffer. In AllocationInfo, this buffer will + // have `before` = node_idx and `after` = node_idx. + int node_idx; +} ScratchBufferRequest; + +} // namespace internal + +typedef struct { + TfLiteNode node; + const TfLiteRegistration* registration; +} NodeAndRegistration; + +// Holds a pointer to a buffer for a scratch buffer requested by a kernel during +// the model prepare stage. This struct is allocated in-place and allows for +// quick pointer-indexed lookup for speed during model inference. +typedef struct { + // Pointer to location of the scratch buffer: + uint8_t* data; +} ScratchBufferHandle; + +// Allocator responsible for allocating memory for all intermediate tensors +// necessary to invoke a model. +// +// The lifetime of the model, tensor arena and error reporter must be at +// least as long as that of the allocator object, since the allocator needs +// them to be accessible during its entire lifetime. +// +// The MicroAllocator simply plans out additional allocations that are required +// to standup a model for inference in TF Micro. This class currently relies on +// an additional allocator - SimpleMemoryAllocator - for all allocations from an +// arena. These allocations are divided into head (non-persistent) and tail +// (persistent) regions: +// +// Memory layout to help understand how it works +// This information could change in the future version. +// ************** .memory_allocator->GetBuffer() +// Tensors/Scratch buffers (head) +// ************** .head_watermark +// unused memory +// ************** .memory_allocator->GetBuffer() + ->GetMaxBufferSize() +// - ->GetDataSize() +// persistent area (tail) +// ************** .memory_allocator->GetBuffer() + ->GetMaxBufferSize() +class MicroAllocator { + public: + // Creates a MicroAllocator instance from a given tensor arena. This arena + // will be managed by the created instance. + // Note: Please use __declspec(align(16)) to make sure tensor_arena is 16 + // bytes aligned, otherwise some head room will be wasted. + // TODO(b/157615197): Cleanup constructor + factory usage. + static MicroAllocator* Create(uint8_t* tensor_arena, size_t arena_size, + ErrorReporter* error_reporter); + + // Creates a MicroAllocator instance using the provided SimpleMemoryAllocator + // intance. This allocator instance will use the SimpleMemoryAllocator + // instance to manage allocations internally. + static MicroAllocator* Create(SimpleMemoryAllocator* memory_allocator, + ErrorReporter* error_reporter); + + // Begin allocating internal resources required for model inference. + // This method will run through the flatbuffer data supplied in the model to + // properly allocate tensor, node, and op registration data. This method is + // expected to be followed with a call to FinishModelAllocation() before + // resuming allocation with another model. All persistent tensor buffers are + // stored in the out-param eval_tensors. This value is allocated from the + // persistent memory arena and will be used to host runtime tensor buffers. + TfLiteStatus StartModelAllocation( + const Model* model, const MicroOpResolver& op_resolver, + NodeAndRegistration** node_and_registrations, + TfLiteEvalTensor** eval_tensors); + + // Finish allocating internal resources required for model inference. + // This method will plan non-persistent buffers and commit a memory plan to + // the 'head' section of the memory arena. All variable tensor data will also + // be allocated. This method should be called after assigning model resources + // in StartModelAllocation(). The eval_tensors pointer should be the value + // passed into this class during StartModelAllocation(). Scratch buffer + // handles are stored in the out-param `scratch_buffer_handles`. This value + // will be used in `GetScratchBuffer` call to retrieve scratch buffers. + TfLiteStatus FinishModelAllocation( + const Model* model, TfLiteEvalTensor* eval_tensors, + ScratchBufferHandle** scratch_buffer_handles); + + // Allocates a TfLiteTensor struct and populates the returned value with + // properties from the model flatbuffer. This struct is allocated from + // persistent arena memory is only guaranteed for the lifetime of the + // application. The eval_tensors pointer should be the value passed into this + // class during StartModelAllocation() and contains the source-of-truth for + // buffers. + virtual TfLiteTensor* AllocatePersistentTfLiteTensor( + const Model* model, TfLiteEvalTensor* eval_tensors, int tensor_index); + + // Allocates a TfLiteTensor struct and populates the returned value with + // properties from the model flatbuffer. This struct is allocated from + // temporary arena memory is only guaranteed until a call is made to + // ResetTempAllocations(). The eval_tensors pointer should be the value passed + // into this class during StartModelAllocation() and contains the + // source-of-truth for buffers. + virtual TfLiteTensor* AllocateTempTfLiteTensor(const Model* model, + TfLiteEvalTensor* eval_tensors, + int tensor_index); + + // Resets all temporary allocations. This method should be called after a + // chain of temp allocations (e.g. chain of TfLiteTensor objects via + // AllocateTfLiteTensor()). + virtual void ResetTempAllocations(); + + // Allocates persistent buffer which has the same life time as the allocator. + // The memory is immediately available and is allocated from the tail of the + // arena. + virtual void* AllocatePersistentBuffer(size_t bytes); + + // Register a scratch buffer of size `bytes` for Node with `node_id`. + // This method only requests a buffer with a given size to be used after a + // model has finished allocation via FinishModelAllocation(). All requested + // buffers will be accessible by the out-param in that method. + TfLiteStatus RequestScratchBufferInArena(size_t bytes, int* buffer_idx); + + // Finish allocating a specific NodeAndRegistration prepare block (kernel + // entry for a model) with a given node ID. This call ensures that any scratch + // buffer requests and temporary allocations are handled and ready for the + // next node prepare block. + TfLiteStatus FinishPrepareNodeAllocations(int node_id); + + // Returns the arena usage in bytes, only available after + // `FinishModelAllocation`. Otherwise, it will return 0. + size_t used_bytes() const; + + protected: + MicroAllocator(SimpleMemoryAllocator* memory_allocator, + ErrorReporter* error_reporter); + virtual ~MicroAllocator(); + + // Allocates an array in the arena to hold pointers to the node and + // registration pointers required to represent the inference graph of the + // model. + virtual TfLiteStatus AllocateNodeAndRegistrations( + const Model* model, NodeAndRegistration** node_and_registrations); + + // Populates node and registration pointers representing the inference graph + // of the model from values inside the flatbuffer (loaded from the TfLiteModel + // instance). Persistent data (e.g. operator data) is allocated from the + // arena. + virtual TfLiteStatus PrepareNodeAndRegistrationDataFromFlatbuffer( + const Model* model, const MicroOpResolver& op_resolver, + NodeAndRegistration* node_and_registrations); + + // Allocates the list of persistent TfLiteEvalTensors that are used for the + // "eval" phase of model inference. These structs will be the source of truth + // for all tensor buffers. Allocation results are stored in the out-param + // eval_tensors. + virtual TfLiteStatus AllocateTfLiteEvalTensors( + const Model* model, TfLiteEvalTensor** eval_tensors); + + // Allocates persistent tensor buffers for variable tensors in the subgraph. + virtual TfLiteStatus AllocateVariables(const SubGraph* subgraph, + TfLiteEvalTensor* eval_tensors); + + // Allocate and return a persistent TfLiteTensor. + // TODO(b/162311891): Drop this method when the interpreter has an API for + // accessing TfLiteEvalTensor structs. + virtual TfLiteTensor* AllocatePersistentTfLiteTensorInternal( + const Model* model, TfLiteEvalTensor* eval_tensors, int tensor_index); + + // Populates a TfLiteTensor struct with data from the model flatbuffer. Any + // quantization data is allocated from either the tail (persistent) or temp + // sections of the arena based on the allocation flag. + virtual TfLiteStatus PopulateTfLiteTensorFromFlatbuffer( + const Model* model, const SubGraph* subgraph, TfLiteTensor* tensor, + int tensor_index, bool allocate_temp); + + ErrorReporter* error_reporter() const; + + // Returns the first subgraph from the model. + const SubGraph* GetSubGraphFromModel(const Model* model); + + private: + // Commits a memory plan for all non-persistent buffer allocations in the + // 'head' section of the memory arena. The eval_tensors pointer is the list of + // pre-allocated TfLiteEvalTensor structs that will point to the buffers that + // will be allocated into the head section in this function call. The + // scratch_buffer_handles pointer is the array of pre-allocated + // ScratchBufferHandle structs that will point to allocated buffers also in + // the head section. + virtual TfLiteStatus CommitStaticMemoryPlan( + const Model* model, const SubGraph* subgraph, + TfLiteEvalTensor* eval_tensors, + ScratchBufferHandle* scratch_buffer_handles); + + // Allocates an array of ScratchBufferHandle structs in the tail section for a + // given number of handles. + virtual TfLiteStatus AllocateScratchBufferHandles( + ScratchBufferHandle** scratch_buffer_handles, size_t handle_count); + + // Clears all internal scratch buffer request counts and resets the head to + // prepare for kernels to request scratch buffer data when a model is + // preparing. + TfLiteStatus InitScratchBufferData(); + + // Returns the pointer for the array of ScratchBufferRequest allocations in + // the head section. + internal::ScratchBufferRequest* GetScratchBufferRequests(); + + // A simple memory allocator that always allocate from the arena tail or head. + SimpleMemoryAllocator* memory_allocator_; + + ErrorReporter* error_reporter_; + bool model_is_allocating_; + + // Holds the number of ScratchBufferRequest instances stored in the head + // section when a model is allocating. + size_t scratch_buffer_request_count_ = 0; + + // Holds the byte length of the memory plan with the largest head usage. Used + // to ensure that multi-tenant allocations can share the head for buffers. + size_t max_head_buffer_usage_ = 0; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite +#endif // TENSORFLOW_LITE_MICRO_MICRO_ALLOCATOR_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.cpp b/src/tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.cpp new file mode 100644 index 0000000..8bc3813 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.cpp @@ -0,0 +1,44 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.h" + +#include + +#ifndef TF_LITE_STRIP_ERROR_STRINGS +#include "tensorflow_arm/tensorflow/lite/micro/debug_log.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_string.h" +#endif + +namespace tflite { + +int MicroErrorReporter::Report(const char* format, va_list args) { +#ifndef TF_LITE_STRIP_ERROR_STRINGS + // Only pulling in the implementation of this function for builds where we + // expect to make use of it to be extra cautious about not increasing the code + // size. + static constexpr int kMaxLogLen = 256; + char log_buffer[kMaxLogLen]; + MicroVsnprintf(log_buffer, kMaxLogLen, format, args); + DebugLog(log_buffer); + DebugLog("\r\n"); +#endif + return 0; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_error_reporter.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.h similarity index 80% rename from src/tensorflow/lite/micro/micro_error_reporter.h rename to src/tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.h index b3542e4..999e8aa 100644 --- a/src/tensorflow/lite/micro/micro_error_reporter.h +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,16 +16,16 @@ limitations under the License. #ifndef TENSORFLOW_LITE_MICRO_MICRO_ERROR_REPORTER_H_ #define TENSORFLOW_LITE_MICRO_MICRO_ERROR_REPORTER_H_ -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/micro/compatibility.h" -#include "tensorflow/lite/micro/debug_log.h" -#include "tensorflow/lite/micro/debug_log_numbers.h" +#include + +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" namespace tflite { class MicroErrorReporter : public ErrorReporter { public: - ~MicroErrorReporter() {} + ~MicroErrorReporter() override {} int Report(const char* format, va_list args) override; private: @@ -34,3 +35,5 @@ class MicroErrorReporter : public ErrorReporter { } // namespace tflite #endif // TENSORFLOW_LITE_MICRO_MICRO_ERROR_REPORTER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.cpp b/src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.cpp new file mode 100644 index 0000000..9b687a2 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.cpp @@ -0,0 +1,456 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/micro/micro_interpreter.h" + +#include +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/tensor_utils.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_allocator.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_profiler.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { +namespace { + +#ifndef TF_LITE_STRIP_ERROR_STRINGS +const char* OpNameFromRegistration(const TfLiteRegistration* registration) { + if (registration->builtin_code == BuiltinOperator_CUSTOM) { + return registration->custom_name; + } else { + return EnumNameBuiltinOperator(BuiltinOperator(registration->builtin_code)); + } +} +#endif // !defined(TF_LITE_STRIP_ERROR_STRINGS) + +} // namespace + +namespace internal { + +ContextHelper::ContextHelper(ErrorReporter* error_reporter, + MicroAllocator* allocator, const Model* model) + : allocator_(allocator), error_reporter_(error_reporter), model_(model) {} + +void* ContextHelper::AllocatePersistentBuffer(TfLiteContext* ctx, + size_t bytes) { + return reinterpret_cast(ctx->impl_) + ->allocator_->AllocatePersistentBuffer(bytes); +} + +TfLiteStatus ContextHelper::RequestScratchBufferInArena(TfLiteContext* ctx, + size_t bytes, + int* buffer_idx) { + ContextHelper* helper = reinterpret_cast(ctx->impl_); + return helper->allocator_->RequestScratchBufferInArena(bytes, buffer_idx); +} + +void* ContextHelper::GetScratchBuffer(TfLiteContext* ctx, int buffer_idx) { + ContextHelper* helper = reinterpret_cast(ctx->impl_); + ScratchBufferHandle* handle = helper->scratch_buffer_handles_ + buffer_idx; + return handle->data; +} + +void ContextHelper::ReportOpError(struct TfLiteContext* context, + const char* format, ...) { +#ifndef TF_LITE_STRIP_ERROR_STRINGS + ContextHelper* helper = static_cast(context->impl_); + va_list args; + va_start(args, format); + TF_LITE_REPORT_ERROR(helper->error_reporter_, format, args); + va_end(args); +#endif +} + +TfLiteTensor* ContextHelper::GetTensor(const struct TfLiteContext* context, + int tensor_idx) { + ContextHelper* helper = static_cast(context->impl_); + return helper->allocator_->AllocateTempTfLiteTensor( + helper->model_, helper->eval_tensors_, tensor_idx); +} + +TfLiteEvalTensor* ContextHelper::GetEvalTensor( + const struct TfLiteContext* context, int tensor_idx) { + ContextHelper* helper = reinterpret_cast(context->impl_); + return &helper->eval_tensors_[tensor_idx]; +} + +void ContextHelper::SetTfLiteEvalTensors(TfLiteEvalTensor* eval_tensors) { + eval_tensors_ = eval_tensors; +} + +void ContextHelper::SetScratchBufferHandles( + ScratchBufferHandle* scratch_buffer_handles) { + scratch_buffer_handles_ = scratch_buffer_handles; +} + +} // namespace internal + +MicroInterpreter::MicroInterpreter(const Model* model, + const MicroOpResolver& op_resolver, + uint8_t* tensor_arena, + size_t tensor_arena_size, + ErrorReporter* error_reporter, + tflite::Profiler* profiler) + : model_(model), + op_resolver_(op_resolver), + error_reporter_(error_reporter), + allocator_(*MicroAllocator::Create(tensor_arena, tensor_arena_size, + error_reporter)), + tensors_allocated_(false), + initialization_status_(kTfLiteError), + eval_tensors_(nullptr), + context_helper_(error_reporter_, &allocator_, model), + input_tensor_(nullptr), + output_tensor_(nullptr) { + Init(profiler); +} + +MicroInterpreter::MicroInterpreter(const Model* model, + const MicroOpResolver& op_resolver, + MicroAllocator* allocator, + ErrorReporter* error_reporter, + tflite::Profiler* profiler) + : model_(model), + op_resolver_(op_resolver), + error_reporter_(error_reporter), + allocator_(*allocator), + tensors_allocated_(false), + initialization_status_(kTfLiteError), + eval_tensors_(nullptr), + context_helper_(error_reporter_, &allocator_, model), + input_tensor_(nullptr), + output_tensor_(nullptr) { + Init(profiler); +} + +MicroInterpreter::~MicroInterpreter() { + if (node_and_registrations_ != nullptr) { + for (size_t i = 0; i < subgraph_->operators()->size(); ++i) { + TfLiteNode* node = &(node_and_registrations_[i].node); + const TfLiteRegistration* registration = + node_and_registrations_[i].registration; + // registration is allocated outside the interpreter, so double check to + // make sure it's not nullptr; + if (registration != nullptr && registration->free != nullptr) { + registration->free(&context_, node->user_data); + } + } + } +} + +void MicroInterpreter::Init(tflite::Profiler* profiler) { + const flatbuffers::Vector>* subgraphs = + model_->subgraphs(); + if (subgraphs->size() != 1) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Only 1 subgraph is currently supported.\n"); + initialization_status_ = kTfLiteError; + return; + } + subgraph_ = (*subgraphs)[0]; + + context_.impl_ = static_cast(&context_helper_); + context_.ReportError = context_helper_.ReportOpError; + context_.GetTensor = context_helper_.GetTensor; + context_.GetEvalTensor = context_helper_.GetEvalTensor; + context_.recommended_num_threads = 1; + context_.profiler = profiler; + + initialization_status_ = kTfLiteOk; +} + +void MicroInterpreter::CorrectTensorEndianness(TfLiteEvalTensor* tensorCorr) { + int32_t tensorSize = 1; + for (int d = 0; d < tensorCorr->dims->size; ++d) + tensorSize *= reinterpret_cast(tensorCorr->dims->data)[d]; + + switch (tensorCorr->type) { + case TfLiteType::kTfLiteFloat32: + CorrectTensorDataEndianness(tensorCorr->data.f, tensorSize); + break; + case TfLiteType::kTfLiteFloat16: + CorrectTensorDataEndianness(tensorCorr->data.f16, tensorSize); + break; + case TfLiteType::kTfLiteInt64: + CorrectTensorDataEndianness(tensorCorr->data.i64, tensorSize); + break; + case TfLiteType::kTfLiteUInt64: + CorrectTensorDataEndianness(tensorCorr->data.u64, tensorSize); + break; + case TfLiteType::kTfLiteInt32: + CorrectTensorDataEndianness(tensorCorr->data.i32, tensorSize); + break; + case TfLiteType::kTfLiteInt16: + CorrectTensorDataEndianness(tensorCorr->data.i16, tensorSize); + break; + case TfLiteType::kTfLiteComplex64: + CorrectTensorDataEndianness(tensorCorr->data.c64, tensorSize); + break; + case TfLiteType::kTfLiteComplex128: + CorrectTensorDataEndianness(tensorCorr->data.c128, tensorSize); + break; + default: + // Do nothing for other data types. + break; + } +} + +template +void MicroInterpreter::CorrectTensorDataEndianness(T* data, int32_t size) { + for (int32_t i = 0; i < size; ++i) { + data[i] = flatbuffers::EndianScalar(data[i]); + } +} + +TfLiteStatus MicroInterpreter::AllocateTensors() { + if (allocator_.StartModelAllocation(model_, op_resolver_, + &node_and_registrations_, + &eval_tensors_) != kTfLiteOk) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Failed starting model allocation.\n"); + initialization_status_ = kTfLiteError; + return kTfLiteError; + } + + // Update the pointer now that TfLiteEvalTensor allocation has completed on + // the context helper. + // TODO(b/16157777): This call would not be needed if ContextHelper rolled + // into the interpreter. + context_helper_.SetTfLiteEvalTensors(eval_tensors_); + context_.tensors_size = subgraph_->tensors()->size(); + + // If the system is big endian then convert weights from the flatbuffer from + // little to big endian on startup so that it does not need to be done during + // inference. + // NOTE: This requires that the flatbuffer is held in memory which can be + // modified by this process. + if (!FLATBUFFERS_LITTLEENDIAN) { + for (size_t t = 0; t < subgraph_->tensors()->size(); ++t) { + if (auto* buffer = + (*model_->buffers())[subgraph_->tensors()->Get(t)->buffer()]) { + // If we've found a buffer, does it have any data? + if (auto* array = buffer->data()) { + // If it has any data, is the data size larger than zero? + if (array->size()) { + // Update the endianness of the corresponding eval tensor since that + // struct holds the buffer used at inference time. + CorrectTensorEndianness(&eval_tensors_[t]); + } + } + } + } + } + + // Only allow AllocatePersistentBuffer in Init stage. + context_.AllocatePersistentBuffer = context_helper_.AllocatePersistentBuffer; + context_.RequestScratchBufferInArena = nullptr; + context_.GetScratchBuffer = nullptr; + + for (size_t i = 0; i < subgraph_->operators()->size(); ++i) { + auto* node = &(node_and_registrations_[i].node); + auto* registration = node_and_registrations_[i].registration; + size_t init_data_size; + const char* init_data; + if (registration->builtin_code == BuiltinOperator_CUSTOM) { + init_data = reinterpret_cast(node->custom_initial_data); + init_data_size = node->custom_initial_data_size; + } else { + init_data = reinterpret_cast(node->builtin_data); + init_data_size = 0; + } + if (registration->init) { + node->user_data = + registration->init(&context_, init_data, init_data_size); + } + } + + // Both AllocatePersistentBuffer and RequestScratchBufferInArena is + // available in Prepare stage. + context_.RequestScratchBufferInArena = + context_helper_.RequestScratchBufferInArena; + for (size_t i = 0; i < subgraph_->operators()->size(); ++i) { + auto* node = &(node_and_registrations_[i].node); + auto* registration = node_and_registrations_[i].registration; + if (registration->prepare) { + TfLiteStatus prepare_status = registration->prepare(&context_, node); + if (prepare_status != kTfLiteOk) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Node %s (number %df) failed to prepare with status %d", + OpNameFromRegistration(registration), i, prepare_status); + return kTfLiteError; + } + } + allocator_.FinishPrepareNodeAllocations(/*node_id=*/i); + } + + // Prepare is done, we're ready for Invoke. Memory allocation is no longer + // allowed. Kernels can only fetch scratch buffers via GetScratchBuffer. + context_.AllocatePersistentBuffer = nullptr; + context_.RequestScratchBufferInArena = nullptr; + context_.GetScratchBuffer = context_helper_.GetScratchBuffer; + + TF_LITE_ENSURE_OK(&context_, + allocator_.FinishModelAllocation(model_, eval_tensors_, + &scratch_buffer_handles_)); + // TODO(b/16157777): Remove this when ContextHelper is rolled into this class. + context_helper_.SetScratchBufferHandles(scratch_buffer_handles_); + + TF_LITE_ENSURE_STATUS(ResetVariableTensors()); + + tensors_allocated_ = true; + return kTfLiteOk; +} + +TfLiteStatus MicroInterpreter::Invoke() { + if (initialization_status_ != kTfLiteOk) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Invoke() called after initialization failed\n"); + return kTfLiteError; + } + + // Ensure tensors are allocated before the interpreter is invoked to avoid + // difficult to debug segfaults. + if (!tensors_allocated_) { + TF_LITE_ENSURE_OK(&context_, AllocateTensors()); + } + + for (size_t i = 0; i < subgraph_->operators()->size(); ++i) { + auto* node = &(node_and_registrations_[i].node); + auto* registration = node_and_registrations_[i].registration; + + if (registration->invoke) { + TfLiteStatus invoke_status; +#ifndef NDEBUG // Omit profiler overhead from release builds. + // The case where profiler == nullptr is handled by + // ScopedOperatorProfile. + tflite::Profiler* profiler = + reinterpret_cast(context_.profiler); + ScopedOperatorProfile scoped_profiler( + profiler, OpNameFromRegistration(registration), i); +#endif + invoke_status = registration->invoke(&context_, node); + + // All TfLiteTensor structs used in the kernel are allocated from temp + // memory in the allocator. This creates a chain of allocations in the + // temp section. The call below resets the chain of allocations to + // prepare for the next call. + allocator_.ResetTempAllocations(); + + if (invoke_status == kTfLiteError) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Node %s (number %d) failed to invoke with status %d", + OpNameFromRegistration(registration), i, invoke_status); + return kTfLiteError; + } else if (invoke_status != kTfLiteOk) { + return invoke_status; + } + } + } + return kTfLiteOk; +} + +TfLiteTensor* MicroInterpreter::input(size_t index) { + const size_t length = inputs_size(); + if (index >= length) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Input index %d out of range (length is %d)", index, + length); + return nullptr; + } + if (index != 0) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Input tensors not at index 0 are allocated from the " + "persistent memory arena. Repeat calls will cause excess " + "allocation!"); + return allocator_.AllocatePersistentTfLiteTensor(model_, eval_tensors_, + inputs().Get(index)); + } + if (input_tensor_ == nullptr) { + input_tensor_ = allocator_.AllocatePersistentTfLiteTensor( + model_, eval_tensors_, inputs().Get(index)); + } + return input_tensor_; +} + +TfLiteTensor* MicroInterpreter::output(size_t index) { + const size_t length = outputs_size(); + if (index >= length) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Output index %d out of range (length is %d)", index, + length); + return nullptr; + } + if (index != 0) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Output tensors not at index 0 are allocated from the " + "persistent memory arena. Repeat calls will cause excess " + "allocation!"); + return allocator_.AllocatePersistentTfLiteTensor(model_, eval_tensors_, + outputs().Get(index)); + } + if (output_tensor_ == nullptr) { + // TODO(b/162311891): Drop these allocations when the interpreter supports + // handling buffers from TfLiteEvalTensor. + output_tensor_ = allocator_.AllocatePersistentTfLiteTensor( + model_, eval_tensors_, outputs().Get(index)); + } + return output_tensor_; +} + +TfLiteTensor* MicroInterpreter::tensor(size_t index) { + const size_t length = tensors_size(); + if (index >= length) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Tensor index %d out of range (length is %d)", index, + length); + return nullptr; + } + return allocator_.AllocatePersistentTfLiteTensor(model_, eval_tensors_, + index); +} + +TfLiteStatus MicroInterpreter::ResetVariableTensors() { + for (size_t i = 0; i < subgraph_->tensors()->size(); ++i) { + auto* tensor = subgraph_->tensors()->Get(i); + if (tensor->is_variable()) { + size_t buffer_size; + TF_LITE_ENSURE_STATUS( + TfLiteEvalTensorByteLength(&eval_tensors_[i], &buffer_size)); + + int value = 0; + if (tensor->type() == tflite::TensorType_INT8) { + value = tensor->quantization()->zero_point()->Get(0); + } + memset(eval_tensors_[i].data.raw, value, buffer_size); + } + } + + return kTfLiteOk; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.h new file mode 100644 index 0000000..12e1593 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_interpreter.h @@ -0,0 +1,214 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_MICRO_INTERPRETER_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_INTERPRETER_H_ + +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/profiler.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_allocator.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h" +#include "tensorflow_arm/tensorflow/lite/portable_type_to_tflitetype.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +namespace internal { + +// A helper class to encapsulate the implementation of APIs in Context. +// context->impl_ points to an instance of this class. +// Check tensorflow/lite/c/common.h for detailed descriptions. +// TODO(b/16157777): Consider rolling this class into MicroInterpreter. +class ContextHelper { + public: + explicit ContextHelper(ErrorReporter* error_reporter, + MicroAllocator* allocator, const Model* model); + + // Functions that will be assigned to function pointers on TfLiteContext: + static void* AllocatePersistentBuffer(TfLiteContext* ctx, size_t bytes); + static TfLiteStatus RequestScratchBufferInArena(TfLiteContext* ctx, + size_t bytes, + int* buffer_idx); + static void* GetScratchBuffer(TfLiteContext* ctx, int buffer_idx); + static void ReportOpError(struct TfLiteContext* context, const char* format, + ...); + static TfLiteTensor* GetTensor(const struct TfLiteContext* context, + int tensor_idx); + static TfLiteEvalTensor* GetEvalTensor(const struct TfLiteContext* context, + int tensor_idx); + + // Sets the pointer to a list of TfLiteEvalTensor instances. + void SetTfLiteEvalTensors(TfLiteEvalTensor* eval_tensors); + + // Sets the pointer to a list of ScratchBufferHandle instances. + void SetScratchBufferHandles(ScratchBufferHandle* scratch_buffer_handles); + + private: + MicroAllocator* allocator_ = nullptr; + ErrorReporter* error_reporter_ = nullptr; + const Model* model_ = nullptr; + TfLiteEvalTensor* eval_tensors_ = nullptr; + ScratchBufferHandle* scratch_buffer_handles_ = nullptr; +}; + +} // namespace internal + +class MicroInterpreter { + public: + // The lifetime of the model, op resolver, tensor arena, error reporter and + // profiler must be at least as long as that of the interpreter object, since + // the interpreter may need to access them at any time. This means that you + // should usually create them with the same scope as each other, for example + // having them all allocated on the stack as local variables through a + // top-level function. The interpreter doesn't do any deallocation of any of + // the pointed-to objects, ownership remains with the caller. + MicroInterpreter(const Model* model, const MicroOpResolver& op_resolver, + uint8_t* tensor_arena, size_t tensor_arena_size, + ErrorReporter* error_reporter, + tflite::Profiler* profiler = nullptr); + + // Create an interpreter instance using an existing MicroAllocator instance. + // This constructor should be used when creating an allocator that needs to + // have allocation handled in more than one interpreter or for recording + // allocations inside the interpreter. The lifetime of the allocator must be + // as long as that of the interpreter object. + MicroInterpreter(const Model* model, const MicroOpResolver& op_resolver, + MicroAllocator* allocator, ErrorReporter* error_reporter, + tflite::Profiler* profiler = nullptr); + + ~MicroInterpreter(); + + // Runs through the model and allocates all necessary input, output and + // intermediate tensors. + TfLiteStatus AllocateTensors(); + + // In order to support partial graph runs for strided models, this can return + // values other than kTfLiteOk and kTfLiteError. + // TODO(b/149795762): Add this to the TfLiteStatus enum. + TfLiteStatus Invoke(); + + size_t tensors_size() const { return context_.tensors_size; } + TfLiteTensor* tensor(size_t tensor_index); + template + T* typed_tensor(int tensor_index) { + if (TfLiteTensor* tensor_ptr = tensor(tensor_index)) { + if (tensor_ptr->type == typeToTfLiteType()) { + return GetTensorData(tensor_ptr); + } + } + return nullptr; + } + + TfLiteTensor* input(size_t index); + size_t inputs_size() const { return subgraph_->inputs()->Length(); } + const flatbuffers::Vector& inputs() const { + return *subgraph_->inputs(); + } + TfLiteTensor* input_tensor(size_t index) { return input(index); } + template + T* typed_input_tensor(int tensor_index) { + if (TfLiteTensor* tensor_ptr = input_tensor(tensor_index)) { + if (tensor_ptr->type == typeToTfLiteType()) { + return GetTensorData(tensor_ptr); + } + } + return nullptr; + } + + TfLiteTensor* output(size_t index); + size_t outputs_size() const { return subgraph_->outputs()->Length(); } + const flatbuffers::Vector& outputs() const { + return *subgraph_->outputs(); + } + TfLiteTensor* output_tensor(size_t index) { return output(index); } + template + T* typed_output_tensor(int tensor_index) { + if (TfLiteTensor* tensor_ptr = output_tensor(tensor_index)) { + if (tensor_ptr->type == typeToTfLiteType()) { + return GetTensorData(tensor_ptr); + } + } + return nullptr; + } + + // Reset all variable tensors to the default value. + TfLiteStatus ResetVariableTensors(); + + TfLiteStatus initialization_status() const { return initialization_status_; } + + size_t operators_size() const { return subgraph_->operators()->size(); } + + // For debugging only. + const NodeAndRegistration node_and_registration(int node_index) const { + return node_and_registrations_[node_index]; + } + + // For debugging only. + // Returns the actual used arena in bytes. This method gives the optimal arena + // size. It's only available after `AllocateTensors` has been called. + // Note that normally `tensor_arena` requires 16 bytes alignment to fully + // utilize the space. If it's not the case, the optimial arena size would be + // arena_used_bytes() + 16. + size_t arena_used_bytes() const { return allocator_.used_bytes(); } + + protected: + const MicroAllocator& allocator() const { return allocator_; } + const TfLiteContext& context() const { return context_; } + + private: + // TODO(b/158263161): Consider switching to Create() function to enable better + // error reporting during initialization. + void Init(tflite::Profiler* profiler); + + void CorrectTensorEndianness(TfLiteEvalTensor* tensorCorr); + + template + void CorrectTensorDataEndianness(T* data, int32_t size); + + NodeAndRegistration* node_and_registrations_ = nullptr; + + const Model* model_; + const MicroOpResolver& op_resolver_; + ErrorReporter* error_reporter_; + TfLiteContext context_ = {}; + MicroAllocator& allocator_; + bool tensors_allocated_; + + TfLiteStatus initialization_status_; + + const SubGraph* subgraph_ = nullptr; + TfLiteEvalTensor* eval_tensors_ = nullptr; + ScratchBufferHandle* scratch_buffer_handles_ = nullptr; + + // TODO(b/16157777): Drop this reference: + internal::ContextHelper context_helper_; + + // TODO(b/162311891): Clean these pointers up when this class supports buffers + // from TfLiteEvalTensor. + TfLiteTensor* input_tensor_; + TfLiteTensor* output_tensor_; +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MICRO_INTERPRETER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_mutable_op_resolver.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_mutable_op_resolver.h new file mode 100644 index 0000000..69d2ccc --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_mutable_op_resolver.h @@ -0,0 +1,488 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/ethosu.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/fully_connected.h" +#include "tensorflow_arm/tensorflow/lite/micro/kernels/micro_ops.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { +TfLiteRegistration* Register_DETECTION_POSTPROCESS(); + +template +class MicroMutableOpResolver : public MicroOpResolver { + public: + TF_LITE_REMOVE_VIRTUAL_DELETE + + explicit MicroMutableOpResolver(ErrorReporter* error_reporter = nullptr) + : error_reporter_(error_reporter) {} + + const TfLiteRegistration* FindOp(tflite::BuiltinOperator op) const override { + if (op == BuiltinOperator_CUSTOM) return nullptr; + + for (unsigned int i = 0; i < registrations_len_; ++i) { + const TfLiteRegistration& registration = registrations_[i]; + if (registration.builtin_code == op) { + return ®istration; + } + } + return nullptr; + } + + const TfLiteRegistration* FindOp(const char* op) const override { + for (unsigned int i = 0; i < registrations_len_; ++i) { + const TfLiteRegistration& registration = registrations_[i]; + if ((registration.builtin_code == BuiltinOperator_CUSTOM) && + (strcmp(registration.custom_name, op) == 0)) { + return ®istration; + } + } + return nullptr; + } + + MicroOpResolver::BuiltinParseFunction GetOpDataParser( + BuiltinOperator op) const override { + TFLITE_DCHECK(num_buitin_ops_ <= tOpCount); + for (unsigned int i = 0; i < num_buitin_ops_; ++i) { + if (builtin_codes_[i] == op) return builtin_parsers_[i]; + } + return nullptr; + } + + // Registers a Custom Operator with the MicroOpResolver. + // + // Only the first call for a given name will be successful. i.e. if this + // function is called again for a previously added Custom Operator, the + // MicroOpResolver will be unchanged and this function will return + // kTfLiteError. + TfLiteStatus AddCustom(const char* name, TfLiteRegistration* registration) { + if (registrations_len_ >= tOpCount) { + if (error_reporter_) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Couldn't register custom op '%s', resolver size is too small (%d)", + name, tOpCount); + } + return kTfLiteError; + } + + if (FindOp(name) != nullptr) { + if (error_reporter_ != nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Calling AddCustom for the same op more than once " + "is not supported (Op: %s).", + name); + } + return kTfLiteError; + } + + TfLiteRegistration* new_registration = ®istrations_[registrations_len_]; + registrations_len_ += 1; + + *new_registration = *registration; + new_registration->builtin_code = BuiltinOperator_CUSTOM; + new_registration->custom_name = name; + return kTfLiteOk; + } + + // The Add* functions below add the various Builtin operators to the + // MicroMutableOpResolver object. + + TfLiteStatus AddAbs() { + return AddBuiltin(BuiltinOperator_ABS, tflite::ops::micro::Register_ABS(), + ParseAbs); + } + + TfLiteStatus AddAdd() { + return AddBuiltin(BuiltinOperator_ADD, tflite::ops::micro::Register_ADD(), + ParseAdd); + } + + TfLiteStatus AddArgMax() { + return AddBuiltin(BuiltinOperator_ARG_MAX, + tflite::ops::micro::Register_ARG_MAX(), ParseArgMax); + } + + TfLiteStatus AddArgMin() { + return AddBuiltin(BuiltinOperator_ARG_MIN, + tflite::ops::micro::Register_ARG_MIN(), ParseArgMin); + } + + TfLiteStatus AddAveragePool2D() { + return AddBuiltin(BuiltinOperator_AVERAGE_POOL_2D, + tflite::ops::micro::Register_AVERAGE_POOL_2D(), + ParsePool); + } + + TfLiteStatus AddCeil() { + return AddBuiltin(BuiltinOperator_CEIL, tflite::ops::micro::Register_CEIL(), + ParseCeil); + } + + TfLiteStatus AddCircularBuffer() { + return AddCustom("CIRCULAR_BUFFER", + tflite::ops::micro::Register_CIRCULAR_BUFFER()); + } + + TfLiteStatus AddConcatenation() { + return AddBuiltin(BuiltinOperator_CONCATENATION, + tflite::ops::micro::Register_CONCATENATION(), + ParseConcatenation); + } + + TfLiteStatus AddConv2D() { + return AddBuiltin(BuiltinOperator_CONV_2D, Register_CONV_2D(), ParseConv2D); + } + + TfLiteStatus AddCos() { + return AddBuiltin(BuiltinOperator_COS, tflite::ops::micro::Register_COS(), + ParseCos); + } + + TfLiteStatus AddDepthwiseConv2D() { + return AddBuiltin(BuiltinOperator_DEPTHWISE_CONV_2D, + Register_DEPTHWISE_CONV_2D(), ParseDepthwiseConv2D); + } + + TfLiteStatus AddDequantize() { + return AddBuiltin(BuiltinOperator_DEQUANTIZE, + tflite::ops::micro::Register_DEQUANTIZE(), + ParseDequantize); + } + + TfLiteStatus AddDetectionPostprocess() { + return AddCustom("TFLite_Detection_PostProcess", + tflite::Register_DETECTION_POSTPROCESS()); + } + + TfLiteStatus AddEqual() { + return AddBuiltin(BuiltinOperator_EQUAL, + tflite::ops::micro::Register_EQUAL(), ParseEqual); + } + + TfLiteStatus AddEthosU() { + TfLiteRegistration* registration = tflite::Register_ETHOSU(); + if (registration) { + return AddCustom(tflite::GetString_ETHOSU(), registration); + } + return kTfLiteOk; + } + + TfLiteStatus AddFloor() { + return AddBuiltin(BuiltinOperator_FLOOR, + tflite::ops::micro::Register_FLOOR(), ParseFloor); + } + + TfLiteStatus AddFullyConnected( + const TfLiteRegistration& registration = Register_FULLY_CONNECTED()) { + return AddBuiltin(BuiltinOperator_FULLY_CONNECTED, registration, + ParseFullyConnected); + } + + TfLiteStatus AddGreater() { + return AddBuiltin(BuiltinOperator_GREATER, + tflite::ops::micro::Register_GREATER(), ParseGreater); + } + + TfLiteStatus AddGreaterEqual() { + return AddBuiltin(BuiltinOperator_GREATER_EQUAL, + tflite::ops::micro::Register_GREATER_EQUAL(), + ParseGreaterEqual); + } + + TfLiteStatus AddHardSwish() { + return AddBuiltin(BuiltinOperator_HARD_SWISH, + tflite::ops::micro::Register_HARD_SWISH(), + ParseHardSwish); + } + + TfLiteStatus AddL2Normalization() { + return AddBuiltin(BuiltinOperator_L2_NORMALIZATION, + tflite::ops::micro::Register_L2_NORMALIZATION(), + ParseL2Normalization); + } + + TfLiteStatus AddLess() { + return AddBuiltin(BuiltinOperator_LESS, tflite::ops::micro::Register_LESS(), + ParseLess); + } + + TfLiteStatus AddLessEqual() { + return AddBuiltin(BuiltinOperator_LESS_EQUAL, + tflite::ops::micro::Register_LESS_EQUAL(), + ParseLessEqual); + } + + TfLiteStatus AddLog() { + return AddBuiltin(BuiltinOperator_LOG, tflite::ops::micro::Register_LOG(), + ParseLog); + } + + TfLiteStatus AddLogicalAnd() { + return AddBuiltin(BuiltinOperator_LOGICAL_AND, + tflite::ops::micro::Register_LOGICAL_AND(), + ParseLogicalAnd); + } + + TfLiteStatus AddLogicalNot() { + return AddBuiltin(BuiltinOperator_LOGICAL_NOT, + tflite::ops::micro::Register_LOGICAL_NOT(), + ParseLogicalNot); + } + + TfLiteStatus AddLogicalOr() { + return AddBuiltin(BuiltinOperator_LOGICAL_OR, + tflite::ops::micro::Register_LOGICAL_OR(), + ParseLogicalOr); + } + + TfLiteStatus AddLogistic() { + return AddBuiltin(BuiltinOperator_LOGISTIC, + tflite::ops::micro::Register_LOGISTIC(), ParseLogistic); + } + + TfLiteStatus AddMaximum() { + return AddBuiltin(BuiltinOperator_MAXIMUM, + tflite::ops::micro::Register_MAXIMUM(), ParseMaximum); + } + + TfLiteStatus AddMaxPool2D() { + return AddBuiltin(BuiltinOperator_MAX_POOL_2D, + tflite::ops::micro::Register_MAX_POOL_2D(), ParsePool); + } + + TfLiteStatus AddMean() { + return AddBuiltin(BuiltinOperator_MEAN, tflite::ops::micro::Register_MEAN(), + ParseReducer); + } + + TfLiteStatus AddMinimum() { + return AddBuiltin(BuiltinOperator_MINIMUM, + tflite::ops::micro::Register_MINIMUM(), ParseMinimum); + } + + TfLiteStatus AddMul() { + return AddBuiltin(BuiltinOperator_MUL, tflite::ops::micro::Register_MUL(), + ParseMul); + } + + TfLiteStatus AddNeg() { + return AddBuiltin(BuiltinOperator_NEG, tflite::ops::micro::Register_NEG(), + ParseNeg); + } + + TfLiteStatus AddNotEqual() { + return AddBuiltin(BuiltinOperator_NOT_EQUAL, + tflite::ops::micro::Register_NOT_EQUAL(), ParseNotEqual); + } + + TfLiteStatus AddPack() { + return AddBuiltin(BuiltinOperator_PACK, tflite::ops::micro::Register_PACK(), + ParsePack); + } + + TfLiteStatus AddPad() { + return AddBuiltin(BuiltinOperator_PAD, tflite::ops::micro::Register_PAD(), + ParsePad); + } + + TfLiteStatus AddPadV2() { + return AddBuiltin(BuiltinOperator_PADV2, + tflite::ops::micro::Register_PADV2(), ParsePadV2); + } + + TfLiteStatus AddPrelu() { + return AddBuiltin(BuiltinOperator_PRELU, + tflite::ops::micro::Register_PRELU(), ParsePrelu); + } + + TfLiteStatus AddQuantize() { + return AddBuiltin(BuiltinOperator_QUANTIZE, Register_QUANTIZE(), + ParseQuantize); + } + + TfLiteStatus AddReduceMax() { + return AddBuiltin(BuiltinOperator_REDUCE_MAX, + tflite::ops::micro::Register_REDUCE_MAX(), ParseReducer); + } + + TfLiteStatus AddRelu() { + return AddBuiltin(BuiltinOperator_RELU, tflite::ops::micro::Register_RELU(), + ParseRelu); + } + + TfLiteStatus AddRelu6() { + return AddBuiltin(BuiltinOperator_RELU6, + tflite::ops::micro::Register_RELU6(), ParseRelu6); + } + + TfLiteStatus AddReshape() { + return AddBuiltin(BuiltinOperator_RESHAPE, + tflite::ops::micro::Register_RESHAPE(), ParseReshape); + } + + TfLiteStatus AddResizeNearestNeighbor() { + return AddBuiltin(BuiltinOperator_RESIZE_NEAREST_NEIGHBOR, + tflite::ops::micro::Register_RESIZE_NEAREST_NEIGHBOR(), + ParseResizeNearestNeighbor); + } + + TfLiteStatus AddRound() { + return AddBuiltin(BuiltinOperator_ROUND, + tflite::ops::micro::Register_ROUND(), ParseRound); + } + + TfLiteStatus AddRsqrt() { + return AddBuiltin(BuiltinOperator_RSQRT, + tflite::ops::micro::Register_RSQRT(), ParseRsqrt); + } + + TfLiteStatus AddShape() { + return AddBuiltin(BuiltinOperator_SHAPE, Register_SHAPE(), ParseShape); + } + + TfLiteStatus AddSin() { + return AddBuiltin(BuiltinOperator_SIN, tflite::ops::micro::Register_SIN(), + ParseSin); + } + + TfLiteStatus AddSoftmax() { + return AddBuiltin(BuiltinOperator_SOFTMAX, Register_SOFTMAX(), + ParseSoftmax); + } + + TfLiteStatus AddSplit() { + return AddBuiltin(BuiltinOperator_SPLIT, + tflite::ops::micro::Register_SPLIT(), ParseSplit); + } + + TfLiteStatus AddSplitV() { + return AddBuiltin(BuiltinOperator_SPLIT_V, + tflite::ops::micro::Register_SPLIT_V(), ParseSplitV); + } + + TfLiteStatus AddSqrt() { + return AddBuiltin(BuiltinOperator_SQRT, tflite::ops::micro::Register_SQRT(), + ParseSqrt); + } + + TfLiteStatus AddSquare() { + return AddBuiltin(BuiltinOperator_SQUARE, + tflite::ops::micro::Register_SQUARE(), ParseSquare); + } + + TfLiteStatus AddStridedSlice() { + return AddBuiltin(BuiltinOperator_STRIDED_SLICE, + tflite::ops::micro::Register_STRIDED_SLICE(), + ParseStridedSlice); + } + + TfLiteStatus AddSub() { + return AddBuiltin(BuiltinOperator_SUB, tflite::ops::micro::Register_SUB(), + ParseSub); + } + + TfLiteStatus AddSvdf() { + return AddBuiltin(BuiltinOperator_SVDF, Register_SVDF(), ParseSvdf); + } + + TfLiteStatus AddTanh() { + return AddBuiltin(BuiltinOperator_TANH, tflite::ops::micro::Register_TANH(), + ParseTanh); + } + + TfLiteStatus AddUnpack() { + return AddBuiltin(BuiltinOperator_UNPACK, + tflite::ops::micro::Register_UNPACK(), ParseUnpack); + } + + unsigned int GetRegistrationLength() { return registrations_len_; } + + private: + TfLiteStatus AddBuiltin(tflite::BuiltinOperator op, + const TfLiteRegistration& registration, + MicroOpResolver::BuiltinParseFunction parser) { + if (op == BuiltinOperator_CUSTOM) { + if (error_reporter_ != nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Invalid parameter BuiltinOperator_CUSTOM to the " + "AddBuiltin function."); + } + return kTfLiteError; + } + + if (FindOp(op) != nullptr) { + if (error_reporter_ != nullptr) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Calling AddBuiltin with the same op more than " + "once is not supported (Op: #%d).", + op); + } + return kTfLiteError; + } + + if (registrations_len_ >= tOpCount) { + if (error_reporter_) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Couldn't register builtin op #%d, resolver size " + "is too small (%d).", + op, tOpCount); + } + return kTfLiteError; + } + + registrations_[registrations_len_] = registration; + // Strictly speaking, the builtin_code is not necessary for TFLM but filling + // it in regardless. + registrations_[registrations_len_].builtin_code = op; + registrations_len_++; + + builtin_codes_[num_buitin_ops_] = op; + builtin_parsers_[num_buitin_ops_] = parser; + num_buitin_ops_++; + + return kTfLiteOk; + } + + TfLiteRegistration registrations_[tOpCount]; + unsigned int registrations_len_ = 0; + + // Arrays (and counter) to store the builtin codes and their corresponding + // parse functions as these are registered with the Op Resolver. + BuiltinOperator builtin_codes_[tOpCount]; + MicroOpResolver::BuiltinParseFunction builtin_parsers_[tOpCount]; + unsigned int num_buitin_ops_ = 0; + + ErrorReporter* error_reporter_; +}; + +}; // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h new file mode 100644 index 0000000..3a415cb --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_op_resolver.h @@ -0,0 +1,76 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_MICRO_OP_RESOLVER_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_OP_RESOLVER_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_arm/tensorflow/lite/core/api/op_resolver.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +// This is an interface for the OpResolver for TFLiteMicro. The differences from +// the TFLite OpResolver base class are to: +// * explicitly remove support for Op versions +// * allow for finer grained registration of the Builtin Ops to reduce code +// size for TFLiteMicro. +// +// We need an interface class instead of directly using MicroMutableOpResolver +// because MicroMutableOpResolver is a class template with the number of +// registered Ops as the template parameter. +class MicroOpResolver : public OpResolver { + public: + typedef TfLiteStatus (*BuiltinParseFunction)(const Operator* op, + ErrorReporter* error_reporter, + BuiltinDataAllocator* allocator, + void** builtin_data); + + // Returns the Op registration struct corresponding to the enum code from the + // flatbuffer schema. Returns nullptr if the op is not found or if op == + // BuiltinOperator_CUSTOM. + virtual const TfLiteRegistration* FindOp(BuiltinOperator op) const = 0; + + // Returns the Op registration struct corresponding to the custom operator by + // name. + virtual const TfLiteRegistration* FindOp(const char* op) const = 0; + + // This implementation exists for compatibility with the OpResolver base class + // and disregards the version parameter. + const TfLiteRegistration* FindOp(BuiltinOperator op, + int version) const final { + return FindOp(op); + } + + // This implementation exists for compatibility with the OpResolver base class + // and disregards the version parameter. + const TfLiteRegistration* FindOp(const char* op, int version) const final { + return FindOp(op); + } + + // Returns the operator specific parsing function for the OpData for a + // BuiltinOperator (if registered), else nullptr. + virtual BuiltinParseFunction GetOpDataParser(BuiltinOperator op) const = 0; + + ~MicroOpResolver() override {} +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MICRO_OP_RESOLVER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.cpp b/src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.cpp new file mode 100644 index 0000000..acc2d20 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.cpp @@ -0,0 +1,45 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/micro_profiler.h" + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_time.h" + +namespace tflite { + +MicroProfiler::MicroProfiler(tflite::ErrorReporter* reporter) + : reporter_(reporter) {} + +uint32_t MicroProfiler::BeginEvent(const char* tag, EventType event_type, + int64_t event_metadata1, + int64_t event_metadata2) { + start_time_ = GetCurrentTimeTicks(); + TFLITE_DCHECK(tag != nullptr); + event_tag_ = tag; + return 0; +} + +void MicroProfiler::EndEvent(uint32_t event_handle) { +#ifndef TF_LITE_STRIP_ERROR_STRINGS + int32_t end_time = GetCurrentTimeTicks(); + TF_LITE_REPORT_ERROR(reporter_, "%s took %d cycles\n", event_tag_, + end_time - start_time_); +#endif +} +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.h new file mode 100644 index 0000000..673e3ef --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_profiler.h @@ -0,0 +1,74 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_MICRO_PROFILER_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_PROFILER_H_ + +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/core/api/profiler.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" + +namespace tflite { + +// MicroProfiler creates a common way to gain fine-grained insight into runtime +// performance. Bottleck operators can be identified along with slow code +// sections. This can be used in conjunction with running the relevant micro +// benchmark to evaluate end-to-end performance. +// +// Usage example: +// MicroProfiler profiler(error_reporter); +// { +// ScopedProfile scoped_profile(profiler, tag); +// work_to_profile(); +// } +// +// This will call the following methods in order: +// int event_handle = profiler->BeginEvent(op_name, EventType::DEFAULT, 0) +// work_to_profile(); +// profiler->EndEvent(event_handle) +class MicroProfiler : public tflite::Profiler { + public: + explicit MicroProfiler(tflite::ErrorReporter* reporter); + ~MicroProfiler() override = default; + + // AddEvent is unused for Tf Micro. + void AddEvent(const char* tag, EventType event_type, uint64_t start, + uint64_t end, int64_t event_metadata1, + int64_t event_metadata2) override{}; + + // BeginEvent followed by code followed by EndEvent will profile the code + // enclosed. Multiple concurrent events are unsupported, so the return value + // is always 0. Event_metadata1 and event_metadata2 are unused. The tag + // pointer must be valid until EndEvent is called. + uint32_t BeginEvent(const char* tag, EventType event_type, + int64_t event_metadata1, + int64_t event_metadata2) override; + + // Event_handle is ignored since TF Micro does not support concurrent events. + void EndEvent(uint32_t event_handle) override; + + private: + tflite::ErrorReporter* reporter_; + int32_t start_time_; + const char* event_tag_; + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MICRO_PROFILER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_string.cpp b/src/tensorflow_arm/tensorflow/lite/micro/micro_string.cpp new file mode 100644 index 0000000..1a8f09c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_string.cpp @@ -0,0 +1,312 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// Implements debug logging for numbers by converting them into strings and then +// calling the main DebugLog(char*) function. These are separated into a +// different file so that platforms can just implement the string output version +// of DebugLog() and then get the numerical variations without requiring any +// more code. + +#include "tensorflow_arm/tensorflow/lite/micro/micro_string.h" + +#include +#include +#include + +namespace { + +// Int formats can need up to 10 bytes for the value plus a single byte for the +// sign. +constexpr int kMaxIntCharsNeeded = 10 + 1; +// Hex formats can need up to 8 bytes for the value plus two bytes for the "0x". +constexpr int kMaxHexCharsNeeded = 8 + 2; + +// Float formats can need up to 7 bytes for the fraction plus 3 bytes for "x2^" +// plus 3 bytes for the exponent and a single sign bit. +constexpr float kMaxFloatCharsNeeded = 7 + 3 + 3 + 1; + +// All input buffers to the number conversion functions must be this long. +const int kFastToBufferSize = 48; + +// Reverses a zero-terminated string in-place. +char* ReverseStringInPlace(char* start, char* end) { + char* p1 = start; + char* p2 = end - 1; + while (p1 < p2) { + char tmp = *p1; + *p1++ = *p2; + *p2-- = tmp; + } + return start; +} + +// Appends a string to a string, in-place. You need to pass in the maximum +// string length as the second argument. +char* StrCatStr(char* main, int main_max_length, const char* to_append) { + char* current = main; + while (*current != 0) { + ++current; + } + char* current_end = main + (main_max_length - 1); + while ((*to_append != 0) && (current < current_end)) { + *current = *to_append; + ++current; + ++to_append; + } + *current = 0; + return current; +} + +// Populates the provided buffer with an ASCII representation of the number. +char* FastUInt32ToBufferLeft(uint32_t i, char* buffer, int base) { + char* start = buffer; + do { + int32_t digit = i % base; + char character; + if (digit < 10) { + character = '0' + digit; + } else { + character = 'a' + (digit - 10); + } + *buffer++ = character; + i /= base; + } while (i > 0); + *buffer = 0; + ReverseStringInPlace(start, buffer); + return buffer; +} + +// Populates the provided buffer with an ASCII representation of the number. +char* FastInt32ToBufferLeft(int32_t i, char* buffer) { + uint32_t u = i; + if (i < 0) { + *buffer++ = '-'; + u = -u; + } + return FastUInt32ToBufferLeft(u, buffer, 10); +} + +// Converts a number to a string and appends it to another. +char* StrCatInt32(char* main, int main_max_length, int32_t number) { + char number_string[kFastToBufferSize]; + FastInt32ToBufferLeft(number, number_string); + return StrCatStr(main, main_max_length, number_string); +} + +// Converts a number to a string and appends it to another. +char* StrCatUInt32(char* main, int main_max_length, uint32_t number, int base) { + char number_string[kFastToBufferSize]; + FastUInt32ToBufferLeft(number, number_string, base); + return StrCatStr(main, main_max_length, number_string); +} + +// Populates the provided buffer with ASCII representation of the float number. +// Avoids the use of any floating point instructions (since these aren't +// supported on many microcontrollers) and as a consequence prints values with +// power-of-two exponents. +char* FastFloatToBufferLeft(float f, char* buffer) { + char* current = buffer; + char* current_end = buffer + (kFastToBufferSize - 1); + // Access the bit fields of the floating point value to avoid requiring any + // float instructions. These constants are derived from IEEE 754. + const uint32_t sign_mask = 0x80000000; + const uint32_t exponent_mask = 0x7f800000; + const int32_t exponent_shift = 23; + const int32_t exponent_bias = 127; + const uint32_t fraction_mask = 0x007fffff; + uint32_t u; + memcpy(&u, &f, sizeof(int32_t)); + const int32_t exponent = + ((u & exponent_mask) >> exponent_shift) - exponent_bias; + const uint32_t fraction = (u & fraction_mask); + // Expect ~0x2B1B9D3 for fraction. + if (u & sign_mask) { + *current = '-'; + current += 1; + } + *current = 0; + // These are special cases for infinities and not-a-numbers. + if (exponent == 128) { + if (fraction == 0) { + current = StrCatStr(current, (current_end - current), "Inf"); + return current; + } else { + current = StrCatStr(current, (current_end - current), "NaN"); + return current; + } + } + // 0x007fffff (8388607) represents 0.99... for the fraction, so to print the + // correct decimal digits we need to scale our value before passing it to the + // conversion function. This scale should be 10000000/8388608 = 1.1920928955. + // We can approximate this using multiply-adds and right-shifts using the + // values in this array. The 1. portion of the number string is printed out + // in a fixed way before the fraction, below. + const int32_t scale_shifts_size = 13; + const int8_t scale_shifts[13] = {3, 4, 8, 11, 13, 14, 17, + 18, 19, 20, 21, 22, 23}; + uint32_t scaled_fraction = fraction; + for (int i = 0; i < scale_shifts_size; ++i) { + scaled_fraction += (fraction >> scale_shifts[i]); + } + *current = '1'; + current += 1; + *current = '.'; + current += 1; + *current = 0; + + // Prepend leading zeros to fill in all 7 bytes of the fraction. Truncate + // zeros off the end of the fraction. Every fractional value takes 7 bytes. + // For example, 2500 would be written into the buffer as 0002500 since it + // represents .00025. + constexpr int kMaxFractionalDigits = 7; + + // Abort early if there is not enough space in the buffer. + if (current_end - current <= kMaxFractionalDigits) { + return current; + } + + // Pre-fill buffer with zeros to ensure zero-truncation works properly. + for (int i = 1; i < kMaxFractionalDigits; i++) { + *(current + i) = '0'; + } + + // Track how large the fraction is to add leading zeros. + char* previous = current; + current = StrCatUInt32(current, (current_end - current), scaled_fraction, 10); + int fraction_digits = current - previous; + int leading_zeros = kMaxFractionalDigits - fraction_digits; + + // Overwrite the null terminator from StrCatUInt32 to ensure zero-trunctaion + // works properly. + *current = '0'; + + // Shift fraction values and prepend zeros if necessary. + if (leading_zeros != 0) { + for (int i = 0; i < fraction_digits; i++) { + current--; + *(current + leading_zeros) = *current; + *current = '0'; + } + current += kMaxFractionalDigits; + } + + // Truncate trailing zeros for cleaner logs. Ensure we leave at least one + // fractional character for the case when scaled_fraction is 0. + while (*(current - 1) == '0' && (current - 1) > previous) { + current--; + } + *current = 0; + current = StrCatStr(current, (current_end - current), "*2^"); + current = StrCatInt32(current, (current_end - current), exponent); + return current; +} + +int FormatInt32(char* output, int32_t i) { + return static_cast(FastInt32ToBufferLeft(i, output) - output); +} + +int FormatUInt32(char* output, uint32_t i) { + return static_cast(FastUInt32ToBufferLeft(i, output, 10) - output); +} + +int FormatHex(char* output, uint32_t i) { + return static_cast(FastUInt32ToBufferLeft(i, output, 16) - output); +} + +int FormatFloat(char* output, float i) { + return static_cast(FastFloatToBufferLeft(i, output) - output); +} + +} // namespace + +extern "C" int MicroVsnprintf(char* output, int len, const char* format, + va_list args) { + int output_index = 0; + const char* current = format; + // One extra character must be left for the null terminator. + const int usable_length = len - 1; + while (*current != '\0' && output_index < usable_length) { + if (*current == '%') { + current++; + switch (*current) { + case 'd': + // Cut off log message if format could exceed log buffer length. + if (usable_length - output_index < kMaxIntCharsNeeded) { + output[output_index++] = '\0'; + return output_index; + } + output_index += + FormatInt32(&output[output_index], va_arg(args, int32_t)); + current++; + break; + case 'u': + if (usable_length - output_index < kMaxIntCharsNeeded) { + output[output_index++] = '\0'; + return output_index; + } + output_index += + FormatUInt32(&output[output_index], va_arg(args, uint32_t)); + current++; + break; + case 'x': + if (usable_length - output_index < kMaxHexCharsNeeded) { + output[output_index++] = '\0'; + return output_index; + } + output[output_index++] = '0'; + output[output_index++] = 'x'; + output_index += + FormatHex(&output[output_index], va_arg(args, uint32_t)); + current++; + break; + case 'f': + if (usable_length - output_index < kMaxFloatCharsNeeded) { + output[output_index++] = '\0'; + return output_index; + } + output_index += + FormatFloat(&output[output_index], va_arg(args, double)); + current++; + break; + case '%': + output[output_index++] = *current++; + break; + case 's': + char* string = va_arg(args, char*); + int string_idx = 0; + while (string_idx + output_index < usable_length && + string[string_idx] != '\0') { + output[output_index++] = string[string_idx++]; + } + current++; + } + } else { + output[output_index++] = *current++; + } + } + output[output_index++] = '\0'; + return output_index; +} + +extern "C" int MicroSnprintf(char* output, int len, const char* format, ...) { + va_list args; + va_start(args, format); + int bytes_written = MicroVsnprintf(output, len, format, args); + va_end(args); + return bytes_written; +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_string.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_string.h new file mode 100644 index 0000000..17903cb --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_string.h @@ -0,0 +1,36 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_MICRO_STRING_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_STRING_H_ + +#include + +// Implements simple string formatting for numeric types. Returns the number of +// bytes written to output. +extern "C" { +// Functionally equivalent to vsnprintf, trimmed down for TFLite Micro. +// MicroSnprintf() is implemented using MicroVsnprintf(). +int MicroVsnprintf(char* output, int len, const char* format, va_list args); +// Functionally equavalent to snprintf, trimmed down for TFLite Micro. +// For example, MicroSnprintf(buffer, 10, "int %d", 10) will put the string +// "int 10" in the buffer. +// Floating point values are logged in exponent notation (1.XXX*2^N). +int MicroSnprintf(char* output, int len, const char* format, ...); +} + +#endif // TENSORFLOW_LITE_MICRO_MICRO_STRING_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_time.cpp b/src/tensorflow_arm/tensorflow/lite/micro/micro_time.cpp new file mode 100644 index 0000000..15633c3 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_time.cpp @@ -0,0 +1,62 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// Reference implementation of timer functions. Platforms are not required to +// implement these timer methods, but they are required to enable profiling. + +// On platforms that have a POSIX stack or C library, it can be written using +// methods from or clock() from . + +// To add an equivalent function for your own platform, create your own +// implementation file, and place it in a subfolder with named after the OS +// you're targeting. For example, see the Cortex M bare metal version in +// tensorflow/lite/micro/bluepill/micro_time.cc or the mbed one on +// tensorflow/lite/micro/mbed/micro_time.cc. + +#include "tensorflow_arm/tensorflow/lite/micro/micro_time.h" + +#if defined(TF_LITE_USE_CTIME) +#include +#endif + +namespace tflite { + +#if !defined(TF_LITE_USE_CTIME) + +// Reference implementation of the ticks_per_second() function that's required +// for a platform to support Tensorflow Lite for Microcontrollers profiling. +// This returns 0 by default because timing is an optional feature that builds +// without errors on platforms that do not need it. +int32_t ticks_per_second() { return 0; } + +// Reference implementation of the GetCurrentTimeTicks() function that's +// required for a platform to support Tensorflow Lite for Microcontrollers +// profiling. This returns 0 by default because timing is an optional feature +// that builds without errors on platforms that do not need it. +int32_t GetCurrentTimeTicks() { return 0; } + +#else // defined(TF_LITE_USE_CTIME) + +// For platforms that support ctime, we implment the micro_time interface in +// this central location. +int32_t ticks_per_second() { return CLOCKS_PER_SEC; } + +int32_t GetCurrentTimeTicks() { return clock(); } +#endif + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_time.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_time.h new file mode 100644 index 0000000..85cea8c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_time.h @@ -0,0 +1,34 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_MICRO_MICRO_TIME_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_TIME_H_ + +#include + +namespace tflite { + +// These functions should be implemented by each target platform, and provide an +// accurate tick count along with how many ticks there are per second. +int32_t ticks_per_second(); + +// Return time in ticks. The meaning of a tick varies per platform. +int32_t GetCurrentTimeTicks(); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MICRO_TIME_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_utils.cpp b/src/tensorflow_arm/tensorflow/lite/micro/micro_utils.cpp new file mode 100644 index 0000000..2fbea22 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_utils.cpp @@ -0,0 +1,83 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" + +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/op_macros.h" + +namespace tflite { + +int ElementCount(const TfLiteIntArray& dims) { + int result = 1; + for (int i = 0; i < dims.size; ++i) { + result *= dims.data[i]; + } + return result; +} + +void SignedSymmetricPerChannelQuantize(const float* values, + TfLiteIntArray* dims, + int quantized_dimension, + int8_t* quantized_values, + float* scaling_factors) { + int input_size = ElementCount(*dims); + int channel_count = dims->data[quantized_dimension]; + int per_channel_size = input_size / channel_count; + + int stride; + int channel_stride; + if (quantized_dimension == 0) { + stride = 1; + channel_stride = per_channel_size; + } else if (quantized_dimension == 3) { + stride = channel_count; + channel_stride = 1; + } else { + TF_LITE_FATAL("quantized dimension must be 0 or 3"); + } + + // Calculate scales for each channel. + for (int channel = 0; channel < channel_count; channel++) { + float min = 0; + float max = 0; + + for (int i = 0; i < per_channel_size; i++) { + int idx = channel * channel_stride + i * stride; + min = fminf(min, values[idx]); + max = fmaxf(max, values[idx]); + } + scaling_factors[channel] = + fmaxf(fabs(min), fabs(max)) / std::numeric_limits::max(); + for (int i = 0; i < per_channel_size; i++) { + int idx = channel * channel_stride + i * stride; + const int32_t quantized_value = + static_cast(roundf(values[idx] / scaling_factors[channel])); + // Clamp: just in case some odd numeric offset. + quantized_values[idx] = + fminf(std::numeric_limits::max(), + fmaxf(std::numeric_limits::min() + 1, quantized_value)); + } + } +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/micro_utils.h b/src/tensorflow_arm/tensorflow/lite/micro/micro_utils.h new file mode 100644 index 0000000..c4e2a6c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/micro_utils.h @@ -0,0 +1,137 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_MICRO_UTILS_H_ +#define TENSORFLOW_LITE_MICRO_MICRO_UTILS_H_ + +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +// Returns number of elements in the shape array. + +int ElementCount(const TfLiteIntArray& dims); + +// Converts a float value into a quantized value. Note that large values (close +// to max int and min int) may see significant error due to a lack of floating +// point granularity for large values. +template +T FloatToQuantizedType(const float value, const float scale, int zero_point) { + int32_t result = round(value / scale) + zero_point; + result = + std::max(static_cast(std::numeric_limits::min()), result); + result = + std::min(static_cast(std::numeric_limits::max()), result); + return result; +} + +template +T FloatToSymmetricQuantizedType(const float value, const float scale) { + int32_t result = round(value / scale); + result = + std::max(static_cast(std::numeric_limits::min() + 1), result); + result = + std::min(static_cast(std::numeric_limits::max()), result); + return result; +} + +// Helper methods to quantize arrays of floats to the desired format. +// +// There are several key flavors of quantization in TfLite: +// asymmetric symmetric per channel +// int8_t | X | X | X | +// uint8_t | X | X | | +// int16_t | X | | | +// int32_t | | X | X | +// +// The per-op quantization spec can be found here: +// https://www.tensorflow.org/lite/performance/quantization_spec +template +void Quantize(const float* input, T* output, int num_elements, float scale, + int zero_point) { + for (int i = 0; i < num_elements; i++) { + output[i] = FloatToQuantizedType(input[i], scale, zero_point); + } +} + +template +void SymmetricQuantize(const float* input, T* output, int num_elements, + float scale) { + for (int i = 0; i < num_elements; i++) { + output[i] = FloatToSymmetricQuantizedType(input[i], scale); + } +} + +template +void SymmetricPerChannelQuantize(const float* input, T* output, + int num_elements, int num_channels, + float* scales) { + int elements_per_channel = num_elements / num_channels; + for (int i = 0; i < num_channels; i++) { + for (int j = 0; j < elements_per_channel; j++) { + output[i * elements_per_channel + j] = FloatToSymmetricQuantizedType( + input[i * elements_per_channel + j], scales[i]); + } + } +} + +void SignedSymmetricPerChannelQuantize(const float* values, + TfLiteIntArray* dims, + int quantized_dimension, + int8_t* quantized_values, + float* scaling_factor); + +// Quantizes inputs based on the values provided, choosing the smallest range +// which includes all input values. +template +void SymmetricQuantizeCalculateScales(const float* values, TfLiteIntArray* dims, + T* output, float* scale) { + int input_size = ElementCount(*dims); + + float min = 0; + float max = 0; + for (int i = 0; i < input_size; i++) { + min = fminf(min, values[i]); + max = fmaxf(max, values[i]); + } + *scale = fmaxf(std::abs(min), std::abs(max)) / std::numeric_limits::max(); + for (int i = 0; i < input_size; i++) { + const int32_t quantized_value = + static_cast(roundf(values[i] / *scale)); + // Clamp: just in case some odd numeric offset. + quantized_value = fminf(std::numeric_limits::max(), quantized_value); + quantized_value = fmaxf(std::numeric_limits::min() + 1, quantized_value); + output[i] = quantized_value; + } +} + +template +void Dequantize(const T* values, const int size, const float scale, + int zero_point, float* dequantized_values) { + for (int i = 0; i < size; ++i) { + dequantized_values[i] = (values[i] - zero_point) * scale; + } +} + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_MICRO_UTILS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.cpp b/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.cpp new file mode 100644 index 0000000..e4c63d2 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.cpp @@ -0,0 +1,247 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.h" + +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_allocator.h" +#include "tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.h" + +namespace tflite { + +RecordingMicroAllocator::RecordingMicroAllocator( + RecordingSimpleMemoryAllocator* recording_memory_allocator, + ErrorReporter* error_reporter) + : MicroAllocator(recording_memory_allocator, error_reporter), + recording_memory_allocator_(recording_memory_allocator) {} + +RecordingMicroAllocator* RecordingMicroAllocator::Create( + uint8_t* tensor_arena, size_t arena_size, ErrorReporter* error_reporter) { + TFLITE_DCHECK(error_reporter != nullptr); + + RecordingSimpleMemoryAllocator* simple_memory_allocator = + RecordingSimpleMemoryAllocator::Create(error_reporter, tensor_arena, + arena_size); + TFLITE_DCHECK(simple_memory_allocator != nullptr); + + uint8_t* allocator_buffer = simple_memory_allocator->AllocateFromTail( + sizeof(RecordingMicroAllocator), alignof(RecordingMicroAllocator)); + RecordingMicroAllocator* allocator = new (allocator_buffer) + RecordingMicroAllocator(simple_memory_allocator, error_reporter); + return allocator; +} + +RecordedAllocation RecordingMicroAllocator::GetRecordedAllocation( + RecordedAllocationType allocation_type) const { + switch (allocation_type) { + case RecordedAllocationType::kTfLiteEvalTensorData: + return recorded_tflite_eval_tensor_data_; + case RecordedAllocationType::kPersistentTfLiteTensorData: + return recorded_persistent_tflite_tensor_data_; + case RecordedAllocationType::kPersistentTfLiteTensorQuantizationData: + return recorded_persistent_tflite_tensor_quantization_data_; + case RecordedAllocationType::kPersistentBufferData: + return recorded_persistent_buffer_data_; + case RecordedAllocationType::kTfLiteTensorVariableBufferData: + return recorded_tflite_tensor_variable_buffer_data_; + case RecordedAllocationType::kNodeAndRegistrationArray: + return recorded_node_and_registration_array_data_; + case RecordedAllocationType::kOpData: + return recorded_op_data_; + } + TF_LITE_REPORT_ERROR(error_reporter(), "Invalid allocation type supplied: %d", + allocation_type); + return RecordedAllocation(); +} + +const RecordingSimpleMemoryAllocator* +RecordingMicroAllocator::GetSimpleMemoryAllocator() const { + return recording_memory_allocator_; +} + +void RecordingMicroAllocator::PrintAllocations() const { + TF_LITE_REPORT_ERROR( + error_reporter(), + "[RecordingMicroAllocator] Arena allocation total %d bytes", + recording_memory_allocator_->GetUsedBytes()); + TF_LITE_REPORT_ERROR( + error_reporter(), + "[RecordingMicroAllocator] Arena allocation head %d bytes", + recording_memory_allocator_->GetHeadUsedBytes()); + TF_LITE_REPORT_ERROR( + error_reporter(), + "[RecordingMicroAllocator] Arena allocation tail %d bytes", + recording_memory_allocator_->GetTailUsedBytes()); + PrintRecordedAllocation(RecordedAllocationType::kTfLiteEvalTensorData, + "TfLiteEvalTensor data", "allocations"); + PrintRecordedAllocation(RecordedAllocationType::kPersistentTfLiteTensorData, + "Persistent TfLiteTensor data", "tensors"); + PrintRecordedAllocation( + RecordedAllocationType::kPersistentTfLiteTensorQuantizationData, + "Persistent TfLiteTensor quantization data", "allocations"); + PrintRecordedAllocation(RecordedAllocationType::kPersistentBufferData, + "Persistent buffer data", "allocations"); + PrintRecordedAllocation( + RecordedAllocationType::kTfLiteTensorVariableBufferData, + "TfLiteTensor variable buffer data", "allocations"); + PrintRecordedAllocation(RecordedAllocationType::kNodeAndRegistrationArray, + "NodeAndRegistration struct", + "NodeAndRegistration structs"); + PrintRecordedAllocation(RecordedAllocationType::kOpData, + "Operator runtime data", "OpData structs"); +} + +void* RecordingMicroAllocator::AllocatePersistentBuffer(size_t bytes) { + RecordedAllocation allocations = SnapshotAllocationUsage(); + void* buffer = MicroAllocator::AllocatePersistentBuffer(bytes); + RecordAllocationUsage(allocations, recorded_persistent_buffer_data_); + + return buffer; +} + +void RecordingMicroAllocator::PrintRecordedAllocation( + RecordedAllocationType allocation_type, const char* allocation_name, + const char* allocation_description) const { +#ifndef TF_LITE_STRIP_ERROR_STRINGS + RecordedAllocation allocation = GetRecordedAllocation(allocation_type); + if (allocation.used_bytes > 0 || allocation.requested_bytes > 0) { + TF_LITE_REPORT_ERROR( + error_reporter(), + "[RecordingMicroAllocator] '%s' used %d bytes with alignment overhead " + "(requested %d bytes for %d %s)", + allocation_name, allocation.used_bytes, allocation.requested_bytes, + allocation.count, allocation_description); + } +#endif +} + +TfLiteStatus RecordingMicroAllocator::AllocateNodeAndRegistrations( + const Model* model, NodeAndRegistration** node_and_registrations) { + RecordedAllocation allocations = SnapshotAllocationUsage(); + + TfLiteStatus status = MicroAllocator::AllocateNodeAndRegistrations( + model, node_and_registrations); + + RecordAllocationUsage(allocations, + recorded_node_and_registration_array_data_); + // The allocation count in SimpleMemoryAllocator will only be 1. To provide + // better logging, decrement by 1 and add in the actual number of operators + // used in the graph: + // The allocation for this recording will always be 1. This is because the + // parent class mallocs one large allocation for the number of nodes in the + // graph (e.g. sizeof(NodeAndRegistration) * num_nodes). + // To prevent extra overhead and potential for fragmentation, manually adjust + // the accounting by decrementing by 1 and adding the actual number of nodes + // used in the graph: + recorded_node_and_registration_array_data_.count += + GetSubGraphFromModel(model)->operators()->size() - 1; + return status; +} + +TfLiteStatus +RecordingMicroAllocator::PrepareNodeAndRegistrationDataFromFlatbuffer( + const Model* model, const MicroOpResolver& op_resolver, + NodeAndRegistration* node_and_registrations) { + RecordedAllocation allocations = SnapshotAllocationUsage(); + + TfLiteStatus status = + MicroAllocator::PrepareNodeAndRegistrationDataFromFlatbuffer( + model, op_resolver, node_and_registrations); + + RecordAllocationUsage(allocations, recorded_op_data_); + return status; +} + +TfLiteStatus RecordingMicroAllocator::AllocateTfLiteEvalTensors( + const Model* model, TfLiteEvalTensor** eval_tensors) { + RecordedAllocation allocations = SnapshotAllocationUsage(); + + TfLiteStatus status = + MicroAllocator::AllocateTfLiteEvalTensors(model, eval_tensors); + + RecordAllocationUsage(allocations, recorded_tflite_eval_tensor_data_); + // The allocation for this recording will always be 1. This is because the + // parent class mallocs one large allocation for the number of tensors in the + // graph (e.g. sizeof(TfLiteEvalTensor) * num_tensors). + // To prevent extra overhead and potential for fragmentation, manually adjust + // the accounting by decrementing by 1 and adding the actual number of tensors + // used in the graph: + recorded_tflite_eval_tensor_data_.count += + GetSubGraphFromModel(model)->tensors()->size() - 1; + return status; +} + +TfLiteStatus RecordingMicroAllocator::AllocateVariables( + const SubGraph* subgraph, TfLiteEvalTensor* eval_tensors) { + RecordedAllocation allocations = SnapshotAllocationUsage(); + + TfLiteStatus status = + MicroAllocator::AllocateVariables(subgraph, eval_tensors); + + RecordAllocationUsage(allocations, + recorded_tflite_tensor_variable_buffer_data_); + return status; +} + +TfLiteTensor* RecordingMicroAllocator::AllocatePersistentTfLiteTensorInternal( + const Model* model, TfLiteEvalTensor* eval_tensors, int tensor_index) { + RecordedAllocation allocations = SnapshotAllocationUsage(); + + TfLiteTensor* result = MicroAllocator::AllocatePersistentTfLiteTensorInternal( + model, eval_tensors, tensor_index); + + RecordAllocationUsage(allocations, recorded_persistent_tflite_tensor_data_); + return result; +} + +TfLiteStatus RecordingMicroAllocator::PopulateTfLiteTensorFromFlatbuffer( + const Model* model, const SubGraph* subgraph, TfLiteTensor* tensor, + int tensor_index, bool allocate_temp) { + RecordedAllocation allocations = SnapshotAllocationUsage(); + + TfLiteStatus status = MicroAllocator::PopulateTfLiteTensorFromFlatbuffer( + model, subgraph, tensor, tensor_index, allocate_temp); + + RecordAllocationUsage(allocations, + recorded_persistent_tflite_tensor_quantization_data_); + return status; +} + +RecordedAllocation RecordingMicroAllocator::SnapshotAllocationUsage() const { + return {/*requested_bytes=*/recording_memory_allocator_->GetRequestedBytes(), + /*used_bytes=*/recording_memory_allocator_->GetUsedBytes(), + /*count=*/recording_memory_allocator_->GetAllocatedCount()}; +} + +void RecordingMicroAllocator::RecordAllocationUsage( + const RecordedAllocation& snapshotted_allocation, + RecordedAllocation& recorded_allocation) { + recorded_allocation.requested_bytes += + recording_memory_allocator_->GetRequestedBytes() - + snapshotted_allocation.requested_bytes; + recorded_allocation.used_bytes += + recording_memory_allocator_->GetUsedBytes() - + snapshotted_allocation.used_bytes; + recorded_allocation.count += + recording_memory_allocator_->GetAllocatedCount() - + snapshotted_allocation.count; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.h b/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.h new file mode 100644 index 0000000..3690164 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.h @@ -0,0 +1,128 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_RECORDING_MICRO_ALLOCATOR_H_ +#define TENSORFLOW_LITE_MICRO_RECORDING_MICRO_ALLOCATOR_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_allocator.h" +#include "tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.h" + +namespace tflite { + +// List of buckets currently recorded by this class. Each type keeps a list of +// allocated information during model initialization. +// TODO(b/169834511): Add tracking for scratch buffer allocations. +enum class RecordedAllocationType { + kTfLiteEvalTensorData, + kPersistentTfLiteTensorData, + kPersistentTfLiteTensorQuantizationData, + kPersistentBufferData, + kTfLiteTensorVariableBufferData, + kNodeAndRegistrationArray, + kOpData, +}; + +// Container for holding information about allocation recordings by a given +// type. Each recording contains the number of bytes requested, the actual bytes +// allocated (can defer from requested by alignment), and the number of items +// allocated. +struct RecordedAllocation { + size_t requested_bytes; + size_t used_bytes; + size_t count; +}; + +// Utility subclass of MicroAllocator that records all allocations +// inside the arena. A summary of allocations can be logged through the +// ErrorReporter by invoking LogAllocations(). This special allocator requires +// an instance of RecordingSimpleMemoryAllocator to capture allocations in the +// head and tail. Arena allocation recording can be retrieved by type through +// the GetRecordedAllocation() function. This class should only be used for +// auditing memory usage or integration testing. +class RecordingMicroAllocator : public MicroAllocator { + public: + static RecordingMicroAllocator* Create(uint8_t* tensor_arena, + size_t arena_size, + ErrorReporter* error_reporter); + + // Returns the recorded allocations information for a given allocation type. + RecordedAllocation GetRecordedAllocation( + RecordedAllocationType allocation_type) const; + + const RecordingSimpleMemoryAllocator* GetSimpleMemoryAllocator() const; + + // Logs out through the ErrorReporter all allocation recordings by type + // defined in RecordedAllocationType. + void PrintAllocations() const; + + void* AllocatePersistentBuffer(size_t bytes) override; + + protected: + TfLiteStatus AllocateNodeAndRegistrations( + const Model* model, + NodeAndRegistration** node_and_registrations) override; + TfLiteStatus PrepareNodeAndRegistrationDataFromFlatbuffer( + const Model* model, const MicroOpResolver& op_resolver, + NodeAndRegistration* node_and_registrations) override; + TfLiteStatus AllocateTfLiteEvalTensors( + const Model* model, TfLiteEvalTensor** eval_tensors) override; + TfLiteStatus AllocateVariables(const SubGraph* subgraph, + TfLiteEvalTensor* eval_tensors) override; + // TODO(b/162311891): Once all kernels have been updated to the new API drop + // this method. It is only used to record TfLiteTensor persistent allocations. + TfLiteTensor* AllocatePersistentTfLiteTensorInternal( + const Model* model, TfLiteEvalTensor* eval_tensors, + int tensor_index) override; + // TODO(b/162311891): Once all kernels have been updated to the new API drop + // this function since all allocations for quantized data will take place in + // the temp section. + TfLiteStatus PopulateTfLiteTensorFromFlatbuffer(const Model* model, + const SubGraph* subgraph, + TfLiteTensor* tensor, + int tensor_index, + bool allocate_temp) override; + + private: + RecordingMicroAllocator(RecordingSimpleMemoryAllocator* memory_allocator, + ErrorReporter* error_reporter); + + void PrintRecordedAllocation(RecordedAllocationType allocation_type, + const char* allocation_name, + const char* allocation_description) const; + + RecordedAllocation SnapshotAllocationUsage() const; + void RecordAllocationUsage(const RecordedAllocation& snapshotted_allocation, + RecordedAllocation& recorded_allocation); + + const RecordingSimpleMemoryAllocator* recording_memory_allocator_; + + RecordedAllocation recorded_tflite_eval_tensor_data_ = {}; + RecordedAllocation recorded_persistent_tflite_tensor_data_ = {}; + RecordedAllocation recorded_persistent_tflite_tensor_quantization_data_ = {}; + RecordedAllocation recorded_persistent_buffer_data_ = {}; + RecordedAllocation recorded_tflite_tensor_variable_buffer_data_ = {}; + RecordedAllocation recorded_node_and_registration_array_data_ = {}; + RecordedAllocation recorded_op_data_ = {}; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_RECORDING_MICRO_ALLOCATOR_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_interpreter.h b/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_interpreter.h new file mode 100644 index 0000000..69a366b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/recording_micro_interpreter.h @@ -0,0 +1,68 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_RECORDING_MICRO_INTERPRETER_H_ +#define TENSORFLOW_LITE_MICRO_RECORDING_MICRO_INTERPRETER_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/micro_interpreter.h" +#include "tensorflow_arm/tensorflow/lite/micro/recording_micro_allocator.h" + +namespace tflite { + +// Utility subclass that enables internal recordings of the MicroInterpreter. +// This class should be used to audit and analyze memory arena usage for a given +// model and interpreter. +// +// After construction and the first Invoke() or AllocateTensors() call - the +// memory usage is recorded and available through the GetMicroAllocator() +// function. See RecordingMicroAlloctor for more details on what is currently +// recorded from arena allocations. +// +// It is recommended for users to increase the tensor arena size by at least 1kb +// to ensure enough additional memory is available for internal recordings. +class RecordingMicroInterpreter : public MicroInterpreter { + public: + RecordingMicroInterpreter(const Model* model, + const MicroOpResolver& op_resolver, + uint8_t* tensor_arena, size_t tensor_arena_size, + ErrorReporter* error_reporter) + : MicroInterpreter(model, op_resolver, + RecordingMicroAllocator::Create( + tensor_arena, tensor_arena_size, error_reporter), + error_reporter), + recording_micro_allocator_( + static_cast(allocator())) {} + + RecordingMicroInterpreter(const Model* model, + const MicroOpResolver& op_resolver, + RecordingMicroAllocator* allocator, + ErrorReporter* error_reporter) + : MicroInterpreter(model, op_resolver, allocator, error_reporter), + recording_micro_allocator_(*allocator) {} + + const RecordingMicroAllocator& GetMicroAllocator() const { + return recording_micro_allocator_; + } + + private: + const RecordingMicroAllocator& recording_micro_allocator_; +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_RECORDING_MICRO_INTERPRETER_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.cpp b/src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.cpp new file mode 100644 index 0000000..0411e28 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.cpp @@ -0,0 +1,87 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.h" + +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" + +namespace tflite { + +RecordingSimpleMemoryAllocator::RecordingSimpleMemoryAllocator( + ErrorReporter* error_reporter, uint8_t* buffer_head, size_t buffer_size) + : SimpleMemoryAllocator(error_reporter, buffer_head, buffer_size), + requested_head_bytes_(0), + requested_tail_bytes_(0), + used_bytes_(0), + alloc_count_(0) {} + +RecordingSimpleMemoryAllocator::~RecordingSimpleMemoryAllocator() {} + +RecordingSimpleMemoryAllocator* RecordingSimpleMemoryAllocator::Create( + ErrorReporter* error_reporter, uint8_t* buffer_head, size_t buffer_size) { + TFLITE_DCHECK(error_reporter != nullptr); + TFLITE_DCHECK(buffer_head != nullptr); + RecordingSimpleMemoryAllocator tmp = + RecordingSimpleMemoryAllocator(error_reporter, buffer_head, buffer_size); + + uint8_t* allocator_buffer = + tmp.AllocateFromTail(sizeof(RecordingSimpleMemoryAllocator), + alignof(RecordingSimpleMemoryAllocator)); + // Use the default copy constructor to populate internal states. + return new (allocator_buffer) RecordingSimpleMemoryAllocator(tmp); +} + +size_t RecordingSimpleMemoryAllocator::GetRequestedBytes() const { + return requested_head_bytes_ + requested_tail_bytes_; +} + +size_t RecordingSimpleMemoryAllocator::GetUsedBytes() const { + return used_bytes_; +} + +size_t RecordingSimpleMemoryAllocator::GetAllocatedCount() const { + return alloc_count_; +} + +TfLiteStatus RecordingSimpleMemoryAllocator::SetHeadBufferSize( + size_t size, size_t alignment) { + const uint8_t* previous_head = head(); + TfLiteStatus status = + SimpleMemoryAllocator::SetHeadBufferSize(size, alignment); + if (status == kTfLiteOk) { + used_bytes_ += head() - previous_head; + requested_head_bytes_ = size; + } + return status; +} + +uint8_t* RecordingSimpleMemoryAllocator::AllocateFromTail(size_t size, + size_t alignment) { + const uint8_t* previous_tail = tail(); + uint8_t* result = SimpleMemoryAllocator::AllocateFromTail(size, alignment); + if (result != nullptr) { + used_bytes_ += previous_tail - tail(); + requested_tail_bytes_ += size; + alloc_count_++; + } + return result; +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.h b/src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.h new file mode 100644 index 0000000..dc42565 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/recording_simple_memory_allocator.h @@ -0,0 +1,67 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_RECORDING_SIMPLE_MEMORY_ALLOCATOR_H_ +#define TENSORFLOW_LITE_MICRO_RECORDING_SIMPLE_MEMORY_ALLOCATOR_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h" + +namespace tflite { + +// Utility class used to log allocations of a SimpleMemoryAllocator. Should only +// be used in debug/evaluation settings or unit tests to evaluate allocation +// usage. +class RecordingSimpleMemoryAllocator : public SimpleMemoryAllocator { + public: + RecordingSimpleMemoryAllocator(ErrorReporter* error_reporter, + uint8_t* buffer_head, size_t buffer_size); + // TODO(b/157615197): Cleanup constructors/destructor and use factory + // functions. + ~RecordingSimpleMemoryAllocator() override; + + static RecordingSimpleMemoryAllocator* Create(ErrorReporter* error_reporter, + uint8_t* buffer_head, + size_t buffer_size); + + // Returns the number of bytes requested from the head or tail. + size_t GetRequestedBytes() const; + + // Returns the number of bytes actually allocated from the head or tail. This + // value will be >= to the number of requested bytes due to padding and + // alignment. + size_t GetUsedBytes() const; + + // Returns the number of alloc calls from the head or tail. + size_t GetAllocatedCount() const; + + TfLiteStatus SetHeadBufferSize(size_t size, size_t alignment) override; + uint8_t* AllocateFromTail(size_t size, size_t alignment) override; + + private: + size_t requested_head_bytes_; + size_t requested_tail_bytes_; + size_t used_bytes_; + size_t alloc_count_; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_RECORDING_SIMPLE_MEMORY_ALLOCATOR_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.cpp b/src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.cpp new file mode 100644 index 0000000..a5c75b3 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.cpp @@ -0,0 +1,152 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h" + +#include +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/memory_helpers.h" + +namespace tflite { + +SimpleMemoryAllocator::SimpleMemoryAllocator(ErrorReporter* error_reporter, + uint8_t* buffer_head, + uint8_t* buffer_tail) + : error_reporter_(error_reporter), + buffer_head_(buffer_head), + buffer_tail_(buffer_tail), + head_(buffer_head), + tail_(buffer_tail), + temp_(buffer_head_) {} + +SimpleMemoryAllocator::SimpleMemoryAllocator(ErrorReporter* error_reporter, + uint8_t* buffer, + size_t buffer_size) + : SimpleMemoryAllocator(error_reporter, buffer, buffer + buffer_size) {} + +/* static */ +SimpleMemoryAllocator* SimpleMemoryAllocator::Create( + ErrorReporter* error_reporter, uint8_t* buffer_head, size_t buffer_size) { + TFLITE_DCHECK(error_reporter != nullptr); + TFLITE_DCHECK(buffer_head != nullptr); + SimpleMemoryAllocator tmp = + SimpleMemoryAllocator(error_reporter, buffer_head, buffer_size); + + // Allocate enough bytes from the buffer to create a SimpleMemoryAllocator. + // The new instance will use the current adjusted tail buffer from the tmp + // allocator instance. + uint8_t* allocator_buffer = tmp.AllocateFromTail( + sizeof(SimpleMemoryAllocator), alignof(SimpleMemoryAllocator)); + // Use the default copy constructor to populate internal states. + return new (allocator_buffer) SimpleMemoryAllocator(tmp); +} + +SimpleMemoryAllocator::~SimpleMemoryAllocator() {} + +TfLiteStatus SimpleMemoryAllocator::SetHeadBufferSize(size_t size, + size_t alignment) { + if (head_ != temp_) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Internal error: SetHeadBufferSize() needs to be called " + "after ResetTempAllocations()."); + return kTfLiteError; + } + + uint8_t* const aligned_result = AlignPointerUp(buffer_head_, alignment); + const size_t available_memory = tail_ - aligned_result; + if (available_memory < size) { + TF_LITE_REPORT_ERROR( + error_reporter_, + "Failed to set head size. Requested: %u, available %u, missing: %u", + size, available_memory, size - available_memory); + return kTfLiteError; + } + head_ = aligned_result + size; + temp_ = head_; + + return kTfLiteOk; +} + +uint8_t* SimpleMemoryAllocator::AllocateFromTail(size_t size, + size_t alignment) { + uint8_t* const aligned_result = AlignPointerDown(tail_ - size, alignment); + if (aligned_result < head_) { +#ifndef TF_LITE_STRIP_ERROR_STRINGS + const size_t missing_memory = head_ - aligned_result; + TF_LITE_REPORT_ERROR(error_reporter_, + "Failed to allocate tail memory. Requested: %u, " + "available %u, missing: %u", + size, size - missing_memory, missing_memory); +#endif + return nullptr; + } + tail_ = aligned_result; + return aligned_result; +} + +uint8_t* SimpleMemoryAllocator::AllocateTemp(size_t size, size_t alignment) { + uint8_t* const aligned_result = AlignPointerUp(temp_, alignment); + const size_t available_memory = tail_ - aligned_result; + if (available_memory < size) { + TF_LITE_REPORT_ERROR(error_reporter_, + "Failed to allocate temp memory. Requested: %u, " + "available %u, missing: %u", + size, available_memory, size - available_memory); + return nullptr; + } + temp_ = aligned_result + size; + return aligned_result; +} + +void SimpleMemoryAllocator::ResetTempAllocations() { temp_ = head_; } + +uint8_t* SimpleMemoryAllocator::GetHeadBuffer() const { return buffer_head_; } + +size_t SimpleMemoryAllocator::GetHeadUsedBytes() const { + return head_ - buffer_head_; +} + +size_t SimpleMemoryAllocator::GetTailUsedBytes() const { + return buffer_tail_ - tail_; +} + +size_t SimpleMemoryAllocator::GetAvailableMemory(size_t alignment) const { + uint8_t* const aligned_temp = AlignPointerUp(temp_, alignment); + uint8_t* const aligned_tail = AlignPointerDown(tail_, alignment); + return aligned_tail - aligned_temp; +} + +size_t SimpleMemoryAllocator::GetUsedBytes() const { + return GetBufferSize() - (tail_ - temp_); +} + +size_t SimpleMemoryAllocator::GetBufferSize() const { + return buffer_tail_ - buffer_head_; +} + +uint8_t* SimpleMemoryAllocator::head() const { return head_; } + +uint8_t* SimpleMemoryAllocator::tail() const { return tail_; } + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h b/src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h new file mode 100644 index 0000000..b78fdec --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/simple_memory_allocator.h @@ -0,0 +1,115 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ +#define TENSORFLOW_LITE_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ + +#include +#include + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/micro/compatibility.h" + +namespace tflite { + +// TODO(petewarden): This allocator never frees up or reuses any memory, even +// though we have enough information about lifetimes of the tensors to do so. +// This makes it pretty wasteful, so we should use a more intelligent method. +class SimpleMemoryAllocator { + public: + // TODO(b/157615197): Cleanup constructors/destructor and use factory + // functions. + SimpleMemoryAllocator(ErrorReporter* error_reporter, uint8_t* buffer_head, + uint8_t* buffer_tail); + SimpleMemoryAllocator(ErrorReporter* error_reporter, uint8_t* buffer, + size_t buffer_size); + virtual ~SimpleMemoryAllocator(); + + // Creates a new SimpleMemoryAllocator from a given buffer head and size. + static SimpleMemoryAllocator* Create(ErrorReporter* error_reporter, + uint8_t* buffer_head, + size_t buffer_size); + + // Adjusts the head (lowest address and moving upwards) memory allocation to a + // given size. Calls to this method will also invalidate all temporary + // allocation values (it sets the location of temp space at the end of the + // head section). This call will fail if a chain of allocations through + // AllocateTemp() have not been cleaned up with a call to + // ResetTempAllocations(). + virtual TfLiteStatus SetHeadBufferSize(size_t size, size_t alignment); + + // Allocates memory starting at the tail of the arena (highest address and + // moving downwards). + virtual uint8_t* AllocateFromTail(size_t size, size_t alignment); + + // Allocates a temporary buffer from the head of the arena (lowest address and + // moving upwards) but does not update the actual head allocation size or + // position. The returned buffer is guaranteed until either + // ResetTempAllocations() is called or another call to AllocateFromHead(). + // Repeat calls to this function will create a chain of temp allocations. All + // calls to AllocateTemp() must end with a call to ResetTempAllocations(). If + // AllocateFromHead() is called before a call to ResetTempAllocations(), it + // will fail with an error message. + virtual uint8_t* AllocateTemp(size_t size, size_t alignment); + + // Resets a chain of temporary allocations back to the current head of the + // arena (lowest address). + virtual void ResetTempAllocations(); + + // Returns a pointer to the buffer currently assigned to the head section. + // This buffer is set by calling SetHeadSize(). + uint8_t* GetHeadBuffer() const; + + // Returns the size of the head section in bytes. + size_t GetHeadUsedBytes() const; + + // Returns the size of all allocations in the tail section in bytes. + size_t GetTailUsedBytes() const; + + // Returns the number of bytes available with a given alignment. This number + // takes in account any temporary allocations. + size_t GetAvailableMemory(size_t alignment) const; + + // Returns the number of used bytes in the allocator. This number takes in + // account any temporary allocations. + size_t GetUsedBytes() const; + + protected: + // Returns a pointer to the current end of the head buffer. + uint8_t* head() const; + + // Returns a pointer to the current end of the tail buffer. + uint8_t* tail() const; + + private: + size_t GetBufferSize() const; + + ErrorReporter* error_reporter_; + uint8_t* buffer_head_; + uint8_t* buffer_tail_; + uint8_t* head_; + uint8_t* tail_; + uint8_t* temp_; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/test_helpers.cpp b/src/tensorflow_arm/tensorflow/lite/micro/test_helpers.cpp new file mode 100644 index 0000000..a4e8bb9 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/test_helpers.cpp @@ -0,0 +1,1082 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/test_helpers.h" + +#include +#include +#include +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +// TODO(b/170464050): Use TFLM test only version of schema_utils. + +namespace tflite { +namespace testing { +namespace { + +class StackAllocator : public flatbuffers::Allocator { + public: + StackAllocator() : data_(data_backing_), data_size_(0) {} + + uint8_t* allocate(size_t size) override { + TFLITE_DCHECK((data_size_ + size) <= kStackAllocatorSize); + uint8_t* result = data_; + data_ += size; + data_size_ += size; + return result; + } + + void deallocate(uint8_t* p, size_t) override {} + + static StackAllocator& instance() { + // Avoid using true dynamic memory allocation to be portable to bare metal. + static char inst_memory[sizeof(StackAllocator)]; + static StackAllocator* inst = new (inst_memory) StackAllocator; + return *inst; + } + + static constexpr size_t kStackAllocatorSize = 8192; + + private: + uint8_t data_backing_[kStackAllocatorSize]; + uint8_t* data_; + int data_size_; +}; + +flatbuffers::FlatBufferBuilder* BuilderInstance() { + static char inst_memory[sizeof(flatbuffers::FlatBufferBuilder)]; + static flatbuffers::FlatBufferBuilder* inst = + new (inst_memory) flatbuffers::FlatBufferBuilder( + StackAllocator::kStackAllocatorSize, &StackAllocator::instance()); + return inst; +} + +// A wrapper around FlatBuffer API to help build model easily. +class ModelBuilder { + public: + typedef int32_t Tensor; + typedef int Operator; + typedef int Node; + + // `builder` needs to be available until BuildModel is called. + explicit ModelBuilder(flatbuffers::FlatBufferBuilder* builder) + : builder_(builder) {} + + // Registers an operator that will be used in the model. + Operator RegisterOp(BuiltinOperator op, const char* custom_code); + + // Adds a tensor to the model. + Tensor AddTensor(TensorType type, std::initializer_list shape) { + return AddTensorImpl(type, /* is_variable */ false, shape); + } + + // Adds a variable tensor to the model. + Tensor AddVariableTensor(TensorType type, + std::initializer_list shape) { + return AddTensorImpl(type, /* is_variable */ true, shape); + } + + // Adds a node to the model with given input and output Tensors. + Node AddNode(Operator op, std::initializer_list inputs, + std::initializer_list outputs); + + void AddMetadata(const char* description_string, + const int32_t* metadata_buffer_data, size_t num_elements); + + // Constructs the flatbuffer model using `builder_` and return a pointer to + // it. The returned model has the same lifetime as `builder_`. + // Note the default value of 0 for num_subgraph_inputs means all tensor inputs + // are in subgraph input list. + const Model* BuildModel(std::initializer_list inputs, + std::initializer_list outputs, + size_t num_subgraph_inputs = 0); + + private: + // Adds a tensor to the model. + Tensor AddTensorImpl(TensorType type, bool is_variable, + std::initializer_list shape); + + flatbuffers::FlatBufferBuilder* builder_; + + static constexpr int kMaxOperatorCodes = 10; + flatbuffers::Offset operator_codes_[kMaxOperatorCodes]; + int next_operator_code_id_ = 0; + + static constexpr int kMaxOperators = 50; + flatbuffers::Offset operators_[kMaxOperators]; + int next_operator_id_ = 0; + + static constexpr int kMaxTensors = 50; + flatbuffers::Offset tensors_[kMaxTensors]; + + static constexpr int kMaxMetadataBuffers = 10; + + static constexpr int kMaxMetadatas = 10; + flatbuffers::Offset metadata_[kMaxMetadatas]; + + flatbuffers::Offset metadata_buffers_[kMaxMetadataBuffers]; + + int nbr_of_metadata_buffers_ = 0; + + int next_tensor_id_ = 0; +}; + +ModelBuilder::Operator ModelBuilder::RegisterOp(BuiltinOperator op, + const char* custom_code) { + TFLITE_DCHECK(next_operator_code_id_ <= kMaxOperatorCodes); + operator_codes_[next_operator_code_id_] = tflite::CreateOperatorCodeDirect( + *builder_, /*deprecated_builtin_code=*/0, custom_code, /*version=*/0, op); + next_operator_code_id_++; + return next_operator_code_id_ - 1; +} + +ModelBuilder::Node ModelBuilder::AddNode( + ModelBuilder::Operator op, + std::initializer_list inputs, + std::initializer_list outputs) { + TFLITE_DCHECK(next_operator_id_ <= kMaxOperators); + operators_[next_operator_id_] = tflite::CreateOperator( + *builder_, op, builder_->CreateVector(inputs.begin(), inputs.size()), + builder_->CreateVector(outputs.begin(), outputs.size()), + BuiltinOptions_NONE); + next_operator_id_++; + return next_operator_id_ - 1; +} + +void ModelBuilder::AddMetadata(const char* description_string, + const int32_t* metadata_buffer_data, + size_t num_elements) { + metadata_[ModelBuilder::nbr_of_metadata_buffers_] = + CreateMetadata(*builder_, builder_->CreateString(description_string), + 1 + ModelBuilder::nbr_of_metadata_buffers_); + + metadata_buffers_[nbr_of_metadata_buffers_] = tflite::CreateBuffer( + *builder_, builder_->CreateVector((uint8_t*)metadata_buffer_data, + sizeof(uint32_t) * num_elements)); + + ModelBuilder::nbr_of_metadata_buffers_++; +} + +const Model* ModelBuilder::BuildModel( + std::initializer_list inputs, + std::initializer_list outputs, + size_t num_subgraph_inputs) { + // Model schema requires an empty buffer at idx 0. + size_t buffer_size = 1 + ModelBuilder::nbr_of_metadata_buffers_; + flatbuffers::Offset buffers[kMaxMetadataBuffers]; + buffers[0] = tflite::CreateBuffer(*builder_); + + // Place the metadata buffers first in the buffer since the indices for them + // have already been set in AddMetadata() + for (int i = 1; i < ModelBuilder::nbr_of_metadata_buffers_ + 1; ++i) { + buffers[i] = metadata_buffers_[i - 1]; + } + + // TFLM only supports single subgraph. + constexpr size_t subgraphs_size = 1; + + // Find out number of subgraph inputs. + if (num_subgraph_inputs == 0) { + // This is the default case. + num_subgraph_inputs = inputs.size(); + } else { + // A non-zero value of num_subgraph_inputs means that some of + // the operator input tensors are not subgraph inputs. + TFLITE_DCHECK(num_subgraph_inputs <= inputs.size()); + } + + const flatbuffers::Offset subgraphs[subgraphs_size] = { + tflite::CreateSubGraph( + *builder_, builder_->CreateVector(tensors_, next_tensor_id_), + builder_->CreateVector(inputs.begin(), num_subgraph_inputs), + builder_->CreateVector(outputs.begin(), outputs.size()), + builder_->CreateVector(operators_, next_operator_id_), + builder_->CreateString("test_subgraph"))}; + + flatbuffers::Offset model_offset; + if (ModelBuilder::nbr_of_metadata_buffers_ > 0) { + model_offset = tflite::CreateModel( + *builder_, 0, + builder_->CreateVector(operator_codes_, next_operator_code_id_), + builder_->CreateVector(subgraphs, subgraphs_size), + builder_->CreateString("teset_model"), + builder_->CreateVector(buffers, buffer_size), 0, + builder_->CreateVector(metadata_, + ModelBuilder::nbr_of_metadata_buffers_)); + } else { + model_offset = tflite::CreateModel( + *builder_, 0, + builder_->CreateVector(operator_codes_, next_operator_code_id_), + builder_->CreateVector(subgraphs, subgraphs_size), + builder_->CreateString("teset_model"), + builder_->CreateVector(buffers, buffer_size)); + } + + tflite::FinishModelBuffer(*builder_, model_offset); + void* model_pointer = builder_->GetBufferPointer(); + const Model* model = flatbuffers::GetRoot(model_pointer); + return model; +} + +ModelBuilder::Tensor ModelBuilder::AddTensorImpl( + TensorType type, bool is_variable, std::initializer_list shape) { + TFLITE_DCHECK(next_tensor_id_ <= kMaxTensors); + tensors_[next_tensor_id_] = tflite::CreateTensor( + *builder_, builder_->CreateVector(shape.begin(), shape.size()), type, + /* buffer */ 0, /* name */ 0, /* quantization */ 0, + /* is_variable */ is_variable, + /* sparsity */ 0); + next_tensor_id_++; + return next_tensor_id_ - 1; +} + +const Model* BuildSimpleStatefulModel() { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* fb_builder = BuilderInstance(); + + ModelBuilder model_builder(fb_builder); + + const int op_id = + model_builder.RegisterOp(BuiltinOperator_CUSTOM, "simple_stateful_op"); + const int input_tensor = model_builder.AddTensor(TensorType_UINT8, {3}); + const int median_tensor = model_builder.AddTensor(TensorType_UINT8, {3}); + const int invoke_count_tensor = + model_builder.AddTensor(TensorType_INT32, {1}); + + model_builder.AddNode(op_id, {input_tensor}, + {median_tensor, invoke_count_tensor}); + return model_builder.BuildModel({input_tensor}, + {median_tensor, invoke_count_tensor}); +} + +const Model* BuildSimpleModelWithBranch() { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* fb_builder = BuilderInstance(); + + ModelBuilder model_builder(fb_builder); + /* Model structure + | t0 + +------| + | v + | +---------+ + | | n0 | + | | | + | +---------+ + v + + | + +---------+ | t1 + | n1 | | + | | | + +---------+ | + | | + t2 | v + | +---------+ + +-->| n2 | + | | + +-------|-+ + |t3 + v + */ + const int op_id = + model_builder.RegisterOp(BuiltinOperator_CUSTOM, "mock_custom"); + const int t0 = model_builder.AddTensor(TensorType_FLOAT32, {2, 2, 3}); + const int t1 = model_builder.AddTensor(TensorType_FLOAT32, {2, 2, 3}); + const int t2 = model_builder.AddTensor(TensorType_FLOAT32, {2, 2, 3}); + const int t3 = model_builder.AddTensor(TensorType_FLOAT32, {2, 2, 3}); + model_builder.AddNode(op_id, {t0}, {t1}); // n0 + model_builder.AddNode(op_id, {t0}, {t2}); // n1 + model_builder.AddNode(op_id, {t1, t2}, {t3}); // n2 + return model_builder.BuildModel({t0}, {t3}); +} + +const Model* BuildModelWithOfflinePlanning(int number_of_tensors, + const int32_t* metadata_buffer, + NodeConnection* node_conn, + int num_conns, + int num_subgraph_inputs) { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* fb_builder = BuilderInstance(); + + ModelBuilder model_builder(fb_builder); + + const int op_id = + model_builder.RegisterOp(BuiltinOperator_CUSTOM, "mock_custom"); + + for (int i = 0; i < number_of_tensors; ++i) { + model_builder.AddTensor(TensorType_FLOAT32, {2, 2, 3}); + } + + for (int i = 0; i < num_conns; ++i) { + model_builder.AddNode(op_id, node_conn[i].input, node_conn[i].output); + } + + model_builder.AddMetadata( + "OfflineMemoryAllocation", metadata_buffer, + number_of_tensors + tflite::testing::kOfflinePlannerHeaderSize); + + return model_builder.BuildModel( + node_conn[0].input, node_conn[num_conns - 1].output, num_subgraph_inputs); +} + +const Model* BuildSimpleMockModel() { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); + + constexpr size_t buffer_data_size = 1; + const uint8_t buffer_data[buffer_data_size] = {21}; + constexpr size_t buffers_size = 2; + const Offset buffers[buffers_size] = { + CreateBuffer(*builder), + CreateBuffer(*builder, + builder->CreateVector(buffer_data, buffer_data_size))}; + constexpr size_t tensor_shape_size = 1; + const int32_t tensor_shape[tensor_shape_size] = {1}; + constexpr size_t tensors_size = 4; + const Offset tensors[tensors_size] = { + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, + builder->CreateString("test_input_tensor"), 0, false), + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_UINT8, 1, + builder->CreateString("test_weight_tensor"), 0, false), + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, + builder->CreateString("test_output_tensor"), 0, false), + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, + builder->CreateString("test_output2_tensor"), 0, false), + }; + constexpr size_t inputs_size = 1; + const int32_t inputs[inputs_size] = {0}; + constexpr size_t outputs_size = 2; + const int32_t outputs[outputs_size] = {2, 3}; + constexpr size_t operator_inputs_size = 2; + const int32_t operator_inputs[operator_inputs_size] = {0, 1}; + constexpr size_t operator_outputs_size = 1; + const int32_t operator_outputs[operator_outputs_size] = {2}; + const int32_t operator2_outputs[operator_outputs_size] = {3}; + constexpr size_t operators_size = 2; + const Offset operators[operators_size] = { + CreateOperator( + *builder, 0, + builder->CreateVector(operator_inputs, operator_inputs_size), + builder->CreateVector(operator_outputs, operator_outputs_size), + BuiltinOptions_NONE), + CreateOperator( + *builder, 0, + builder->CreateVector(operator_inputs, operator_inputs_size), + builder->CreateVector(operator2_outputs, operator_outputs_size), + BuiltinOptions_NONE), + }; + constexpr size_t subgraphs_size = 1; + const Offset subgraphs[subgraphs_size] = { + CreateSubGraph(*builder, builder->CreateVector(tensors, tensors_size), + builder->CreateVector(inputs, inputs_size), + builder->CreateVector(outputs, outputs_size), + builder->CreateVector(operators, operators_size), + builder->CreateString("test_subgraph"))}; + constexpr size_t operator_codes_size = 1; + const Offset operator_codes[operator_codes_size] = { + CreateOperatorCodeDirect(*builder, /*deprecated_builtin_code=*/0, + "mock_custom", + /*version=*/0, BuiltinOperator_CUSTOM)}; + const Offset model_offset = CreateModel( + *builder, 0, builder->CreateVector(operator_codes, operator_codes_size), + builder->CreateVector(subgraphs, subgraphs_size), + builder->CreateString("test_model"), + builder->CreateVector(buffers, buffers_size)); + FinishModelBuffer(*builder, model_offset); + void* model_pointer = builder->GetBufferPointer(); + const Model* model = flatbuffers::GetRoot(model_pointer); + return model; +} + +const Model* BuildComplexMockModel() { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); + + constexpr size_t buffer_data_size = 1; + const uint8_t buffer_data_1[buffer_data_size] = {21}; + const uint8_t buffer_data_2[buffer_data_size] = {21}; + const uint8_t buffer_data_3[buffer_data_size] = {21}; + constexpr size_t buffers_size = 7; + const Offset buffers[buffers_size] = { + // Op 1 buffers: + CreateBuffer(*builder), + CreateBuffer(*builder), + CreateBuffer(*builder, + builder->CreateVector(buffer_data_1, buffer_data_size)), + // Op 2 buffers: + CreateBuffer(*builder), + CreateBuffer(*builder, + builder->CreateVector(buffer_data_2, buffer_data_size)), + // Op 3 buffers: + CreateBuffer(*builder), + CreateBuffer(*builder, + builder->CreateVector(buffer_data_3, buffer_data_size)), + }; + constexpr size_t tensor_shape_size = 1; + const int32_t tensor_shape[tensor_shape_size] = {1}; + + constexpr size_t tensors_size = 10; + const Offset tensors[tensors_size] = { + // Op 1 inputs: + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, builder->CreateString("test_input_tensor_1"), 0, + false /* is_variable */), + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 1, builder->CreateString("test_variable_tensor_1"), + 0, true /* is_variable */), + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_UINT8, 2, builder->CreateString("test_weight_tensor_1"), 0, + false /* is_variable */), + // Op 1 output / Op 2 input: + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, builder->CreateString("test_output_tensor_1"), 0, + false /* is_variable */), + // Op 2 inputs: + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 1, builder->CreateString("test_variable_tensor_2"), + 0, true /* is_variable */), + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_UINT8, 2, builder->CreateString("test_weight_tensor_2"), 0, + false /* is_variable */), + // Op 2 output / Op 3 input: + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, builder->CreateString("test_output_tensor_2"), 0, + false /* is_variable */), + // Op 3 inputs: + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 1, builder->CreateString("test_variable_tensor_3"), + 0, true /* is_variable */), + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_UINT8, 2, builder->CreateString("test_weight_tensor_3"), 0, + false /* is_variable */), + // Op 3 output: + CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, builder->CreateString("test_output_tensor_3"), 0, + false /* is_variable */), + }; + + constexpr size_t operators_size = 3; + Offset operators[operators_size]; + { + // Set Op 1 attributes: + constexpr size_t operator_inputs_size = 3; + const int32_t operator_inputs[operator_inputs_size] = {0, 1, 2}; + constexpr size_t operator_outputs_size = 1; + const int32_t operator_outputs[operator_outputs_size] = {3}; + + operators[0] = {CreateOperator( + *builder, 0, + builder->CreateVector(operator_inputs, operator_inputs_size), + builder->CreateVector(operator_outputs, operator_outputs_size), + BuiltinOptions_NONE)}; + } + + { + // Set Op 2 attributes + constexpr size_t operator_inputs_size = 3; + const int32_t operator_inputs[operator_inputs_size] = {3, 4, 5}; + constexpr size_t operator_outputs_size = 1; + const int32_t operator_outputs[operator_outputs_size] = {6}; + + operators[1] = {CreateOperator( + *builder, 0, + builder->CreateVector(operator_inputs, operator_inputs_size), + builder->CreateVector(operator_outputs, operator_outputs_size), + BuiltinOptions_NONE)}; + } + + { + // Set Op 3 attributes + constexpr size_t operator_inputs_size = 3; + const int32_t operator_inputs[operator_inputs_size] = {6, 7, 8}; + constexpr size_t operator_outputs_size = 1; + const int32_t operator_outputs[operator_outputs_size] = {9}; + + operators[2] = {CreateOperator( + *builder, 0, + builder->CreateVector(operator_inputs, operator_inputs_size), + builder->CreateVector(operator_outputs, operator_outputs_size), + BuiltinOptions_NONE)}; + } + + constexpr size_t inputs_size = 1; + const int32_t inputs[inputs_size] = {0}; + constexpr size_t outputs_size = 1; + const int32_t outputs[outputs_size] = {9}; + + constexpr size_t subgraphs_size = 1; + const Offset subgraphs[subgraphs_size] = { + CreateSubGraph(*builder, builder->CreateVector(tensors, tensors_size), + builder->CreateVector(inputs, inputs_size), + builder->CreateVector(outputs, outputs_size), + builder->CreateVector(operators, operators_size), + builder->CreateString("test_subgraph"))}; + + constexpr size_t operator_codes_size = 1; + const Offset operator_codes[operator_codes_size] = { + CreateOperatorCodeDirect(*builder, /*deprecated_builtin_code=*/0, + "mock_custom", + /*version=*/0, BuiltinOperator_CUSTOM)}; + + const Offset model_offset = CreateModel( + *builder, 0, builder->CreateVector(operator_codes, operator_codes_size), + builder->CreateVector(subgraphs, subgraphs_size), + builder->CreateString("test_model"), + builder->CreateVector(buffers, buffers_size)); + + FinishModelBuffer(*builder, model_offset); + void* model_pointer = builder->GetBufferPointer(); + const Model* model = flatbuffers::GetRoot(model_pointer); + return model; +} + +const Model* BuildSimpleMultipleInputsModel() { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); + + constexpr size_t buffers_size = 1; + const Offset buffers[buffers_size] = { + CreateBuffer(*builder), + }; + constexpr size_t tensor_shape_size = 1; + const int32_t tensor_shape[tensor_shape_size] = {1}; + constexpr size_t tensors_size = 4; + const Offset tensors[tensors_size] = { + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, + builder->CreateString("test_input_tensor1"), 0, false), + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT8, 0, + builder->CreateString("test_input_tensor2"), 0, false), + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, + builder->CreateString("test_input_tensor3"), 0, false), + CreateTensor(*builder, + builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, + builder->CreateString("test_output_tensor"), 0, false), + }; + constexpr size_t inputs_size = 3; + const int32_t inputs[inputs_size] = {0, 1, 2}; + constexpr size_t outputs_size = 1; + const int32_t outputs[outputs_size] = {3}; + constexpr size_t operator_inputs_size = 3; + const int32_t operator_inputs[operator_inputs_size] = {0, 1, 2}; + constexpr size_t operator_outputs_size = 1; + const int32_t operator_outputs[operator_outputs_size] = {3}; + constexpr size_t operators_size = 1; + const Offset operators[operators_size] = { + CreateOperator( + *builder, 0, + builder->CreateVector(operator_inputs, operator_inputs_size), + builder->CreateVector(operator_outputs, operator_outputs_size), + BuiltinOptions_NONE), + }; + constexpr size_t subgraphs_size = 1; + const Offset subgraphs[subgraphs_size] = { + CreateSubGraph(*builder, builder->CreateVector(tensors, tensors_size), + builder->CreateVector(inputs, inputs_size), + builder->CreateVector(outputs, outputs_size), + builder->CreateVector(operators, operators_size), + builder->CreateString("test_subgraph"))}; + constexpr size_t operator_codes_size = 1; + const Offset operator_codes[operator_codes_size] = { + CreateOperatorCodeDirect(*builder, /*deprecated_builtin_code=*/0, + "multiple_inputs_op", + /*version=*/0, BuiltinOperator_CUSTOM)}; + const Offset model_offset = CreateModel( + *builder, 0, builder->CreateVector(operator_codes, operator_codes_size), + builder->CreateVector(subgraphs, subgraphs_size), + builder->CreateString("test_model"), + builder->CreateVector(buffers, buffers_size)); + FinishModelBuffer(*builder, model_offset); + void* model_pointer = builder->GetBufferPointer(); + const Model* model = flatbuffers::GetRoot(model_pointer); + return model; +} + +} // namespace + +const TfLiteRegistration* SimpleStatefulOp::getRegistration() { + return GetMutableRegistration(); +} + +TfLiteRegistration* SimpleStatefulOp::GetMutableRegistration() { + static TfLiteRegistration r; + r.init = Init; + r.prepare = Prepare; + r.invoke = Invoke; + return &r; +} + +void* SimpleStatefulOp::Init(TfLiteContext* context, const char* buffer, + size_t length) { + TFLITE_DCHECK(context->AllocateBufferForEval == nullptr); + TFLITE_DCHECK(context->GetScratchBuffer == nullptr); + TFLITE_DCHECK(context->RequestScratchBufferInArena == nullptr); + + void* raw = context->AllocatePersistentBuffer(context, sizeof(OpData)); + OpData* data = reinterpret_cast(raw); + *data = {}; + return raw; +} + +TfLiteStatus SimpleStatefulOp::Prepare(TfLiteContext* context, + TfLiteNode* node) { + OpData* data = reinterpret_cast(node->user_data); + + // Make sure that the input is in uint8_t with at least 1 data entry. + const TfLiteTensor* input; + TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTensor, &input)); + if (input->type != kTfLiteUInt8) return kTfLiteError; + if (NumElements(input->dims) == 0) return kTfLiteError; + + // Allocate a temporary buffer with the same size of input for sorting. + TF_LITE_ENSURE_STATUS(context->RequestScratchBufferInArena( + context, sizeof(uint8_t) * NumElements(input->dims), + &data->sorting_buffer)); + // We can interleave scratch / persistent buffer allocation. + data->invoke_count = reinterpret_cast( + context->AllocatePersistentBuffer(context, sizeof(int))); + *data->invoke_count = 0; + + return kTfLiteOk; +} + +TfLiteStatus SimpleStatefulOp::Invoke(TfLiteContext* context, + TfLiteNode* node) { + OpData* data = reinterpret_cast(node->user_data); + *data->invoke_count += 1; + + const TfLiteTensor* input; + TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTensor, &input)); + const uint8_t* input_data = GetTensorData(input); + int size = NumElements(input->dims); + + uint8_t* sorting_buffer = reinterpret_cast( + context->GetScratchBuffer(context, data->sorting_buffer)); + // Copy inputs data to the sorting buffer. We don't want to mutate the input + // tensor as it might be used by a another node. + for (int i = 0; i < size; i++) { + sorting_buffer[i] = input_data[i]; + } + + // In place insertion sort on `sorting_buffer`. + for (int i = 1; i < size; i++) { + for (int j = i; j > 0 && sorting_buffer[j] < sorting_buffer[j - 1]; j--) { + std::swap(sorting_buffer[j], sorting_buffer[j - 1]); + } + } + + TfLiteTensor* median; + TF_LITE_ENSURE_OK(context, + GetOutputSafe(context, node, kMedianTensor, &median)); + uint8_t* median_data = GetTensorData(median); + TfLiteTensor* invoke_count; + TF_LITE_ENSURE_OK(context, + GetOutputSafe(context, node, kInvokeCount, &invoke_count)); + int32_t* invoke_count_data = GetTensorData(invoke_count); + + median_data[0] = sorting_buffer[size / 2]; + invoke_count_data[0] = *data->invoke_count; + return kTfLiteOk; +} + +const TfLiteRegistration* MockCustom::getRegistration() { + return GetMutableRegistration(); +} + +TfLiteRegistration* MockCustom::GetMutableRegistration() { + static TfLiteRegistration r; + r.init = Init; + r.prepare = Prepare; + r.invoke = Invoke; + r.free = Free; + return &r; +} + +void* MockCustom::Init(TfLiteContext* context, const char* buffer, + size_t length) { + // We don't support delegate in TFL micro. This is a weak check to test if + // context struct being zero-initialized. + TFLITE_DCHECK(context->ReplaceNodeSubsetsWithDelegateKernels == nullptr); + freed_ = false; + // Do nothing. + return nullptr; +} + +void MockCustom::Free(TfLiteContext* context, void* buffer) { freed_ = true; } + +TfLiteStatus MockCustom::Prepare(TfLiteContext* context, TfLiteNode* node) { + return kTfLiteOk; +} + +TfLiteStatus MockCustom::Invoke(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input; + TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, 0, &input)); + const int32_t* input_data = input->data.i32; + const TfLiteTensor* weight; + TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, 1, &weight)); + const uint8_t* weight_data = weight->data.uint8; + TfLiteTensor* output; + TF_LITE_ENSURE_OK(context, GetOutputSafe(context, node, 0, &output)); + int32_t* output_data = output->data.i32; + output_data[0] = + 0; // Catch output tensor sharing memory with an input tensor + output_data[0] = input_data[0] + weight_data[0]; + return kTfLiteOk; +} + +bool MockCustom::freed_ = false; + +const TfLiteRegistration* MultipleInputs::getRegistration() { + return GetMutableRegistration(); +} + +TfLiteRegistration* MultipleInputs::GetMutableRegistration() { + static TfLiteRegistration r; + r.init = Init; + r.prepare = Prepare; + r.invoke = Invoke; + r.free = Free; + return &r; +} + +void* MultipleInputs::Init(TfLiteContext* context, const char* buffer, + size_t length) { + // We don't support delegate in TFL micro. This is a weak check to test if + // context struct being zero-initialized. + TFLITE_DCHECK(context->ReplaceNodeSubsetsWithDelegateKernels == nullptr); + freed_ = false; + // Do nothing. + return nullptr; +} + +void MultipleInputs::Free(TfLiteContext* context, void* buffer) { + freed_ = true; +} + +TfLiteStatus MultipleInputs::Prepare(TfLiteContext* context, TfLiteNode* node) { + return kTfLiteOk; +} + +TfLiteStatus MultipleInputs::Invoke(TfLiteContext* context, TfLiteNode* node) { + const TfLiteTensor* input; + TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, 0, &input)); + const int32_t* input_data = input->data.i32; + const TfLiteTensor* input1; + TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, 1, &input1)); + const int32_t* input_data1 = input1->data.i32; + const TfLiteTensor* input2; + TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, 2, &input2)); + const int32_t* input_data2 = input2->data.i32; + + TfLiteTensor* output; + TF_LITE_ENSURE_OK(context, GetOutputSafe(context, node, 0, &output)); + int32_t* output_data = output->data.i32; + output_data[0] = + 0; // Catch output tensor sharing memory with an input tensor + output_data[0] = input_data[0] + input_data1[0] + input_data2[0]; + return kTfLiteOk; +} + +bool MultipleInputs::freed_ = false; + +AllOpsResolver GetOpResolver() { + AllOpsResolver op_resolver; + op_resolver.AddCustom("mock_custom", MockCustom::GetMutableRegistration()); + op_resolver.AddCustom("simple_stateful_op", + SimpleStatefulOp::GetMutableRegistration()); + op_resolver.AddCustom("multiple_inputs_op", + MultipleInputs::GetMutableRegistration()); + return op_resolver; +} + +const Model* GetSimpleMockModel() { + static Model* model = nullptr; + if (!model) { + model = const_cast(BuildSimpleMockModel()); + } + return model; +} + +const Model* GetSimpleMultipleInputsModel() { + static Model* model = nullptr; + if (!model) { + model = const_cast(BuildSimpleMultipleInputsModel()); + } + return model; +} + +const Model* GetComplexMockModel() { + static Model* model = nullptr; + if (!model) { + model = const_cast(BuildComplexMockModel()); + } + return model; +} + +const Model* GetSimpleModelWithBranch() { + static Model* model = nullptr; + if (!model) { + model = const_cast(BuildSimpleModelWithBranch()); + } + return model; +} + +const Model* GetModelWithOfflinePlanning(int num_tensors, + const int32_t* metadata_buffer, + NodeConnection* node_conn, + int num_conns, + int num_subgraph_inputs) { + const Model* model = BuildModelWithOfflinePlanning( + num_tensors, metadata_buffer, node_conn, num_conns, num_subgraph_inputs); + return model; +} + +const Model* GetSimpleStatefulModel() { + static Model* model = nullptr; + if (!model) { + model = const_cast(BuildSimpleStatefulModel()); + } + return model; +} + +const Tensor* Create1dFlatbufferTensor(int size, bool is_variable) { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); + constexpr size_t tensor_shape_size = 1; + const int32_t tensor_shape[tensor_shape_size] = {size}; + const Offset tensor_offset = CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, builder->CreateString("test_tensor"), 0, + is_variable); + builder->Finish(tensor_offset); + void* tensor_pointer = builder->GetBufferPointer(); + const Tensor* tensor = flatbuffers::GetRoot(tensor_pointer); + return tensor; +} + +const Tensor* CreateQuantizedFlatbufferTensor(int size) { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); + const Offset quant_params = + CreateQuantizationParameters( + *builder, + /*min=*/builder->CreateVector({0.1f}), + /*max=*/builder->CreateVector({0.2f}), + /*scale=*/builder->CreateVector({0.3f}), + /*zero_point=*/builder->CreateVector({100ll})); + + constexpr size_t tensor_shape_size = 1; + const int32_t tensor_shape[tensor_shape_size] = {size}; + const Offset tensor_offset = CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, builder->CreateString("test_tensor"), quant_params, + false); + builder->Finish(tensor_offset); + void* tensor_pointer = builder->GetBufferPointer(); + const Tensor* tensor = flatbuffers::GetRoot(tensor_pointer); + return tensor; +} + +const Tensor* CreateMissingQuantizationFlatbufferTensor(int size) { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); + const Offset quant_params = + CreateQuantizationParameters(*builder, 0, 0, 0, 0, + QuantizationDetails_NONE, 0, 0); + constexpr size_t tensor_shape_size = 1; + const int32_t tensor_shape[tensor_shape_size] = {size}; + const Offset tensor_offset = CreateTensor( + *builder, builder->CreateVector(tensor_shape, tensor_shape_size), + TensorType_INT32, 0, builder->CreateString("test_tensor"), quant_params, + false); + builder->Finish(tensor_offset); + void* tensor_pointer = builder->GetBufferPointer(); + const Tensor* tensor = flatbuffers::GetRoot(tensor_pointer); + return tensor; +} + +const flatbuffers::Vector>* +CreateFlatbufferBuffers() { + using flatbuffers::Offset; + flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); + constexpr size_t buffers_size = 1; + const Offset buffers[buffers_size] = { + CreateBuffer(*builder), + }; + const flatbuffers::Offset>> + buffers_offset = builder->CreateVector(buffers, buffers_size); + builder->Finish(buffers_offset); + void* buffers_pointer = builder->GetBufferPointer(); + const flatbuffers::Vector>* result = + flatbuffers::GetRoot>>( + buffers_pointer); + return result; +} + +int TestStrcmp(const char* a, const char* b) { + if ((a == nullptr) || (b == nullptr)) { + return -1; + } + while ((*a != 0) && (*a == *b)) { + a++; + b++; + } + return *reinterpret_cast(a) - + *reinterpret_cast(b); +} + +// Wrapper to forward kernel errors to the interpreter's error reporter. +void ReportOpError(struct TfLiteContext* context, const char* format, ...) { +#ifndef TF_LITE_STRIP_ERROR_STRINGS + ErrorReporter* error_reporter = static_cast(context->impl_); + va_list args; + va_start(args, format); + TF_LITE_REPORT_ERROR(error_reporter, format, args); + va_end(args); +#endif +} + +// Create a TfLiteIntArray from an array of ints. The first element in the +// supplied array must be the size of the array expressed as an int. +TfLiteIntArray* IntArrayFromInts(const int* int_array) { + return const_cast( + reinterpret_cast(int_array)); +} + +// Create a TfLiteFloatArray from an array of floats. The first element in the +// supplied array must be the size of the array expressed as a float. +TfLiteFloatArray* FloatArrayFromFloats(const float* floats) { + static_assert(sizeof(float) == sizeof(int), + "assumes sizeof(float) == sizeof(int) to perform casting"); + int size = static_cast(floats[0]); + *reinterpret_cast(const_cast(floats)) = size; + return reinterpret_cast(const_cast(floats)); +} + +TfLiteTensor CreateQuantizedBiasTensor(const float* data, int32_t* quantized, + TfLiteIntArray* dims, float input_scale, + float weights_scale, bool is_variable) { + float bias_scale = input_scale * weights_scale; + tflite::SymmetricQuantize(data, quantized, ElementCount(*dims), bias_scale); + + // Quantized int32_t tensors always have a zero point of 0, since the range of + // int32_t values is large, and because zero point costs extra cycles during + // processing. + TfLiteTensor result = + CreateQuantizedTensor(quantized, dims, bias_scale, 0, is_variable); + return result; +} + +// Quantizes int32_t bias tensor with per-channel weights determined by input +// scale multiplied by weight scale for each channel. +TfLiteTensor CreatePerChannelQuantizedBiasTensor( + const float* input, int32_t* quantized, TfLiteIntArray* dims, + float input_scale, float* weight_scales, float* scales, int* zero_points, + TfLiteAffineQuantization* affine_quant, int quantized_dimension, + bool is_variable) { + int input_size = ElementCount(*dims); + int num_channels = dims->data[quantized_dimension]; + // First element is reserved for array length + zero_points[0] = num_channels; + scales[0] = static_cast(num_channels); + float* scales_array = &scales[1]; + for (int i = 0; i < num_channels; i++) { + scales_array[i] = input_scale * weight_scales[i]; + zero_points[i + 1] = 0; + } + + SymmetricPerChannelQuantize(input, quantized, input_size, + num_channels, scales_array); + + affine_quant->scale = FloatArrayFromFloats(scales); + affine_quant->zero_point = IntArrayFromInts(zero_points); + affine_quant->quantized_dimension = quantized_dimension; + + TfLiteTensor result = CreateTensor(quantized, dims, is_variable); + result.quantization = {kTfLiteAffineQuantization, affine_quant}; + return result; +} + +TfLiteTensor CreateSymmetricPerChannelQuantizedTensor( + const float* input, int8_t* quantized, TfLiteIntArray* dims, float* scales, + int* zero_points, TfLiteAffineQuantization* affine_quant, + int quantized_dimension, bool is_variable) { + int channel_count = dims->data[quantized_dimension]; + scales[0] = static_cast(channel_count); + zero_points[0] = channel_count; + + SignedSymmetricPerChannelQuantize(input, dims, quantized_dimension, quantized, + &scales[1]); + + for (int i = 0; i < channel_count; i++) { + zero_points[i + 1] = 0; + } + + affine_quant->scale = FloatArrayFromFloats(scales); + affine_quant->zero_point = IntArrayFromInts(zero_points); + affine_quant->quantized_dimension = quantized_dimension; + + TfLiteTensor result = CreateTensor(quantized, dims, is_variable); + result.quantization = {kTfLiteAffineQuantization, affine_quant}; + return result; +} + +size_t GetModelTensorCount(const Model* model) { + auto* subgraphs = model->subgraphs(); + if (subgraphs) { + return (*subgraphs)[0]->tensors()->size(); + } + return 0; +} + +} // namespace testing +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/test_helpers.h b/src/tensorflow_arm/tensorflow/lite/micro/test_helpers.h new file mode 100644 index 0000000..5956175 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/test_helpers.h @@ -0,0 +1,244 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_TEST_HELPERS_H_ +#define TENSORFLOW_LITE_MICRO_TEST_HELPERS_H_ + +// Useful functions for writing tests. + +#include +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite//kernels/internal/tensor_ctypes.h" +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_arm/tensorflow/lite/micro/all_ops_resolver.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_utils.h" +#include "tensorflow_arm/tensorflow/lite/portable_type_to_tflitetype.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { +namespace testing { + +constexpr int kOfflinePlannerHeaderSize = 3; + +struct NodeConnection_ { + std::initializer_list input; + std::initializer_list output; +}; +typedef struct NodeConnection_ NodeConnection; + +// A simple operator that returns the median of the input with the number of +// times the kernel was invoked. The implementation below is deliberately +// complicated, just to demonstrate how kernel memory planning works. +class SimpleStatefulOp { + static constexpr int kBufferNotAllocated = 0; + // Inputs: + static constexpr int kInputTensor = 0; + // Outputs: + static constexpr int kMedianTensor = 0; + static constexpr int kInvokeCount = 1; + struct OpData { + int* invoke_count = nullptr; + int sorting_buffer = kBufferNotAllocated; + }; + + public: + static const TfLiteRegistration* getRegistration(); + static TfLiteRegistration* GetMutableRegistration(); + static void* Init(TfLiteContext* context, const char* buffer, size_t length); + static TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node); + static TfLiteStatus Invoke(TfLiteContext* context, TfLiteNode* node); +}; + +class MockCustom { + public: + static const TfLiteRegistration* getRegistration(); + static TfLiteRegistration* GetMutableRegistration(); + static void* Init(TfLiteContext* context, const char* buffer, size_t length); + static void Free(TfLiteContext* context, void* buffer); + static TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node); + static TfLiteStatus Invoke(TfLiteContext* context, TfLiteNode* node); + + static bool freed_; +}; + +// A simple operator with the purpose of testing multiple inputs. It returns +// the sum of the inputs. +class MultipleInputs { + public: + static const TfLiteRegistration* getRegistration(); + static TfLiteRegistration* GetMutableRegistration(); + static void* Init(TfLiteContext* context, const char* buffer, size_t length); + static void Free(TfLiteContext* context, void* buffer); + static TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node); + static TfLiteStatus Invoke(TfLiteContext* context, TfLiteNode* node); + + static bool freed_; +}; + +// Returns an Op Resolver that can be used in the testing code. +AllOpsResolver GetOpResolver(); + +// Returns a simple example flatbuffer TensorFlow Lite model. Contains 1 input, +// 1 layer of weights, 1 output Tensor, and 1 operator. +const Model* GetSimpleMockModel(); + +// Returns a flatbuffer TensorFlow Lite model with more inputs, variable +// tensors, and operators. +const Model* GetComplexMockModel(); + +// Returns a simple flatbuffer model with two branches. +const Model* GetSimpleModelWithBranch(); + +// Returns a simple example flatbuffer TensorFlow Lite model. Contains 3 inputs, +// 1 output Tensor, and 1 operator. +const Model* GetSimpleMultipleInputsModel(); + +// Returns a simple flatbuffer model with offline planned tensors +// @param[in] num_tensors Number of tensors in the model. +// @param[in] metadata_buffer Metadata for offline planner. +// @param[in] node_con List of connections, i.e. operators +// in the model. +// @param[in] num_conns Number of connections. +// @param[in] num_subgraph_inputs How many of the input tensors are in +// the subgraph inputs. The default value +// of 0 means all of the input tensors +// are in the subgraph input list. There +// must be at least 1 input tensor in the +// subgraph input list. +const Model* GetModelWithOfflinePlanning(int num_tensors, + const int32_t* metadata_buffer, + NodeConnection* node_conn, + int num_conns, + int num_subgraph_inputs = 0); + +// Returns a flatbuffer model with `simple_stateful_op` +const Model* GetSimpleStatefulModel(); + +// Builds a one-dimensional flatbuffer tensor of the given size. +const Tensor* Create1dFlatbufferTensor(int size, bool is_variable = false); + +// Builds a one-dimensional flatbuffer tensor of the given size with +// quantization metadata. +const Tensor* CreateQuantizedFlatbufferTensor(int size); + +// Creates a one-dimensional tensor with no quantization metadata. +const Tensor* CreateMissingQuantizationFlatbufferTensor(int size); + +// Creates a vector of flatbuffer buffers. +const flatbuffers::Vector>* +CreateFlatbufferBuffers(); + +// Performs a simple string comparison without requiring standard C library. +int TestStrcmp(const char* a, const char* b); + +// Wrapper to forward kernel errors to the interpreter's error reporter. +void ReportOpError(struct TfLiteContext* context, const char* format, ...); + +void PopulateContext(TfLiteTensor* tensors, int tensors_size, + TfLiteContext* context); + +// Create a TfLiteIntArray from an array of ints. The first element in the +// supplied array must be the size of the array expressed as an int. +TfLiteIntArray* IntArrayFromInts(const int* int_array); + +// Create a TfLiteFloatArray from an array of floats. The first element in the +// supplied array must be the size of the array expressed as a float. +TfLiteFloatArray* FloatArrayFromFloats(const float* floats); + +template +TfLiteTensor CreateTensor(const T* data, TfLiteIntArray* dims, + const bool is_variable = false) { + TfLiteTensor result; + result.dims = dims; + result.params = {}; + result.quantization = {kTfLiteNoQuantization, nullptr}; + result.is_variable = is_variable; + result.allocation_type = kTfLiteMemNone; + result.type = typeToTfLiteType(); + // Const cast is used to allow passing in const and non-const arrays within a + // single CreateTensor method. A Const array should be used for immutable + // input tensors and non-const array should be used for mutable and output + // tensors. + result.data.data = const_cast(data); + result.quantization = {kTfLiteAffineQuantization, nullptr}; + result.bytes = ElementCount(*dims) * sizeof(T); + return result; +} + +template +TfLiteTensor CreateQuantizedTensor(const T* data, TfLiteIntArray* dims, + const float scale, const int zero_point = 0, + const bool is_variable = false) { + TfLiteTensor result = CreateTensor(data, dims, is_variable); + result.params = {scale, zero_point}; + result.quantization = {kTfLiteAffineQuantization, nullptr}; + return result; +} + +template +TfLiteTensor CreateQuantizedTensor(const float* input, T* quantized, + TfLiteIntArray* dims, float scale, + int zero_point, bool is_variable = false) { + int input_size = ElementCount(*dims); + tflite::Quantize(input, quantized, input_size, scale, zero_point); + return CreateQuantizedTensor(quantized, dims, scale, zero_point, is_variable); +} + +TfLiteTensor CreateQuantizedBiasTensor(const float* data, int32_t* quantized, + TfLiteIntArray* dims, float input_scale, + float weights_scale, + bool is_variable = false); + +// Quantizes int32_t bias tensor with per-channel weights determined by input +// scale multiplied by weight scale for each channel. +TfLiteTensor CreatePerChannelQuantizedBiasTensor( + const float* input, int32_t* quantized, TfLiteIntArray* dims, + float input_scale, float* weight_scales, float* scales, int* zero_points, + TfLiteAffineQuantization* affine_quant, int quantized_dimension, + bool is_variable = false); + +TfLiteTensor CreateSymmetricPerChannelQuantizedTensor( + const float* input, int8_t* quantized, TfLiteIntArray* dims, float* scales, + int* zero_points, TfLiteAffineQuantization* affine_quant, + int quantized_dimension, bool is_variable = false); + +// Returns the number of tensors in the default subgraph for a tflite::Model. +size_t GetModelTensorCount(const Model* model); + +// Derives the quantization scaling factor from a min and max range. +template +inline float ScaleFromMinMax(const float min, const float max) { + return (max - min) / + static_cast((std::numeric_limits::max() * 1.0) - + std::numeric_limits::min()); +} + +// Derives the quantization zero point from a min and max range. +template +inline int ZeroPointFromMinMax(const float min, const float max) { + return static_cast(std::numeric_limits::min()) + + static_cast(-min / ScaleFromMinMax(min, max) + 0.5f); +} + +} // namespace testing +} // namespace tflite + +#endif // TENSORFLOW_LITE_MICRO_TEST_HELPERS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/testing/micro_test.h b/src/tensorflow_arm/tensorflow/lite/micro/testing/micro_test.h new file mode 100644 index 0000000..4764d3b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/testing/micro_test.h @@ -0,0 +1,241 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// An ultra-lightweight testing framework designed for use with microcontroller +// applications. Its only dependency is on TensorFlow Lite's ErrorReporter +// interface, where log messages are output. This is designed to be usable even +// when no standard C or C++ libraries are available, and without any dynamic +// memory allocation or reliance on global constructors. +// +// To build a test, you use syntax similar to gunit, but with some extra +// decoration to create a hidden 'main' function containing each of the tests to +// be run. Your code should look something like: +// ---------------------------------------------------------------------------- +// #include "path/to/this/header" +// +// TF_LITE_MICRO_TESTS_BEGIN +// +// TF_LITE_MICRO_TEST(SomeTest) { +// TF_LITE_LOG_EXPECT_EQ(true, true); +// } +// +// TF_LITE_MICRO_TESTS_END +// ---------------------------------------------------------------------------- +// If you compile this for your platform, you'll get a normal binary that you +// should be able to run. Executing it will output logging information like this +// to stderr (or whatever equivalent is available and written to by +// ErrorReporter): +// ---------------------------------------------------------------------------- +// Testing SomeTest +// 1/1 tests passed +// ~~~ALL TESTS PASSED~~~ +// ---------------------------------------------------------------------------- +// This is designed to be human-readable, so you can just run tests manually, +// but the string "~~~ALL TESTS PASSED~~~" should only appear if all of the +// tests do pass. This makes it possible to integrate with automated test +// systems by scanning the output logs and looking for that magic value. +// +// This framework is intended to be a rudimentary alternative to no testing at +// all on systems that struggle to run more conventional approaches, so use with +// caution! + +#ifndef TENSORFLOW_LITE_MICRO_TESTING_MICRO_TEST_H_ +#define TENSORFLOW_LITE_MICRO_TESTING_MICRO_TEST_H_ + +#include "tensorflow_arm/tensorflow/lite/c/common.h" +#include "tensorflow_arm/tensorflow/lite/micro/micro_error_reporter.h" + +namespace micro_test { +extern int tests_passed; +extern int tests_failed; +extern bool is_test_complete; +extern bool did_test_fail; +extern tflite::ErrorReporter* reporter; +} // namespace micro_test + +#define TF_LITE_MICRO_TESTS_BEGIN \ + namespace micro_test { \ + int tests_passed; \ + int tests_failed; \ + bool is_test_complete; \ + bool did_test_fail; \ + tflite::ErrorReporter* reporter; \ + } \ + \ + int tflite_micro_main(int argc, char** argv) { \ + micro_test::tests_passed = 0; \ + micro_test::tests_failed = 0; \ + tflite::MicroErrorReporter error_reporter; \ + micro_test::reporter = &error_reporter; + +#define TF_LITE_MICRO_TESTS_END \ + micro_test::reporter->Report( \ + "%d/%d tests passed", micro_test::tests_passed, \ + (micro_test::tests_failed + micro_test::tests_passed)); \ + if (micro_test::tests_failed == 0) { \ + micro_test::reporter->Report("~~~ALL TESTS PASSED~~~\n"); \ + return kTfLiteOk; \ + } else { \ + micro_test::reporter->Report("~~~SOME TESTS FAILED~~~\n"); \ + return kTfLiteError; \ + } \ + } + +// TODO(petewarden): I'm going to hell for what I'm doing to this poor for loop. +#define TF_LITE_MICRO_TEST(name) \ + micro_test::reporter->Report("Testing " #name); \ + for (micro_test::is_test_complete = false, \ + micro_test::did_test_fail = false; \ + !micro_test::is_test_complete; micro_test::is_test_complete = true, \ + micro_test::tests_passed += (micro_test::did_test_fail) ? 0 : 1, \ + micro_test::tests_failed += (micro_test::did_test_fail) ? 1 : 0) + +#define TF_LITE_MICRO_EXPECT(x) \ + do { \ + if (!(x)) { \ + micro_test::reporter->Report(#x " failed at %s:%d", __FILE__, __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +// TODO(b/139142772): this macro is used with types other than ints even though +// the printf specifier is %d. +#define TF_LITE_MICRO_EXPECT_EQ(x, y) \ + do { \ + auto vx = x; \ + auto vy = y; \ + if ((vx) != (vy)) { \ + micro_test::reporter->Report(#x " == " #y " failed at %s:%d (%d vs %d)", \ + __FILE__, __LINE__, static_cast(vx), \ + static_cast(vy)); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_NE(x, y) \ + do { \ + if ((x) == (y)) { \ + micro_test::reporter->Report(#x " != " #y " failed at %s:%d", __FILE__, \ + __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +// TODO(wangtz): Making it more generic once needed. +#define TF_LITE_MICRO_ARRAY_ELEMENT_EXPECT_NEAR(arr1, idx1, arr2, idx2, \ + epsilon) \ + do { \ + auto delta = ((arr1)[(idx1)] > (arr2)[(idx2)]) \ + ? ((arr1)[(idx1)] - (arr2)[(idx2)]) \ + : ((arr2)[(idx2)] - (arr1)[(idx1)]); \ + if (delta > epsilon) { \ + micro_test::reporter->Report( \ + #arr1 "[%d] (%f) near " #arr2 "[%d] (%f) failed at %s:%d", \ + static_cast(idx1), static_cast((arr1)[(idx1)]), \ + static_cast(idx2), static_cast((arr2)[(idx2)]), \ + __FILE__, __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_NEAR(x, y, epsilon) \ + do { \ + auto vx = (x); \ + auto vy = (y); \ + auto delta = ((vx) > (vy)) ? ((vx) - (vy)) : ((vy) - (vx)); \ + if (delta > epsilon) { \ + micro_test::reporter->Report( \ + #x " (%f) near " #y " (%f) failed at %s:%d", \ + static_cast(vx), static_cast(vy), __FILE__, \ + __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_GT(x, y) \ + do { \ + if ((x) <= (y)) { \ + micro_test::reporter->Report(#x " > " #y " failed at %s:%d", __FILE__, \ + __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_LT(x, y) \ + do { \ + if ((x) >= (y)) { \ + micro_test::reporter->Report(#x " < " #y " failed at %s:%d", __FILE__, \ + __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_GE(x, y) \ + do { \ + if ((x) < (y)) { \ + micro_test::reporter->Report(#x " >= " #y " failed at %s:%d", __FILE__, \ + __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_LE(x, y) \ + do { \ + if ((x) > (y)) { \ + micro_test::reporter->Report(#x " <= " #y " failed at %s:%d", __FILE__, \ + __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_TRUE(x) \ + do { \ + if (!(x)) { \ + micro_test::reporter->Report(#x " was not true failed at %s:%d", \ + __FILE__, __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_FALSE(x) \ + do { \ + if (x) { \ + micro_test::reporter->Report(#x " was not false failed at %s:%d", \ + __FILE__, __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } while (false) + +#define TF_LITE_MICRO_FAIL(msg) \ + do { \ + micro_test::reporter->Report("FAIL: %s", msg, __FILE__, __LINE__); \ + micro_test::did_test_fail = true; \ + } while (false) + +#define TF_LITE_MICRO_EXPECT_STRING_EQ(string1, string2) \ + do { \ + for (int i = 0; string1[i] != '\0' && string2[i] != '\0'; i++) { \ + if (string1[i] != string2[i]) { \ + micro_test::reporter->Report("FAIL: %s did not match %s", string1, \ + string2, __FILE__, __LINE__); \ + micro_test::did_test_fail = true; \ + } \ + } \ + } while (false) + +#endif // TENSORFLOW_LITE_MICRO_TESTING_MICRO_TEST_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.cpp b/src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.cpp new file mode 100644 index 0000000..cdf3604 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.cpp @@ -0,0 +1,1802 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.h" + +extern const unsigned char kTestConvModelData[] = { + 0x24, 0x00, 0x00, 0x00, 0x54, 0x46, 0x4c, 0x33, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x12, 0x00, 0x1c, 0x00, 0x04, 0x00, + 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00, 0x00, 0x00, 0x18, 0x00, + 0x12, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0xb4, 0x52, 0x00, 0x00, + 0x3c, 0x42, 0x00, 0x00, 0x24, 0x42, 0x00, 0x00, 0x3c, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x0c, 0x00, 0x04, 0x00, 0x08, 0x00, 0x08, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, + 0x6d, 0x69, 0x6e, 0x5f, 0x72, 0x75, 0x6e, 0x74, 0x69, 0x6d, 0x65, 0x5f, + 0x76, 0x65, 0x72, 0x73, 0x69, 0x6f, 0x6e, 0x00, 0x0f, 0x00, 0x00, 0x00, + 0xd4, 0x41, 0x00, 0x00, 0xc0, 0x41, 0x00, 0x00, 0x64, 0x41, 0x00, 0x00, + 0xc0, 0x40, 0x00, 0x00, 0x7c, 0x40, 0x00, 0x00, 0x58, 0x40, 0x00, 0x00, + 0x44, 0x13, 0x00, 0x00, 0xa0, 0x12, 0x00, 0x00, 0x8c, 0x00, 0x00, 0x00, + 0x80, 0x00, 0x00, 0x00, 0x6c, 0x00, 0x00, 0x00, 0x58, 0x00, 0x00, 0x00, + 0x44, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0xd6, 0xbe, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x31, 0x2e, 0x35, 0x2e, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x94, 0xb2, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xa4, 0xb2, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0xb4, 0xb2, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0xc4, 0xb2, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xd4, 0xb2, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x00, 0x46, 0xbf, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x12, 0x00, 0x00, 0x7d, 0x6a, 0x24, 0xa1, 0xf6, 0xca, 0x70, 0x2f, + 0x8e, 0xb1, 0xe8, 0x15, 0x42, 0x08, 0x32, 0xf6, 0xe9, 0xfb, 0xa0, 0xda, + 0xe4, 0xf1, 0x0a, 0x9d, 0x72, 0x66, 0x88, 0x37, 0xe9, 0x9e, 0x08, 0x54, + 0x61, 0x51, 0x40, 0x93, 0x4d, 0xcf, 0xe2, 0x08, 0x36, 0xad, 0xb1, 0x8e, + 0xfc, 0xe4, 0x02, 0xd1, 0x9a, 0x1e, 0x05, 0x67, 0xa3, 0x3b, 0xa6, 0xde, + 0x5d, 0x2a, 0xcc, 0x8c, 0x3c, 0x2e, 0xd2, 0x15, 0xc2, 0x60, 0xab, 0xea, + 0x73, 0xe4, 0x88, 0xc1, 0x66, 0x21, 0xb0, 0xe5, 0x5b, 0x55, 0xda, 0x69, + 0x2d, 0x0c, 0x66, 0x07, 0x74, 0x36, 0xcd, 0x79, 0x81, 0xf9, 0x5c, 0x2c, + 0xb5, 0x93, 0xab, 0x76, 0xa1, 0x1f, 0x20, 0x90, 0x89, 0xe1, 0x41, 0xc7, + 0x32, 0xc2, 0xa3, 0x03, 0x77, 0x86, 0x79, 0xf7, 0x89, 0xc1, 0xb1, 0x42, + 0x2a, 0x75, 0xc7, 0xc1, 0x2f, 0xbb, 0xf6, 0xe8, 0x23, 0x99, 0x9b, 0x74, + 0x9c, 0xe5, 0x91, 0x15, 0xc6, 0x08, 0x0e, 0xae, 0x7c, 0xd3, 0x27, 0x54, + 0xfb, 0xa7, 0x49, 0x65, 0x52, 0x2f, 0x63, 0x33, 0x8b, 0x5f, 0x67, 0x21, + 0x25, 0xe0, 0xcf, 0x95, 0x03, 0x05, 0x19, 0x0c, 0x3d, 0xfc, 0x95, 0x42, + 0xa9, 0x26, 0x27, 0x54, 0xa3, 0x71, 0xb4, 0x70, 0x7a, 0x40, 0x0d, 0xc1, + 0x72, 0x04, 0x81, 0x3b, 0xb9, 0xb7, 0xd2, 0xc1, 0x4e, 0xf8, 0xff, 0xca, + 0x66, 0xc1, 0xbe, 0xb9, 0x09, 0xbd, 0xb9, 0x2c, 0x5b, 0x97, 0xc3, 0xa8, + 0xf6, 0xc4, 0x23, 0x93, 0x2e, 0xf6, 0xce, 0x2e, 0xdb, 0xfb, 0x8f, 0xb0, + 0xc8, 0xba, 0xfa, 0x97, 0xfd, 0xc0, 0x0a, 0xc8, 0x2c, 0xf3, 0x4c, 0x4d, + 0x8b, 0x3b, 0x47, 0x11, 0xfb, 0xe8, 0x96, 0xe3, 0xcc, 0xef, 0xe4, 0xb5, + 0x07, 0xa1, 0xb7, 0xa9, 0xf7, 0x98, 0x71, 0x59, 0x9b, 0x5a, 0x7b, 0x88, + 0xe4, 0xcf, 0x9b, 0x55, 0x26, 0xce, 0x59, 0x73, 0x66, 0x17, 0x9c, 0x74, + 0x02, 0xfc, 0x24, 0x01, 0xde, 0x44, 0x98, 0xe3, 0x8b, 0x18, 0x02, 0x42, + 0xf5, 0x0f, 0xbc, 0xcb, 0xf7, 0x37, 0xb1, 0xd5, 0xb4, 0x7c, 0x0a, 0x6a, + 0x59, 0x59, 0xc9, 0x11, 0xd8, 0x0f, 0xf9, 0xab, 0x40, 0xdd, 0x14, 0xf9, + 0x30, 0xaa, 0xf1, 0x8c, 0x6d, 0xbc, 0x4c, 0x5b, 0x71, 0x95, 0xfd, 0x41, + 0x4c, 0xf3, 0xb4, 0x7f, 0x1c, 0xb6, 0x4b, 0x12, 0x3b, 0x6e, 0xc1, 0xce, + 0x6f, 0xf8, 0x57, 0xb7, 0x5e, 0x2a, 0x36, 0x32, 0x3d, 0x85, 0xc6, 0xbf, + 0xd7, 0xab, 0x95, 0x45, 0x62, 0xae, 0xb8, 0xa6, 0x03, 0xcc, 0x21, 0x25, + 0x18, 0x5a, 0xa8, 0x03, 0x27, 0x33, 0x47, 0xb1, 0x7e, 0x0e, 0xbd, 0xc3, + 0x24, 0x25, 0x78, 0x28, 0xa4, 0xe3, 0x5b, 0x08, 0xbf, 0x04, 0xa2, 0xae, + 0x90, 0x4c, 0x96, 0x78, 0xa8, 0xb1, 0xb8, 0x54, 0x89, 0x25, 0x2d, 0x35, + 0x93, 0x95, 0xa5, 0xd3, 0x1a, 0xe6, 0x00, 0x8b, 0xfe, 0x36, 0x0f, 0xd2, + 0x6e, 0xff, 0x86, 0x93, 0x48, 0xb8, 0x08, 0x39, 0x1f, 0x3a, 0x2d, 0xe7, + 0x47, 0x5e, 0x05, 0x66, 0x7a, 0xb8, 0xe4, 0xda, 0xbc, 0x5b, 0x57, 0xdf, + 0xd9, 0x0a, 0xb9, 0x48, 0x5d, 0x0c, 0x57, 0xed, 0x8d, 0xbb, 0x8d, 0x4b, + 0x0e, 0xb8, 0xea, 0x02, 0x06, 0x2f, 0xfd, 0x28, 0x0d, 0x0b, 0xf4, 0xf4, + 0x52, 0x81, 0x77, 0x15, 0x87, 0x53, 0x28, 0xef, 0xbe, 0xc6, 0x4c, 0x45, + 0x3e, 0x1a, 0x6e, 0xbd, 0x10, 0xd8, 0x9a, 0x72, 0x1f, 0x14, 0xe2, 0x37, + 0x08, 0xaf, 0xfa, 0xce, 0xd3, 0x84, 0x23, 0x43, 0x8c, 0x5c, 0xce, 0x1b, + 0xf7, 0xf3, 0xb0, 0x3b, 0xfd, 0x33, 0xf8, 0x09, 0xf1, 0x41, 0xa5, 0xa8, + 0x86, 0x8d, 0x56, 0xde, 0xf6, 0x68, 0xe3, 0x4c, 0x97, 0xa6, 0xc3, 0x66, + 0x9b, 0xa9, 0x8a, 0xbd, 0x59, 0x45, 0xfb, 0xdf, 0xa1, 0x42, 0x10, 0x1c, + 0x55, 0x22, 0x53, 0xe1, 0x32, 0x33, 0xf9, 0xfa, 0xc2, 0x70, 0x0f, 0x49, + 0x15, 0xa7, 0x21, 0xbc, 0x56, 0x35, 0x09, 0x06, 0xe6, 0x5e, 0xc4, 0xc1, + 0x64, 0x93, 0x59, 0x3b, 0x8e, 0xb7, 0x52, 0x6c, 0x4d, 0xa1, 0xb7, 0xee, + 0x14, 0xc2, 0x01, 0x25, 0xbb, 0x5e, 0xe0, 0xc6, 0xa4, 0x4f, 0xb5, 0x20, + 0x88, 0xe0, 0xd7, 0x5e, 0x26, 0x5b, 0x9f, 0xf7, 0xb5, 0x26, 0x5b, 0xfc, + 0xf3, 0x3e, 0xf3, 0x57, 0x6f, 0x9e, 0x9e, 0x51, 0x07, 0x6e, 0xc0, 0x53, + 0x17, 0x89, 0x79, 0xf0, 0x91, 0xb2, 0x54, 0x30, 0x1f, 0x97, 0x95, 0xfc, + 0x02, 0x2d, 0x0c, 0x06, 0xb0, 0x82, 0xad, 0x20, 0xc2, 0xdc, 0x78, 0xbc, + 0xbe, 0x5b, 0x88, 0xa0, 0xdd, 0x45, 0x49, 0x26, 0xec, 0xb4, 0xa5, 0x8b, + 0x7f, 0xdd, 0x40, 0xcf, 0x9e, 0xbe, 0x46, 0x4d, 0x36, 0xab, 0x0a, 0x34, + 0x1a, 0x2a, 0xd0, 0xd3, 0x83, 0x96, 0xff, 0x88, 0xa4, 0xd8, 0x48, 0x75, + 0x2f, 0xcb, 0x3c, 0xc3, 0xbb, 0xc7, 0x2f, 0xe9, 0xf9, 0xa3, 0xde, 0x9d, + 0xbb, 0x5e, 0x37, 0x29, 0xf6, 0x75, 0xcc, 0x85, 0xeb, 0xf9, 0x73, 0xf7, + 0xdc, 0x31, 0x8c, 0x56, 0x52, 0x4a, 0x44, 0xa4, 0x2a, 0x2a, 0x51, 0x49, + 0x77, 0x6d, 0x35, 0x0a, 0xf9, 0x44, 0xaa, 0x36, 0x05, 0xef, 0x1e, 0x6b, + 0xe5, 0x65, 0x6b, 0xaa, 0xc1, 0x41, 0x9c, 0x62, 0xd0, 0x70, 0x78, 0xff, + 0x88, 0xe8, 0x5f, 0x3c, 0x2e, 0x00, 0x6c, 0xe3, 0xdb, 0xc3, 0x54, 0x66, + 0xa9, 0xf4, 0xe2, 0x4c, 0x91, 0x11, 0xc8, 0x3c, 0x39, 0x9b, 0x31, 0x81, + 0xc7, 0x11, 0x22, 0x62, 0xb7, 0x26, 0xa0, 0x0c, 0x2e, 0x6c, 0xe7, 0x34, + 0x3b, 0x1f, 0x27, 0xb3, 0xe5, 0x4f, 0xc9, 0x71, 0xb2, 0x18, 0x99, 0x59, + 0x95, 0xc6, 0x35, 0x4c, 0x5d, 0xa3, 0x59, 0xd1, 0x8b, 0x71, 0xea, 0xe7, + 0x30, 0x3f, 0xe7, 0x8c, 0x1a, 0x59, 0xeb, 0xc5, 0x5d, 0xbd, 0xe6, 0x00, + 0x67, 0x02, 0xfb, 0xca, 0x8d, 0xdf, 0x71, 0xb6, 0xed, 0xc7, 0xd2, 0xf2, + 0x72, 0x1b, 0xd3, 0x63, 0x51, 0x1f, 0x04, 0xe9, 0xf9, 0xe2, 0x38, 0x13, + 0x48, 0x63, 0x19, 0x66, 0x2b, 0x48, 0xc8, 0x1b, 0x9d, 0x19, 0x5a, 0x57, + 0x44, 0x2d, 0x30, 0xb5, 0xce, 0x3b, 0xcc, 0xae, 0xc4, 0x5e, 0x4e, 0x96, + 0x62, 0x5c, 0x53, 0x1f, 0xbf, 0xbd, 0xc8, 0x9d, 0xcf, 0x81, 0xb3, 0x1e, + 0xb0, 0x22, 0xd5, 0xbe, 0x60, 0x65, 0xd9, 0xeb, 0x11, 0x74, 0x8c, 0x24, + 0x18, 0x67, 0x45, 0xd3, 0xf8, 0x3f, 0xc5, 0xdf, 0xac, 0x65, 0xd4, 0x0c, + 0x82, 0x63, 0xd6, 0x43, 0x94, 0xa0, 0x3b, 0xff, 0x03, 0x0f, 0xbb, 0xe4, + 0x4d, 0x3b, 0x41, 0x9f, 0xf4, 0x1a, 0xa9, 0xdb, 0x15, 0x5b, 0x9a, 0x92, + 0xcb, 0xd5, 0xb8, 0x33, 0x5e, 0xea, 0x28, 0x3d, 0x2d, 0x30, 0x20, 0xcd, + 0xb6, 0x23, 0x18, 0x0e, 0x10, 0x2a, 0xa9, 0xe1, 0xad, 0xbc, 0x96, 0xd1, + 0xf9, 0xf3, 0x95, 0x4f, 0x2a, 0x0b, 0x91, 0xff, 0xf0, 0x96, 0x14, 0x00, + 0xaa, 0xfb, 0x1a, 0x44, 0x21, 0x9b, 0xe8, 0x71, 0x31, 0x9e, 0xd6, 0x58, + 0x7f, 0x02, 0x36, 0x5e, 0x92, 0x8d, 0x93, 0x99, 0xac, 0xb6, 0x87, 0x39, + 0xda, 0x47, 0xef, 0x70, 0xd4, 0xf7, 0x8d, 0x2a, 0xbd, 0x08, 0x40, 0x4d, + 0xec, 0xeb, 0x4e, 0x1b, 0x85, 0x5d, 0x55, 0x64, 0x4c, 0xf3, 0x5e, 0x8f, + 0x68, 0x1e, 0x5e, 0x64, 0xc3, 0xb8, 0x92, 0x24, 0x41, 0x98, 0x78, 0x09, + 0x85, 0x87, 0x17, 0x2c, 0x88, 0x9e, 0x62, 0x86, 0x4f, 0x44, 0x71, 0x9c, + 0xa8, 0x73, 0xb3, 0x14, 0x1f, 0x3c, 0x96, 0x6b, 0xab, 0xad, 0x43, 0xdf, + 0x67, 0x34, 0x66, 0x30, 0x1d, 0x15, 0xd3, 0xe7, 0xd5, 0x8b, 0x00, 0xaa, + 0x11, 0x77, 0xea, 0x36, 0xc9, 0x49, 0x99, 0x93, 0x01, 0x6e, 0x00, 0x4a, + 0x93, 0x08, 0x2c, 0x44, 0x01, 0x91, 0xe0, 0x91, 0xdd, 0xab, 0x70, 0x4b, + 0xe7, 0xbf, 0x2d, 0x0f, 0xd4, 0x52, 0xa0, 0xf1, 0x5d, 0xa0, 0xcc, 0xb9, + 0x1b, 0xa2, 0x62, 0xeb, 0x23, 0x1e, 0x8e, 0xbb, 0x2b, 0xb6, 0xc5, 0x3a, + 0xdf, 0x32, 0x99, 0xde, 0x2e, 0x94, 0xcf, 0x98, 0x99, 0x34, 0x59, 0x60, + 0xcf, 0x57, 0xe0, 0xb0, 0xd9, 0x89, 0xaa, 0xc2, 0x4f, 0x1e, 0x38, 0x88, + 0xca, 0x32, 0x93, 0x9b, 0xa3, 0x2b, 0x17, 0x0b, 0x40, 0x5e, 0x69, 0xbd, + 0x14, 0x15, 0xca, 0x1a, 0x21, 0xdf, 0xa8, 0x4e, 0x14, 0x5e, 0x18, 0x40, + 0xe3, 0x4e, 0x04, 0x1f, 0xe5, 0x81, 0x53, 0x11, 0xae, 0x5e, 0x30, 0xe5, + 0xda, 0xd7, 0xf1, 0x3b, 0x72, 0x1b, 0xa5, 0xe3, 0x13, 0xad, 0x40, 0x54, + 0xae, 0xf0, 0xbc, 0x2b, 0xc1, 0x1a, 0x9c, 0xdd, 0xe1, 0xd0, 0x12, 0x10, + 0xfd, 0x59, 0xce, 0x36, 0x60, 0x86, 0xa0, 0xa7, 0xee, 0xe1, 0x02, 0xe6, + 0xf8, 0xf0, 0x5c, 0x4f, 0xa3, 0xa4, 0xe4, 0x09, 0xb9, 0xc3, 0x84, 0xe3, + 0x8d, 0x97, 0x21, 0x62, 0xf3, 0x11, 0x47, 0xb1, 0x4a, 0xce, 0x5b, 0x89, + 0xde, 0x86, 0xb5, 0x0e, 0xba, 0xbc, 0x8c, 0xcf, 0x54, 0x38, 0x3a, 0xc6, + 0xaf, 0x8c, 0x4d, 0x9d, 0xff, 0x58, 0x9b, 0xe8, 0x32, 0xb7, 0xa2, 0x29, + 0xad, 0x91, 0x3a, 0xa5, 0xc7, 0x54, 0xff, 0xd8, 0x47, 0x4f, 0x8f, 0x38, + 0x91, 0x12, 0x76, 0xa3, 0x2e, 0xf7, 0xdd, 0xba, 0xa7, 0xd4, 0x49, 0xe5, + 0xd1, 0x74, 0xe9, 0x2a, 0x29, 0xe4, 0x64, 0xb9, 0x58, 0x98, 0x0c, 0xe5, + 0x1f, 0xb2, 0x0e, 0x33, 0xea, 0xf8, 0x2e, 0xb1, 0x22, 0x46, 0xc2, 0x67, + 0x2d, 0xfe, 0x2e, 0xd3, 0xcf, 0xbc, 0x64, 0x7b, 0x75, 0x24, 0x53, 0x1c, + 0x42, 0x8c, 0x0b, 0x99, 0x9e, 0xa7, 0xa6, 0xb9, 0xfb, 0x5d, 0x86, 0x9f, + 0xe9, 0x04, 0x62, 0xb2, 0x42, 0x81, 0xa2, 0x0d, 0x60, 0x83, 0x40, 0xbb, + 0x21, 0x10, 0xdf, 0xaa, 0xe6, 0x6c, 0x72, 0xc5, 0xb1, 0xad, 0x9f, 0xd2, + 0x91, 0xf8, 0xb6, 0x56, 0xfb, 0x2e, 0xb3, 0xc4, 0x12, 0xd9, 0x86, 0x29, + 0x6c, 0x55, 0x88, 0x72, 0xba, 0xfb, 0x9b, 0xb9, 0x6f, 0x2d, 0x7d, 0x75, + 0xd0, 0x9d, 0xaf, 0x44, 0xb6, 0xbd, 0x7b, 0xec, 0x78, 0xf1, 0xbf, 0x66, + 0xe8, 0x79, 0x66, 0x16, 0x5e, 0xf9, 0x68, 0x89, 0x5b, 0xde, 0x8f, 0xf9, + 0xeb, 0x04, 0x0b, 0x6a, 0x71, 0xa1, 0x3b, 0x46, 0x03, 0xb4, 0x29, 0xa9, + 0x31, 0xf4, 0xc5, 0xd3, 0x43, 0x6d, 0x88, 0x43, 0xa8, 0xef, 0xb7, 0xd7, + 0x75, 0x6b, 0x83, 0x35, 0xb6, 0x2f, 0xe0, 0x5f, 0xf2, 0x14, 0xcd, 0xd0, + 0x06, 0xb3, 0x5e, 0x8b, 0xdb, 0x86, 0x11, 0x94, 0x2f, 0xfb, 0x92, 0x19, + 0x52, 0x7f, 0xcb, 0xe5, 0x22, 0x27, 0x5f, 0xe4, 0x68, 0xb2, 0xcb, 0xc7, + 0xb8, 0xec, 0xfd, 0x9e, 0x39, 0x9c, 0x5b, 0xe4, 0xae, 0xca, 0x83, 0x19, + 0xcf, 0xf0, 0x01, 0xe3, 0xfc, 0xb0, 0x28, 0xda, 0x79, 0x84, 0xfb, 0xfe, + 0xa5, 0xb6, 0xb3, 0xd2, 0x73, 0xd3, 0x11, 0xe5, 0xdf, 0x7a, 0xd7, 0x82, + 0x78, 0x25, 0x06, 0x5b, 0x0f, 0x89, 0x9d, 0x0b, 0x9b, 0xd1, 0x1b, 0xc5, + 0xb7, 0x67, 0xef, 0x7c, 0xa2, 0xa3, 0xca, 0x27, 0xd0, 0x59, 0xb9, 0x99, + 0x86, 0xa9, 0xf6, 0x9a, 0x28, 0xf0, 0xbb, 0x42, 0xd2, 0xa0, 0xa8, 0x01, + 0x29, 0xa1, 0x0c, 0x1b, 0x33, 0x1b, 0x9c, 0xcb, 0xe4, 0x6c, 0x61, 0x0a, + 0xc4, 0xd7, 0x6c, 0xec, 0x86, 0xb3, 0xd2, 0xaa, 0x8c, 0xab, 0x1a, 0xf4, + 0x03, 0x2e, 0x2b, 0x42, 0xbe, 0xc1, 0x31, 0x1d, 0x57, 0x47, 0xdc, 0x7b, + 0xb5, 0x8f, 0x8b, 0xdf, 0x06, 0xad, 0x3f, 0xf4, 0x4f, 0xb5, 0x52, 0x07, + 0x4e, 0x25, 0xb3, 0x73, 0x34, 0x92, 0x6a, 0x89, 0x93, 0x28, 0x8b, 0x96, + 0x9d, 0xdb, 0xb4, 0x77, 0x81, 0x76, 0x86, 0xd2, 0xa5, 0x94, 0x76, 0x35, + 0xc9, 0x66, 0x4e, 0xd8, 0xc5, 0xc3, 0xc9, 0x34, 0xaf, 0xad, 0x4a, 0x7c, + 0x92, 0x24, 0xb1, 0x7d, 0x7d, 0xac, 0xf6, 0xcb, 0x8f, 0x36, 0xc1, 0xb2, + 0x63, 0x78, 0x99, 0x33, 0x23, 0x68, 0x6e, 0x71, 0x6a, 0xcc, 0x05, 0xf9, + 0x41, 0x92, 0x30, 0xf0, 0xb1, 0xb4, 0xa6, 0x46, 0x86, 0x62, 0xd9, 0xd9, + 0x94, 0x8a, 0xb2, 0x9c, 0x68, 0xff, 0xf4, 0x3a, 0x2e, 0xaf, 0xee, 0xcf, + 0x04, 0x94, 0x53, 0x35, 0x25, 0xf9, 0xaa, 0x74, 0x93, 0xf3, 0x63, 0xc0, + 0xd2, 0x22, 0x30, 0x8c, 0xde, 0xa6, 0xb1, 0xb4, 0xa1, 0x56, 0x07, 0x06, + 0x71, 0xa2, 0x9e, 0x42, 0x31, 0xa3, 0x1e, 0xa6, 0x9a, 0xbc, 0x9f, 0x5b, + 0x12, 0x3c, 0xc2, 0x74, 0xf9, 0x61, 0x71, 0xef, 0x73, 0x86, 0xc2, 0x3b, + 0x25, 0x8a, 0x31, 0x72, 0x27, 0xac, 0xa4, 0x72, 0xf3, 0xbb, 0x78, 0x2c, + 0x94, 0xed, 0xa8, 0x3a, 0x42, 0x98, 0x34, 0xda, 0x3e, 0x60, 0x1c, 0x4a, + 0xec, 0x6b, 0x4e, 0x5f, 0x2a, 0x62, 0xb9, 0xad, 0xc9, 0xd9, 0x38, 0x90, + 0xa7, 0x3b, 0xd3, 0x1a, 0xbb, 0x81, 0x0d, 0x33, 0xd9, 0x16, 0x35, 0x8e, + 0xc3, 0x88, 0x36, 0xfa, 0x3e, 0xa8, 0x4f, 0x30, 0x9d, 0xf1, 0x08, 0xea, + 0x40, 0x1b, 0x87, 0x4d, 0x23, 0x8e, 0x8e, 0xb0, 0xe2, 0xf0, 0x27, 0xc1, + 0xdc, 0x0d, 0xe2, 0x8f, 0x93, 0xef, 0x8b, 0xd1, 0x19, 0xa5, 0xbe, 0xd7, + 0x5a, 0x8a, 0x38, 0x62, 0x43, 0xba, 0x74, 0xf8, 0xae, 0x11, 0x1f, 0x1d, + 0xa4, 0x6e, 0x70, 0x94, 0x91, 0x14, 0xf4, 0xff, 0xbe, 0x39, 0xb4, 0x33, + 0xc2, 0x87, 0x74, 0x1b, 0xfd, 0x9a, 0xa8, 0x64, 0x09, 0x4b, 0x7f, 0x95, + 0x0a, 0xcb, 0x6b, 0x15, 0x54, 0x1d, 0xc6, 0x03, 0x1d, 0x1b, 0x25, 0x56, + 0x15, 0xb5, 0xd7, 0xe5, 0xd6, 0xf3, 0x28, 0xa4, 0xde, 0x1b, 0x39, 0x0d, + 0x59, 0x26, 0x12, 0xe4, 0x32, 0xf2, 0x25, 0xeb, 0xc0, 0xdb, 0x58, 0xe5, + 0xce, 0x64, 0x6f, 0x70, 0x74, 0xc1, 0xc9, 0xbd, 0x75, 0xef, 0x16, 0x02, + 0xdf, 0x27, 0x09, 0xc8, 0xb8, 0x37, 0x8f, 0x44, 0x0d, 0x58, 0x48, 0xf5, + 0xc2, 0x53, 0x21, 0x28, 0x16, 0xa4, 0x56, 0x02, 0xdf, 0xa7, 0x97, 0xa4, + 0x5c, 0x48, 0x75, 0x51, 0x89, 0x0b, 0xa7, 0x4d, 0xd9, 0x9e, 0x04, 0x4e, + 0x5d, 0x6c, 0xe5, 0x1f, 0x68, 0x88, 0xcc, 0xb7, 0x9a, 0x20, 0x05, 0x83, + 0x82, 0x6c, 0xfd, 0xdb, 0x07, 0x6c, 0xec, 0x61, 0xaa, 0x36, 0x57, 0x68, + 0x01, 0xf2, 0x70, 0xfe, 0xe6, 0x4d, 0xe1, 0xa9, 0xb6, 0xb6, 0x52, 0xe6, + 0x20, 0x52, 0x0f, 0x27, 0x9a, 0x1c, 0x2d, 0x20, 0x9b, 0xd4, 0x07, 0xd3, + 0xf6, 0x85, 0x4b, 0xf2, 0x52, 0x4d, 0x4c, 0xd7, 0xf0, 0x32, 0x5d, 0x2e, + 0xef, 0xa2, 0xd0, 0xcd, 0x48, 0x89, 0xbc, 0x9f, 0xcb, 0x37, 0x02, 0x29, + 0xa5, 0xdb, 0xab, 0xfa, 0x1d, 0xf4, 0x53, 0x78, 0x30, 0xde, 0x2c, 0x5c, + 0x35, 0x7f, 0x3d, 0xe1, 0xe0, 0xce, 0xdb, 0x13, 0xca, 0x2a, 0xae, 0xdf, + 0x1c, 0xb1, 0xb6, 0xb9, 0x6a, 0x9f, 0x28, 0xb0, 0x54, 0x5a, 0x00, 0xdd, + 0x76, 0x14, 0xfb, 0x17, 0xc2, 0x2a, 0x45, 0xa2, 0x18, 0xbb, 0x8a, 0x3e, + 0xbe, 0x0e, 0xa5, 0x1b, 0x3c, 0x70, 0x56, 0x10, 0x98, 0xec, 0xc6, 0x3a, + 0x95, 0x2a, 0x96, 0x6a, 0x44, 0xef, 0xd9, 0x9c, 0x2a, 0x45, 0xb4, 0x15, + 0xf8, 0x2e, 0x03, 0x5d, 0x8c, 0x79, 0xfb, 0xb0, 0x53, 0x71, 0xcd, 0x0d, + 0xf4, 0xe2, 0xfc, 0x3b, 0x71, 0xee, 0x30, 0xf2, 0x29, 0xd3, 0xaa, 0x18, + 0x7a, 0x45, 0x1d, 0x99, 0x6d, 0x2f, 0x1f, 0x2d, 0x32, 0x23, 0x48, 0xc2, + 0x69, 0x33, 0x3d, 0x04, 0xa7, 0xa3, 0x96, 0xb5, 0x76, 0x5b, 0x4e, 0xb7, + 0x3c, 0x10, 0x58, 0x17, 0xf4, 0x5f, 0xec, 0x51, 0x6d, 0x5a, 0x3b, 0x7f, + 0x1e, 0x0e, 0xbb, 0xbf, 0x77, 0x43, 0xf7, 0xa4, 0x57, 0xc0, 0x33, 0xac, + 0xc1, 0xe3, 0x3e, 0x1f, 0x65, 0x3c, 0x62, 0x19, 0x46, 0x2d, 0x7b, 0x2d, + 0x07, 0x44, 0x48, 0xf4, 0x91, 0xdf, 0x59, 0x32, 0x10, 0xf7, 0x12, 0xe2, + 0xe5, 0x39, 0x70, 0x37, 0xa4, 0x79, 0x9a, 0x17, 0x19, 0xe8, 0x90, 0xe7, + 0x37, 0x0d, 0xb6, 0x6d, 0x58, 0xe6, 0x7e, 0x57, 0x76, 0x8a, 0xe8, 0xd0, + 0x76, 0x30, 0x25, 0xda, 0xb6, 0xdf, 0x59, 0x3c, 0x6c, 0x20, 0x65, 0x88, + 0xd2, 0x60, 0x5e, 0x39, 0xb6, 0x6b, 0xac, 0xa2, 0x25, 0xc6, 0xa7, 0xb1, + 0x2f, 0xbb, 0x1d, 0x23, 0xee, 0x02, 0x08, 0x1d, 0xd6, 0x6c, 0x0e, 0xbc, + 0xea, 0xd2, 0xc2, 0x70, 0x34, 0xe9, 0x96, 0xd3, 0xf3, 0xf4, 0x8e, 0x94, + 0x6f, 0x86, 0x76, 0xe7, 0x38, 0x08, 0x6f, 0x47, 0xf5, 0xcd, 0xab, 0xad, + 0x7a, 0x39, 0x10, 0x9a, 0xa8, 0x44, 0xba, 0x2d, 0x7f, 0x05, 0x1e, 0xb7, + 0x44, 0xd8, 0x10, 0x05, 0xd1, 0x8d, 0x98, 0x09, 0x14, 0xbb, 0x6b, 0x2b, + 0xf7, 0xeb, 0x9f, 0xa5, 0x65, 0x4b, 0x21, 0xff, 0xaf, 0xe8, 0x2e, 0x34, + 0x52, 0x38, 0xcf, 0xd5, 0x51, 0x29, 0x2c, 0x91, 0x43, 0x3a, 0x49, 0x42, + 0xdd, 0xfb, 0x0e, 0xd2, 0x77, 0x8f, 0x65, 0x93, 0x3e, 0x52, 0x22, 0x58, + 0xd6, 0xf9, 0xd9, 0x58, 0xd4, 0x06, 0xa9, 0x0c, 0x79, 0x9f, 0x1b, 0xa5, + 0x45, 0x61, 0xd8, 0x4e, 0xbf, 0x4b, 0x51, 0xe2, 0xfb, 0x6f, 0x58, 0xee, + 0xc5, 0xa5, 0x11, 0xbd, 0x99, 0x25, 0x14, 0xac, 0x94, 0x0e, 0xd1, 0xf7, + 0x54, 0xb6, 0x05, 0x8c, 0xc3, 0x57, 0xa5, 0x3c, 0x3c, 0xa6, 0x83, 0x47, + 0x38, 0xd1, 0x6a, 0xab, 0x12, 0xc0, 0xd3, 0x7f, 0x96, 0x55, 0xd7, 0xf4, + 0x3a, 0xd0, 0x08, 0x85, 0x5f, 0x3d, 0x65, 0x8e, 0xbb, 0xea, 0x34, 0xf3, + 0x53, 0x96, 0x71, 0x08, 0x9b, 0x50, 0xe9, 0x4b, 0xce, 0x8a, 0x2f, 0xef, + 0xe4, 0xb2, 0x72, 0x68, 0xcb, 0x88, 0xa8, 0xd9, 0xd9, 0xa2, 0xfc, 0x62, + 0xe8, 0x8b, 0x23, 0x2b, 0xbc, 0xf0, 0x9e, 0xb4, 0xd0, 0x40, 0x8b, 0x45, + 0xff, 0x6d, 0x37, 0x01, 0xa6, 0x4b, 0x62, 0xe0, 0x3b, 0x4e, 0x18, 0x67, + 0xb3, 0x97, 0x04, 0xa0, 0x2a, 0xf2, 0x11, 0x79, 0x38, 0xb4, 0xb2, 0xed, + 0x64, 0xc1, 0x1e, 0xfe, 0xc4, 0xf4, 0xe2, 0x4d, 0x94, 0xb4, 0x17, 0x52, + 0x1a, 0x63, 0xe6, 0x56, 0x8a, 0x41, 0x0a, 0x5b, 0xa2, 0x1c, 0x59, 0xef, + 0x17, 0x64, 0xf9, 0xf7, 0x2c, 0xa4, 0xfd, 0x66, 0xf7, 0xe3, 0xae, 0xa0, + 0x54, 0x36, 0x64, 0x26, 0x84, 0x51, 0x49, 0xd5, 0x3a, 0x5e, 0x2c, 0xc5, + 0xca, 0xde, 0x8e, 0xe7, 0x25, 0x59, 0xb3, 0x9a, 0xb2, 0xf0, 0xff, 0xf1, + 0x83, 0xe5, 0x70, 0xc3, 0xef, 0x63, 0x66, 0x31, 0x04, 0x4d, 0x42, 0xf1, + 0xd9, 0x4c, 0x5e, 0x29, 0x92, 0x37, 0x8d, 0xd1, 0x18, 0x2a, 0x9e, 0x3c, + 0xcc, 0x05, 0xb9, 0xc4, 0xb6, 0xe7, 0x2a, 0x09, 0x3a, 0x68, 0xb5, 0x61, + 0x60, 0x36, 0x11, 0x02, 0x92, 0xf8, 0xa0, 0x56, 0x9b, 0xe8, 0xfe, 0xac, + 0x87, 0xcc, 0xaf, 0xb9, 0x62, 0xa7, 0x1e, 0x99, 0xb8, 0x9f, 0x47, 0xf7, + 0xa5, 0x12, 0x47, 0x66, 0xeb, 0xd6, 0x3a, 0x6f, 0xb3, 0x26, 0x63, 0xe2, + 0xec, 0x0c, 0xba, 0x7d, 0xc2, 0x9b, 0xb2, 0x10, 0x62, 0x03, 0x3f, 0x20, + 0xed, 0x7a, 0xce, 0x47, 0xd0, 0x50, 0x5b, 0x5c, 0x66, 0xbf, 0x01, 0x09, + 0x84, 0x0b, 0x71, 0xa8, 0x1f, 0x8d, 0xe1, 0x05, 0x09, 0xb4, 0xd5, 0x34, + 0xf1, 0xba, 0x31, 0xc6, 0x76, 0x8e, 0x00, 0x96, 0x3d, 0x6b, 0xe4, 0x66, + 0x3a, 0x22, 0xcd, 0x7f, 0x9d, 0xf8, 0x64, 0xfc, 0x76, 0x42, 0x88, 0x0e, + 0x32, 0xa5, 0xd0, 0x69, 0x56, 0xe2, 0xa5, 0x6f, 0xbb, 0xfa, 0xd8, 0xde, + 0xb4, 0x23, 0xa9, 0xc7, 0x9a, 0xc1, 0x99, 0xa7, 0x7f, 0x79, 0x58, 0xe1, + 0xe7, 0xc5, 0x56, 0x36, 0xc0, 0xfb, 0x8d, 0x8f, 0xe4, 0x6c, 0x96, 0x89, + 0xcb, 0xb0, 0xb0, 0x6e, 0xee, 0x20, 0x46, 0xd3, 0x43, 0x83, 0xac, 0x39, + 0x7c, 0x25, 0xba, 0x69, 0x3a, 0x58, 0x8a, 0x48, 0x0a, 0xf7, 0xb7, 0xfc, + 0x58, 0x7b, 0x93, 0x8b, 0xcd, 0x81, 0x7e, 0x94, 0xe0, 0xdf, 0xb1, 0xca, + 0xf6, 0x60, 0x54, 0xa9, 0x6e, 0xc6, 0x7f, 0xac, 0xfb, 0x62, 0xfe, 0xd9, + 0xd5, 0xf4, 0x6c, 0x62, 0x65, 0xf6, 0x0b, 0x24, 0x49, 0x1d, 0x55, 0xd6, + 0x4c, 0x0b, 0x5a, 0xf1, 0x2e, 0x78, 0x7a, 0x4e, 0xc1, 0xd0, 0xdb, 0xfe, + 0xd2, 0x84, 0x60, 0x68, 0x51, 0x8e, 0x3f, 0xf1, 0xa8, 0x90, 0xbf, 0xda, + 0x86, 0xda, 0x41, 0xd8, 0x90, 0x7b, 0xc3, 0xc8, 0x9e, 0xa5, 0x77, 0x06, + 0x56, 0x02, 0x13, 0x59, 0xaa, 0x89, 0xf9, 0xd5, 0x3c, 0x1d, 0xe2, 0xa9, + 0xb1, 0xc8, 0x02, 0x5a, 0x1c, 0xae, 0x72, 0x66, 0xdf, 0xb4, 0x1a, 0xb7, + 0xd2, 0x4d, 0xda, 0x4f, 0xc9, 0xed, 0x88, 0x7d, 0x9b, 0xc4, 0x4a, 0x8c, + 0x5e, 0x77, 0xaf, 0xd6, 0xd3, 0xbb, 0x38, 0xd2, 0xfa, 0x85, 0xe4, 0xdd, + 0xe7, 0x6e, 0xcb, 0x0b, 0x34, 0x1e, 0xa8, 0xfd, 0xf4, 0xd2, 0xc3, 0xdd, + 0xe0, 0xa6, 0xb1, 0x78, 0x16, 0x85, 0x2b, 0x1b, 0x22, 0xa6, 0xd5, 0x93, + 0x4f, 0xa1, 0xd5, 0x10, 0x96, 0xab, 0x38, 0xa7, 0x3c, 0xf2, 0xbd, 0xd9, + 0x7c, 0x59, 0x71, 0x25, 0x6f, 0x7c, 0xce, 0x73, 0x8e, 0x4e, 0xfb, 0x5a, + 0x30, 0x24, 0x53, 0xc5, 0xa3, 0x20, 0x13, 0x03, 0xfc, 0x7a, 0xaf, 0x1f, + 0x71, 0x5d, 0x6b, 0xce, 0x2e, 0x92, 0x16, 0x4d, 0xab, 0x96, 0x10, 0xc0, + 0xf6, 0x3c, 0xfe, 0x51, 0x89, 0x4d, 0x39, 0x45, 0x2c, 0x92, 0x5a, 0x86, + 0x24, 0xce, 0xbc, 0x75, 0xc6, 0x7f, 0x0e, 0xc2, 0xd1, 0xe7, 0x6a, 0x75, + 0x30, 0x59, 0xfb, 0xbf, 0x6b, 0xcf, 0x60, 0x90, 0x07, 0x73, 0xb1, 0x47, + 0x6e, 0x5d, 0xcd, 0x44, 0xac, 0xee, 0x2a, 0xdb, 0x16, 0x5a, 0x1a, 0xaf, + 0xba, 0xf8, 0x64, 0xdd, 0xdd, 0xed, 0x46, 0x4b, 0x67, 0xf3, 0xf8, 0x2d, + 0x22, 0xe9, 0x25, 0x74, 0x4c, 0x70, 0xe0, 0x3d, 0xbc, 0x11, 0xd3, 0x56, + 0xec, 0x86, 0x39, 0x89, 0x4c, 0xf2, 0xbc, 0x39, 0xdc, 0xde, 0x5f, 0x3b, + 0x42, 0xcb, 0xf6, 0x0c, 0x49, 0x8c, 0x66, 0x76, 0x58, 0x28, 0xe8, 0x47, + 0x59, 0x40, 0x11, 0xef, 0xb5, 0x9d, 0x93, 0xe5, 0x39, 0x56, 0x62, 0x0d, + 0xd0, 0xdd, 0xbb, 0x51, 0xff, 0x87, 0xa3, 0xd1, 0x9e, 0x0e, 0x0c, 0xbd, + 0x8e, 0xfc, 0xa5, 0x44, 0xc7, 0x6d, 0x35, 0x1d, 0x69, 0x14, 0x5b, 0x0d, + 0x45, 0xff, 0x85, 0x2d, 0xd1, 0x14, 0xf4, 0x5e, 0x5b, 0x49, 0x85, 0xad, + 0x69, 0xf1, 0x34, 0x9e, 0x7a, 0xf3, 0xed, 0x2d, 0xf2, 0x5f, 0x70, 0x5a, + 0xc1, 0xca, 0x63, 0xb5, 0xec, 0x49, 0xfc, 0x88, 0xcb, 0x0f, 0x81, 0x1d, + 0xd4, 0x2f, 0x18, 0xf6, 0xfe, 0x71, 0x51, 0xe2, 0x25, 0x71, 0x48, 0xa4, + 0xb2, 0x9f, 0x4f, 0xc0, 0xa5, 0x24, 0x12, 0x5b, 0xf8, 0xf2, 0xcf, 0x6e, + 0x52, 0x52, 0x6a, 0xee, 0x7d, 0xa5, 0x9b, 0xdb, 0x9c, 0xc9, 0x35, 0x30, + 0x1a, 0xf0, 0x7d, 0xcc, 0x98, 0x73, 0x09, 0x16, 0x8c, 0x05, 0x8d, 0x70, + 0xa3, 0x15, 0xd6, 0x7a, 0xa0, 0x7c, 0xd5, 0xcc, 0xd3, 0x29, 0x32, 0x2e, + 0xa5, 0xde, 0xf6, 0xd3, 0xa4, 0x03, 0x59, 0x6c, 0x05, 0x2d, 0x0e, 0x8b, + 0xb7, 0x1f, 0xa0, 0x57, 0x5c, 0x76, 0xde, 0x81, 0xcb, 0x64, 0xb9, 0x73, + 0xc1, 0x3b, 0x26, 0xba, 0x16, 0xdb, 0xe6, 0x40, 0x23, 0xa4, 0xe9, 0x24, + 0x48, 0xb8, 0x73, 0x23, 0x67, 0xbf, 0x26, 0xca, 0x95, 0x4f, 0xa0, 0x60, + 0x95, 0xa2, 0x0f, 0x29, 0xed, 0x5d, 0x71, 0x66, 0x94, 0xa3, 0xd0, 0x2a, + 0x4e, 0x17, 0x32, 0x18, 0xe6, 0xd6, 0x75, 0x84, 0xa5, 0x2a, 0x72, 0x18, + 0x60, 0x85, 0xde, 0x66, 0x22, 0x52, 0xf6, 0x45, 0xd6, 0xf0, 0xed, 0x93, + 0x0f, 0x5a, 0xa9, 0x12, 0x2a, 0xc4, 0xa8, 0x3d, 0x97, 0xc9, 0xc7, 0x84, + 0x71, 0x14, 0xb3, 0x54, 0xb6, 0xf7, 0x92, 0x7a, 0xc0, 0x6e, 0x02, 0xf7, + 0x48, 0xdb, 0x7c, 0xc1, 0x45, 0x21, 0xdb, 0x1b, 0x51, 0xc3, 0xea, 0xc0, + 0x19, 0x31, 0xe4, 0x6c, 0x20, 0x5f, 0x08, 0xe7, 0x88, 0xf7, 0xc0, 0x6e, + 0xee, 0x5f, 0x20, 0x33, 0x68, 0xef, 0xc5, 0x33, 0x1b, 0x40, 0x66, 0xc5, + 0xa3, 0x68, 0xdb, 0xbc, 0x8a, 0xb7, 0x54, 0xdb, 0xc7, 0xc5, 0x2c, 0x42, + 0x65, 0x51, 0xab, 0x56, 0x94, 0x73, 0xec, 0xd9, 0x95, 0xfa, 0x6a, 0x56, + 0xef, 0x22, 0x95, 0xa4, 0x75, 0x46, 0xee, 0x60, 0x8b, 0x25, 0xa6, 0x92, + 0x0a, 0x8e, 0xc1, 0x39, 0x97, 0x69, 0xa9, 0x19, 0x97, 0xf1, 0x0f, 0x61, + 0xc2, 0x40, 0x7d, 0x62, 0xe9, 0x5e, 0x22, 0x1f, 0x27, 0xe5, 0xc7, 0xe7, + 0xa4, 0x35, 0x5d, 0x90, 0xc7, 0x38, 0x38, 0x2d, 0xb0, 0x1e, 0x29, 0x0f, + 0x4f, 0x08, 0x8b, 0xdd, 0x69, 0x3c, 0x5c, 0x03, 0xbe, 0x9a, 0x76, 0xba, + 0x91, 0xf5, 0x57, 0x07, 0x39, 0xfe, 0x09, 0xfc, 0x01, 0x7b, 0x37, 0xc4, + 0x73, 0x7f, 0x76, 0x50, 0x76, 0xae, 0x6e, 0x4b, 0x22, 0x2c, 0x3b, 0xe7, + 0x77, 0x19, 0x9a, 0x92, 0x26, 0xdf, 0xc4, 0xe6, 0xd8, 0x57, 0xc1, 0x7f, + 0x65, 0x0b, 0xfb, 0xfa, 0xdd, 0xd2, 0x8c, 0xc7, 0xb1, 0x72, 0x2a, 0xb2, + 0x5a, 0xfa, 0xb2, 0x84, 0xb1, 0xec, 0x79, 0x9e, 0xde, 0xd8, 0x2f, 0xdf, + 0x3b, 0x39, 0x0b, 0xac, 0xfa, 0xb8, 0x07, 0x38, 0xff, 0x2e, 0x22, 0x2b, + 0xc9, 0x31, 0x3b, 0x09, 0x05, 0xd2, 0x06, 0xc4, 0x2d, 0x22, 0x1c, 0x21, + 0x70, 0x03, 0x93, 0xd1, 0x3a, 0x8d, 0x94, 0x60, 0xfe, 0x99, 0x13, 0xc3, + 0x00, 0x03, 0x41, 0xfa, 0x50, 0x79, 0x31, 0xeb, 0xf0, 0xf4, 0x06, 0x7a, + 0x19, 0xe8, 0x90, 0xdf, 0x61, 0x4d, 0x5f, 0xe3, 0x99, 0x1b, 0xca, 0xbf, + 0xcf, 0xae, 0xca, 0xfa, 0x84, 0x63, 0x88, 0x56, 0x1d, 0x52, 0x5a, 0x21, + 0xf9, 0xcd, 0xa3, 0x30, 0x16, 0xb9, 0x0d, 0xe1, 0x87, 0x08, 0x78, 0xa2, + 0xdb, 0x7e, 0x16, 0x82, 0x48, 0x48, 0x17, 0x1a, 0xa8, 0x3f, 0xc7, 0x4d, + 0xfd, 0x99, 0x2b, 0x36, 0xbf, 0x08, 0xb9, 0xeb, 0xa6, 0xbf, 0xb6, 0xa0, + 0x9e, 0x26, 0x15, 0xac, 0xd2, 0x65, 0xc9, 0x36, 0x41, 0xe3, 0x59, 0x4e, + 0xdc, 0x7b, 0x58, 0x3b, 0x47, 0x0b, 0xc9, 0xf3, 0xb3, 0xf9, 0x81, 0x33, + 0x39, 0xca, 0xf8, 0x97, 0x2d, 0x9b, 0x24, 0x33, 0x69, 0xbe, 0x1b, 0x81, + 0x59, 0x59, 0x17, 0xed, 0x7d, 0x5b, 0xbe, 0xda, 0xeb, 0x4e, 0x5d, 0x5d, + 0x70, 0x13, 0x3c, 0x4b, 0x4a, 0xfc, 0xa4, 0xbe, 0xa0, 0x5d, 0xa2, 0xed, + 0xe8, 0x8d, 0xf8, 0xf2, 0xa5, 0xdd, 0xd4, 0x49, 0x45, 0x04, 0xef, 0x18, + 0x9f, 0xa1, 0xf7, 0xc4, 0x3b, 0xc2, 0x6b, 0xe0, 0x45, 0xa8, 0x76, 0x39, + 0x49, 0x32, 0xec, 0xc3, 0xcb, 0x45, 0x46, 0xd2, 0x4b, 0x3a, 0x55, 0xe5, + 0xce, 0x08, 0xc4, 0x84, 0xe5, 0xd9, 0xb3, 0xf3, 0xc4, 0xa8, 0xe9, 0x88, + 0x83, 0xd5, 0x56, 0xe1, 0xa6, 0xef, 0x41, 0x55, 0xb0, 0x3f, 0xa3, 0xc1, + 0xbe, 0x3b, 0x83, 0xd6, 0x92, 0x90, 0x38, 0xd3, 0xf3, 0x75, 0xf6, 0x49, + 0x95, 0xee, 0xa9, 0xed, 0xaa, 0xf8, 0xb9, 0x14, 0x0e, 0x6a, 0x48, 0x9d, + 0xc5, 0x48, 0x3b, 0x5e, 0x61, 0xd3, 0x8c, 0x4a, 0x10, 0x12, 0x7c, 0x0a, + 0xf7, 0xaf, 0x62, 0x2d, 0xd3, 0x89, 0x8d, 0x75, 0x19, 0x6b, 0x62, 0x4b, + 0x1a, 0x04, 0xc7, 0xd3, 0x32, 0x17, 0x2f, 0x5f, 0x29, 0xfa, 0xb1, 0x8d, + 0x78, 0xe7, 0x27, 0xf6, 0x67, 0x7e, 0x17, 0xa3, 0x18, 0xdc, 0x13, 0x08, + 0x1e, 0x4b, 0xc7, 0x8e, 0xf6, 0xba, 0x90, 0xb3, 0x32, 0x42, 0x37, 0x6b, + 0x60, 0xa9, 0x23, 0xb5, 0x89, 0x57, 0x7b, 0xdb, 0x98, 0x35, 0x1f, 0x95, + 0x86, 0xa5, 0x83, 0x36, 0xd1, 0x8c, 0x8e, 0xc0, 0x77, 0x5c, 0x40, 0x8e, + 0xec, 0xdf, 0x25, 0x69, 0x0a, 0x83, 0x8f, 0xdf, 0x91, 0x52, 0x31, 0xab, + 0xd5, 0x61, 0x37, 0xbd, 0x83, 0x1d, 0x4c, 0x8b, 0xa1, 0x4a, 0x81, 0x8b, + 0xa0, 0xf4, 0x41, 0xbd, 0x54, 0x36, 0x36, 0x56, 0x6d, 0x4c, 0xe7, 0xd9, + 0xc7, 0x09, 0xd9, 0x4b, 0xf0, 0x54, 0x45, 0x3c, 0x62, 0x47, 0x17, 0x54, + 0x1f, 0x55, 0x2f, 0x74, 0xdc, 0x11, 0xe9, 0xa3, 0xb5, 0x75, 0xe9, 0x10, + 0xde, 0x62, 0xa9, 0x24, 0x39, 0xd4, 0x17, 0xbb, 0x15, 0xe4, 0x48, 0x09, + 0x26, 0x6a, 0xbd, 0x3b, 0x10, 0xa1, 0x55, 0xe5, 0x99, 0x53, 0x1e, 0xd2, + 0xee, 0x7c, 0x54, 0xd8, 0x06, 0x8b, 0x1e, 0xe7, 0x3f, 0x08, 0x38, 0x9b, + 0x2e, 0x41, 0xdf, 0x0b, 0x7e, 0x83, 0x7f, 0x04, 0x38, 0xa5, 0x1f, 0x46, + 0x8b, 0x94, 0x28, 0x9f, 0xb8, 0x8c, 0x41, 0xfe, 0x96, 0xe2, 0x24, 0xd1, + 0x97, 0xa4, 0xcb, 0xba, 0xfa, 0x19, 0xc9, 0x57, 0x30, 0x0f, 0x88, 0x58, + 0xa9, 0x67, 0x31, 0x74, 0x51, 0x34, 0x03, 0xbc, 0xff, 0x3b, 0x12, 0x61, + 0x84, 0x63, 0x74, 0xec, 0x4d, 0xda, 0xa3, 0x56, 0xc3, 0xe5, 0x5e, 0x4a, + 0x03, 0x26, 0x88, 0x1a, 0x1d, 0x7f, 0xe8, 0x3f, 0x61, 0x78, 0xb6, 0xc5, + 0x66, 0xb7, 0xb4, 0xc1, 0xe7, 0x82, 0xc1, 0x44, 0xdf, 0xf9, 0x30, 0x30, + 0xe1, 0xd0, 0xf8, 0xf5, 0x40, 0x5a, 0x72, 0x29, 0xef, 0x30, 0xe1, 0x01, + 0xca, 0x1b, 0xb0, 0xa6, 0xa3, 0x17, 0x2b, 0x58, 0x03, 0xda, 0x25, 0x0f, + 0xdc, 0x49, 0x7c, 0xc5, 0x8f, 0x2d, 0x83, 0xca, 0x43, 0x08, 0xc0, 0x36, + 0x70, 0x1e, 0x42, 0xfd, 0xac, 0x4d, 0x31, 0xcf, 0x68, 0x4a, 0xda, 0xd8, + 0xcb, 0xee, 0xaa, 0xfc, 0xcf, 0xcc, 0xe6, 0xb2, 0x77, 0x8b, 0x83, 0x5b, + 0xd5, 0x3d, 0x55, 0xba, 0x03, 0x45, 0xce, 0x51, 0x78, 0x36, 0xcb, 0xcd, + 0x9a, 0x0f, 0x58, 0xbe, 0x15, 0x10, 0xdb, 0x3f, 0x1d, 0x28, 0x27, 0x11, + 0x69, 0xca, 0x95, 0x68, 0xa8, 0xc8, 0xff, 0x0c, 0x3f, 0xd5, 0x11, 0x91, + 0x35, 0x45, 0x35, 0x9d, 0x1c, 0x58, 0xa2, 0xe5, 0xab, 0x83, 0x95, 0x10, + 0x44, 0xd4, 0xc0, 0x27, 0xf4, 0xc2, 0x72, 0x0f, 0x1a, 0x3d, 0x1c, 0xf2, + 0x7f, 0xb9, 0x54, 0xf2, 0x41, 0x24, 0xa8, 0x67, 0x30, 0xa0, 0x57, 0x67, + 0x00, 0xa8, 0x06, 0x60, 0xc3, 0x74, 0x6d, 0x54, 0x90, 0x5e, 0xad, 0x71, + 0x41, 0x50, 0xab, 0x9d, 0xba, 0x34, 0x1a, 0xfd, 0x19, 0x21, 0x0e, 0x87, + 0xb7, 0x22, 0xe6, 0xca, 0xb9, 0x0d, 0x3c, 0x4f, 0xad, 0x16, 0xf1, 0xa5, + 0x6d, 0xba, 0x6d, 0x7b, 0xbe, 0x7b, 0xe3, 0x95, 0xec, 0x1b, 0x8b, 0x6e, + 0xb0, 0xdc, 0x5c, 0xfd, 0x31, 0x73, 0x85, 0x02, 0x63, 0xc6, 0xcc, 0x04, + 0x29, 0xa5, 0xf4, 0x1f, 0xcb, 0x90, 0xf7, 0x83, 0x0d, 0x36, 0xbf, 0x31, + 0xc0, 0xfc, 0x26, 0x15, 0x87, 0xc8, 0x15, 0x88, 0xc9, 0x79, 0x11, 0x67, + 0x23, 0x53, 0xca, 0x03, 0x7a, 0x02, 0xe5, 0xfc, 0xb3, 0x38, 0xf3, 0x5d, + 0xfc, 0x91, 0x6f, 0x59, 0x26, 0xae, 0xd8, 0x45, 0xfa, 0xc4, 0x5b, 0xa2, + 0xfb, 0x2c, 0xc5, 0x36, 0xc6, 0x0d, 0x7b, 0x4e, 0xd2, 0x7f, 0x61, 0xc5, + 0xcc, 0x74, 0xd3, 0x41, 0xd4, 0x8a, 0xaf, 0xcb, 0x32, 0x50, 0xca, 0xeb, + 0x59, 0x0a, 0x05, 0x25, 0xe0, 0x5f, 0x30, 0x2b, 0x5d, 0x9b, 0xf7, 0xe8, + 0x14, 0x14, 0xb5, 0xfe, 0xd5, 0x2f, 0x94, 0x84, 0x5b, 0xc7, 0x4f, 0x82, + 0x01, 0x50, 0xbf, 0x54, 0xe2, 0x7d, 0xeb, 0x0c, 0x85, 0xc8, 0x99, 0x45, + 0x50, 0x8e, 0x4e, 0x10, 0x12, 0x01, 0x17, 0x41, 0xf3, 0x21, 0x4a, 0xee, + 0xaf, 0x0f, 0x76, 0x44, 0xe2, 0x8e, 0xf8, 0x36, 0x25, 0xab, 0x0d, 0x8f, + 0xb1, 0x0a, 0xbf, 0x63, 0x0e, 0xf2, 0x0c, 0x9d, 0x39, 0xa1, 0x98, 0x98, + 0x69, 0x91, 0xd1, 0x9b, 0xe8, 0xcf, 0x16, 0x65, 0x02, 0xc9, 0x67, 0x72, + 0x71, 0x7c, 0xfb, 0x41, 0x2d, 0xe4, 0xd3, 0xfb, 0x44, 0x8a, 0x7a, 0x88, + 0x32, 0x62, 0x26, 0x63, 0xfe, 0x5b, 0x0c, 0x4f, 0x6c, 0xad, 0x2f, 0x64, + 0x6f, 0xc9, 0xda, 0x95, 0x10, 0xbe, 0xd1, 0xfa, 0x8b, 0x67, 0x64, 0x35, + 0x2d, 0xed, 0xca, 0xf3, 0x12, 0xb7, 0x06, 0xc3, 0xa9, 0x8e, 0x3f, 0x09, + 0x4d, 0x1f, 0x50, 0x3a, 0x97, 0xb7, 0xa7, 0xce, 0x4d, 0x46, 0xf1, 0x61, + 0xc1, 0x06, 0x95, 0x0d, 0x07, 0xa2, 0xbc, 0xed, 0xeb, 0x45, 0xb4, 0x69, + 0x05, 0x7a, 0x30, 0x47, 0xa3, 0xbf, 0x81, 0xa9, 0xa7, 0xf0, 0x53, 0x36, + 0x31, 0x37, 0x13, 0xe5, 0x0e, 0xd6, 0xe6, 0xc7, 0x17, 0x17, 0x21, 0x6d, + 0x36, 0xd0, 0xf6, 0x2a, 0xea, 0x2d, 0x32, 0x0e, 0x90, 0x03, 0x30, 0x4d, + 0x30, 0x31, 0xaa, 0x79, 0x2d, 0xae, 0x2e, 0xb0, 0x13, 0xad, 0x63, 0x69, + 0x67, 0xd8, 0xf3, 0x6e, 0xa4, 0x34, 0xcf, 0x02, 0x10, 0xdd, 0x76, 0xfa, + 0xa7, 0xb0, 0x92, 0xea, 0x47, 0xbd, 0xff, 0xf9, 0xac, 0x8a, 0x1f, 0x31, + 0xf8, 0x05, 0xd4, 0xce, 0x23, 0xad, 0x32, 0x8c, 0x6c, 0x92, 0x85, 0xb9, + 0x74, 0xa6, 0xab, 0x6e, 0x76, 0xfd, 0x3e, 0x8a, 0xac, 0xa3, 0xd1, 0xb7, + 0x40, 0x53, 0x87, 0x28, 0xfc, 0xbc, 0x8a, 0x52, 0x8e, 0x2e, 0x59, 0x2c, + 0x5f, 0x3f, 0xcb, 0xd8, 0xbe, 0x37, 0xfd, 0xdc, 0xc0, 0x34, 0x85, 0x67, + 0x28, 0x9f, 0x1d, 0x05, 0x05, 0x94, 0xed, 0x6f, 0x54, 0x7a, 0x51, 0x9a, + 0xaa, 0xca, 0xe1, 0x41, 0x10, 0xf0, 0x9d, 0x38, 0x9c, 0x5e, 0x95, 0xe3, + 0x7e, 0x62, 0xe2, 0x31, 0x81, 0x28, 0x4a, 0x3c, 0x5e, 0x04, 0x11, 0xe2, + 0x6a, 0x45, 0x6f, 0x68, 0x96, 0x5b, 0xbf, 0x22, 0xd8, 0x29, 0x91, 0x76, + 0xe1, 0xb2, 0x5f, 0xfc, 0x89, 0x90, 0x87, 0xf8, 0xb8, 0x3f, 0xd5, 0x11, + 0xe7, 0x36, 0x47, 0x71, 0xb9, 0x52, 0x97, 0x8e, 0x62, 0x8b, 0x05, 0x31, + 0xe5, 0xd9, 0xa2, 0xc3, 0x1a, 0xb5, 0xda, 0xc7, 0xa5, 0x37, 0x06, 0x67, + 0x41, 0x1f, 0x6e, 0xa3, 0xc2, 0xb4, 0x96, 0x64, 0xfc, 0x46, 0x85, 0x95, + 0x4e, 0xd8, 0x2a, 0x4b, 0xaa, 0x1e, 0xec, 0xd5, 0xed, 0x81, 0x23, 0x68, + 0x0f, 0x5d, 0x0b, 0x95, 0x29, 0xd4, 0x36, 0x4d, 0x8c, 0x32, 0x73, 0x6a, + 0xb7, 0xad, 0xb8, 0x9c, 0xad, 0x76, 0x09, 0xad, 0xb9, 0xea, 0x2d, 0x17, + 0x3c, 0x33, 0x87, 0x7f, 0x62, 0x74, 0x77, 0xc9, 0xd6, 0x3d, 0x17, 0xbc, + 0xff, 0x57, 0x10, 0xec, 0x7a, 0xb7, 0x89, 0x05, 0x26, 0xf1, 0xb2, 0x53, + 0xa1, 0x91, 0xc5, 0x2a, 0xfb, 0x5a, 0xce, 0x5d, 0xd1, 0x6b, 0xbc, 0xb7, + 0x39, 0x09, 0x43, 0xdf, 0x20, 0xd3, 0xc1, 0x74, 0x8d, 0xf4, 0x0b, 0x2a, + 0xc7, 0xe8, 0xa1, 0x5f, 0xb2, 0xfe, 0x1a, 0x96, 0x3a, 0x92, 0xbc, 0x8f, + 0x85, 0xe2, 0x22, 0x73, 0x3f, 0x49, 0xb3, 0x6b, 0x90, 0xbd, 0xcb, 0x3f, + 0x36, 0x6c, 0x3d, 0xe3, 0x00, 0x00, 0x00, 0x00, 0x56, 0xd1, 0xff, 0xff, + 0x04, 0x00, 0x00, 0x00, 0x90, 0x00, 0x00, 0x00, 0x1f, 0x05, 0x81, 0x3f, + 0x25, 0x68, 0xde, 0x72, 0x88, 0x26, 0x66, 0x2d, 0xe4, 0xc8, 0x81, 0xf8, + 0x5d, 0x98, 0xa2, 0xc2, 0x02, 0x62, 0x63, 0x47, 0xe6, 0x61, 0x7f, 0xee, + 0xca, 0x3f, 0x81, 0xd7, 0x1e, 0xa9, 0xbf, 0x66, 0x59, 0x7f, 0xc3, 0x35, + 0x03, 0xae, 0xe5, 0xf2, 0x4d, 0x81, 0x82, 0x78, 0x5e, 0xaf, 0xaa, 0xd1, + 0x27, 0x41, 0x19, 0x93, 0xa8, 0x9b, 0x78, 0x4e, 0x95, 0x89, 0x7f, 0xce, + 0x49, 0xd0, 0x45, 0xb5, 0x7f, 0x1d, 0xe9, 0xee, 0x7f, 0x91, 0xf4, 0x0a, + 0x67, 0x7d, 0x75, 0xff, 0x38, 0x81, 0x27, 0x90, 0x14, 0xa5, 0x99, 0x40, + 0x5b, 0xe6, 0x9a, 0x81, 0x75, 0x22, 0x5f, 0x18, 0x81, 0x34, 0xb7, 0x54, + 0x2e, 0x8d, 0x81, 0x36, 0x0e, 0x5e, 0xc0, 0x5f, 0xd4, 0xc6, 0x34, 0x81, + 0xc8, 0xb9, 0xe2, 0xa9, 0x77, 0x81, 0x44, 0xb4, 0x06, 0x24, 0x81, 0x74, + 0x1c, 0xeb, 0xfb, 0xdd, 0x25, 0x81, 0x14, 0x09, 0x2d, 0xba, 0x11, 0x4b, + 0x07, 0x13, 0xf1, 0xae, 0x81, 0xaf, 0xa3, 0x87, 0x00, 0x00, 0x00, 0x00, + 0xf6, 0xd1, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x00, 0x2d, 0x00, 0x00, + 0x8a, 0x29, 0x03, 0xe6, 0x24, 0x2a, 0xd6, 0x21, 0xb6, 0xb1, 0x2d, 0x3a, + 0xff, 0xd6, 0x27, 0xd7, 0x18, 0x42, 0xc1, 0xb4, 0xf8, 0xfd, 0xdf, 0x45, + 0x09, 0x91, 0xcb, 0xfe, 0xe9, 0xb5, 0x24, 0xf1, 0xc0, 0x69, 0xd0, 0x64, + 0xa8, 0xeb, 0x12, 0x71, 0xe3, 0xb4, 0xbe, 0xb4, 0x93, 0xbf, 0x8a, 0x8b, + 0xf3, 0x4d, 0x13, 0x3b, 0x6f, 0x6f, 0x32, 0x12, 0x98, 0x95, 0xb9, 0x63, + 0xcd, 0xa5, 0x23, 0xa4, 0xb8, 0x2e, 0x74, 0x75, 0xbc, 0xe4, 0xc7, 0x46, + 0x96, 0xd4, 0x47, 0xa0, 0x65, 0xec, 0xea, 0xcf, 0xd0, 0xdc, 0xe9, 0x8b, + 0xcc, 0x1d, 0x2f, 0x0d, 0x0a, 0x9c, 0x6e, 0x99, 0x97, 0x97, 0xcc, 0x00, + 0xd2, 0x8e, 0xbc, 0x3c, 0x9a, 0xf1, 0x32, 0x0e, 0xf3, 0xd6, 0x27, 0x1c, + 0xea, 0xab, 0xca, 0x4d, 0x69, 0x32, 0x30, 0x5f, 0x18, 0xd7, 0xb7, 0x4a, + 0xcb, 0x8e, 0xb2, 0x96, 0x39, 0xa3, 0xc7, 0x42, 0xca, 0x60, 0x9b, 0xad, + 0x8e, 0xb7, 0x54, 0x32, 0xea, 0xfd, 0x58, 0xfa, 0xf8, 0x02, 0xef, 0x2f, + 0xec, 0x3c, 0x2a, 0x1a, 0x6a, 0x08, 0xa4, 0x4b, 0xec, 0x30, 0x90, 0xaf, + 0x13, 0x98, 0xcd, 0x48, 0xfd, 0x5f, 0x56, 0x68, 0x17, 0x9e, 0x87, 0xb1, + 0x2b, 0x16, 0xd3, 0x3c, 0xe0, 0xe8, 0x0e, 0xa6, 0xc4, 0x24, 0xd3, 0x05, + 0x75, 0xda, 0x22, 0x44, 0xb5, 0x41, 0xd2, 0xa5, 0x99, 0xf1, 0x5e, 0xbe, + 0x15, 0xb7, 0x33, 0x54, 0x9a, 0x97, 0x5b, 0x35, 0x77, 0x2b, 0x18, 0x46, + 0x2f, 0x92, 0xc5, 0x97, 0x2d, 0x4c, 0xa6, 0xf8, 0x9e, 0xc3, 0xe0, 0x0a, + 0x52, 0xf9, 0x97, 0xc7, 0xd6, 0x36, 0xdd, 0x38, 0xaa, 0xf3, 0x05, 0x30, + 0xc3, 0xe5, 0xaf, 0x54, 0xdc, 0xc4, 0xf2, 0x01, 0x9e, 0xe6, 0xc1, 0x89, + 0xee, 0xd8, 0x5f, 0xfe, 0xf0, 0x70, 0x3c, 0xc4, 0x40, 0xa4, 0xd4, 0xee, + 0xaf, 0x3d, 0xe6, 0xcd, 0x31, 0x16, 0x31, 0x3b, 0xa0, 0x0e, 0xc4, 0x71, + 0xbf, 0xbd, 0x39, 0x89, 0x0f, 0x36, 0xba, 0xd8, 0xa2, 0x49, 0x01, 0xab, + 0xf4, 0x07, 0x99, 0xc7, 0xb1, 0x0c, 0x33, 0x9d, 0x71, 0xf1, 0x15, 0x4b, + 0x60, 0xe0, 0xed, 0x59, 0x0a, 0x34, 0xd9, 0xa2, 0x45, 0x99, 0x4a, 0x60, + 0xd3, 0xdc, 0x37, 0x56, 0x32, 0x4c, 0xea, 0xdc, 0xcf, 0xe6, 0x22, 0x27, + 0x17, 0xea, 0x75, 0x3f, 0x69, 0xd4, 0xcf, 0x53, 0x92, 0x98, 0xf4, 0xfe, + 0x13, 0xa8, 0xe2, 0xb2, 0x48, 0x5f, 0x64, 0xab, 0x2b, 0x61, 0x97, 0xf5, + 0xc5, 0xb6, 0xef, 0x32, 0x4e, 0x47, 0x26, 0x42, 0x48, 0x9c, 0x5b, 0x24, + 0xa3, 0xcb, 0x70, 0xc7, 0x31, 0x6c, 0xc8, 0x4d, 0x5c, 0x02, 0xca, 0x71, + 0x1e, 0x56, 0xdb, 0x27, 0x66, 0x5d, 0x4f, 0x0b, 0x09, 0x57, 0xbe, 0x72, + 0x17, 0x3b, 0xce, 0xdd, 0xd2, 0x20, 0x13, 0x67, 0x32, 0x04, 0xee, 0xc4, + 0x66, 0x23, 0x0e, 0x97, 0x5e, 0x21, 0x30, 0xb2, 0xe4, 0x16, 0x06, 0x57, + 0xc3, 0x9b, 0x29, 0x5b, 0x76, 0xd0, 0x36, 0xac, 0xe6, 0xa2, 0x91, 0x57, + 0x96, 0x4e, 0x1c, 0x6f, 0x4a, 0x03, 0x50, 0x55, 0x6d, 0xaf, 0x9a, 0x29, + 0xc9, 0x61, 0x6c, 0x18, 0x4c, 0xb9, 0xd5, 0x41, 0xf8, 0x75, 0x2b, 0xc3, + 0x0e, 0x69, 0x9f, 0x45, 0x93, 0x2f, 0xa6, 0xf9, 0x30, 0x65, 0x05, 0x13, + 0xe3, 0x00, 0x54, 0x0e, 0xa4, 0xb5, 0x89, 0x6d, 0x4d, 0x11, 0x3d, 0x2a, + 0x29, 0x99, 0xd9, 0xdf, 0x75, 0xce, 0x01, 0x21, 0xbc, 0x26, 0xb3, 0x22, + 0xf9, 0xb0, 0x45, 0x5c, 0xf8, 0xea, 0xb2, 0x08, 0x1a, 0xf7, 0xa0, 0x70, + 0x65, 0xa8, 0xab, 0xe1, 0x92, 0xcc, 0xcc, 0x1f, 0x0e, 0x36, 0x60, 0xb7, + 0xea, 0xcb, 0x3d, 0xf6, 0x98, 0xbf, 0xcd, 0x00, 0xc9, 0x16, 0x1e, 0xdb, + 0x58, 0x24, 0xb1, 0xd8, 0xaf, 0x01, 0x00, 0xfa, 0x15, 0xf4, 0x37, 0x05, + 0xd7, 0x17, 0x2a, 0xd2, 0xe8, 0xe4, 0x0c, 0x50, 0xfa, 0xe8, 0xd6, 0x99, + 0xa9, 0x58, 0x61, 0x38, 0xee, 0x22, 0x3c, 0x53, 0xcf, 0x64, 0x8e, 0xad, + 0x4d, 0xd6, 0xc3, 0xc3, 0xdd, 0xb0, 0xb3, 0xf7, 0xdd, 0x37, 0xfd, 0xf3, + 0x2b, 0x6a, 0xe2, 0xd4, 0xfc, 0x0c, 0x74, 0xca, 0x37, 0x2f, 0xd2, 0xf8, + 0x5b, 0xf1, 0x8c, 0x32, 0xa0, 0xdc, 0x2c, 0xa8, 0x36, 0x2f, 0xbe, 0x45, + 0x9b, 0x42, 0x95, 0x15, 0x5e, 0x08, 0xb1, 0x61, 0xec, 0xa2, 0xdf, 0x5f, + 0xca, 0xf8, 0x62, 0x73, 0xfd, 0x66, 0xc8, 0x51, 0x2a, 0x69, 0x3c, 0x8f, + 0x75, 0xa4, 0x6f, 0xbe, 0xc1, 0x5c, 0x66, 0xe2, 0x60, 0x92, 0xd7, 0x0e, + 0xee, 0x1b, 0xc7, 0x39, 0x8b, 0x56, 0x6c, 0xc6, 0x20, 0xfa, 0xec, 0x96, + 0xa5, 0x0f, 0x74, 0x42, 0x32, 0x12, 0x11, 0xdf, 0x02, 0xfe, 0x42, 0x1c, + 0xfe, 0xf1, 0x72, 0xaf, 0x47, 0x3b, 0x62, 0xe3, 0x27, 0x29, 0xf0, 0xec, + 0x39, 0xd2, 0xdd, 0xb6, 0xe9, 0xbe, 0x5f, 0x66, 0x67, 0x6c, 0xc9, 0xa1, + 0xf0, 0x25, 0x9a, 0x1b, 0xa8, 0xa0, 0x15, 0xcb, 0x61, 0x98, 0x98, 0xfd, + 0xef, 0xba, 0x74, 0x9b, 0x54, 0xf3, 0x6d, 0xe1, 0xa4, 0xcf, 0xb5, 0xe7, + 0xba, 0x0f, 0xd1, 0x41, 0xd8, 0x63, 0x94, 0x09, 0xcd, 0x4f, 0xb1, 0x31, + 0x49, 0x5e, 0x54, 0xb1, 0x28, 0x39, 0x8e, 0x13, 0x48, 0x2e, 0x20, 0xb0, + 0xf7, 0x18, 0x9a, 0xea, 0xf2, 0x9b, 0xde, 0x8f, 0x16, 0xc8, 0x9e, 0x31, + 0xca, 0x94, 0x28, 0x26, 0x0d, 0x8c, 0x0f, 0x09, 0x69, 0xc5, 0x2a, 0x38, + 0xae, 0x6b, 0xfb, 0x4f, 0xbb, 0xf4, 0x14, 0xea, 0x8d, 0x13, 0xc0, 0x09, + 0xe2, 0xfb, 0xfb, 0x09, 0xa1, 0xfc, 0x49, 0xff, 0x0f, 0x52, 0x3e, 0xe8, + 0xda, 0xfe, 0xe1, 0x67, 0x8f, 0x21, 0xcf, 0xaf, 0xb7, 0xe2, 0xcf, 0x09, + 0x15, 0x10, 0x51, 0x72, 0x8f, 0x42, 0x09, 0x9d, 0xea, 0x27, 0x2d, 0x25, + 0x9f, 0x54, 0x50, 0xfa, 0xdf, 0x9f, 0x41, 0xe8, 0xd2, 0x66, 0xd8, 0x28, + 0xfb, 0x8b, 0xe4, 0x42, 0x03, 0x92, 0xf9, 0xcd, 0xcc, 0xb0, 0xc0, 0x52, + 0x53, 0x6d, 0xcd, 0xed, 0x16, 0xad, 0x3c, 0x3d, 0xf9, 0x3b, 0x05, 0xbb, + 0xac, 0x9e, 0xa3, 0x4b, 0x17, 0xb4, 0xc7, 0xdd, 0xd4, 0xd3, 0x0c, 0x10, + 0x0d, 0xd8, 0x9c, 0xdb, 0xa4, 0x60, 0x06, 0x89, 0x4b, 0x06, 0x4c, 0x9f, + 0xc4, 0x47, 0xc8, 0xaf, 0xab, 0x02, 0x23, 0x89, 0x6e, 0xf2, 0x9d, 0x2b, + 0x6b, 0x9a, 0xa4, 0xee, 0x16, 0x0b, 0x3c, 0x76, 0xd4, 0xf0, 0x17, 0x90, + 0xca, 0xf5, 0xc8, 0xbf, 0xcb, 0xb1, 0x02, 0x69, 0x34, 0x71, 0x59, 0x5d, + 0x0e, 0x56, 0xd8, 0x41, 0x0a, 0xa5, 0x0a, 0x16, 0xbc, 0x93, 0x63, 0xf9, + 0xd9, 0xab, 0x3e, 0x75, 0x1e, 0xd3, 0xf3, 0x56, 0xf5, 0x14, 0xee, 0x65, + 0xf3, 0x2f, 0x72, 0x03, 0xcb, 0x69, 0x90, 0x91, 0x0d, 0x31, 0x8e, 0x3e, + 0xe9, 0xb0, 0xe6, 0x2e, 0x37, 0x5d, 0xb0, 0x38, 0x52, 0xe6, 0x23, 0x24, + 0x36, 0xb2, 0xe9, 0xa5, 0xa0, 0xae, 0xed, 0xfd, 0x95, 0xa5, 0xcf, 0x4a, + 0xe3, 0xbd, 0xe7, 0x29, 0xd0, 0x57, 0x3e, 0xf1, 0xdf, 0xc8, 0xc7, 0x26, + 0xf6, 0xc7, 0x4b, 0xc8, 0x6a, 0x4a, 0xed, 0x49, 0x60, 0x2d, 0x1c, 0xe3, + 0x8b, 0x10, 0x24, 0xfc, 0xef, 0xbb, 0x1e, 0x24, 0xbb, 0x40, 0xeb, 0x99, + 0xba, 0xe1, 0x4a, 0xd4, 0x1f, 0x69, 0x47, 0xa4, 0x8f, 0x48, 0x05, 0x17, + 0xcb, 0xee, 0x55, 0xca, 0xe5, 0xe3, 0x60, 0xec, 0xfa, 0xe6, 0xd1, 0x28, + 0xc5, 0xa8, 0x04, 0xd8, 0xce, 0x13, 0x2b, 0x99, 0x2b, 0xc7, 0x94, 0x9d, + 0xda, 0xd7, 0x6f, 0x31, 0xfe, 0xee, 0x6c, 0x9b, 0xf1, 0x70, 0xd2, 0xee, + 0xc4, 0xba, 0xb7, 0xbe, 0xd3, 0x37, 0xdc, 0x43, 0x4e, 0x30, 0x4a, 0x67, + 0xf2, 0x45, 0x29, 0xe1, 0x8b, 0xb8, 0x6d, 0xca, 0xec, 0xb9, 0xd6, 0xd3, + 0xdd, 0xcb, 0xde, 0xdb, 0xa9, 0x4d, 0xdd, 0x3d, 0x41, 0xae, 0x99, 0x89, + 0xce, 0x70, 0x50, 0x61, 0x07, 0xf3, 0xca, 0x24, 0x56, 0x76, 0x3f, 0xe0, + 0x6e, 0xbe, 0xa7, 0xc6, 0xac, 0x6c, 0xf1, 0x8c, 0xa2, 0x0e, 0xc4, 0x2a, + 0x48, 0x30, 0x8b, 0xc9, 0xc0, 0x5a, 0xb2, 0x2b, 0xbd, 0xa2, 0xcc, 0xf7, + 0x25, 0x16, 0xc3, 0xde, 0x1b, 0x8d, 0x23, 0x8c, 0xb6, 0xc4, 0xaa, 0x4a, + 0x0b, 0x66, 0x25, 0x35, 0xb3, 0x9a, 0x74, 0x27, 0x63, 0xea, 0xef, 0x92, + 0x12, 0x8c, 0x58, 0xd9, 0x3a, 0x55, 0xd6, 0x61, 0x29, 0x9f, 0xbc, 0x28, + 0xbd, 0x30, 0xcd, 0x43, 0xe6, 0x36, 0x36, 0x66, 0x20, 0x8c, 0x9e, 0x23, + 0xfe, 0x6d, 0xf0, 0xbc, 0x61, 0xcd, 0x58, 0xd8, 0xe0, 0x2e, 0xe4, 0xcf, + 0x61, 0xf7, 0xd5, 0x6b, 0x54, 0x33, 0xb3, 0x2c, 0x60, 0xa8, 0x59, 0x21, + 0x5d, 0xaa, 0x65, 0x9e, 0xdc, 0xa3, 0xc9, 0xc4, 0x9d, 0x4d, 0x95, 0x29, + 0xf6, 0x2b, 0xcd, 0xc9, 0xb9, 0x9d, 0x46, 0xa0, 0x89, 0xf4, 0x4e, 0x52, + 0x55, 0xe2, 0x13, 0x98, 0xf0, 0xef, 0x27, 0xc3, 0xc9, 0xd1, 0xe1, 0xee, + 0x07, 0x1b, 0x9d, 0x8a, 0x5b, 0x9d, 0x06, 0x26, 0x61, 0x2a, 0x55, 0x6f, + 0x54, 0x22, 0xd5, 0x06, 0x20, 0xed, 0x06, 0x4d, 0xa2, 0xb3, 0xaa, 0x4f, + 0x1f, 0x3e, 0xd2, 0x0d, 0x6a, 0xab, 0x6d, 0xee, 0x8f, 0x09, 0xb2, 0xd9, + 0x39, 0x46, 0x0f, 0xe7, 0x51, 0x70, 0x51, 0xdb, 0x09, 0xf8, 0x8e, 0xbb, + 0x06, 0x98, 0x49, 0x69, 0xb7, 0x9e, 0xa0, 0xbc, 0x16, 0x5f, 0x96, 0xad, + 0xe9, 0x76, 0x9f, 0x71, 0xe2, 0x1b, 0x91, 0x73, 0xd9, 0x74, 0x6a, 0x70, + 0x48, 0x71, 0x47, 0x3b, 0x0c, 0xd5, 0x96, 0xe3, 0x6e, 0xdb, 0xbb, 0x9c, + 0x44, 0x5c, 0xe5, 0x07, 0x73, 0x31, 0xd1, 0x55, 0x07, 0xff, 0x5f, 0xb1, + 0x55, 0x9d, 0x0d, 0xbf, 0x32, 0x53, 0xf9, 0xfe, 0xcd, 0xc8, 0xe0, 0x56, + 0x18, 0x8f, 0x4b, 0x51, 0xd1, 0x23, 0x2e, 0x9f, 0xb9, 0xee, 0xf3, 0xfd, + 0x26, 0x02, 0xf6, 0x54, 0xd5, 0x3e, 0x13, 0xc1, 0xc1, 0xe4, 0xa8, 0xb4, + 0x5f, 0x5c, 0xa0, 0x9f, 0xb5, 0x19, 0xbb, 0x4e, 0xd6, 0xf8, 0x18, 0x9b, + 0xeb, 0x9e, 0x58, 0x9d, 0x00, 0x51, 0x24, 0x28, 0x70, 0x55, 0xf7, 0xb9, + 0x5a, 0x59, 0x50, 0xc5, 0x72, 0xab, 0x6b, 0x13, 0x95, 0xfb, 0xe4, 0xc2, + 0x05, 0x96, 0xf3, 0x48, 0xef, 0x02, 0x67, 0xd5, 0x8f, 0x5b, 0x8e, 0xb6, + 0xbe, 0xc1, 0x3d, 0x8e, 0x22, 0xee, 0x49, 0xc7, 0xbe, 0xfb, 0x2d, 0x51, + 0x45, 0x44, 0xca, 0x94, 0x8e, 0xce, 0xb5, 0x9a, 0x29, 0xc7, 0x52, 0xde, + 0x2c, 0xdf, 0xcc, 0x43, 0xc7, 0xd7, 0x51, 0xb7, 0x07, 0xf0, 0x9b, 0x9d, + 0x33, 0x98, 0x62, 0xfa, 0xc9, 0x13, 0x0b, 0xcd, 0xdf, 0xbd, 0xff, 0x8e, + 0x13, 0x44, 0xda, 0x62, 0xc0, 0xd1, 0x8d, 0x57, 0x0e, 0xec, 0x53, 0x8a, + 0x04, 0xcf, 0x0f, 0x5a, 0xd7, 0x3c, 0x4b, 0x17, 0xda, 0x3b, 0xf0, 0x30, + 0xbf, 0xea, 0x40, 0xa6, 0x36, 0xed, 0xda, 0xf7, 0x40, 0x6b, 0xf1, 0x1e, + 0x61, 0xa0, 0x8b, 0x5d, 0xfa, 0xa8, 0x6a, 0xca, 0xfd, 0x6a, 0x06, 0xb4, + 0xf5, 0xb6, 0xc7, 0xbe, 0xdf, 0xac, 0x17, 0x00, 0x4a, 0x91, 0x8d, 0x97, + 0x5b, 0xc8, 0xcb, 0xd4, 0xc8, 0x20, 0x0b, 0x53, 0xee, 0x2b, 0x25, 0xb8, + 0xa1, 0x24, 0xa1, 0xa0, 0x17, 0x60, 0xd9, 0xf7, 0x2d, 0x00, 0x6c, 0x70, + 0x44, 0x0d, 0x60, 0xe7, 0x95, 0x1e, 0x8a, 0x1b, 0x29, 0xcf, 0xb5, 0xc1, + 0xbe, 0xd0, 0xe5, 0xeb, 0xd8, 0x71, 0x88, 0x34, 0xcb, 0xbd, 0x32, 0x52, + 0xa7, 0xcf, 0x6d, 0x9b, 0xef, 0xf2, 0xe4, 0x68, 0x6f, 0xfe, 0xb9, 0x17, + 0x31, 0xa0, 0x3e, 0xfc, 0xae, 0xf6, 0x54, 0xe3, 0x33, 0x24, 0xd1, 0xfc, + 0xb7, 0x37, 0x8f, 0xd3, 0x4f, 0xf2, 0x59, 0x53, 0xea, 0xaf, 0x71, 0xc5, + 0xb1, 0xdb, 0xf9, 0xed, 0xc0, 0x46, 0x56, 0xfc, 0x09, 0x90, 0xf7, 0x09, + 0x5a, 0x12, 0x71, 0xad, 0xa6, 0x0f, 0xba, 0x4c, 0x2f, 0xd7, 0x61, 0xcb, + 0xf2, 0xab, 0x44, 0x67, 0x43, 0xd0, 0x41, 0xd5, 0xba, 0xff, 0x26, 0x50, + 0x5b, 0x97, 0x91, 0xc4, 0x8f, 0x2a, 0x64, 0x3c, 0x06, 0x2e, 0x26, 0x8e, + 0x5f, 0xb1, 0xba, 0x74, 0x16, 0xeb, 0xee, 0x6e, 0xe1, 0x68, 0xcc, 0x09, + 0xed, 0xa5, 0x5d, 0xf7, 0xef, 0xd6, 0xfa, 0x9f, 0x39, 0xe1, 0x5c, 0x38, + 0xbd, 0x1b, 0xe6, 0x8a, 0xfa, 0xea, 0xbc, 0x14, 0x4c, 0x31, 0xa8, 0x9d, + 0x64, 0xa6, 0xec, 0xf0, 0xf8, 0xa2, 0x0a, 0x6c, 0xb9, 0xc5, 0x3d, 0x40, + 0x48, 0x41, 0x1d, 0xf2, 0xab, 0xd4, 0xdf, 0xfb, 0x55, 0x9e, 0xa5, 0xac, + 0xe9, 0xf0, 0x46, 0x96, 0xc5, 0x4d, 0x5f, 0x5f, 0x64, 0x00, 0x69, 0x48, + 0x0e, 0xa3, 0xb5, 0x5d, 0x45, 0xce, 0x57, 0xc4, 0x45, 0xdb, 0xc6, 0x13, + 0x4b, 0xa7, 0xa0, 0xd5, 0x31, 0xb4, 0xd4, 0x0f, 0x4f, 0x29, 0x40, 0xc0, + 0xaa, 0xb7, 0x54, 0x21, 0xd5, 0x3a, 0x01, 0xbc, 0xa8, 0x58, 0xb5, 0x3f, + 0xa6, 0x1a, 0x06, 0xb5, 0x07, 0xd3, 0xb6, 0xff, 0x6e, 0x74, 0x08, 0x16, + 0x45, 0xaf, 0xd9, 0xc5, 0x4a, 0x0d, 0xd2, 0x8a, 0xd1, 0x6c, 0xba, 0x5a, + 0xd0, 0xee, 0x57, 0x10, 0xa4, 0x1a, 0xf4, 0x92, 0x97, 0xe0, 0xd7, 0xa8, + 0xff, 0x47, 0xed, 0x56, 0x6b, 0x91, 0x77, 0x5d, 0xa6, 0xcf, 0xed, 0x96, + 0xc5, 0x5a, 0xe3, 0x0b, 0x1d, 0xc0, 0xcc, 0xa1, 0x71, 0x95, 0xa8, 0xec, + 0xef, 0x33, 0x91, 0xd6, 0x53, 0x1f, 0xef, 0x43, 0xa9, 0x42, 0x2a, 0xc7, + 0xf6, 0x15, 0x60, 0xc2, 0xde, 0xeb, 0xac, 0xf8, 0x55, 0x27, 0x14, 0xf1, + 0xf8, 0x69, 0x55, 0xc8, 0x69, 0x1f, 0xf3, 0xc2, 0x71, 0xe8, 0x75, 0xa9, + 0x1a, 0x91, 0xc5, 0x1e, 0xe3, 0x52, 0x24, 0x5f, 0x60, 0xb5, 0xf1, 0xe6, + 0xdd, 0x4b, 0x1b, 0xdd, 0x3a, 0xad, 0x58, 0x36, 0x9c, 0xb3, 0x25, 0x9e, + 0x28, 0xd4, 0x3b, 0x6a, 0x64, 0xe7, 0x57, 0x54, 0xad, 0x4d, 0x44, 0xfc, + 0x54, 0xd3, 0xa3, 0x96, 0x4e, 0xee, 0xde, 0x23, 0x30, 0x30, 0x1f, 0x57, + 0x2f, 0xd6, 0xb4, 0xfa, 0x5c, 0x1b, 0x4a, 0x1b, 0x96, 0x58, 0x9a, 0xc7, + 0x25, 0xd0, 0x9c, 0xf3, 0x2b, 0x16, 0x58, 0x62, 0x0c, 0x5b, 0x45, 0x96, + 0xb0, 0xc2, 0x3e, 0xca, 0x0a, 0xb5, 0x0f, 0x06, 0xa8, 0xa3, 0xb2, 0x0a, + 0x6a, 0xc5, 0xb7, 0xf8, 0x69, 0xfa, 0xc1, 0xa8, 0xbc, 0x17, 0x6c, 0x92, + 0x06, 0x50, 0x74, 0x4b, 0x02, 0xc8, 0x4d, 0x9c, 0x3e, 0x94, 0x6f, 0xef, + 0x3e, 0xd9, 0x71, 0xa6, 0x3a, 0x70, 0x6a, 0x14, 0x0e, 0x06, 0xbe, 0x40, + 0x2b, 0xa1, 0xbb, 0x05, 0x71, 0x05, 0xbd, 0xd5, 0x2d, 0xd9, 0xe2, 0xf6, + 0xb4, 0x32, 0x33, 0xac, 0x0f, 0x9a, 0xe3, 0xaf, 0xf4, 0x44, 0x21, 0x59, + 0x91, 0x0d, 0xd0, 0xf1, 0x47, 0x9e, 0x00, 0x38, 0xa2, 0x1d, 0x61, 0x54, + 0xd2, 0x18, 0x9d, 0xe4, 0x4f, 0xf3, 0xbd, 0x04, 0xdb, 0x4d, 0x59, 0x8c, + 0xfa, 0x12, 0xdd, 0xe4, 0xb5, 0x32, 0x3b, 0xf8, 0x93, 0xae, 0x3b, 0xa9, + 0xb3, 0xe9, 0x57, 0x30, 0x49, 0x6d, 0xaa, 0x35, 0x12, 0xce, 0x16, 0x98, + 0x3c, 0xd0, 0xed, 0xe8, 0xa6, 0xbc, 0xa6, 0xe6, 0x66, 0x0f, 0xb3, 0x12, + 0x95, 0x19, 0x56, 0x23, 0xb1, 0x30, 0x5d, 0xb3, 0x4c, 0x5f, 0x0c, 0xef, + 0x24, 0x12, 0xe0, 0x97, 0xf3, 0x3e, 0x9c, 0x49, 0xff, 0xa6, 0x6f, 0xa6, + 0xd2, 0x58, 0xbe, 0x3f, 0x30, 0xdd, 0x65, 0xd0, 0x40, 0xe1, 0xaf, 0x09, + 0xf1, 0xf4, 0x0f, 0x1a, 0xe5, 0xef, 0x51, 0x50, 0x38, 0x5d, 0xb0, 0x1e, + 0xed, 0x19, 0x8d, 0x4e, 0x20, 0xa1, 0x65, 0x07, 0x5b, 0x23, 0x0c, 0x14, + 0xd3, 0x18, 0xa3, 0xda, 0x58, 0x9f, 0x10, 0x00, 0xbd, 0xb5, 0x95, 0x07, + 0x1d, 0x0f, 0xf9, 0x2a, 0xe4, 0x35, 0x3c, 0x60, 0xad, 0xb2, 0x13, 0x3b, + 0xd5, 0x9e, 0xeb, 0xc7, 0x09, 0x6e, 0x53, 0xff, 0x95, 0xf3, 0xc1, 0x9b, + 0xcd, 0x21, 0x15, 0x3b, 0x5f, 0xfe, 0x4e, 0xaf, 0x3f, 0xf8, 0xe3, 0xa8, + 0x35, 0xee, 0x44, 0x33, 0xc7, 0x8c, 0x9c, 0x1c, 0x33, 0x55, 0x3c, 0x4a, + 0xa4, 0x35, 0xf6, 0xf0, 0x32, 0x8e, 0xed, 0x6d, 0x06, 0xff, 0x8d, 0x24, + 0x05, 0x72, 0x4c, 0xa2, 0x97, 0x25, 0x93, 0x3d, 0x79, 0x18, 0x22, 0x15, + 0xec, 0x5c, 0xc4, 0x10, 0x65, 0xec, 0x90, 0x6d, 0x28, 0xba, 0x93, 0xb5, + 0x2f, 0x53, 0xe4, 0x00, 0x9c, 0x39, 0xf5, 0x4c, 0xde, 0x51, 0x39, 0xc3, + 0xd8, 0x03, 0xc3, 0x97, 0xe1, 0xa8, 0x3e, 0x06, 0x26, 0x4d, 0xd9, 0x49, + 0x75, 0xbb, 0xd5, 0x69, 0x20, 0xfb, 0x85, 0x12, 0xc9, 0xac, 0xfc, 0x05, + 0xad, 0x57, 0xa9, 0x58, 0xcd, 0xfd, 0xbe, 0x64, 0x31, 0x50, 0x4d, 0xa4, + 0x93, 0xb6, 0x23, 0x3b, 0xfd, 0xd9, 0xdb, 0x46, 0xdd, 0x1f, 0x07, 0x54, + 0xc2, 0xc2, 0xd6, 0xad, 0xf6, 0x21, 0x39, 0xa1, 0x96, 0x53, 0x12, 0x46, + 0x5a, 0xc8, 0xf3, 0xf8, 0xe2, 0xa3, 0xd0, 0x29, 0x3f, 0x30, 0xca, 0x0b, + 0x57, 0xab, 0xcf, 0x1e, 0x08, 0x59, 0x3d, 0x41, 0x6a, 0xf7, 0xb2, 0xfc, + 0xff, 0x33, 0x46, 0xd1, 0x1a, 0xa6, 0x91, 0x54, 0xca, 0x27, 0x5a, 0x94, + 0x13, 0xf4, 0xf0, 0xcf, 0x58, 0xe0, 0x96, 0x50, 0xda, 0xe6, 0x91, 0xc7, + 0x8d, 0x14, 0x5b, 0xc1, 0xeb, 0x4a, 0x96, 0xf1, 0xa5, 0x43, 0xf6, 0x29, + 0x91, 0xb9, 0xb9, 0x67, 0x3f, 0x31, 0xd7, 0x08, 0xe6, 0x2b, 0xfb, 0x43, + 0x56, 0x39, 0x4e, 0xf9, 0x02, 0x8e, 0x96, 0x1f, 0xa3, 0x3c, 0xae, 0x55, + 0x03, 0x05, 0x9a, 0x39, 0xbe, 0xf7, 0x67, 0xa1, 0x6b, 0x2f, 0x42, 0x45, + 0x9b, 0x45, 0x8f, 0x53, 0x1f, 0x96, 0x42, 0x54, 0xd2, 0x5b, 0xf0, 0x17, + 0x94, 0x41, 0xaf, 0xd4, 0xc6, 0x37, 0x5f, 0xc0, 0xbd, 0xe3, 0x44, 0x8d, + 0xc1, 0x69, 0x64, 0x2a, 0xe7, 0x08, 0xe5, 0x18, 0x92, 0x53, 0xfc, 0xed, + 0xd3, 0x69, 0x94, 0x6b, 0x10, 0x0b, 0x5e, 0x91, 0x38, 0x4b, 0xa5, 0x19, + 0x3a, 0x6a, 0x2e, 0x5a, 0xa2, 0x6f, 0x34, 0x2c, 0x7b, 0x5d, 0x53, 0x33, + 0x77, 0x46, 0xf8, 0x4a, 0xa2, 0x8d, 0x55, 0x67, 0xa8, 0xbd, 0xc6, 0x3c, + 0x5d, 0x47, 0xeb, 0x99, 0xed, 0xdc, 0xae, 0xcf, 0xec, 0xbe, 0x40, 0x60, + 0xfc, 0x36, 0x5c, 0x93, 0x95, 0x64, 0xd8, 0x47, 0x14, 0xe2, 0x1e, 0xa2, + 0xd4, 0xd4, 0xdf, 0xd9, 0x23, 0x18, 0xf2, 0x99, 0xe8, 0xe4, 0x2a, 0x3b, + 0xec, 0x2e, 0x28, 0xa8, 0x04, 0x74, 0x04, 0xa4, 0x32, 0xa6, 0x49, 0xf9, + 0x33, 0x6c, 0xa8, 0x1d, 0xb2, 0xbb, 0x57, 0xe4, 0xcf, 0xf2, 0x9e, 0x74, + 0x8d, 0xf7, 0x22, 0xaa, 0x0d, 0x8a, 0x2f, 0x34, 0x72, 0x33, 0xec, 0xdf, + 0x46, 0x57, 0x6c, 0x97, 0x94, 0xad, 0x06, 0x88, 0xeb, 0x20, 0xec, 0x79, + 0x44, 0xe1, 0xbc, 0xf8, 0xbd, 0xeb, 0x99, 0xe3, 0xaf, 0xfe, 0xc5, 0xb5, + 0xfa, 0x31, 0x75, 0x62, 0xff, 0x2a, 0x2a, 0x1b, 0xce, 0xad, 0xa8, 0xc8, + 0x3c, 0x54, 0x23, 0xf9, 0x9e, 0x2d, 0xe2, 0xa4, 0x4f, 0x5b, 0x4d, 0xb8, + 0x4f, 0xc6, 0xb3, 0xc6, 0xef, 0x66, 0x54, 0x31, 0xab, 0xd3, 0xf0, 0xb9, + 0xfa, 0xb6, 0x15, 0xe6, 0xdb, 0x4b, 0x51, 0x4d, 0x77, 0xa5, 0x3d, 0x4e, + 0xd9, 0xc9, 0xdb, 0x95, 0x31, 0x1d, 0x4d, 0x37, 0xe0, 0x34, 0xd3, 0xf3, + 0x20, 0x6b, 0xb8, 0x16, 0x0b, 0x4e, 0x55, 0x96, 0x56, 0x1e, 0xa7, 0xe8, + 0xc6, 0x3a, 0x08, 0x49, 0xa1, 0x16, 0x46, 0xc9, 0x43, 0xcb, 0x8f, 0x28, + 0x4a, 0x78, 0xaa, 0xf9, 0x6c, 0x74, 0xc8, 0x0b, 0xce, 0x13, 0x2c, 0xef, + 0xfe, 0x73, 0x42, 0xa7, 0xbc, 0x3d, 0xc9, 0xf2, 0xaf, 0x1c, 0x32, 0xdb, + 0xb2, 0x15, 0x70, 0x6b, 0x9b, 0x6e, 0x6f, 0x6e, 0xf7, 0x95, 0xea, 0x3e, + 0xd0, 0xb1, 0x2a, 0xbe, 0x8c, 0x66, 0x4e, 0xe9, 0x29, 0xe3, 0x35, 0xde, + 0xbf, 0x44, 0xbc, 0x5e, 0x56, 0x8b, 0xb3, 0xd4, 0xdf, 0xf5, 0x4e, 0x2e, + 0xeb, 0xe6, 0x8e, 0x58, 0xe2, 0xfd, 0xe7, 0x27, 0xff, 0x07, 0x49, 0x20, + 0xdd, 0xcf, 0xe4, 0xd7, 0x5c, 0x5f, 0x1f, 0xcc, 0xeb, 0x29, 0xeb, 0x34, + 0xac, 0xd6, 0xb6, 0xf8, 0xae, 0xdf, 0x11, 0x58, 0xd5, 0xea, 0xf1, 0x76, + 0xe5, 0x4d, 0x51, 0x72, 0xd4, 0x5e, 0x1e, 0x0f, 0xfd, 0x2e, 0xbe, 0x8e, + 0x07, 0x1a, 0x1f, 0x99, 0x4d, 0x73, 0x70, 0xe1, 0x41, 0xb4, 0x20, 0x10, + 0x75, 0x0f, 0xc8, 0x69, 0x5f, 0x6c, 0x20, 0x2b, 0xc8, 0xfd, 0xe9, 0x4c, + 0xf4, 0x6f, 0x6a, 0xe0, 0x1a, 0xb5, 0xec, 0x2e, 0xf5, 0x25, 0x6d, 0x56, + 0x56, 0xb9, 0x42, 0xca, 0x70, 0x72, 0xe5, 0x41, 0x07, 0x4f, 0x41, 0x25, + 0xea, 0x0a, 0x5d, 0xe1, 0x0a, 0xd5, 0x6f, 0x35, 0x50, 0xcc, 0x27, 0x53, + 0x5f, 0x31, 0x1c, 0xee, 0xae, 0x26, 0xc8, 0xc4, 0x4f, 0x9b, 0xf5, 0xf6, + 0x4d, 0x19, 0xb9, 0xc4, 0x55, 0xcd, 0xe5, 0x8a, 0xe9, 0x45, 0xec, 0xf2, + 0xf9, 0x33, 0x4d, 0xba, 0x57, 0x8f, 0xd6, 0xf5, 0xf7, 0x92, 0xb3, 0xd3, + 0x65, 0x39, 0x07, 0x04, 0x92, 0x2f, 0x70, 0x99, 0x97, 0x96, 0x60, 0xe5, + 0x92, 0x60, 0xc3, 0x72, 0x1e, 0xc7, 0xe6, 0x1d, 0xbb, 0x5b, 0xd5, 0x64, + 0x1b, 0x36, 0x45, 0xb8, 0xcb, 0x42, 0xe7, 0x26, 0x45, 0x65, 0xc8, 0x04, + 0x1c, 0x05, 0x9b, 0x48, 0xe3, 0x93, 0x8e, 0xb2, 0x1c, 0x6a, 0xab, 0x60, + 0xc2, 0xa6, 0x1a, 0x71, 0xd5, 0x2c, 0xb8, 0xe9, 0x9e, 0x66, 0x8d, 0xb6, + 0xb1, 0x99, 0x90, 0x9c, 0x1b, 0xc9, 0x44, 0x6d, 0x31, 0xbb, 0x62, 0x6e, + 0x46, 0xcc, 0xd7, 0x47, 0x3a, 0x40, 0x63, 0x33, 0x34, 0x4f, 0x50, 0x3c, + 0x94, 0x97, 0xe9, 0xe8, 0x3a, 0xf7, 0x2d, 0x2d, 0x9c, 0xb6, 0x5d, 0x52, + 0xbd, 0xa9, 0x2d, 0x42, 0xfc, 0xe8, 0x70, 0x09, 0x48, 0xd0, 0x36, 0x0b, + 0x3d, 0x2b, 0x9f, 0xe2, 0x4c, 0xdf, 0xf3, 0x57, 0x73, 0x55, 0xf7, 0x34, + 0xb8, 0x6b, 0x44, 0x6f, 0xf6, 0x6d, 0xcf, 0x93, 0x09, 0x14, 0xac, 0x8f, + 0xde, 0xce, 0x5f, 0x05, 0x04, 0x9f, 0xc7, 0x05, 0x5f, 0xdd, 0x2e, 0xfc, + 0x53, 0xec, 0x9e, 0xdb, 0xa8, 0xa2, 0xc7, 0x53, 0x5c, 0x9a, 0x4d, 0xb6, + 0x6f, 0xa5, 0xc6, 0xf3, 0xc5, 0xa4, 0x56, 0x62, 0xdc, 0x75, 0xe4, 0x0b, + 0xb0, 0xcc, 0x38, 0xde, 0x2d, 0xbb, 0xbc, 0x0b, 0xc6, 0xab, 0xac, 0xac, + 0x46, 0xce, 0x1e, 0xe6, 0x47, 0x6c, 0x6e, 0x8e, 0x00, 0x00, 0xa0, 0xae, + 0x1e, 0x1d, 0xaa, 0x22, 0xaf, 0x34, 0xc7, 0x26, 0x37, 0x01, 0x46, 0x25, + 0x9c, 0x5f, 0x92, 0xef, 0xda, 0x07, 0x64, 0x62, 0xe4, 0xf7, 0x4c, 0xa2, + 0x41, 0xf1, 0x10, 0xe0, 0xe5, 0x73, 0x72, 0xe1, 0xf8, 0x66, 0x19, 0x58, + 0xa9, 0xdf, 0xb1, 0x41, 0xcb, 0xb3, 0xc4, 0xe6, 0x21, 0xbe, 0x17, 0x26, + 0xa9, 0x68, 0x96, 0xde, 0x5d, 0xba, 0x8f, 0x1b, 0x09, 0x00, 0x39, 0x0e, + 0xc2, 0x8d, 0x31, 0x61, 0xfe, 0x9e, 0x60, 0x05, 0xf3, 0x72, 0xdf, 0x78, + 0x14, 0x5a, 0x1b, 0x74, 0xa1, 0x23, 0xa7, 0x6e, 0x93, 0x76, 0xfa, 0x4a, + 0x73, 0xa1, 0x3b, 0xda, 0x0b, 0x06, 0xdd, 0xfc, 0x2f, 0xef, 0x0a, 0x38, + 0x03, 0xbf, 0xbb, 0x12, 0x29, 0x6b, 0xec, 0x68, 0xc7, 0xa6, 0xf9, 0x72, + 0xbc, 0xdb, 0xeb, 0x4e, 0x8f, 0x5f, 0x3a, 0xa9, 0x06, 0x4e, 0x3c, 0xf4, + 0x3b, 0xe0, 0x98, 0x9b, 0x77, 0x57, 0x0f, 0x39, 0x08, 0x43, 0x3f, 0x9b, + 0x76, 0x11, 0xd3, 0x38, 0xb6, 0x1f, 0x1e, 0xfe, 0xbb, 0x16, 0x37, 0x24, + 0x15, 0xf7, 0x8e, 0x61, 0x3d, 0xf5, 0x60, 0xab, 0x46, 0x49, 0xd6, 0xb2, + 0x8e, 0x35, 0xd5, 0x66, 0x20, 0x1f, 0xad, 0xf5, 0x95, 0xc3, 0x3e, 0xaa, + 0xda, 0x12, 0x1f, 0x33, 0xf4, 0xc0, 0xd9, 0x9e, 0x09, 0x76, 0x8b, 0x2f, + 0x35, 0xe2, 0x58, 0x09, 0x36, 0xf1, 0x03, 0xbc, 0xc2, 0x54, 0x67, 0x29, + 0x00, 0x3b, 0xf0, 0x24, 0xdf, 0xa0, 0x92, 0x71, 0xc3, 0x98, 0xe8, 0x5d, + 0xbe, 0xc7, 0xe8, 0x6f, 0x2f, 0x05, 0x89, 0x9f, 0xa1, 0x63, 0x29, 0x12, + 0x94, 0xff, 0xc7, 0x4c, 0xec, 0x98, 0x0e, 0xb8, 0xeb, 0x9e, 0x6d, 0x1e, + 0x4f, 0x4a, 0x1e, 0x41, 0xb0, 0xf9, 0x40, 0x8b, 0xdd, 0xd9, 0xa6, 0x1b, + 0xd4, 0x6d, 0xaf, 0x5b, 0x14, 0x68, 0xfd, 0x96, 0x5d, 0x0d, 0xad, 0x46, + 0x03, 0xf8, 0xd7, 0x13, 0x1d, 0xf3, 0x47, 0xbe, 0x46, 0x3d, 0xc7, 0xdd, + 0xa9, 0x60, 0x05, 0x15, 0xef, 0x9d, 0xa4, 0xb8, 0xde, 0xf2, 0x41, 0xe2, + 0x07, 0x1d, 0xcb, 0xe8, 0xf3, 0x9c, 0x9c, 0x5e, 0xcd, 0xec, 0x53, 0x39, + 0xf2, 0x62, 0x3b, 0x69, 0x3a, 0x29, 0xc7, 0xb3, 0x57, 0xce, 0x58, 0xd6, + 0x55, 0xf8, 0xc2, 0xf1, 0x16, 0xf3, 0x33, 0x3f, 0xf2, 0xaa, 0x63, 0x42, + 0x27, 0x01, 0x22, 0x5a, 0x1e, 0x8d, 0xa5, 0x33, 0x34, 0x29, 0x12, 0xf6, + 0x07, 0x22, 0xfd, 0xbb, 0x72, 0x60, 0x2a, 0xf5, 0xec, 0x71, 0xfe, 0xd7, + 0xc1, 0xf5, 0xdf, 0x97, 0x3e, 0x4a, 0x9a, 0x97, 0x6f, 0x56, 0xf1, 0xd4, + 0xba, 0x29, 0x09, 0x46, 0x3f, 0x10, 0xdc, 0x2d, 0xb2, 0x04, 0x32, 0x38, + 0xa3, 0xc7, 0x75, 0x95, 0x16, 0xd6, 0x12, 0x44, 0x7a, 0xd3, 0x18, 0xb3, + 0x51, 0x72, 0x63, 0xb8, 0xae, 0x9b, 0xf1, 0xec, 0x17, 0xe4, 0x2d, 0xed, + 0x29, 0x05, 0x63, 0xd7, 0x01, 0xf4, 0xf5, 0xc1, 0x6d, 0x13, 0x5f, 0x5c, + 0x73, 0x11, 0xc9, 0x53, 0xf4, 0xda, 0x90, 0xa2, 0x1c, 0x0b, 0x1d, 0x37, + 0x28, 0xa1, 0x06, 0x65, 0xd3, 0x49, 0x5d, 0x07, 0x1f, 0x93, 0xa9, 0x98, + 0xc5, 0xa5, 0x13, 0xc5, 0xac, 0xda, 0x64, 0x25, 0x77, 0x9a, 0xd5, 0xa9, + 0xe9, 0x3a, 0x77, 0x62, 0xac, 0xf2, 0x76, 0xf4, 0x03, 0xb6, 0x03, 0x6e, + 0xef, 0x97, 0x13, 0x1c, 0xd1, 0xb9, 0x73, 0x12, 0xf7, 0x10, 0xbd, 0x1c, + 0xa1, 0xe7, 0xed, 0xd7, 0xa0, 0xd7, 0x53, 0xa1, 0x21, 0xf1, 0x5f, 0x1e, + 0xec, 0x36, 0x0d, 0x2c, 0xce, 0x74, 0x4a, 0x0c, 0x97, 0x5a, 0x76, 0x62, + 0x18, 0x9c, 0xc3, 0xc1, 0xc4, 0x5e, 0xf1, 0xfa, 0xe6, 0x4b, 0x15, 0xda, + 0xfa, 0xfd, 0xe9, 0x98, 0x09, 0xc3, 0x67, 0x63, 0x1f, 0x28, 0x37, 0xf0, + 0x59, 0x4b, 0x4b, 0xa3, 0xd1, 0x41, 0x94, 0xa6, 0x05, 0xb0, 0x93, 0xee, + 0x41, 0xa4, 0xce, 0xee, 0xea, 0xc4, 0x43, 0x6e, 0xab, 0x65, 0x70, 0xe3, + 0x4d, 0xf1, 0x02, 0xf5, 0x0f, 0xd5, 0x5e, 0xfd, 0x03, 0xcd, 0x22, 0x27, + 0x90, 0xf4, 0x98, 0xa2, 0xc0, 0xb4, 0xd5, 0x04, 0xfa, 0x75, 0x22, 0x4c, + 0xe7, 0xdd, 0xef, 0x3a, 0x1d, 0xb6, 0x00, 0x58, 0xcd, 0x5a, 0xbc, 0x12, + 0xea, 0x5a, 0xda, 0xa9, 0x18, 0x0e, 0xff, 0x51, 0xc4, 0xaf, 0xc8, 0x95, + 0xfb, 0x92, 0xdf, 0x99, 0xc9, 0x4e, 0xfe, 0xb1, 0xb0, 0xca, 0xa1, 0xba, + 0x90, 0xc8, 0x07, 0x34, 0x52, 0x6d, 0xd8, 0x05, 0x72, 0x2e, 0xee, 0x98, + 0xc0, 0x1e, 0x25, 0xb3, 0xa2, 0xb4, 0x9c, 0xa5, 0xdc, 0xd3, 0xb1, 0xdf, + 0x17, 0xd9, 0xda, 0xe9, 0x5d, 0x41, 0xca, 0xc7, 0xe4, 0x94, 0x0d, 0x67, + 0xba, 0x9c, 0xcf, 0x52, 0xf0, 0x00, 0x54, 0xe0, 0xbd, 0x3c, 0xc7, 0xb9, + 0x6a, 0x11, 0xc6, 0xd1, 0x62, 0xc3, 0xcf, 0xc2, 0x6a, 0x44, 0xeb, 0x41, + 0x43, 0x54, 0xe2, 0xf5, 0xc4, 0x11, 0xd7, 0x6a, 0xf2, 0x76, 0xa9, 0x16, + 0xae, 0xe2, 0x11, 0xfb, 0x04, 0x3d, 0xee, 0xd1, 0x98, 0x30, 0x0b, 0x6b, + 0x8a, 0x6f, 0x45, 0xb7, 0x01, 0x64, 0x46, 0x32, 0x61, 0xd5, 0x05, 0xfa, + 0xb1, 0x14, 0x54, 0x39, 0x13, 0x9b, 0xd5, 0x1d, 0x5c, 0xad, 0xd0, 0x5e, + 0x6d, 0xb3, 0xa1, 0xb3, 0xc5, 0x8d, 0xf8, 0x12, 0xd9, 0x5f, 0x94, 0x27, + 0xdf, 0x30, 0xc8, 0x0e, 0x3a, 0x46, 0x70, 0x5c, 0x4c, 0xaa, 0x24, 0xc3, + 0x50, 0x62, 0x52, 0xc8, 0x63, 0x64, 0xc9, 0x49, 0x74, 0x1c, 0xd2, 0x49, + 0x0f, 0x20, 0x69, 0x53, 0x97, 0x34, 0xc0, 0x92, 0x48, 0x28, 0x7b, 0x64, + 0xca, 0xea, 0x07, 0x6c, 0x63, 0x3e, 0xb6, 0xdb, 0xd5, 0x52, 0x9d, 0x7a, + 0x5f, 0x46, 0xc1, 0xb9, 0x3e, 0xe2, 0xe9, 0xeb, 0x04, 0x65, 0xc0, 0x74, + 0x4b, 0x07, 0x6a, 0x19, 0x4a, 0x9d, 0x05, 0xa0, 0xba, 0xae, 0x74, 0xef, + 0x62, 0x09, 0x57, 0x36, 0xe5, 0x9c, 0x54, 0x59, 0x3d, 0x04, 0xf0, 0xfb, + 0x6f, 0x89, 0x13, 0x1f, 0x1f, 0x88, 0x03, 0x6b, 0x0c, 0xeb, 0x53, 0xac, + 0x3a, 0x18, 0xa4, 0x93, 0xcc, 0x4f, 0xf5, 0x92, 0x44, 0x23, 0x9e, 0x67, + 0xf0, 0xf5, 0x2f, 0xb9, 0xc9, 0x34, 0x76, 0x97, 0x1d, 0x94, 0x75, 0x3f, + 0x47, 0x97, 0xe0, 0x30, 0xcc, 0xff, 0xd2, 0x7a, 0x3b, 0x04, 0xa7, 0xa5, + 0x62, 0x9e, 0xe4, 0x8f, 0xd8, 0x62, 0xee, 0x1d, 0x1c, 0xff, 0xad, 0x18, + 0xc9, 0x66, 0x47, 0x36, 0xfb, 0x2e, 0x74, 0x2a, 0xe7, 0x5f, 0xb2, 0x12, + 0xd2, 0x9e, 0xae, 0x2b, 0x92, 0xb8, 0x53, 0x66, 0x22, 0x5c, 0xa8, 0xaf, + 0x4f, 0x29, 0xab, 0x64, 0x50, 0x09, 0xe9, 0x2f, 0x2e, 0x62, 0x2e, 0x0e, + 0x8a, 0xd6, 0xeb, 0xa7, 0x5d, 0x3e, 0x9e, 0xe1, 0x39, 0x52, 0x13, 0x57, + 0x54, 0x5c, 0x78, 0xed, 0xb3, 0xfc, 0x5f, 0xa1, 0xf3, 0x2a, 0x77, 0x90, + 0xa9, 0x09, 0xa1, 0x05, 0x3b, 0xa9, 0x6a, 0xf5, 0xc4, 0xfa, 0x97, 0x79, + 0x64, 0x57, 0x1a, 0xf1, 0x74, 0xe5, 0x16, 0x93, 0xa9, 0xef, 0xe6, 0xdf, + 0x36, 0xd2, 0xd0, 0xe6, 0xb8, 0xdd, 0xe9, 0x13, 0x4c, 0xcd, 0x22, 0x98, + 0xc1, 0x94, 0xbb, 0x04, 0x2a, 0x4a, 0x69, 0x10, 0x5a, 0xcb, 0x1d, 0x9e, + 0xc4, 0x3d, 0x6d, 0x0e, 0xe0, 0x12, 0xb4, 0xe1, 0x6c, 0x55, 0x6f, 0xa3, + 0xf5, 0x1b, 0x0c, 0xe5, 0x1c, 0x99, 0x8b, 0x23, 0x23, 0xbc, 0x33, 0xe4, + 0xd4, 0x15, 0xfd, 0xcc, 0x90, 0x87, 0xb5, 0x0e, 0x24, 0xba, 0x20, 0x1b, + 0xcf, 0x67, 0x98, 0x1a, 0x35, 0xe7, 0xc3, 0x95, 0x29, 0xd6, 0xd2, 0x4f, + 0xe4, 0x14, 0xd5, 0xa1, 0x93, 0xff, 0x24, 0x0e, 0xfc, 0xb7, 0xd6, 0xde, + 0x05, 0xc5, 0x2f, 0xaa, 0x92, 0xd4, 0xd8, 0xac, 0x8f, 0x67, 0x45, 0xdb, + 0x36, 0x19, 0x15, 0x09, 0x9a, 0x3f, 0x2a, 0x56, 0xd5, 0xa9, 0x26, 0xb6, + 0xcb, 0x19, 0xf3, 0x6a, 0xbb, 0xba, 0xba, 0xa3, 0x68, 0x90, 0x0f, 0xb1, + 0x98, 0x14, 0x33, 0xd8, 0x12, 0xdf, 0xef, 0xe5, 0x01, 0x93, 0xab, 0xf8, + 0x93, 0x40, 0xbd, 0xa0, 0x01, 0x34, 0x54, 0xfd, 0xa0, 0xc4, 0xc3, 0xf3, + 0x6b, 0x90, 0x30, 0xc1, 0xbe, 0xd8, 0xbb, 0xab, 0x71, 0xaa, 0xe5, 0x3b, + 0x2d, 0x5d, 0x6e, 0x00, 0x34, 0xa8, 0x02, 0x34, 0xa9, 0x67, 0x95, 0xcd, + 0xed, 0xa2, 0x25, 0x55, 0xc9, 0x03, 0x1c, 0x30, 0xe7, 0xdf, 0xe6, 0xe7, + 0x2b, 0x5a, 0x9a, 0xcd, 0xa8, 0xf0, 0x4e, 0xe4, 0xd7, 0x90, 0x5f, 0x4e, + 0xbf, 0x5d, 0x68, 0x12, 0x1c, 0x4c, 0x68, 0x03, 0x9c, 0x49, 0xcb, 0xe6, + 0xc4, 0xfd, 0xad, 0xd5, 0xa8, 0xd8, 0xda, 0x2f, 0x13, 0xbc, 0x42, 0x61, + 0xa5, 0x0a, 0x1a, 0xe9, 0x5e, 0x5c, 0x01, 0x7c, 0xca, 0x73, 0x6f, 0x32, + 0xc1, 0x96, 0x24, 0x9d, 0x12, 0x20, 0x11, 0x6a, 0xf6, 0xbc, 0xff, 0x6a, + 0xc1, 0x58, 0x0d, 0xb9, 0xad, 0xc5, 0xde, 0x69, 0x37, 0xbe, 0xd9, 0x93, + 0xcc, 0x2b, 0xe9, 0x13, 0x45, 0xa0, 0x6c, 0x3f, 0x44, 0x34, 0xaf, 0x43, + 0x6d, 0xae, 0xef, 0xb2, 0x65, 0x03, 0xc1, 0xef, 0x10, 0x1e, 0xd8, 0x6e, + 0xb5, 0xb9, 0x03, 0xd8, 0x6e, 0x2f, 0x53, 0xe6, 0xc0, 0xaf, 0x44, 0xd2, + 0xd8, 0x15, 0x56, 0x15, 0x59, 0xd6, 0xd4, 0xe4, 0x1a, 0x25, 0xd5, 0xcf, + 0xe7, 0x6a, 0x55, 0xd4, 0xf8, 0x42, 0x4c, 0xcb, 0x9a, 0x48, 0x4d, 0x27, + 0x61, 0x4c, 0x36, 0x2b, 0xcb, 0x10, 0xba, 0xf7, 0xe3, 0x23, 0x27, 0xc5, + 0x6a, 0x1b, 0x94, 0x69, 0x64, 0xb1, 0x8c, 0xdb, 0xd4, 0x0d, 0x32, 0x3e, + 0x58, 0x73, 0xa8, 0x2f, 0x3d, 0x22, 0xd9, 0x0d, 0x2a, 0x52, 0xf0, 0xdd, + 0xeb, 0x21, 0x42, 0xc7, 0x59, 0x96, 0x09, 0x93, 0x5a, 0x70, 0xc3, 0x21, + 0x5f, 0xce, 0xc2, 0xdd, 0xcf, 0x61, 0xed, 0x1c, 0xfb, 0x2f, 0x57, 0xf7, + 0x31, 0xb8, 0x3e, 0x92, 0x29, 0xd4, 0x47, 0x6a, 0x19, 0x66, 0x00, 0xc2, + 0xc4, 0x6c, 0xb5, 0xc5, 0x68, 0x24, 0xa8, 0x64, 0x26, 0x72, 0x43, 0x20, + 0x9f, 0xf1, 0x3f, 0xac, 0x64, 0xb5, 0x12, 0x26, 0x13, 0x76, 0x52, 0x05, + 0xda, 0x57, 0xe3, 0x53, 0x73, 0x30, 0x21, 0x27, 0x75, 0x8d, 0x37, 0xd1, + 0x77, 0x40, 0x97, 0x2a, 0xb7, 0x0b, 0x2e, 0x9e, 0x4c, 0x36, 0x75, 0x44, + 0x15, 0xdb, 0x96, 0x70, 0xf9, 0x33, 0x9a, 0x1e, 0x6e, 0x13, 0x05, 0x38, + 0x2c, 0xbf, 0x0a, 0xdd, 0x2b, 0x2b, 0x38, 0x77, 0xa9, 0x00, 0x2d, 0x5e, + 0xee, 0x4b, 0xf3, 0x20, 0x7a, 0x90, 0x97, 0x44, 0xdf, 0x55, 0xfd, 0x50, + 0xe3, 0x24, 0x25, 0xa9, 0xd9, 0x3f, 0x6d, 0x09, 0x32, 0x67, 0xb5, 0x43, + 0xf1, 0xc7, 0xa7, 0xfb, 0x92, 0xde, 0xc3, 0xbf, 0x64, 0x6b, 0x35, 0xda, + 0x08, 0x94, 0x68, 0xb0, 0xc8, 0x3f, 0xb5, 0x9f, 0x15, 0x05, 0xff, 0x6c, + 0xbc, 0x22, 0x61, 0xf4, 0x67, 0xf8, 0x1f, 0x2e, 0x91, 0xc8, 0x12, 0xdc, + 0xcb, 0x22, 0x05, 0xb8, 0xab, 0x0d, 0x0e, 0xd7, 0x04, 0x8e, 0x32, 0x0e, + 0xfe, 0x72, 0x79, 0xc3, 0xba, 0xd8, 0x68, 0x3e, 0x5d, 0xab, 0xa0, 0xf8, + 0x26, 0x57, 0xe4, 0x20, 0x91, 0x0a, 0xde, 0x52, 0x95, 0xbc, 0xb7, 0x71, + 0x50, 0xe4, 0x3f, 0x07, 0x4c, 0xa8, 0x6a, 0xb6, 0xa0, 0x95, 0xe2, 0x31, + 0x8f, 0x5f, 0xfa, 0xdd, 0xee, 0x02, 0x23, 0x56, 0xf1, 0xdd, 0x1a, 0xa6, + 0xa0, 0x2d, 0x46, 0x36, 0x6c, 0x79, 0xe8, 0x67, 0x43, 0xdd, 0xe7, 0x2e, + 0x25, 0xda, 0x35, 0x6f, 0x63, 0xf1, 0x2c, 0x6c, 0x61, 0xaa, 0xb7, 0x51, + 0x91, 0xa1, 0x7c, 0x54, 0x9a, 0xf6, 0x3c, 0x3f, 0xa8, 0xba, 0x4d, 0xee, + 0xb6, 0xab, 0xa5, 0x05, 0xc6, 0xb6, 0xe8, 0x2f, 0x1b, 0x99, 0xb0, 0x45, + 0x3e, 0xc3, 0x50, 0x26, 0x0b, 0x10, 0x61, 0x5a, 0xc6, 0x25, 0x2d, 0x07, + 0xb6, 0x28, 0x59, 0xf3, 0xb4, 0x02, 0x61, 0xa0, 0xd0, 0x0a, 0xae, 0xd6, + 0x3c, 0xcc, 0x5f, 0xfb, 0xc0, 0xfd, 0xeb, 0x7b, 0xe2, 0x66, 0xc5, 0x98, + 0x70, 0x50, 0x31, 0x3a, 0x12, 0x45, 0xf4, 0x1c, 0xba, 0xa6, 0x92, 0x51, + 0xae, 0x68, 0xec, 0xb0, 0x1a, 0xd9, 0x45, 0x00, 0xd6, 0x9e, 0xad, 0x64, + 0xfe, 0xd9, 0xfb, 0xcc, 0x57, 0xff, 0x9e, 0xa3, 0x71, 0xe7, 0x7a, 0xaf, + 0x26, 0x31, 0x31, 0x6a, 0x41, 0xa4, 0x4d, 0x68, 0xbc, 0xcb, 0xfa, 0xb4, + 0x3a, 0x1c, 0x3a, 0x8f, 0xcd, 0xc1, 0x95, 0xb2, 0x46, 0x72, 0xf7, 0xfc, + 0x20, 0xe2, 0x2f, 0x0f, 0xbd, 0x74, 0xe1, 0x2a, 0xd5, 0xf6, 0xe9, 0xe1, + 0x45, 0x7d, 0x95, 0xb0, 0x49, 0xce, 0xe8, 0x53, 0x69, 0x46, 0x9d, 0x03, + 0x5f, 0x15, 0x2e, 0x92, 0x4c, 0xb7, 0xf1, 0x43, 0x67, 0x8a, 0x43, 0xc6, + 0x90, 0xec, 0xb5, 0x5d, 0xd5, 0x64, 0x16, 0x6e, 0xf0, 0xad, 0x4e, 0xf0, + 0x56, 0xe8, 0x77, 0xd5, 0x47, 0x47, 0x41, 0xc9, 0x98, 0x3a, 0xcb, 0xe0, + 0x01, 0x77, 0x93, 0x15, 0xe0, 0xd3, 0x93, 0xbe, 0xe1, 0x97, 0xe0, 0x21, + 0x60, 0x2b, 0xf1, 0x4a, 0x62, 0x29, 0x11, 0xe9, 0x61, 0x55, 0xc4, 0x57, + 0x04, 0xa8, 0xb3, 0xb3, 0x61, 0xd7, 0xa6, 0xce, 0x50, 0xd2, 0xc3, 0x38, + 0xda, 0xc2, 0x23, 0x67, 0x37, 0x09, 0xa7, 0xfd, 0x29, 0xdc, 0xcc, 0x52, + 0x65, 0xea, 0x3f, 0xcc, 0x67, 0x5e, 0x3b, 0xd4, 0x59, 0x59, 0x12, 0x9b, + 0xf1, 0xd2, 0x43, 0x46, 0x54, 0xcd, 0xb9, 0xbe, 0x71, 0xb6, 0x6d, 0x6a, + 0x62, 0xc5, 0x59, 0xc1, 0x21, 0xf7, 0x4c, 0x91, 0x64, 0xe0, 0xd7, 0xd9, + 0x34, 0x60, 0x0d, 0xb2, 0x93, 0xd8, 0xd3, 0x01, 0x8b, 0xf3, 0x9c, 0x6c, + 0xff, 0x63, 0xca, 0xd2, 0xf4, 0x76, 0xe3, 0x60, 0x52, 0x5c, 0x0e, 0xa3, + 0x13, 0xc8, 0xd9, 0xa7, 0x13, 0x6d, 0x1b, 0x29, 0xc0, 0xb1, 0x54, 0x31, + 0x33, 0x55, 0x44, 0x0a, 0x0a, 0x96, 0x3f, 0xf0, 0xb2, 0x64, 0x23, 0xa1, + 0xc8, 0x08, 0x01, 0x94, 0x2f, 0xc8, 0x0a, 0xfb, 0x93, 0x38, 0xe4, 0xc1, + 0xd9, 0xea, 0x46, 0x96, 0xdd, 0x5d, 0x62, 0xfc, 0xb0, 0x4d, 0x17, 0xe8, + 0xa0, 0xd4, 0x35, 0x98, 0x65, 0xb0, 0x27, 0x97, 0xbc, 0xe8, 0x48, 0x38, + 0x90, 0x9b, 0x6e, 0xf1, 0xd2, 0x17, 0x1b, 0xbf, 0x03, 0xc6, 0xa3, 0x42, + 0xaf, 0xdc, 0x44, 0x9d, 0x9e, 0x69, 0x67, 0x33, 0x61, 0xfb, 0x96, 0xfa, + 0xff, 0xf4, 0xa8, 0x3c, 0xb6, 0x42, 0xd2, 0x4c, 0xc0, 0xa8, 0x2a, 0x4b, + 0x37, 0x78, 0x41, 0x94, 0xf6, 0x04, 0xb9, 0x54, 0xe4, 0x2b, 0xfc, 0xed, + 0xf5, 0xf7, 0x62, 0x23, 0x44, 0xc4, 0xd7, 0x5a, 0xeb, 0xc2, 0x3d, 0x4c, + 0x41, 0x22, 0xa0, 0xe3, 0x22, 0xbc, 0x91, 0x69, 0x37, 0x3f, 0x94, 0xfd, + 0x07, 0xa7, 0x6e, 0x53, 0x27, 0xdc, 0xb0, 0x14, 0x8d, 0x0a, 0x08, 0x31, + 0xba, 0xf0, 0xd0, 0xda, 0xa6, 0x7a, 0xc0, 0x4c, 0x9d, 0x3b, 0x8f, 0xee, + 0x11, 0xc7, 0x9f, 0xc9, 0xcc, 0x4c, 0x26, 0x51, 0xb4, 0x10, 0xde, 0xc2, + 0xa3, 0xe0, 0xaa, 0x7c, 0x9c, 0x27, 0x8d, 0x04, 0x8e, 0xfc, 0xe4, 0x68, + 0x93, 0xf9, 0x67, 0x28, 0xa0, 0xe6, 0xca, 0xbd, 0x5a, 0x64, 0x98, 0x9f, + 0xe3, 0x7b, 0x16, 0x5d, 0x61, 0xcc, 0x4c, 0x64, 0x04, 0x1b, 0xcc, 0xa6, + 0xa2, 0x31, 0x28, 0xa2, 0xac, 0xd0, 0xce, 0x40, 0x19, 0xe7, 0xf9, 0xea, + 0xc5, 0x98, 0x50, 0x16, 0x38, 0xad, 0x58, 0x21, 0x2e, 0x10, 0x48, 0x4f, + 0xe7, 0xc0, 0xc0, 0x6c, 0xcd, 0xe2, 0xc3, 0xcd, 0xc5, 0xfc, 0x26, 0x91, + 0xea, 0xcf, 0x52, 0x97, 0x9f, 0xdc, 0x2c, 0x45, 0xd8, 0x50, 0xf8, 0x75, + 0xa2, 0x93, 0x52, 0x2b, 0x23, 0xd3, 0x30, 0x9d, 0xa7, 0xf7, 0xbb, 0xc2, + 0xd2, 0xb7, 0x9d, 0xec, 0xf9, 0x9a, 0xec, 0x3e, 0xc0, 0xce, 0x64, 0xb8, + 0xf5, 0x41, 0x4e, 0x06, 0xa1, 0x25, 0xf2, 0x40, 0xee, 0x07, 0xec, 0x6d, + 0x9a, 0xd0, 0x5c, 0xdd, 0xe9, 0xf5, 0x56, 0xf9, 0x2e, 0xf5, 0xdb, 0x69, + 0xc9, 0x3e, 0xb5, 0x0c, 0xbc, 0x29, 0xa4, 0xa9, 0x55, 0x9b, 0xf6, 0xab, + 0x1f, 0x55, 0x9d, 0x25, 0xd2, 0xde, 0x3f, 0xa0, 0xe5, 0x1c, 0xb3, 0x90, + 0x2f, 0x6c, 0xaf, 0xb5, 0x6d, 0x23, 0x15, 0xab, 0x91, 0x55, 0x5f, 0x02, + 0x20, 0x22, 0x8e, 0xc1, 0x4a, 0x63, 0xa6, 0x5e, 0x85, 0x99, 0x58, 0xdc, + 0xde, 0xb0, 0x76, 0x9f, 0x21, 0x4d, 0xe9, 0x47, 0xcc, 0x3f, 0x02, 0x91, + 0x75, 0x67, 0xe5, 0x6a, 0x2c, 0xc3, 0x69, 0x95, 0x2d, 0x74, 0x77, 0xf7, + 0x1d, 0xe1, 0x12, 0x2b, 0xcf, 0x4c, 0x7b, 0xcf, 0xbe, 0x24, 0x1d, 0x07, + 0x34, 0xd3, 0x67, 0xa8, 0xb9, 0x76, 0x2a, 0x3e, 0xfd, 0xb5, 0xcd, 0xf6, + 0x29, 0x07, 0x4e, 0x17, 0xcf, 0x28, 0xdd, 0x90, 0x4b, 0x17, 0x24, 0x55, + 0xdc, 0x78, 0xe5, 0xf4, 0x97, 0x31, 0x3d, 0xfa, 0x96, 0xe2, 0x99, 0x61, + 0xb1, 0xcb, 0xa4, 0x7b, 0x4e, 0x5d, 0x6a, 0xf8, 0xb2, 0x79, 0xfc, 0xa9, + 0xd9, 0x27, 0x46, 0xdd, 0x52, 0xdf, 0x24, 0x66, 0x1c, 0xa6, 0xbc, 0x18, + 0x13, 0x72, 0x38, 0x53, 0xac, 0x1b, 0x67, 0x1f, 0x30, 0xae, 0x5a, 0xf3, + 0x55, 0xd0, 0xe1, 0x23, 0x9a, 0x46, 0xa4, 0xbb, 0x68, 0x73, 0x30, 0xda, + 0xb7, 0x3b, 0xff, 0xd1, 0x0d, 0xe0, 0xf7, 0xda, 0x36, 0x3a, 0x7a, 0x19, + 0xf5, 0x2e, 0xf4, 0xda, 0xa4, 0x09, 0x94, 0xb8, 0x18, 0xad, 0x6b, 0xf6, + 0x64, 0xbf, 0x2a, 0x04, 0xc6, 0xde, 0x0f, 0x45, 0x27, 0x3a, 0x3d, 0x61, + 0xf5, 0xde, 0x38, 0x1d, 0x23, 0x23, 0x70, 0x00, 0xfc, 0x0c, 0x5c, 0x96, + 0xc1, 0x21, 0x78, 0x25, 0x24, 0x71, 0xd1, 0xe2, 0xe9, 0x1a, 0x2f, 0x48, + 0x4d, 0x09, 0x24, 0x27, 0xe4, 0xe7, 0x42, 0x76, 0x92, 0x93, 0x7a, 0x62, + 0x76, 0xc6, 0xd7, 0xdf, 0xe4, 0x5e, 0x0e, 0xfc, 0x4e, 0x0a, 0x65, 0x63, + 0x51, 0x90, 0xfd, 0x92, 0x5f, 0x9a, 0x49, 0xa9, 0x6c, 0xb1, 0xb6, 0xe6, + 0xab, 0xf7, 0xb9, 0x39, 0xc0, 0xed, 0x1d, 0x65, 0x9c, 0x24, 0x21, 0xc1, + 0x0d, 0xd6, 0x9a, 0xbe, 0xd4, 0x74, 0xa2, 0x70, 0xab, 0x0b, 0x45, 0xf0, + 0xc9, 0xaa, 0xf1, 0x49, 0x0b, 0x6c, 0x20, 0xdc, 0x37, 0x2b, 0x13, 0x68, + 0x48, 0x0e, 0xd8, 0xd1, 0x67, 0xd8, 0xa3, 0x7e, 0xd7, 0xb7, 0x50, 0xc8, + 0x14, 0x58, 0x6a, 0x04, 0xa5, 0x70, 0x22, 0x2d, 0x41, 0xea, 0x28, 0xb7, + 0xf0, 0xde, 0xc4, 0xe4, 0x5b, 0x4d, 0xc1, 0x33, 0x9e, 0x14, 0x32, 0xa8, + 0x9b, 0xc8, 0xd9, 0x5b, 0x95, 0x2a, 0x91, 0x9d, 0xe8, 0x15, 0x19, 0x9b, + 0x38, 0xf3, 0x35, 0x69, 0x3e, 0xd3, 0x4b, 0xcc, 0xf2, 0x94, 0x5a, 0xaf, + 0x91, 0xa4, 0xa1, 0x03, 0x48, 0x5f, 0x6d, 0x16, 0x56, 0x03, 0x5a, 0xcb, + 0x99, 0x19, 0x45, 0x9c, 0xba, 0xc9, 0xbc, 0x5b, 0x0f, 0xf5, 0xde, 0x70, + 0xa3, 0x70, 0x0d, 0x3f, 0x3e, 0x5c, 0x4d, 0x5a, 0x1a, 0x46, 0x1b, 0x44, + 0x4a, 0x73, 0xfa, 0xb1, 0xc4, 0x42, 0x7b, 0x0c, 0x15, 0x0d, 0x35, 0xc4, + 0xa3, 0xea, 0x17, 0xa0, 0x0b, 0xfb, 0x4d, 0x1b, 0x2f, 0x96, 0x1f, 0xaa, + 0xc0, 0xad, 0xdc, 0xf3, 0xb2, 0xb1, 0x44, 0x1f, 0x39, 0xc7, 0x33, 0x18, + 0xad, 0xe1, 0x50, 0x7d, 0xf9, 0x2a, 0x90, 0xf2, 0x06, 0xce, 0x07, 0xae, + 0x9f, 0xbc, 0x4d, 0xae, 0x30, 0xdd, 0x47, 0xa2, 0xd3, 0x6d, 0x0c, 0xc6, + 0xb7, 0xae, 0xf5, 0x38, 0xa3, 0x00, 0x59, 0x6a, 0x00, 0x04, 0xd2, 0x77, + 0x0a, 0x58, 0xc9, 0xaf, 0x1b, 0x59, 0x29, 0xf3, 0xdd, 0x58, 0xcf, 0xa1, + 0x6d, 0xb4, 0x66, 0x23, 0x9f, 0x9b, 0x41, 0x2a, 0xc8, 0x28, 0x34, 0x77, + 0x3a, 0x1f, 0xa5, 0xde, 0x4b, 0x3f, 0xc7, 0x19, 0xf5, 0xdb, 0x98, 0xc4, + 0x6c, 0x2f, 0x34, 0x20, 0xc9, 0x52, 0x16, 0x60, 0xbc, 0x04, 0xd5, 0xff, + 0x4b, 0x07, 0x28, 0x5a, 0x3a, 0x48, 0x5b, 0x96, 0xee, 0x1f, 0xf1, 0xb4, + 0x9b, 0xb5, 0x64, 0xde, 0x1c, 0xd5, 0x3c, 0x1b, 0x98, 0x11, 0xc7, 0x0b, + 0x97, 0x00, 0x2f, 0x8f, 0xf9, 0x24, 0x4d, 0xba, 0x75, 0x6a, 0xce, 0xd8, + 0x7a, 0xee, 0x02, 0xd5, 0x19, 0xd6, 0x26, 0x40, 0xa7, 0x78, 0x76, 0x1a, + 0x17, 0xc2, 0xe6, 0x5a, 0x6e, 0x24, 0xb1, 0x17, 0xf8, 0x9f, 0xdc, 0x64, + 0xf0, 0x59, 0xc5, 0xfc, 0x4c, 0xbb, 0x3d, 0x3f, 0x70, 0x2c, 0x0d, 0xf5, + 0x6c, 0x96, 0x46, 0x1a, 0x1e, 0x5f, 0xd1, 0x3a, 0x00, 0x9a, 0x9d, 0x63, + 0xe6, 0xd1, 0xa2, 0x5a, 0x4a, 0x50, 0xa8, 0xd5, 0x91, 0x90, 0x69, 0x58, + 0x65, 0x00, 0xc7, 0xf1, 0xa6, 0x45, 0xfd, 0x5a, 0xe6, 0x05, 0x4b, 0xb2, + 0x3a, 0xdf, 0xa9, 0xd9, 0xe5, 0xa6, 0xe5, 0xe2, 0x5b, 0x3b, 0x2f, 0x57, + 0x6c, 0xc4, 0x06, 0xe1, 0x8e, 0x15, 0x98, 0xc8, 0x5e, 0x63, 0xba, 0x37, + 0xe6, 0x91, 0x5f, 0x1c, 0x5b, 0x77, 0xb5, 0x91, 0x07, 0x3a, 0xa6, 0x67, + 0x6d, 0xdf, 0x15, 0x62, 0x6b, 0x3b, 0xed, 0xa2, 0xc7, 0x46, 0x52, 0x8f, + 0xf2, 0x9f, 0x69, 0x00, 0xb8, 0x49, 0xcf, 0xd4, 0xf0, 0x95, 0x51, 0xda, + 0x0f, 0x4e, 0x0d, 0x11, 0x2f, 0x27, 0x73, 0xe9, 0x13, 0xcb, 0xa1, 0xfc, + 0x6b, 0x45, 0xf0, 0xfd, 0xc7, 0x17, 0xaa, 0x0c, 0xac, 0x98, 0xc4, 0x6c, + 0xf0, 0x32, 0x45, 0x67, 0xfe, 0x6f, 0x2e, 0xfb, 0xec, 0x19, 0xda, 0xbd, + 0x93, 0x5f, 0x50, 0xc2, 0x22, 0x9a, 0x3a, 0x5b, 0x31, 0xf5, 0x4e, 0x91, + 0xa6, 0xea, 0x67, 0xdd, 0x69, 0xf4, 0xd7, 0xea, 0x02, 0xbe, 0x55, 0x52, + 0xb9, 0x30, 0x21, 0xe5, 0xfc, 0x9a, 0x93, 0xd6, 0x6c, 0x33, 0x06, 0xb9, + 0xe3, 0xb0, 0x6a, 0xff, 0x9e, 0xc2, 0x5e, 0x1d, 0xd6, 0xdb, 0xa1, 0x60, + 0x34, 0x5d, 0x08, 0xf9, 0xeb, 0xd6, 0x1f, 0x90, 0xf1, 0xf4, 0x07, 0x47, + 0xbf, 0xd9, 0xc9, 0xe8, 0xcf, 0xce, 0xa5, 0x1d, 0xb0, 0xd9, 0xbe, 0xc7, + 0xfb, 0xcc, 0xac, 0x3e, 0x92, 0x59, 0x0d, 0x1d, 0x65, 0x16, 0xa3, 0xdc, + 0x9b, 0x72, 0x22, 0x46, 0x04, 0xca, 0xb3, 0x5a, 0x2f, 0x3d, 0x99, 0x5c, + 0xb5, 0xb9, 0x30, 0xe3, 0xde, 0x8c, 0xba, 0xc7, 0x4c, 0xe5, 0x34, 0x6e, + 0xf4, 0x75, 0xf4, 0x38, 0x01, 0xf1, 0x61, 0xb8, 0x2b, 0xc3, 0x6f, 0xae, + 0xd1, 0x0a, 0x9d, 0x48, 0xc9, 0xe7, 0xc3, 0xe7, 0xc9, 0xe1, 0x6f, 0x96, + 0xa0, 0xc2, 0x91, 0xfd, 0xad, 0x99, 0x48, 0xde, 0xfc, 0xa3, 0x6e, 0xe3, + 0x94, 0x0e, 0xb5, 0xf6, 0x24, 0x8b, 0xce, 0x70, 0x3c, 0xdc, 0xe2, 0x66, + 0x9f, 0xe3, 0x6b, 0xc5, 0xd1, 0x97, 0x38, 0x12, 0x46, 0x37, 0xd6, 0x9a, + 0x4c, 0x6d, 0x4a, 0x2d, 0xc3, 0x28, 0x20, 0x2f, 0x55, 0x67, 0x17, 0x71, + 0xd3, 0x5c, 0xdc, 0xa3, 0x23, 0x60, 0x25, 0x2d, 0xe0, 0xc2, 0xed, 0xee, + 0x67, 0x9f, 0x26, 0xfb, 0x2f, 0x63, 0xf2, 0x6a, 0x23, 0x45, 0x26, 0x2c, + 0x33, 0x8a, 0xf2, 0xd1, 0xb2, 0x77, 0x99, 0x98, 0xd6, 0x18, 0xfe, 0xf3, + 0xff, 0xa4, 0x36, 0x03, 0xf4, 0xf5, 0xb1, 0xca, 0xa3, 0x5f, 0xe2, 0xc6, + 0xb2, 0x55, 0x2c, 0xaa, 0x64, 0xef, 0x28, 0x3a, 0x9e, 0x98, 0x01, 0x57, + 0x49, 0x98, 0x61, 0x4f, 0x42, 0x57, 0x00, 0x19, 0xb9, 0xa8, 0xec, 0xed, + 0x2b, 0x63, 0xf3, 0x0c, 0x3a, 0x1f, 0x10, 0xab, 0xe9, 0x6e, 0x61, 0x69, + 0xd1, 0x2d, 0xf3, 0x1f, 0xaa, 0x00, 0x57, 0xe2, 0xab, 0x74, 0xcd, 0xff, + 0x97, 0x2c, 0x3b, 0x67, 0xae, 0xa3, 0xfc, 0x69, 0xa9, 0x4e, 0x42, 0x07, + 0xfc, 0xbf, 0x36, 0x1a, 0xef, 0x6d, 0x6d, 0x14, 0x61, 0x30, 0x27, 0x98, + 0xfa, 0xf8, 0xc9, 0x70, 0xb4, 0xaa, 0x53, 0x48, 0x72, 0x3f, 0x58, 0x69, + 0x8d, 0x08, 0xc8, 0x09, 0x2b, 0xfc, 0x1d, 0xa1, 0x92, 0xae, 0x62, 0xa0, + 0xea, 0x05, 0x40, 0xac, 0x9c, 0xaf, 0x0e, 0xf4, 0x1e, 0x45, 0x33, 0xee, + 0x31, 0x39, 0x08, 0x4b, 0x54, 0x02, 0x2d, 0x03, 0x1c, 0xe6, 0x2d, 0x0c, + 0xd0, 0x92, 0x44, 0xd6, 0xa1, 0x57, 0x4e, 0x17, 0xde, 0xe6, 0x4f, 0x6a, + 0x07, 0x9f, 0x58, 0xe2, 0x27, 0xdb, 0xa9, 0x0c, 0x19, 0x56, 0xa3, 0xb4, + 0xc4, 0xe8, 0xa3, 0x52, 0x9f, 0x6a, 0xc9, 0xb1, 0xda, 0xe9, 0xef, 0x12, + 0xc1, 0x6d, 0x5b, 0x04, 0x20, 0x93, 0xac, 0xf4, 0x38, 0x95, 0xdb, 0x50, + 0xa6, 0x2e, 0x5c, 0x3f, 0x2d, 0x32, 0x50, 0x03, 0x73, 0x64, 0x3a, 0xd5, + 0xfd, 0x98, 0x1c, 0x57, 0xc3, 0xe7, 0xf7, 0x14, 0x13, 0x15, 0x2a, 0xa2, + 0x5f, 0xa0, 0x67, 0xdd, 0x67, 0x00, 0x09, 0xc6, 0xfe, 0xad, 0x06, 0x4c, + 0x5e, 0x9a, 0x5b, 0x55, 0x06, 0x8c, 0x9a, 0x2a, 0x51, 0x0e, 0x4f, 0x15, + 0xcc, 0xe1, 0x53, 0x9c, 0x43, 0x37, 0xc1, 0x3e, 0x02, 0x4b, 0x98, 0x6f, + 0x9b, 0x60, 0x31, 0x2c, 0x2b, 0x9d, 0xda, 0xe0, 0x1d, 0xe4, 0x49, 0x66, + 0x65, 0x18, 0xfb, 0x24, 0x97, 0xe0, 0x2d, 0xf5, 0x44, 0x23, 0x09, 0x01, + 0xf9, 0xf5, 0x29, 0xff, 0x01, 0x36, 0xb9, 0x0e, 0x9b, 0xb3, 0x23, 0x1e, + 0xe5, 0x12, 0xbb, 0x3a, 0x04, 0x14, 0xb8, 0x23, 0x43, 0x95, 0xc1, 0x9d, + 0x57, 0x45, 0x46, 0x4c, 0x8f, 0x35, 0x25, 0x5f, 0x2b, 0xd9, 0xc6, 0xdd, + 0x61, 0xb8, 0xbb, 0x4d, 0x49, 0xef, 0x6e, 0x0c, 0x50, 0x07, 0xc9, 0x9b, + 0x2e, 0xb7, 0xbe, 0x23, 0xc3, 0xcf, 0x9d, 0xeb, 0x13, 0xc8, 0xeb, 0x72, + 0x51, 0x71, 0x69, 0x35, 0xf3, 0xce, 0x35, 0x45, 0x02, 0xba, 0x44, 0x5d, + 0xaf, 0xd0, 0xe5, 0x1d, 0x9b, 0x18, 0xbb, 0x62, 0xce, 0xaf, 0x40, 0x48, + 0x40, 0x2a, 0x5d, 0xcd, 0xa7, 0x2b, 0x8f, 0xf4, 0x4a, 0x4c, 0xe1, 0x59, + 0x40, 0x63, 0x33, 0xae, 0xd8, 0x9d, 0x4d, 0x11, 0x3d, 0x2d, 0x11, 0xc6, + 0x8c, 0xa9, 0xab, 0xa2, 0x08, 0xb8, 0xbf, 0x09, 0x66, 0xbc, 0xd7, 0xab, + 0xce, 0x0d, 0xe0, 0x9e, 0x51, 0x2f, 0x5c, 0xc7, 0x21, 0xb9, 0xcf, 0xc4, + 0x8b, 0xc0, 0x4b, 0x04, 0x1b, 0xfd, 0x43, 0xcf, 0xa4, 0x72, 0x62, 0x04, + 0x0b, 0x1f, 0x9f, 0x35, 0x9d, 0xa9, 0x19, 0x71, 0x06, 0xda, 0x03, 0x0f, + 0xcc, 0x3a, 0xf4, 0x3a, 0xaf, 0x07, 0x0f, 0xf2, 0x3e, 0x4a, 0xd3, 0x41, + 0x6a, 0x90, 0x35, 0x39, 0x4c, 0x1d, 0x2f, 0x05, 0xff, 0xcf, 0xc0, 0xbe, + 0x0f, 0xaf, 0x90, 0x4e, 0x45, 0x8c, 0x78, 0x4d, 0x6b, 0xf2, 0x47, 0x26, + 0xe9, 0x0d, 0xee, 0xd3, 0x97, 0x44, 0xaf, 0x6f, 0x95, 0x30, 0x9c, 0x08, + 0xe5, 0x18, 0x9e, 0xad, 0xd2, 0x2a, 0x0c, 0x21, 0x67, 0x50, 0x28, 0x4f, + 0x31, 0x9c, 0xee, 0xb2, 0x95, 0xbd, 0xef, 0xc0, 0xd0, 0x0d, 0xd4, 0x6e, + 0xff, 0x93, 0x12, 0xc3, 0x51, 0x41, 0xe4, 0x6c, 0x19, 0x09, 0xd7, 0x0a, + 0xe0, 0xea, 0x0a, 0xe7, 0xa8, 0x4b, 0x60, 0xd6, 0x0c, 0x4d, 0xb5, 0x29, + 0x01, 0x74, 0xf9, 0x40, 0x8c, 0x6b, 0x11, 0xf6, 0xe4, 0xc9, 0x3c, 0x1a, + 0xf7, 0xce, 0x2c, 0xd8, 0xe3, 0x0e, 0xc5, 0xb9, 0x6c, 0x40, 0x44, 0xc9, + 0x04, 0xf6, 0x5c, 0xe1, 0x9f, 0xc7, 0xe0, 0x68, 0xe7, 0x6a, 0x92, 0xe7, + 0xb2, 0x12, 0x72, 0x3f, 0xfd, 0xc3, 0x06, 0xeb, 0x0a, 0xab, 0x6d, 0xad, + 0x03, 0x0b, 0x5d, 0xcc, 0x49, 0x04, 0x52, 0x19, 0xd4, 0x9d, 0x67, 0xbf, + 0xd3, 0xf4, 0x22, 0x76, 0x99, 0x52, 0xf5, 0xb5, 0x15, 0x38, 0x58, 0x57, + 0x9a, 0xa2, 0xd1, 0xbb, 0x3a, 0x07, 0xe2, 0xd6, 0x8d, 0x69, 0x9e, 0x5c, + 0xf4, 0xba, 0xda, 0x4a, 0x4d, 0x73, 0xdc, 0x32, 0xfd, 0xe1, 0x3a, 0x16, + 0xf1, 0x09, 0x26, 0x3b, 0x2a, 0xa9, 0xa7, 0x2c, 0xd3, 0xcf, 0x6b, 0xc5, + 0xb5, 0xbc, 0x71, 0xb6, 0x9e, 0xa0, 0x6a, 0x69, 0xa5, 0xeb, 0x54, 0x87, + 0xe9, 0x4f, 0x69, 0x39, 0xc5, 0x54, 0x28, 0x55, 0xb9, 0xff, 0x5d, 0x9e, + 0x17, 0x8e, 0x8c, 0xd5, 0x14, 0x5c, 0xa7, 0x33, 0x5a, 0x2f, 0x2d, 0x37, + 0x0e, 0xf2, 0x54, 0x64, 0x9d, 0xdf, 0x49, 0xab, 0xd3, 0x0f, 0xbd, 0xad, + 0x19, 0xb9, 0xcf, 0x0f, 0x40, 0x62, 0x4b, 0x93, 0xd7, 0xf4, 0x3b, 0xee, + 0x2b, 0x97, 0xe3, 0x55, 0xb3, 0x5b, 0x3f, 0x93, 0xa5, 0xf1, 0x40, 0x99, + 0xa1, 0x69, 0xbd, 0xf3, 0xf0, 0xb1, 0x6e, 0x5c, 0xba, 0x4a, 0xc4, 0x51, + 0x8e, 0xe1, 0x5c, 0xb8, 0x92, 0xb5, 0x43, 0xc4, 0x9e, 0x38, 0x0d, 0xfb, + 0x60, 0xb3, 0xe6, 0x0f, 0x55, 0x94, 0x01, 0xaf, 0xaa, 0xc3, 0x6d, 0xea, + 0xb2, 0xfc, 0xb0, 0x06, 0x29, 0x0f, 0xd3, 0x95, 0xb9, 0xf1, 0x8b, 0xce, + 0xd3, 0x5d, 0x16, 0xbf, 0x5c, 0x24, 0xc5, 0x36, 0x98, 0x8c, 0x5b, 0x43, + 0xe7, 0xfe, 0x77, 0xda, 0xc5, 0xd8, 0xf6, 0x72, 0xba, 0xcf, 0x9c, 0x18, + 0x58, 0xb8, 0xe4, 0x1d, 0xf6, 0xfb, 0x3b, 0xb4, 0x1f, 0xea, 0xa3, 0xe3, + 0xd5, 0xbe, 0x3f, 0xd5, 0xf9, 0xc4, 0x00, 0x8e, 0x17, 0x22, 0x3d, 0x96, + 0xd8, 0xb6, 0xa5, 0xf6, 0xcd, 0x55, 0x48, 0x8b, 0x1b, 0x38, 0x9c, 0xd7, + 0x6d, 0x40, 0x2a, 0x5f, 0xcf, 0xcb, 0x67, 0xa4, 0x8c, 0xf4, 0x8f, 0x70, + 0x34, 0xeb, 0x70, 0xcd, 0xee, 0x1c, 0xbd, 0xae, 0xd1, 0xc1, 0xf8, 0x62, + 0x45, 0xb5, 0x5d, 0xe6, 0x0b, 0xd4, 0x3d, 0x23, 0xf0, 0x27, 0x44, 0x56, + 0x32, 0x4d, 0xb1, 0x6c, 0x5d, 0x33, 0x94, 0x77, 0xe3, 0xac, 0x54, 0x56, + 0x24, 0x05, 0x26, 0x4a, 0xf0, 0x59, 0xfb, 0x1f, 0xa4, 0x0f, 0xbe, 0x9e, + 0xbc, 0x76, 0x9d, 0x5a, 0xed, 0x15, 0x97, 0x4e, 0x05, 0x8a, 0x8b, 0xff, + 0xc7, 0x9b, 0x67, 0x32, 0x12, 0x41, 0x04, 0xcb, 0x24, 0xae, 0x9e, 0xcc, + 0xd6, 0xc6, 0x67, 0x53, 0xfa, 0x29, 0x37, 0x73, 0xc6, 0xdf, 0xf2, 0x56, + 0x72, 0x06, 0x03, 0xaa, 0x5d, 0x07, 0xac, 0x38, 0xb9, 0x2a, 0x61, 0x02, + 0x24, 0xcf, 0x54, 0x3f, 0x98, 0xb0, 0x5c, 0xba, 0xe3, 0x15, 0x27, 0x52, + 0x63, 0x43, 0x12, 0x62, 0x33, 0x02, 0xb8, 0x69, 0x52, 0x70, 0x6c, 0xc0, + 0x23, 0x37, 0x65, 0x4b, 0xc9, 0xea, 0x98, 0x06, 0xde, 0x3d, 0x59, 0x72, + 0x94, 0x48, 0x60, 0xeb, 0xe7, 0xaa, 0x68, 0x72, 0x22, 0x15, 0x39, 0xf0, + 0x47, 0x43, 0xeb, 0x37, 0xb1, 0x3b, 0x9e, 0x05, 0x12, 0xdb, 0x74, 0x18, + 0xfe, 0x11, 0xcb, 0xae, 0xe0, 0xed, 0x1c, 0xe3, 0x19, 0x71, 0x56, 0xa6, + 0x04, 0xe6, 0x20, 0x62, 0xfd, 0xb1, 0x57, 0x44, 0xca, 0x3f, 0xdf, 0x51, + 0x23, 0x76, 0x3b, 0x70, 0x27, 0x33, 0x62, 0x74, 0x94, 0xff, 0x70, 0xcc, + 0xd4, 0xbf, 0x67, 0x12, 0x17, 0x5f, 0x71, 0xf8, 0x8f, 0x09, 0xca, 0xb5, + 0x49, 0x38, 0xcf, 0x1f, 0x94, 0x9a, 0xe6, 0x76, 0x0e, 0xa6, 0x5a, 0x2c, + 0x36, 0x61, 0x41, 0x2d, 0x14, 0x2f, 0x35, 0xa2, 0xaa, 0x2d, 0xd5, 0x54, + 0x3c, 0x4e, 0xa0, 0x63, 0xa9, 0x9e, 0xe9, 0x65, 0x62, 0xcf, 0x5a, 0x1a, + 0xb9, 0x70, 0xf7, 0xf1, 0x8a, 0xc7, 0x19, 0x6e, 0x34, 0xa0, 0xbb, 0x1b, + 0x76, 0x9b, 0x60, 0x20, 0xfd, 0xff, 0xe1, 0x40, 0x5e, 0xd7, 0x49, 0xd3, + 0x3c, 0x0f, 0x52, 0xae, 0x37, 0x38, 0x1d, 0xd5, 0xd0, 0xe7, 0xd6, 0xfc, + 0x06, 0x3b, 0x50, 0x06, 0x9c, 0xb4, 0x37, 0x9a, 0x53, 0x09, 0x56, 0xa4, + 0xa8, 0x64, 0x70, 0xa7, 0xaf, 0xb9, 0xd9, 0x19, 0xbc, 0x5b, 0x04, 0x07, + 0x68, 0xc0, 0xa4, 0xc0, 0x3d, 0x32, 0x36, 0x94, 0x24, 0xd3, 0x36, 0x1f, + 0xfc, 0xd8, 0x26, 0x49, 0x94, 0xd2, 0x1e, 0x8b, 0x0c, 0x70, 0x6e, 0xd7, + 0xd2, 0x37, 0x8f, 0x13, 0xef, 0x41, 0xdb, 0x53, 0xb5, 0xba, 0xe5, 0xe3, + 0x0c, 0xcd, 0xa3, 0xfa, 0x74, 0x16, 0xd9, 0x42, 0x10, 0xa3, 0xe6, 0x26, + 0xd6, 0x74, 0xbc, 0x17, 0x9b, 0x2e, 0x4c, 0xe2, 0x13, 0x49, 0x0f, 0xc9, + 0xc2, 0x34, 0xae, 0x5b, 0x6b, 0x46, 0xbc, 0xc4, 0x62, 0xa0, 0x4a, 0x18, + 0x62, 0x69, 0x1c, 0xc3, 0x78, 0x36, 0xfa, 0xd9, 0x8d, 0xd0, 0xf9, 0x4f, + 0x56, 0x90, 0x4b, 0xca, 0xc4, 0xdd, 0x64, 0x2c, 0xd1, 0x3c, 0xa8, 0xbe, + 0x62, 0x8f, 0x2a, 0x11, 0x93, 0x71, 0x75, 0x70, 0x43, 0xd0, 0x5f, 0xfb, + 0x36, 0x2b, 0x35, 0x26, 0xda, 0xda, 0x25, 0x3c, 0x17, 0xf2, 0xb7, 0x36, + 0xd7, 0x8d, 0xd1, 0xbc, 0x2f, 0xe7, 0xf8, 0x55, 0x42, 0x2e, 0xe1, 0xc0, + 0x4a, 0xee, 0x3d, 0x5b, 0xc9, 0x69, 0x15, 0xc5, 0x42, 0x03, 0x2c, 0x46, + 0x02, 0x94, 0x91, 0xfb, 0x0f, 0x98, 0x8d, 0x32, 0xdf, 0x0b, 0x19, 0xda, + 0x9f, 0x96, 0x6e, 0x2d, 0xc4, 0xa1, 0x92, 0xc1, 0x73, 0x2f, 0x23, 0x9f, + 0x55, 0xc5, 0xb4, 0x8c, 0xef, 0xf3, 0xa2, 0x94, 0x8f, 0x6c, 0xd8, 0xb1, + 0x9d, 0x0d, 0x17, 0x93, 0x21, 0xd7, 0xae, 0xa8, 0x41, 0xd3, 0xf1, 0x9a, + 0xe3, 0x36, 0xca, 0x5f, 0xa4, 0xd9, 0xaf, 0x34, 0xbf, 0xe6, 0x9e, 0x4c, + 0xf0, 0xd1, 0xb0, 0x8c, 0x8e, 0x76, 0x3d, 0xb3, 0xf7, 0xd9, 0xfb, 0xbf, + 0x72, 0xae, 0xa8, 0x39, 0x00, 0xe5, 0x53, 0x17, 0x6c, 0x4e, 0x06, 0x22, + 0xc0, 0x10, 0xe7, 0x4d, 0xff, 0x75, 0x03, 0x01, 0x18, 0x46, 0xfd, 0xde, + 0x1e, 0x95, 0x46, 0xb8, 0x5b, 0x36, 0xbc, 0x1d, 0x95, 0x05, 0x8f, 0x5d, + 0x38, 0x41, 0x25, 0x2c, 0x9b, 0x34, 0x75, 0x9b, 0xf0, 0x8b, 0xaf, 0x0d, + 0x2e, 0xc2, 0x1a, 0x03, 0x61, 0xbe, 0xe8, 0x49, 0xbc, 0x9b, 0x45, 0xfb, + 0x35, 0x2b, 0x6c, 0xa1, 0x96, 0xa0, 0x08, 0x0e, 0xca, 0x01, 0xc0, 0x97, + 0xfa, 0xdf, 0x11, 0x1a, 0x0d, 0xf9, 0xc2, 0x5a, 0xe1, 0x4c, 0xb5, 0x37, + 0xff, 0x91, 0xb6, 0x96, 0xbf, 0x62, 0x04, 0x59, 0x69, 0x01, 0x68, 0x66, + 0x52, 0x66, 0x4a, 0x49, 0xe9, 0xe6, 0xe4, 0x44, 0x92, 0x5e, 0xaf, 0xf5, + 0x24, 0xdb, 0x6f, 0x21, 0xf9, 0x21, 0x58, 0x5f, 0xc4, 0xf0, 0x30, 0x90, + 0x68, 0xff, 0x58, 0x5c, 0xbd, 0x6f, 0x58, 0x77, 0xe0, 0x03, 0x68, 0x2a, + 0x1a, 0xa4, 0xd6, 0x9d, 0xd0, 0x38, 0x5a, 0xbd, 0x52, 0xa8, 0xc5, 0xf0, + 0xbc, 0xf2, 0x04, 0x49, 0x0e, 0x1b, 0x1b, 0x93, 0xc0, 0x65, 0xca, 0x05, + 0x42, 0x11, 0x03, 0xd6, 0xd5, 0x2c, 0x4c, 0xcd, 0xed, 0xb4, 0x54, 0xa4, + 0x3d, 0x46, 0x64, 0x4c, 0xc4, 0x8f, 0x0a, 0x95, 0x6a, 0x4f, 0xfb, 0x2e, + 0x1d, 0x5a, 0x8a, 0xcb, 0x31, 0x94, 0x21, 0x54, 0x51, 0xf5, 0x4e, 0x3e, + 0x32, 0x00, 0x12, 0x8e, 0x4c, 0x8c, 0x17, 0x90, 0xea, 0x8d, 0xfe, 0xc3, + 0xfe, 0x69, 0x10, 0xd9, 0x1c, 0x60, 0x91, 0xb6, 0xbb, 0x11, 0xb7, 0x77, + 0x1c, 0x69, 0xec, 0xb5, 0x28, 0x1e, 0x4b, 0xc8, 0xac, 0xe2, 0xe7, 0xe4, + 0xca, 0x1c, 0x6a, 0x16, 0xb8, 0x0a, 0x1c, 0xcb, 0xbd, 0x0e, 0x61, 0xf6, + 0x30, 0xa0, 0xb0, 0x11, 0x57, 0xd0, 0xa0, 0xe5, 0x63, 0xb4, 0x5e, 0x65, + 0x54, 0xbd, 0x2b, 0xcf, 0x92, 0xb3, 0xe2, 0xad, 0xba, 0x6b, 0xd8, 0x8b, + 0xd4, 0xc9, 0x49, 0x6b, 0xe9, 0x6f, 0x30, 0x9a, 0x8d, 0x1a, 0xd2, 0x73, + 0xed, 0x01, 0x20, 0x76, 0x59, 0x3b, 0x63, 0x15, 0xf7, 0x4a, 0x93, 0xf5, + 0xe8, 0xaa, 0x77, 0xf7, 0xee, 0x16, 0x26, 0x6d, 0x6d, 0x1e, 0xb3, 0x04, + 0xd1, 0x36, 0x6d, 0xdb, 0xe1, 0xee, 0xdf, 0x69, 0x0e, 0x28, 0x3b, 0x5a, + 0x37, 0x51, 0x61, 0x10, 0x58, 0xd0, 0x58, 0x75, 0x63, 0x5b, 0x76, 0x3e, + 0x55, 0x0a, 0x07, 0x3e, 0xfe, 0xb9, 0x6e, 0x4c, 0xfc, 0x1b, 0x8a, 0xa5, + 0x03, 0x1a, 0xb9, 0x04, 0x22, 0x60, 0x33, 0x66, 0xda, 0xb7, 0x1c, 0x3a, + 0xb6, 0x92, 0x45, 0x01, 0xc2, 0x73, 0x49, 0x6a, 0x9a, 0x54, 0x10, 0xe2, + 0x36, 0x45, 0xbd, 0x1d, 0x33, 0x2a, 0xd2, 0xc9, 0x70, 0x63, 0x39, 0xcf, + 0xf7, 0x76, 0x70, 0x37, 0xde, 0x23, 0x4c, 0xd2, 0xa1, 0x37, 0x2c, 0x52, + 0xae, 0xa3, 0xfb, 0x45, 0xd0, 0xb9, 0x46, 0x3e, 0x2a, 0xe8, 0xe9, 0x64, + 0xe1, 0x16, 0x30, 0x08, 0x36, 0xcd, 0x9e, 0x15, 0x44, 0xdd, 0x27, 0xa9, + 0x1c, 0x29, 0xf1, 0xa7, 0x20, 0x21, 0x59, 0x61, 0x4c, 0xbe, 0x5e, 0x20, + 0x36, 0xca, 0xb8, 0x6d, 0xb5, 0x0c, 0x29, 0x41, 0xa1, 0xd3, 0x8a, 0x2b, + 0x34, 0xd2, 0x5b, 0x92, 0x12, 0x1f, 0x36, 0x9f, 0x5d, 0x02, 0x2a, 0xca, + 0xac, 0x5b, 0x29, 0x8b, 0x51, 0x3a, 0x65, 0xf5, 0xdf, 0x60, 0x6c, 0x0c, + 0xa7, 0x95, 0x3d, 0x52, 0x13, 0xb4, 0xbd, 0x8c, 0xf1, 0xac, 0xba, 0x3c, + 0x24, 0x6c, 0xc0, 0xdb, 0xa8, 0x5b, 0xd4, 0xdb, 0xf5, 0xcd, 0xaf, 0xdf, + 0x2f, 0xe2, 0x71, 0xcc, 0x00, 0x3a, 0x87, 0xdc, 0x23, 0xdf, 0xa7, 0xb0, + 0xb6, 0xcb, 0xff, 0x1c, 0xe7, 0xfe, 0xa8, 0xa8, 0xea, 0xad, 0x37, 0x58, + 0xfd, 0x58, 0x01, 0xa5, 0xe4, 0x5d, 0xdf, 0x4a, 0x10, 0x0b, 0xc3, 0x5e, + 0xd1, 0x0d, 0x4c, 0x21, 0x0e, 0x51, 0x95, 0x99, 0x58, 0xdf, 0x6d, 0xa8, + 0x8e, 0xf7, 0x51, 0xa6, 0x53, 0x44, 0x6b, 0xb3, 0x00, 0x64, 0xe1, 0x6f, + 0x3d, 0x19, 0x40, 0x30, 0x46, 0x95, 0x9b, 0x39, 0xa5, 0x0d, 0x77, 0xaa, + 0xb1, 0x57, 0x57, 0x08, 0xe0, 0xab, 0xd1, 0xd5, 0x25, 0x59, 0x11, 0x2f, + 0x62, 0xbf, 0x50, 0x95, 0x02, 0x18, 0xdb, 0x2d, 0xbc, 0xdb, 0xfa, 0x3d, + 0x45, 0xab, 0xb5, 0x2e, 0x8e, 0x9b, 0x49, 0xe5, 0x50, 0xbd, 0x1f, 0x1c, + 0x64, 0xd8, 0x9d, 0x0c, 0x0c, 0xe8, 0xf3, 0x54, 0x49, 0x95, 0x3d, 0x71, + 0xa1, 0x16, 0x98, 0x08, 0x16, 0x37, 0x6a, 0x95, 0xa3, 0xaa, 0xb6, 0xf7, + 0x0e, 0x99, 0x2a, 0x0b, 0x68, 0x49, 0xd1, 0xa4, 0x33, 0x3e, 0x57, 0xfc, + 0xc3, 0x5a, 0xa9, 0x1e, 0xbf, 0xf1, 0x19, 0x2d, 0xee, 0xfa, 0x01, 0xa8, + 0x64, 0x0d, 0x74, 0x54, 0xed, 0x4d, 0xab, 0xad, 0x23, 0x25, 0xde, 0xef, + 0xb4, 0x54, 0xfe, 0x3f, 0xba, 0xe0, 0x0e, 0x76, 0x1b, 0x1a, 0xa9, 0xe3, + 0x53, 0xbd, 0xde, 0x65, 0x6b, 0x08, 0x6d, 0x71, 0x45, 0xb4, 0xf8, 0x9a, + 0x06, 0x3d, 0xae, 0x87, 0x25, 0x51, 0x9d, 0x46, 0x33, 0xf3, 0x77, 0x6d, + 0xb6, 0x5d, 0xbe, 0x08, 0xfc, 0xf5, 0x31, 0xa1, 0xd5, 0x22, 0x19, 0xcd, + 0x66, 0x82, 0x19, 0xf5, 0xf5, 0x29, 0x28, 0x83, 0xa5, 0xa3, 0x30, 0x50, + 0xa1, 0xfb, 0xf6, 0x36, 0x31, 0xbf, 0xb5, 0xc4, 0xe7, 0x99, 0xd5, 0x4f, + 0xf5, 0xb0, 0xf5, 0x9a, 0x12, 0x4e, 0x1b, 0xdb, 0x4d, 0x21, 0x6d, 0xda, + 0xeb, 0x6a, 0x11, 0x55, 0xa2, 0xe2, 0x6a, 0xe9, 0xe8, 0x01, 0xa1, 0x97, + 0x68, 0xc2, 0x30, 0xd2, 0xfa, 0x60, 0xec, 0x4d, 0x54, 0x5b, 0x9e, 0x2d, + 0x97, 0xca, 0x1b, 0xc2, 0xb2, 0x14, 0x3f, 0xaf, 0x23, 0x54, 0xe8, 0x0c, + 0x3c, 0xed, 0x50, 0x32, 0xff, 0x3a, 0x8c, 0xe6, 0xdc, 0x17, 0xad, 0x65, + 0x05, 0x35, 0x28, 0xc9, 0x77, 0x21, 0xb1, 0x9a, 0xec, 0xf1, 0xd6, 0x53, + 0xb9, 0xb3, 0xe0, 0x41, 0x11, 0x85, 0x2e, 0x1a, 0xb5, 0xad, 0xab, 0x9b, + 0xae, 0x69, 0xa0, 0xb1, 0xa0, 0x07, 0x72, 0x8f, 0x4a, 0xd9, 0x5e, 0x1f, + 0x29, 0x9e, 0x4d, 0x0b, 0x9a, 0x82, 0xfe, 0x26, 0xc5, 0x17, 0x5b, 0x51, + 0x46, 0xf2, 0xf7, 0x27, 0xba, 0x06, 0x91, 0x0e, 0xc2, 0x07, 0xb3, 0x1b, + 0x54, 0xad, 0xb5, 0xf5, 0x02, 0xc1, 0x39, 0x6a, 0x2a, 0xd7, 0x46, 0xbf, + 0x3d, 0x39, 0x4e, 0x8e, 0xb1, 0x58, 0xf4, 0x90, 0xa7, 0x08, 0x0e, 0x99, + 0x64, 0x33, 0x3e, 0x1e, 0x09, 0xb7, 0x88, 0xa0, 0x29, 0xb2, 0x0b, 0x5c, + 0x15, 0xd4, 0x36, 0x55, 0x42, 0x48, 0xe7, 0x47, 0xf9, 0xb5, 0x05, 0xcd, + 0x40, 0xde, 0x92, 0x27, 0x11, 0x3b, 0xad, 0x3e, 0x9b, 0x95, 0x38, 0xad, + 0x11, 0xd5, 0x9d, 0x1d, 0x38, 0x60, 0xde, 0x31, 0xe3, 0x40, 0xb2, 0xf2, + 0x8e, 0xb4, 0x03, 0xaa, 0x51, 0x15, 0xe4, 0x36, 0x4d, 0x43, 0x05, 0xbc, + 0x36, 0x82, 0xdf, 0xfc, 0xfd, 0x23, 0x4d, 0xad, 0x9f, 0xf4, 0xce, 0xfb, + 0xaf, 0x46, 0xb3, 0x59, 0x98, 0x91, 0x85, 0x4a, 0xa7, 0x67, 0x70, 0xbd, + 0xca, 0x12, 0x9b, 0x6b, 0x00, 0xe5, 0x82, 0x3c, 0x37, 0x99, 0x8d, 0x6b, + 0x32, 0xaf, 0x08, 0x05, 0x36, 0xd6, 0xd7, 0xfb, 0x65, 0xce, 0x4e, 0x9f, + 0xd5, 0xd1, 0x3a, 0x42, 0xb0, 0x31, 0x62, 0xd0, 0xe2, 0xe5, 0x37, 0xc1, + 0x6d, 0x8a, 0x24, 0xa4, 0x19, 0xc2, 0x59, 0x3c, 0x44, 0xef, 0x96, 0xf6, + 0x35, 0x00, 0xe7, 0xe6, 0x2e, 0x82, 0xa5, 0x4a, 0x2f, 0xa2, 0xfe, 0x1f, + 0x53, 0x52, 0x31, 0x97, 0x47, 0x37, 0x15, 0x26, 0xa7, 0x8d, 0xd3, 0x21, + 0x6a, 0x98, 0x6d, 0xf1, 0xe6, 0x29, 0xf8, 0x9d, 0xaf, 0x5f, 0x3e, 0x3a, + 0xbc, 0x65, 0xb2, 0xd8, 0x41, 0xbc, 0xd6, 0x39, 0x3c, 0xc7, 0x2f, 0x2e, + 0xa3, 0x08, 0x9a, 0x21, 0x05, 0xe0, 0x4c, 0x06, 0x4d, 0x82, 0x68, 0x5d, + 0x4a, 0x9e, 0xca, 0xee, 0x3d, 0x28, 0x45, 0x0e, 0xff, 0xdd, 0xe6, 0x46, + 0xbc, 0xf8, 0x19, 0x5b, 0xda, 0xf4, 0x14, 0xd1, 0x4f, 0x02, 0x6e, 0xf6, + 0x01, 0x2d, 0xd6, 0xb6, 0x8b, 0xf5, 0x9c, 0x4e, 0xee, 0xe7, 0xc8, 0x10, + 0x05, 0xb6, 0x6d, 0x8d, 0x49, 0xe2, 0x04, 0xec, 0x4d, 0x61, 0x67, 0xc2, + 0x19, 0x27, 0xab, 0xe1, 0x0d, 0x29, 0xab, 0xf2, 0xa0, 0xf9, 0x69, 0x0d, + 0x81, 0x29, 0x4d, 0x40, 0x6d, 0xd7, 0xda, 0xb7, 0x9e, 0x0b, 0x90, 0x9c, + 0x9b, 0xeb, 0x59, 0x2c, 0xc9, 0xa4, 0x85, 0x95, 0xe2, 0xda, 0x2d, 0xe4, + 0x60, 0x9a, 0x64, 0x21, 0xbf, 0x1d, 0x57, 0x4d, 0x3e, 0xa0, 0x35, 0x0f, + 0xce, 0xd7, 0xe1, 0x44, 0x63, 0x9e, 0xe8, 0x8e, 0xbd, 0xc8, 0xc1, 0x65, + 0xe1, 0xd2, 0x09, 0x45, 0xd3, 0xbd, 0x13, 0xb2, 0x1f, 0x46, 0x32, 0xa6, + 0xcd, 0xa3, 0x44, 0x4c, 0x52, 0xa7, 0xe7, 0x54, 0xea, 0xe6, 0xa0, 0xce, + 0x02, 0x8b, 0x69, 0xdb, 0xde, 0xef, 0x5f, 0xcb, 0x6f, 0x6e, 0x0f, 0xf5, + 0x68, 0x42, 0xf4, 0x37, 0x08, 0x1f, 0x87, 0x55, 0xb4, 0xbc, 0x8a, 0x84, + 0x84, 0x10, 0xc6, 0x36, 0x3e, 0x8a, 0x6b, 0x4e, 0xd5, 0xc8, 0x64, 0xcb, + 0xb5, 0xc0, 0xfe, 0x99, 0x66, 0xaa, 0xb1, 0x50, 0xa7, 0x70, 0xd9, 0xa6, + 0x17, 0x2d, 0xd4, 0xad, 0xdf, 0xf2, 0x2f, 0xac, 0xae, 0xae, 0x12, 0xcf, + 0x5b, 0x09, 0xf2, 0x2d, 0xb4, 0x21, 0xc9, 0xd1, 0x58, 0xdb, 0x4e, 0x9b, + 0xe0, 0x32, 0x08, 0xe4, 0x4a, 0xe6, 0x9c, 0x61, 0x25, 0x90, 0x08, 0xf2, + 0xb1, 0xc1, 0x3c, 0x25, 0x0b, 0x5a, 0x03, 0x40, 0xdb, 0x06, 0x5f, 0xd2, + 0x60, 0x8e, 0x0a, 0x5b, 0xc8, 0xa2, 0xcd, 0xac, 0xb3, 0x54, 0x0b, 0xb6, + 0x05, 0x45, 0xd7, 0xa8, 0x8a, 0xfa, 0x8a, 0xba, 0x09, 0x53, 0x81, 0xd7, + 0xf5, 0x40, 0x61, 0x46, 0xf2, 0x22, 0xe4, 0x21, 0xb4, 0x26, 0x41, 0x10, + 0x25, 0x4d, 0x93, 0xc2, 0xa2, 0xae, 0xc3, 0xaa, 0xbe, 0x71, 0xa6, 0xaa, + 0xf7, 0xb1, 0xbf, 0x02, 0x22, 0xe9, 0xd7, 0xfb, 0xaa, 0x1d, 0x5d, 0xf5, + 0xe7, 0x5b, 0x63, 0xf2, 0xe6, 0x5c, 0xd6, 0x24, 0x6d, 0xb5, 0xca, 0xa3, + 0xe7, 0x57, 0x1a, 0xa5, 0xf7, 0x95, 0xc5, 0x92, 0x51, 0x65, 0x68, 0xc5, + 0xe6, 0x27, 0xa9, 0x94, 0x8a, 0xb6, 0xec, 0x0d, 0x9c, 0x51, 0xdf, 0x22, + 0xca, 0xdf, 0x5a, 0xf5, 0xe4, 0xad, 0xf4, 0xfc, 0x1f, 0x68, 0x9f, 0xdb, + 0x40, 0x4e, 0x6a, 0x1e, 0x5a, 0xd8, 0x6c, 0xd6, 0xef, 0xad, 0x64, 0xe7, + 0xcb, 0xfc, 0x44, 0xae, 0xa5, 0x62, 0x65, 0xad, 0x2e, 0x6a, 0x46, 0xcf, + 0x0d, 0xd0, 0x46, 0x5e, 0x87, 0x37, 0xb6, 0xab, 0x70, 0x52, 0xee, 0x5a, + 0xa7, 0x13, 0xa3, 0xc3, 0x4b, 0x62, 0xe7, 0x31, 0x10, 0xed, 0x39, 0x1c, + 0x4a, 0xe3, 0xc1, 0x57, 0xcb, 0x45, 0xe4, 0x89, 0xee, 0x0e, 0x24, 0xc1, + 0xa6, 0xac, 0xd4, 0x0e, 0x9b, 0xe0, 0x26, 0x28, 0x08, 0x2b, 0xe1, 0xc9, + 0x42, 0x37, 0xa3, 0x46, 0xcc, 0x5d, 0x89, 0x10, 0x1f, 0x23, 0xcb, 0x1c, + 0x67, 0xe2, 0x6d, 0xaa, 0x66, 0xa5, 0xf5, 0xea, 0x94, 0x2b, 0x8c, 0xf6, + 0xf4, 0xd3, 0xfb, 0x9c, 0x96, 0x0a, 0x87, 0xaf, 0x5c, 0x19, 0xb4, 0x3b, + 0x26, 0xb2, 0x48, 0x55, 0x97, 0xfd, 0x3a, 0xec, 0x06, 0xe4, 0x58, 0x99, + 0x9a, 0x26, 0x4f, 0xe0, 0x9c, 0x67, 0x09, 0x05, 0x5b, 0x72, 0x8e, 0xd6, + 0xe4, 0x4e, 0xe2, 0x63, 0xb0, 0x9c, 0xf6, 0x92, 0xd3, 0x05, 0x3f, 0xb0, + 0x04, 0x5f, 0x02, 0x97, 0xf4, 0x42, 0x1d, 0x3b, 0x5c, 0x44, 0x00, 0x95, + 0x8b, 0xf5, 0x06, 0x40, 0xbd, 0xb8, 0xf7, 0x4b, 0x4a, 0xfa, 0xf0, 0x04, + 0x04, 0xd0, 0xa5, 0xb9, 0x3a, 0xa0, 0x2d, 0x0c, 0x1b, 0xec, 0x5a, 0x14, + 0xc8, 0x1d, 0x93, 0x86, 0xfd, 0x16, 0x68, 0xf8, 0x16, 0x9b, 0xb4, 0x88, + 0x99, 0x63, 0x0e, 0xd5, 0x20, 0x07, 0x43, 0x28, 0x26, 0xba, 0xf9, 0x97, + 0xed, 0x6b, 0x40, 0xb8, 0x07, 0x73, 0x59, 0xd5, 0x55, 0xa8, 0x64, 0x14, + 0x1c, 0xc5, 0xc0, 0x1f, 0x8d, 0x09, 0xae, 0x9c, 0x66, 0xa1, 0x94, 0xca, + 0x14, 0x46, 0xed, 0x46, 0x46, 0x25, 0x63, 0x5b, 0x2b, 0x95, 0x85, 0x05, + 0xc2, 0xb7, 0xeb, 0x06, 0x30, 0x5a, 0xf6, 0x22, 0x4e, 0x47, 0x1e, 0x0e, + 0x0c, 0xad, 0xd5, 0x11, 0xa8, 0x6a, 0x89, 0xd5, 0x49, 0xd4, 0xfa, 0x43, + 0xb0, 0x32, 0xb0, 0xb9, 0xb3, 0xda, 0x3f, 0x4f, 0xac, 0x4c, 0xc1, 0xa7, + 0x9f, 0xc2, 0xc2, 0x04, 0x70, 0xa2, 0x08, 0x01, 0xeb, 0x10, 0xa4, 0xa5, + 0x4c, 0xcd, 0xb3, 0x81, 0x4e, 0xbe, 0x6c, 0x51, 0x44, 0xf8, 0x82, 0xbd, + 0x42, 0x34, 0xfb, 0xdb, 0xb4, 0x32, 0xd2, 0x93, 0x63, 0x5e, 0xf6, 0x07, + 0x6e, 0x2c, 0xc2, 0xcf, 0xf4, 0x5d, 0x84, 0xe9, 0x5e, 0x5c, 0xa8, 0x39, + 0x28, 0x4a, 0xed, 0x15, 0x1b, 0xea, 0xe6, 0xde, 0x85, 0x92, 0x86, 0xe7, + 0x83, 0x4b, 0x87, 0xf7, 0x23, 0x60, 0xe2, 0x22, 0xd3, 0x32, 0x16, 0x4e, + 0x2f, 0xde, 0x01, 0x8b, 0x48, 0xea, 0xcd, 0x8a, 0x8b, 0xbc, 0xc6, 0x64, + 0xb2, 0x67, 0x47, 0xf5, 0x98, 0xf8, 0xca, 0xf1, 0x83, 0x66, 0xd7, 0x9a, + 0xef, 0xca, 0x20, 0xc2, 0xec, 0x8c, 0x38, 0xb1, 0x37, 0x13, 0x93, 0x92, + 0xba, 0xa1, 0xee, 0x6a, 0x57, 0x43, 0xaa, 0xdc, 0xdf, 0xa4, 0x3f, 0xc6, + 0xb6, 0xd6, 0x68, 0x54, 0xab, 0x36, 0xe9, 0x0f, 0x6f, 0xd5, 0xa1, 0x1b, + 0xa1, 0x02, 0xc9, 0x41, 0xef, 0x4f, 0x86, 0xcc, 0x1a, 0xfa, 0xd2, 0xdd, + 0x87, 0x04, 0xe0, 0x27, 0x38, 0xcf, 0x91, 0x95, 0xb4, 0x02, 0x10, 0x1d, + 0xc3, 0xcc, 0x6f, 0xaf, 0xbc, 0x94, 0x64, 0x47, 0xbc, 0x37, 0xde, 0xe3, + 0x2e, 0x89, 0x03, 0xb6, 0xd3, 0x28, 0x4a, 0x5e, 0x6d, 0x1e, 0xc5, 0x1a, + 0xa5, 0x0c, 0x92, 0xf7, 0xe2, 0x19, 0xe7, 0x39, 0xf0, 0xf2, 0x49, 0x8b, + 0xe6, 0x99, 0xd8, 0x4b, 0x0d, 0x6e, 0x3f, 0x57, 0x89, 0x9e, 0x0d, 0x34, + 0x4b, 0x52, 0xcd, 0x18, 0x57, 0xc7, 0x8e, 0x48, 0x03, 0x65, 0xd4, 0xdd, + 0xdf, 0x04, 0xf5, 0x39, 0x5e, 0x97, 0xbc, 0xc0, 0xc5, 0x91, 0xe7, 0x9d, + 0xbe, 0x28, 0x4c, 0xe7, 0xf4, 0xa0, 0x34, 0xee, 0xba, 0xa7, 0x8d, 0x52, + 0xc4, 0x07, 0x14, 0xd2, 0x93, 0xb0, 0x1d, 0x61, 0x53, 0x23, 0xc3, 0xe1, + 0xd2, 0xbf, 0xe1, 0xd6, 0x1f, 0x27, 0xcc, 0x8c, 0xe7, 0x0b, 0x09, 0x4f, + 0xe6, 0xa2, 0x41, 0xf4, 0x31, 0xbe, 0x95, 0x17, 0xfb, 0x50, 0xa4, 0xa4, + 0x51, 0x3c, 0x6f, 0xf8, 0x6a, 0xba, 0xac, 0xe4, 0x1e, 0x38, 0x78, 0x18, + 0x58, 0x31, 0x69, 0xc9, 0x52, 0xb0, 0xfc, 0x71, 0x54, 0xad, 0xe2, 0x8e, + 0xa2, 0xf2, 0x8e, 0x58, 0x11, 0x1d, 0xcc, 0x30, 0x74, 0x55, 0x41, 0x02, + 0x9b, 0x2a, 0x2f, 0x17, 0x97, 0xe4, 0x1a, 0xd0, 0xd5, 0x8f, 0x60, 0x10, + 0xdb, 0xc2, 0x69, 0x94, 0x0d, 0xaf, 0x44, 0xd0, 0x95, 0x3d, 0x50, 0xf4, + 0x27, 0x5e, 0xdc, 0x56, 0x5f, 0xa7, 0x4c, 0x41, 0xe5, 0x9e, 0xc8, 0x31, + 0xb0, 0x8e, 0x3f, 0xde, 0xdc, 0x42, 0x24, 0x93, 0x98, 0xce, 0x69, 0x90, + 0x98, 0x73, 0x06, 0xb9, 0x8e, 0xa4, 0x8d, 0x97, 0xb1, 0x41, 0x33, 0x64, + 0x5a, 0xae, 0xe8, 0x2f, 0x5f, 0x99, 0x64, 0x3e, 0xea, 0xd4, 0xbe, 0xa2, + 0x52, 0x2d, 0xc7, 0x56, 0x46, 0xfb, 0x33, 0xd8, 0xde, 0xe6, 0x74, 0xf6, + 0x2e, 0x2a, 0x26, 0xa1, 0x07, 0xcd, 0x3c, 0xca, 0x39, 0x74, 0x61, 0x4a, + 0x53, 0xf7, 0x8c, 0xd7, 0x3c, 0x4f, 0x4f, 0xd9, 0x14, 0x74, 0x56, 0xa8, + 0x3b, 0x3b, 0xe4, 0xe5, 0x70, 0x2e, 0xda, 0xde, 0xcd, 0x65, 0x4f, 0x2e, + 0xb6, 0x76, 0x17, 0x59, 0x6a, 0xaf, 0x0a, 0x24, 0x8c, 0x99, 0x0b, 0x2a, + 0xac, 0x46, 0x74, 0x2c, 0x3b, 0x40, 0x20, 0xad, 0x30, 0xab, 0x63, 0x34, + 0x8f, 0x30, 0x22, 0x50, 0x5c, 0xf8, 0x73, 0x21, 0x3e, 0xeb, 0x16, 0x44, + 0x30, 0xb9, 0x59, 0x0f, 0xf0, 0xe5, 0xb6, 0x6a, 0xde, 0x32, 0x03, 0x28, + 0x3c, 0xc8, 0xc2, 0x8d, 0x6b, 0x72, 0x2f, 0x3e, 0x2b, 0x99, 0xc1, 0xa6, + 0xdf, 0x5a, 0x91, 0x2d, 0x40, 0x39, 0xb2, 0x24, 0x27, 0x25, 0x26, 0x51, + 0xbb, 0xb5, 0x6a, 0x47, 0x38, 0x94, 0x2c, 0x3e, 0xa0, 0x96, 0x19, 0xf7, + 0x99, 0x0c, 0x34, 0x41, 0xb9, 0x0d, 0xad, 0x37, 0xa6, 0x0c, 0x38, 0x9c, + 0xee, 0x03, 0x68, 0x62, 0x76, 0x64, 0x18, 0x63, 0x62, 0x10, 0xd6, 0x2a, + 0xca, 0xdb, 0x73, 0x9b, 0x93, 0x35, 0x29, 0xb0, 0xec, 0x6c, 0xa8, 0x1f, + 0xa6, 0xac, 0xf8, 0xd8, 0xfa, 0x98, 0xc3, 0x02, 0xf0, 0xf5, 0x66, 0x2c, + 0xfc, 0x75, 0xc7, 0xb0, 0x76, 0xfe, 0x0f, 0x92, 0x9b, 0xce, 0xc5, 0xe8, + 0x9a, 0x5e, 0x8f, 0x16, 0x26, 0x8c, 0x97, 0x20, 0x97, 0x36, 0xca, 0x56, + 0xed, 0xf2, 0x05, 0x53, 0xf7, 0x9f, 0x23, 0xbb, 0x1e, 0xdc, 0x5a, 0x94, + 0x0b, 0x1d, 0x0e, 0x55, 0xc7, 0x34, 0xff, 0xd9, 0xa3, 0x37, 0x69, 0x63, + 0x9f, 0x00, 0x0f, 0xa1, 0x5c, 0x1f, 0x50, 0x56, 0x25, 0xf0, 0xb8, 0x0e, + 0x92, 0x70, 0xcd, 0xa0, 0xca, 0x2a, 0xce, 0xa5, 0x21, 0xe7, 0x5b, 0x10, + 0x13, 0xd5, 0x9b, 0x9f, 0x60, 0x1b, 0x3f, 0x21, 0xa9, 0x27, 0xd9, 0xeb, + 0xdc, 0xe8, 0x05, 0x8e, 0x09, 0x27, 0x4b, 0x8b, 0xb1, 0x3b, 0x07, 0xb1, + 0xe9, 0x55, 0xc4, 0xab, 0x5d, 0x74, 0x11, 0xcf, 0x98, 0x5d, 0x47, 0x58, + 0x9d, 0x08, 0xec, 0x0b, 0x31, 0x69, 0x98, 0xad, 0xd0, 0x93, 0x09, 0xc7, + 0xcc, 0xe3, 0x64, 0x67, 0xef, 0xce, 0x98, 0xf3, 0xc2, 0x69, 0xd4, 0x47, + 0x4d, 0xf7, 0x1a, 0x10, 0xa9, 0x18, 0x35, 0x94, 0xc8, 0xe1, 0xd2, 0xf5, + 0xb5, 0xb4, 0x0b, 0xd7, 0x28, 0xa8, 0x97, 0x9b, 0xbf, 0x90, 0xe5, 0xc6, + 0xde, 0xf7, 0x4f, 0x33, 0xaf, 0x36, 0xe2, 0xa8, 0x65, 0x56, 0xdd, 0xe8, + 0x79, 0xae, 0x68, 0xc1, 0xf3, 0x5b, 0x26, 0x59, 0x53, 0x00, 0x43, 0x4c, + 0x3e, 0xf9, 0x24, 0xc4, 0x8d, 0x73, 0x00, 0x6c, 0xb2, 0x97, 0x56, 0x90, + 0x42, 0xde, 0xba, 0xd6, 0x3a, 0x6d, 0x39, 0x9d, 0xbe, 0x1c, 0xca, 0x24, + 0xbb, 0xba, 0x06, 0xf0, 0x59, 0x74, 0x32, 0x99, 0x1b, 0x02, 0xad, 0xc1, + 0x8b, 0xd4, 0x0b, 0xd8, 0xb7, 0xe7, 0xbd, 0xbd, 0x68, 0x56, 0xc1, 0x1e, + 0xda, 0xa4, 0xfe, 0x6b, 0x94, 0xf3, 0xda, 0x9a, 0x33, 0x01, 0x97, 0xb6, + 0x39, 0xc4, 0xe7, 0x57, 0xee, 0xcf, 0x0e, 0xce, 0x40, 0x7a, 0xd4, 0x4d, + 0x30, 0x6a, 0x57, 0x8f, 0x97, 0x92, 0x59, 0xeb, 0xf2, 0x18, 0x8c, 0x77, + 0xd9, 0x8f, 0x72, 0xff, 0xd5, 0xb2, 0x1f, 0x2e, 0xba, 0xb6, 0x46, 0x1a, + 0x33, 0xe0, 0x74, 0x2a, 0xd7, 0xdb, 0xc7, 0x07, 0x37, 0x2f, 0x55, 0xe2, + 0x70, 0x43, 0xc2, 0xbc, 0x33, 0x03, 0xc9, 0xd4, 0x4e, 0x6e, 0x3e, 0xc9, + 0x67, 0x55, 0xf8, 0x6d, 0x63, 0x9f, 0x6b, 0x3f, 0x5b, 0xc7, 0xe9, 0xb8, + 0x31, 0x04, 0x0b, 0x71, 0x15, 0xcd, 0x34, 0xe4, 0xaf, 0x74, 0x73, 0xea, + 0xbf, 0x20, 0x00, 0x75, 0xd7, 0xa7, 0xf7, 0x9c, 0xf5, 0xa1, 0x28, 0xc7, + 0xfe, 0x6b, 0xa2, 0x36, 0xdc, 0xd4, 0xf0, 0xd7, 0x42, 0x4e, 0xe4, 0x3f, + 0x00, 0x09, 0x3c, 0x5e, 0x1f, 0xc8, 0xfd, 0xb9, 0xd8, 0x90, 0xdb, 0xf4, + 0x41, 0x0b, 0xda, 0x68, 0xe1, 0xe4, 0xb9, 0xfb, 0x36, 0x37, 0xa9, 0x5f, + 0xc9, 0xb6, 0xb8, 0xa4, 0xda, 0x41, 0xaa, 0xab, 0xa8, 0xc8, 0xd3, 0xc6, + 0x6a, 0xbe, 0x03, 0x77, 0xcc, 0x1a, 0x8d, 0x0d, 0xe8, 0xcc, 0x58, 0x46, + 0x71, 0x33, 0x19, 0x62, 0xe5, 0xc4, 0xe3, 0x4a, 0x1d, 0xf7, 0x96, 0xd4, + 0x08, 0xe5, 0xa8, 0x18, 0x40, 0x2d, 0xc5, 0xd7, 0xa7, 0x31, 0xa2, 0x5f, + 0x60, 0xde, 0x21, 0xe5, 0xaa, 0x65, 0x93, 0x0d, 0xdb, 0x55, 0x54, 0x88, + 0xbd, 0x53, 0x8e, 0xe0, 0xa6, 0x23, 0xcd, 0x1d, 0xb7, 0xbd, 0x2a, 0x8c, + 0x0e, 0x67, 0x65, 0xab, 0xda, 0xe9, 0x3b, 0x12, 0xf6, 0x97, 0x4b, 0xe8, + 0x16, 0xf7, 0x09, 0xb6, 0x45, 0x97, 0x16, 0xec, 0xd9, 0xdc, 0x8d, 0x01, + 0xba, 0xb0, 0xb6, 0xdd, 0x59, 0x60, 0xbf, 0x92, 0x92, 0xc3, 0x21, 0x41, + 0x46, 0xcb, 0x5e, 0x6e, 0x99, 0x10, 0x41, 0x45, 0x9a, 0xb9, 0xe0, 0x6d, + 0x22, 0x68, 0xd3, 0x5a, 0xaa, 0x6e, 0xb4, 0xc6, 0x42, 0xa2, 0xad, 0xf1, + 0xf7, 0x0b, 0x3d, 0x29, 0x38, 0xa2, 0x11, 0xf8, 0x57, 0x25, 0xb8, 0x8f, + 0xbc, 0x65, 0xac, 0x0d, 0xf0, 0xb7, 0x5c, 0x95, 0xfb, 0x5d, 0xdb, 0x54, + 0x3d, 0x3e, 0xd6, 0x4f, 0x2a, 0xfe, 0x43, 0xfc, 0x1c, 0xca, 0xb9, 0xb3, + 0x95, 0x06, 0x90, 0xd9, 0x5d, 0x43, 0xc4, 0xe9, 0xbb, 0x17, 0xd6, 0xaf, + 0xf2, 0xb0, 0x24, 0x9d, 0x27, 0xdf, 0xaf, 0xf7, 0x6f, 0xd1, 0x4c, 0xbe, + 0xd0, 0x1d, 0x16, 0x3f, 0xf5, 0x23, 0xdb, 0x52, 0xc4, 0x3b, 0x99, 0x3d, + 0xd5, 0xdc, 0x0b, 0x54, 0x3b, 0xfd, 0x9d, 0x36, 0xf6, 0xd9, 0x63, 0xd4, + 0xc0, 0x8f, 0x9d, 0x00, 0xa6, 0x1e, 0x41, 0x72, 0x18, 0xa6, 0xc5, 0xd0, + 0xb6, 0xdd, 0x10, 0x61, 0x45, 0xe0, 0xdc, 0xcc, 0x92, 0xd3, 0x05, 0x54, + 0x26, 0x2c, 0xcf, 0x94, 0x67, 0xa5, 0xae, 0x62, 0x97, 0x4e, 0x10, 0x2b, + 0xf4, 0x65, 0x89, 0x21, 0x98, 0xad, 0x25, 0x6a, 0x01, 0xa9, 0x4f, 0x57, + 0x2b, 0xbe, 0x3b, 0xcc, 0x34, 0x89, 0xc3, 0xd2, 0xa0, 0xc5, 0x72, 0xd9, + 0x39, 0x3f, 0x45, 0x62, 0x73, 0xda, 0xf3, 0xe7, 0xbf, 0xfd, 0xfe, 0x5b, + 0xe0, 0xc5, 0x9f, 0xf9, 0xbe, 0x2b, 0x9a, 0xf7, 0xc2, 0xe9, 0x59, 0x73, + 0xc4, 0x0a, 0xfe, 0x73, 0x5b, 0x34, 0xb9, 0xfc, 0x45, 0xb7, 0x4d, 0x39, + 0xc2, 0xcd, 0x5f, 0x33, 0x91, 0xab, 0x48, 0x57, 0x0a, 0x27, 0xf3, 0xd4, + 0xf3, 0xb4, 0x57, 0x04, 0xeb, 0x8a, 0xb2, 0xd4, 0x06, 0x60, 0x09, 0x48, + 0x58, 0xf8, 0x1f, 0x06, 0x8c, 0x2d, 0x55, 0x2b, 0x8d, 0xbb, 0x37, 0xbb, + 0xc5, 0xa3, 0x05, 0x38, 0xf7, 0x47, 0x0a, 0xd9, 0xa8, 0x5a, 0x5b, 0x75, + 0x58, 0xa3, 0x35, 0x01, 0x1a, 0x5c, 0xe3, 0x97, 0xef, 0x04, 0xd9, 0x28, + 0x93, 0xc9, 0x59, 0xfc, 0xc1, 0x9b, 0x25, 0xe8, 0x44, 0x05, 0x17, 0xdc, + 0xe1, 0xb2, 0x06, 0xd6, 0x08, 0xe0, 0x00, 0xe0, 0x06, 0xaf, 0xb6, 0xf8, + 0x63, 0x6c, 0x54, 0x29, 0x7a, 0x25, 0x0c, 0xc4, 0xe7, 0x6c, 0x2b, 0xe8, + 0xe9, 0x06, 0xa4, 0x9e, 0xb0, 0x38, 0xd4, 0xf1, 0x46, 0xb3, 0x93, 0x54, + 0xa7, 0xa1, 0xcd, 0x65, 0x43, 0xe8, 0xc3, 0x03, 0x60, 0x9c, 0x39, 0x02, + 0xea, 0xc5, 0x0c, 0x96, 0xd2, 0x05, 0x0d, 0x1f, 0xc7, 0x04, 0xc4, 0xa3, + 0xc4, 0xc0, 0xa9, 0x0b, 0xc7, 0xa1, 0x3f, 0xdc, 0x35, 0x51, 0x4d, 0xc8, + 0xc2, 0x87, 0x99, 0x3c, 0x46, 0xb3, 0x4e, 0xc9, 0xbf, 0xb3, 0x34, 0x8b, + 0xb7, 0x6f, 0xe5, 0x95, 0x9b, 0x17, 0x20, 0x56, 0xa6, 0x64, 0x4c, 0x77, + 0xdc, 0x0e, 0x28, 0xc3, 0xef, 0xf4, 0x28, 0x47, 0xd4, 0x0c, 0x6a, 0xe1, + 0x75, 0x63, 0xc9, 0xae, 0xe9, 0x36, 0x57, 0xfd, 0x08, 0x2f, 0xb2, 0x0b, + 0x48, 0xd4, 0x04, 0x24, 0x2f, 0x17, 0x03, 0x9e, 0xfe, 0xfd, 0x67, 0x0e, + 0xbe, 0x66, 0xcf, 0x2c, 0xaa, 0x4f, 0x1c, 0x32, 0x2e, 0xa0, 0xfb, 0x55, + 0x40, 0x15, 0x5d, 0x51, 0xca, 0xbe, 0xff, 0xb2, 0xb2, 0x2b, 0x47, 0xee, + 0x37, 0xc8, 0x65, 0xad, 0xda, 0xb9, 0x3a, 0x75, 0x3a, 0x98, 0x1f, 0xcf, + 0xd7, 0x48, 0x56, 0xa2, 0xed, 0xb4, 0x46, 0x60, 0x30, 0x6a, 0x19, 0x5b, + 0x38, 0xc8, 0x0d, 0x3a, 0xc3, 0xe1, 0x34, 0x6e, 0x39, 0x5f, 0xf2, 0x4d, + 0x78, 0x02, 0xba, 0x3c, 0x71, 0x70, 0x75, 0x6c, 0xb0, 0xfa, 0x38, 0xe3, + 0x6b, 0x42, 0x1e, 0x23, 0xcd, 0xe6, 0xf8, 0xc5, 0x9c, 0x24, 0x3d, 0x98, + 0xa8, 0xbb, 0x4a, 0x07, 0x8c, 0xb6, 0xfa, 0x13, 0xd0, 0xfc, 0xc5, 0xdc, + 0xb2, 0xcd, 0x65, 0x59, 0xc2, 0x3a, 0x24, 0x47, 0x1c, 0x53, 0x92, 0x57, + 0x21, 0xf3, 0x26, 0x9b, 0xe9, 0xa5, 0x95, 0x9a, 0xd6, 0xa5, 0xe2, 0xda, + 0x0e, 0xb7, 0xab, 0x9e, 0xee, 0xe3, 0xef, 0x59, 0xd2, 0x88, 0x32, 0x1f, + 0x0d, 0xbf, 0xf2, 0xa4, 0x3b, 0xd7, 0xd5, 0xf2, 0xa4, 0xae, 0x65, 0xab, + 0xb3, 0x72, 0xf6, 0x3b, 0xe8, 0xc5, 0x2b, 0xad, 0xcc, 0xbe, 0x02, 0x95, + 0x63, 0x95, 0x2c, 0x22, 0x74, 0x3a, 0x1b, 0xd5, 0xd1, 0x1d, 0xf8, 0x69, + 0x03, 0x98, 0x70, 0x66, 0x43, 0xb5, 0x6d, 0xd0, 0x27, 0x6a, 0x1c, 0xfc, + 0xf9, 0xaf, 0x71, 0x9b, 0x8c, 0xcb, 0xf8, 0xbd, 0x18, 0xad, 0x5f, 0xb7, + 0xbc, 0xfb, 0xbd, 0xde, 0xb9, 0xdc, 0x54, 0x65, 0x3b, 0xaf, 0xa7, 0x92, + 0xbe, 0x62, 0xdc, 0x25, 0x50, 0x48, 0x78, 0xd4, 0xed, 0xed, 0x96, 0x3f, + 0x53, 0xc5, 0xb5, 0x5f, 0xac, 0xa7, 0x5c, 0x92, 0xd9, 0xfe, 0x3b, 0xcd, + 0xbb, 0x29, 0xa0, 0xe0, 0x1e, 0xb0, 0x92, 0xad, 0x6b, 0x45, 0x29, 0x59, + 0xff, 0x5d, 0x5a, 0xfe, 0x8f, 0x63, 0x86, 0x6d, 0xa4, 0x4a, 0x53, 0xc4, + 0x3e, 0x39, 0xbf, 0xe5, 0x20, 0xbc, 0xd1, 0xdf, 0x59, 0x9c, 0x3a, 0x72, + 0x3b, 0x8f, 0xb2, 0x40, 0xe5, 0x9e, 0xa5, 0x02, 0x35, 0xd0, 0x4d, 0x6f, + 0x7d, 0xd5, 0x4c, 0xde, 0x51, 0x0a, 0x9a, 0x57, 0x43, 0x43, 0xe5, 0x97, + 0x95, 0x4b, 0xb2, 0x6c, 0xaf, 0x92, 0x4e, 0x52, 0x06, 0x0b, 0x72, 0x60, + 0x9e, 0x5c, 0xa1, 0xe3, 0x9b, 0xb3, 0x8c, 0x32, 0xcd, 0xc1, 0x4a, 0x88, + 0xd6, 0x3d, 0xed, 0xe8, 0x42, 0x5d, 0x53, 0xdd, 0x00, 0x52, 0x26, 0x2e, + 0xd5, 0x41, 0xf2, 0xfc, 0x51, 0x40, 0x45, 0xe4, 0x00, 0xe3, 0x1c, 0xfb, + 0x32, 0x33, 0x22, 0xed, 0x15, 0x12, 0x9b, 0xc4, 0x89, 0xd0, 0x0e, 0x95, + 0xad, 0xfd, 0x04, 0x2e, 0xee, 0x73, 0x06, 0xee, 0x23, 0xe2, 0xd3, 0x3d, + 0x44, 0x62, 0x35, 0xdc, 0x18, 0x9d, 0xf4, 0x9d, 0x92, 0x00, 0x4e, 0x8e, + 0x4e, 0x24, 0xa1, 0x2c, 0xb2, 0xb2, 0x3f, 0xfc, 0xe4, 0x27, 0x43, 0x3b, + 0x59, 0xb4, 0x13, 0xff, 0x57, 0xdf, 0x3d, 0xee, 0x1a, 0xab, 0x8c, 0x51, + 0xd9, 0x96, 0x1f, 0x2b, 0x66, 0x67, 0x42, 0xb6, 0x91, 0xfe, 0x8f, 0x4d, + 0xa6, 0xd3, 0x3b, 0x51, 0x45, 0x35, 0xab, 0xe5, 0x6e, 0x07, 0xed, 0x24, + 0x95, 0x3d, 0x6a, 0x47, 0x3f, 0x4e, 0xe4, 0x13, 0x5f, 0xfc, 0x19, 0xe8, + 0x09, 0x4b, 0x3d, 0xdf, 0x4f, 0xb4, 0xb4, 0xc1, 0x74, 0x31, 0xff, 0x13, + 0x00, 0xaf, 0x07, 0x16, 0xb6, 0x57, 0xfe, 0x6a, 0x37, 0x05, 0x62, 0x01, + 0xa0, 0xfa, 0xe2, 0xe5, 0x57, 0xcb, 0xa4, 0x5a, 0x57, 0xee, 0xd1, 0x5f, + 0x14, 0x23, 0xbe, 0xef, 0x9b, 0x91, 0x0f, 0x97, 0xa8, 0xf2, 0x36, 0xf7, + 0xc3, 0xb6, 0xbe, 0xe5, 0x59, 0x2b, 0x3c, 0xb3, 0x5d, 0x9f, 0x1e, 0x3b, + 0xd3, 0xf7, 0xee, 0x2e, 0xc0, 0x73, 0x6f, 0x2e, 0xfd, 0xc7, 0x3f, 0xfd, + 0x9c, 0xac, 0xbd, 0xa1, 0x8e, 0xcc, 0x59, 0x41, 0xa4, 0x41, 0xd3, 0x39, + 0x28, 0x67, 0x96, 0x14, 0x42, 0xc3, 0x38, 0x96, 0x0d, 0xfc, 0x68, 0x3d, + 0x2e, 0x2f, 0x46, 0x24, 0x66, 0x0d, 0xa6, 0x72, 0xc7, 0x27, 0x66, 0x3c, + 0xad, 0x55, 0xae, 0xbd, 0x34, 0xb4, 0x3b, 0x60, 0x73, 0xa5, 0xaa, 0xd4, + 0x56, 0x0b, 0x61, 0xf5, 0x5c, 0x66, 0x2e, 0x9d, 0x33, 0xfe, 0xfe, 0x7b, + 0x21, 0xbc, 0x36, 0xec, 0x0f, 0x03, 0x28, 0xa4, 0xd6, 0x05, 0x21, 0x30, + 0xf8, 0x3c, 0xd9, 0x3b, 0xaf, 0x5d, 0x92, 0x25, 0xce, 0xac, 0x28, 0xe1, + 0xd1, 0x02, 0x3c, 0x49, 0xe6, 0xed, 0xb7, 0x0e, 0xe7, 0xe7, 0x1e, 0x56, + 0xbf, 0x5d, 0xfd, 0xed, 0xdb, 0x4d, 0x63, 0x03, 0x8c, 0x06, 0x30, 0xfa, + 0x62, 0x78, 0x3f, 0x6e, 0x63, 0x1e, 0xa6, 0x4b, 0x96, 0xe9, 0xe4, 0x2d, + 0x16, 0x51, 0xf2, 0xf1, 0xa7, 0x2a, 0xeb, 0x15, 0xb5, 0xb1, 0x04, 0x9a, + 0xde, 0x77, 0xde, 0xcf, 0xcc, 0x21, 0xd9, 0x30, 0xf1, 0xea, 0xb9, 0xb0, + 0x39, 0xe1, 0x6f, 0xc7, 0x0a, 0xbd, 0x64, 0x75, 0x59, 0xbf, 0x3c, 0xbf, + 0xd0, 0xdb, 0x00, 0xfa, 0x2e, 0x36, 0xcc, 0xb5, 0xd1, 0x20, 0x46, 0xb0, + 0xd7, 0xfc, 0xb1, 0x5b, 0x54, 0x9f, 0xe2, 0xe1, 0xd0, 0x18, 0xa3, 0x51, + 0x62, 0x24, 0x0f, 0xa1, 0xa1, 0x9a, 0x47, 0x33, 0xca, 0xb9, 0x26, 0xb6, + 0x0b, 0x46, 0xd4, 0xb5, 0xc6, 0xbb, 0x72, 0x1e, 0x60, 0xeb, 0xb4, 0x9d, + 0x9f, 0x09, 0x10, 0x12, 0xce, 0x68, 0xa3, 0xb6, 0x8c, 0xce, 0xd7, 0x26, + 0x55, 0xb5, 0x90, 0x08, 0x9f, 0xf2, 0xa8, 0xc0, 0x56, 0xd8, 0xf6, 0x29, + 0x60, 0xe0, 0x73, 0x52, 0x22, 0x6f, 0x35, 0x4e, 0xe7, 0xc5, 0xa3, 0x95, + 0xcd, 0xd0, 0x8e, 0xd3, 0x95, 0xe3, 0x03, 0x04, 0x00, 0x54, 0xeb, 0xef, + 0x27, 0x11, 0xef, 0x38, 0x56, 0x6f, 0xa0, 0xe5, 0x72, 0x2a, 0x97, 0x23, + 0x56, 0xe2, 0x93, 0x21, 0x3f, 0xe2, 0xd6, 0x12, 0xcd, 0x61, 0x50, 0x44, + 0xd3, 0xe3, 0x8d, 0x3f, 0x24, 0x90, 0x6c, 0x53, 0xad, 0x1c, 0xad, 0x03, + 0x0f, 0x89, 0x63, 0xf9, 0xb9, 0xbc, 0xe2, 0x56, 0xdd, 0x16, 0xcf, 0x2d, + 0xa1, 0xda, 0xf9, 0x3f, 0xec, 0xbf, 0xb1, 0xb6, 0xe1, 0xdf, 0x3f, 0x11, + 0x02, 0x76, 0xe9, 0xe2, 0x9f, 0xa2, 0x02, 0xce, 0x3e, 0xf9, 0xcf, 0x4f, + 0xd9, 0x5f, 0x72, 0x5d, 0x51, 0xa7, 0x1d, 0x98, 0xeb, 0x8e, 0x97, 0x98, + 0x39, 0x58, 0x52, 0x11, 0xed, 0x95, 0x3c, 0x94, 0xf0, 0x6c, 0xa2, 0x3e, + 0x5f, 0x5f, 0x05, 0x98, 0xf1, 0x73, 0xab, 0xc7, 0xa8, 0x4b, 0x92, 0x73, + 0xda, 0x59, 0x1d, 0x56, 0x11, 0xc2, 0x38, 0x43, 0xdb, 0x4b, 0xbe, 0x08, + 0xdd, 0xf2, 0x5d, 0x47, 0x26, 0xdc, 0x16, 0xf9, 0x62, 0xf8, 0x92, 0x19, + 0x5c, 0x6f, 0x2b, 0xe1, 0x15, 0x66, 0xfa, 0xdb, 0x3a, 0xe0, 0x92, 0x9c, + 0x70, 0x91, 0x3f, 0xb8, 0xb0, 0x01, 0xc1, 0x44, 0xf6, 0x62, 0x47, 0x37, + 0xe9, 0xd9, 0x4c, 0x0f, 0x99, 0x6a, 0xc4, 0x60, 0x26, 0x2f, 0xc6, 0x43, + 0x50, 0x62, 0xee, 0x44, 0x21, 0xbd, 0xad, 0x50, 0x2d, 0x58, 0x78, 0xea, + 0x5a, 0x5f, 0x5c, 0xf7, 0x28, 0xa9, 0xdf, 0x0e, 0xd3, 0x67, 0xdf, 0x1f, + 0x4c, 0xd3, 0xe9, 0x5e, 0x0f, 0xa3, 0xb7, 0x56, 0xa5, 0x4e, 0x5f, 0x2a, + 0xb6, 0x14, 0x5e, 0x2f, 0x16, 0x71, 0x48, 0x59, 0x77, 0x6b, 0xf9, 0x6c, + 0x79, 0xba, 0xc4, 0x26, 0x30, 0x44, 0x61, 0x62, 0x60, 0xef, 0x35, 0x95, + 0xe3, 0x77, 0xd5, 0xc8, 0x44, 0xa4, 0xf8, 0x95, 0xba, 0xd1, 0x73, 0x6f, + 0x92, 0xf2, 0xd3, 0x98, 0x4c, 0x8f, 0xe0, 0x2e, 0x27, 0xaa, 0x2f, 0x63, + 0x00, 0x00, 0x00, 0x00, 0x06, 0xff, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0x80, 0x04, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x26, 0xff, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, + 0x0e, 0xfe, 0xff, 0xff, 0xbb, 0xfd, 0xff, 0xff, 0xe1, 0x05, 0x00, 0x00, + 0x4b, 0x0f, 0x00, 0x00, 0x8e, 0x15, 0x00, 0x00, 0x7f, 0x04, 0x00, 0x00, + 0x02, 0x02, 0x00, 0x00, 0x53, 0xe6, 0xff, 0xff, 0xa6, 0x04, 0x00, 0x00, + 0xdf, 0x07, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x66, 0xff, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00, + 0x80, 0x00, 0x00, 0x00, 0x7f, 0xfd, 0xff, 0xff, 0x3e, 0xf8, 0xff, 0xff, + 0xae, 0x03, 0x00, 0x00, 0x5c, 0xfe, 0xff, 0xff, 0x82, 0xfa, 0xff, 0xff, + 0xbd, 0xf8, 0xff, 0xff, 0x04, 0xfe, 0xff, 0xff, 0x8c, 0xfe, 0xff, 0xff, + 0x9b, 0xf8, 0xff, 0xff, 0x51, 0x02, 0x00, 0x00, 0x19, 0xfe, 0xff, 0xff, + 0x54, 0xfe, 0xff, 0xff, 0x8f, 0xff, 0xff, 0xff, 0xe7, 0xfd, 0xff, 0xff, + 0xc2, 0x07, 0x00, 0x00, 0x36, 0x06, 0x00, 0x00, 0x57, 0xfd, 0xff, 0xff, + 0xa3, 0x03, 0x00, 0x00, 0x3c, 0x00, 0x00, 0x00, 0x79, 0x03, 0x00, 0x00, + 0x9b, 0xf7, 0xff, 0xff, 0xc7, 0x04, 0x00, 0x00, 0xbf, 0x06, 0x00, 0x00, + 0x86, 0xfe, 0xff, 0xff, 0x20, 0xfb, 0xff, 0xff, 0x90, 0xfc, 0xff, 0xff, + 0x16, 0x00, 0x00, 0x00, 0x8e, 0xff, 0xff, 0xff, 0xa0, 0x03, 0x00, 0x00, + 0xc7, 0xff, 0xff, 0xff, 0x51, 0x01, 0x00, 0x00, 0x24, 0xf8, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x06, 0x00, 0x08, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0xee, 0x01, 0x00, 0x00, + 0xda, 0x02, 0x00, 0x00, 0xa9, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0xc4, 0xfe, 0xff, 0xff, 0xfa, 0xfc, 0xff, 0xff, 0xc0, 0xff, 0xff, 0xff, + 0x6a, 0xff, 0xff, 0xff, 0x92, 0x02, 0x00, 0x00, 0xa4, 0xff, 0xff, 0xff, + 0xfd, 0xfe, 0xff, 0xff, 0x4e, 0xfd, 0xff, 0xff, 0x87, 0x00, 0x00, 0x00, + 0x19, 0xfe, 0xff, 0xff, 0x17, 0xff, 0xff, 0xff, 0xa0, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0xf4, 0xf3, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x04, 0xf4, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0x4d, 0x4c, 0x49, 0x52, + 0x20, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, + 0x18, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00, + 0x0e, 0x00, 0x00, 0x00, 0xf8, 0x01, 0x00, 0x00, 0xec, 0x01, 0x00, 0x00, + 0xe0, 0x01, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x6d, 0x61, 0x69, 0x6e, 0x00, 0x00, 0x00, 0x00, + 0x07, 0x00, 0x00, 0x00, 0xa4, 0x01, 0x00, 0x00, 0x4c, 0x01, 0x00, 0x00, + 0xfc, 0x00, 0x00, 0x00, 0xa8, 0x00, 0x00, 0x00, 0x70, 0x00, 0x00, 0x00, + 0x38, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x82, 0xfe, 0xff, 0xff, + 0x05, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x08, 0x00, + 0x0c, 0x00, 0x10, 0x00, 0x07, 0x00, 0x14, 0x00, 0x0e, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x08, 0x03, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xc8, 0xf4, 0xff, 0xff, + 0x01, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, + 0x0b, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, + 0xe6, 0xfe, 0xff, 0xff, 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x0e, 0x00, 0x1a, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, + 0x07, 0x00, 0x14, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, + 0x01, 0x00, 0x00, 0x00, 0x3c, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x07, 0x00, + 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00, 0x0e, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x01, 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0a, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, + 0xc2, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x01, 0x24, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xb4, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x0e, 0x00, 0x14, 0x00, 0x00, 0x00, 0x08, 0x00, 0x0c, 0x00, + 0x07, 0x00, 0x10, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, + 0x30, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x10, 0x00, 0x06, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x07, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x10, 0x00, 0x04, 0x00, + 0x08, 0x00, 0x0c, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0d, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0d, 0x00, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00, 0xb8, 0x0d, 0x00, 0x00, + 0x64, 0x0c, 0x00, 0x00, 0x64, 0x0a, 0x00, 0x00, 0xe8, 0x09, 0x00, 0x00, + 0x9c, 0x09, 0x00, 0x00, 0x20, 0x09, 0x00, 0x00, 0x6c, 0x07, 0x00, 0x00, + 0x78, 0x04, 0x00, 0x00, 0x74, 0x03, 0x00, 0x00, 0x68, 0x02, 0x00, 0x00, + 0xbc, 0x01, 0x00, 0x00, 0x28, 0x01, 0x00, 0x00, 0xa4, 0x00, 0x00, 0x00, + 0x54, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xc8, 0xff, 0xff, 0xff, + 0x28, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0x0a, 0x00, 0x00, 0x00, + 0x08, 0x00, 0x00, 0x00, 0x49, 0x64, 0x65, 0x6e, 0x74, 0x69, 0x74, 0x79, + 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0a, 0x00, 0x00, 0x00, 0x14, 0x00, 0x10, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x34, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, + 0x10, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x63, 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x5f, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x30, 0xf3, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, + 0x6c, 0x00, 0x00, 0x00, 0x0d, 0x00, 0x00, 0x00, 0x50, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0xff, 0xff, 0xff, 0xff, 0x0a, 0x00, 0x00, 0x00, 0x14, 0xf3, 0xff, 0xff, + 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x13, 0xc2, 0x47, 0x3b, + 0x01, 0x00, 0x00, 0x00, 0x8d, 0xf4, 0xad, 0x3e, 0x01, 0x00, 0x00, 0x00, + 0x15, 0x00, 0xe0, 0xbe, 0x0d, 0x00, 0x00, 0x00, 0x49, 0x64, 0x65, 0x6e, + 0x74, 0x69, 0x74, 0x79, 0x5f, 0x69, 0x6e, 0x74, 0x38, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00, + 0xb0, 0xf3, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, 0x7c, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, + 0x80, 0x04, 0x00, 0x00, 0x94, 0xf3, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00, + 0x24, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x6c, 0x02, 0xa5, 0x3a, + 0x01, 0x00, 0x00, 0x00, 0x6a, 0x5d, 0xa4, 0x3e, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x1a, 0x00, 0x00, 0x00, 0x73, 0x65, 0x71, 0x75, + 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x66, 0x6c, 0x61, 0x74, 0x74, + 0x65, 0x6e, 0x2f, 0x52, 0x65, 0x73, 0x68, 0x61, 0x70, 0x65, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x80, 0x04, 0x00, 0x00, + 0x40, 0xf4, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, 0x8c, 0x00, 0x00, 0x00, + 0x0b, 0x00, 0x00, 0x00, 0x5c, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, + 0x06, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x2c, 0xf4, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x6c, 0x02, 0xa5, 0x3a, 0x01, 0x00, 0x00, 0x00, + 0x6a, 0x5d, 0xa4, 0x3e, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, + 0x61, 0x6c, 0x2f, 0x6d, 0x61, 0x78, 0x5f, 0x70, 0x6f, 0x6f, 0x6c, 0x69, + 0x6e, 0x67, 0x32, 0x64, 0x2f, 0x4d, 0x61, 0x78, 0x50, 0x6f, 0x6f, 0x6c, + 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0xe8, 0xf4, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, 0xec, 0x00, 0x00, 0x00, + 0x0a, 0x00, 0x00, 0x00, 0x5c, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, + 0x0c, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0xd4, 0xf4, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x6c, 0x02, 0xa5, 0x3a, 0x01, 0x00, 0x00, 0x00, + 0x6a, 0x5d, 0xa4, 0x3e, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x83, 0x00, 0x00, 0x00, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, + 0x61, 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x5f, 0x31, 0x2f, + 0x52, 0x65, 0x6c, 0x75, 0x3b, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, + 0x69, 0x61, 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x5f, 0x31, + 0x2f, 0x42, 0x69, 0x61, 0x73, 0x41, 0x64, 0x64, 0x3b, 0x73, 0x65, 0x71, + 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, + 0x32, 0x64, 0x5f, 0x31, 0x2f, 0x43, 0x6f, 0x6e, 0x76, 0x32, 0x44, 0x3b, + 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x63, + 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x5f, 0x31, 0x2f, 0x42, 0x69, 0x61, 0x73, + 0x41, 0x64, 0x64, 0x2f, 0x52, 0x65, 0x61, 0x64, 0x56, 0x61, 0x72, 0x69, + 0x61, 0x62, 0x6c, 0x65, 0x4f, 0x70, 0x2f, 0x72, 0x65, 0x73, 0x6f, 0x75, + 0x72, 0x63, 0x65, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0c, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0xf0, 0xf5, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, 0xe4, 0x00, 0x00, 0x00, + 0x09, 0x00, 0x00, 0x00, 0x5c, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, + 0x0e, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0xdc, 0xf5, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, + 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x15, 0xa1, 0x10, 0x3b, 0x01, 0x00, 0x00, 0x00, + 0x74, 0x10, 0x10, 0x3f, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x7b, 0x00, 0x00, 0x00, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, + 0x61, 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x2f, 0x52, 0x65, + 0x6c, 0x75, 0x3b, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, + 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x2f, 0x42, 0x69, 0x61, + 0x73, 0x41, 0x64, 0x64, 0x3b, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, + 0x69, 0x61, 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x2f, 0x43, + 0x6f, 0x6e, 0x76, 0x32, 0x44, 0x3b, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, + 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x2f, + 0x42, 0x69, 0x61, 0x73, 0x41, 0x64, 0x64, 0x2f, 0x52, 0x65, 0x61, 0x64, + 0x56, 0x61, 0x72, 0x69, 0x61, 0x62, 0x6c, 0x65, 0x4f, 0x70, 0x2f, 0x72, + 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x0e, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x3a, 0xf8, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, + 0xd4, 0x02, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0xac, 0x02, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0xc4, 0xf6, 0xff, 0xff, 0x1c, 0x02, 0x00, 0x00, + 0x94, 0x01, 0x00, 0x00, 0x0c, 0x01, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0xb9, 0x37, 0x74, 0x3a, 0x8b, 0xfe, 0x77, 0x3a, 0x54, 0xc7, 0x75, 0x3a, + 0xc4, 0x11, 0x78, 0x3a, 0xb9, 0x90, 0x74, 0x3a, 0x3b, 0x97, 0x7b, 0x3a, + 0xe8, 0x57, 0x75, 0x3a, 0x0c, 0x0e, 0x74, 0x3a, 0x76, 0x8b, 0x79, 0x3a, + 0x2b, 0x7b, 0x6d, 0x3a, 0x17, 0xad, 0x71, 0x3a, 0xe4, 0x9b, 0x77, 0x3a, + 0x0b, 0xab, 0x7a, 0x3a, 0x9e, 0x12, 0x75, 0x3a, 0x8c, 0xcf, 0x79, 0x3a, + 0xa0, 0x5a, 0x79, 0x3a, 0x74, 0xc3, 0x78, 0x3a, 0x0e, 0xa9, 0x74, 0x3a, + 0x6b, 0xf8, 0x6f, 0x3a, 0x53, 0xeb, 0x72, 0x3a, 0xff, 0xe2, 0x73, 0x3a, + 0x3b, 0x38, 0x78, 0x3a, 0xed, 0x9e, 0x76, 0x3a, 0x77, 0xbc, 0x6d, 0x3a, + 0x4f, 0xf5, 0x71, 0x3a, 0x17, 0xc9, 0x74, 0x3a, 0x87, 0x84, 0x6b, 0x3a, + 0x4b, 0xc5, 0x78, 0x3a, 0xdd, 0x02, 0x75, 0x3a, 0x0e, 0xcf, 0x78, 0x3a, + 0x14, 0x40, 0x75, 0x3a, 0x2e, 0xca, 0x72, 0x3a, 0x20, 0x00, 0x00, 0x00, + 0x95, 0x2f, 0xef, 0x3d, 0x47, 0x1c, 0xf0, 0x3d, 0xc5, 0xdb, 0xf3, 0x3d, + 0x2e, 0x57, 0xe7, 0x3d, 0x98, 0xa7, 0xf2, 0x3d, 0x98, 0x89, 0xe4, 0x3d, + 0x38, 0x6d, 0xf3, 0x3d, 0x3f, 0x38, 0xe2, 0x3d, 0x91, 0x6f, 0xf0, 0x3d, + 0x35, 0xa0, 0xeb, 0x3d, 0x42, 0x3d, 0xeb, 0x3d, 0xed, 0x89, 0xe7, 0x3d, + 0xb5, 0xb5, 0xf8, 0x3d, 0x79, 0x28, 0xf3, 0x3d, 0xed, 0xdb, 0xf7, 0x3d, + 0xeb, 0x67, 0xf7, 0x3d, 0xed, 0xd1, 0xf6, 0x3d, 0xbc, 0xbf, 0xf2, 0x3d, + 0x7a, 0x18, 0xee, 0x3d, 0x7c, 0x05, 0xf1, 0x3d, 0x63, 0x69, 0xe8, 0x3d, + 0xbb, 0xc0, 0xf1, 0x3d, 0xaf, 0xb1, 0xf4, 0x3d, 0xfe, 0xe0, 0xeb, 0x3d, + 0xb6, 0x60, 0xec, 0x3d, 0x8c, 0x32, 0xf0, 0x3d, 0x7e, 0xad, 0xe9, 0x3d, + 0xc0, 0xd3, 0xf6, 0x3d, 0xd7, 0x18, 0xf3, 0x3d, 0x40, 0x53, 0xf0, 0x3d, + 0x2c, 0xdc, 0xf1, 0x3d, 0x9a, 0xe4, 0xf0, 0x3d, 0x20, 0x00, 0x00, 0x00, + 0x4a, 0x4f, 0xf2, 0xbd, 0x8e, 0x0e, 0xf6, 0xbd, 0x74, 0x46, 0xec, 0xbd, + 0xa0, 0x21, 0xf6, 0xbd, 0x8e, 0x27, 0xf0, 0xbd, 0x0d, 0xa0, 0xf9, 0xbd, + 0x0c, 0x97, 0xec, 0xbd, 0xf0, 0x25, 0xf2, 0xbd, 0x5f, 0x98, 0xf7, 0xbd, + 0x27, 0x8d, 0xe8, 0xbd, 0xbd, 0xc9, 0xef, 0xbd, 0xac, 0xac, 0xf5, 0xbd, + 0x5a, 0x94, 0xed, 0xbd, 0x5a, 0x64, 0xf1, 0xbd, 0x2a, 0xa7, 0xe9, 0xbd, + 0x3c, 0x93, 0xf3, 0xbd, 0xf8, 0x2b, 0xf3, 0xbd, 0xf6, 0x35, 0xed, 0xbd, + 0x94, 0xf4, 0xed, 0xbd, 0x70, 0x94, 0xe9, 0xbd, 0x39, 0xfb, 0xf1, 0xbd, + 0xcb, 0x47, 0xf6, 0xbd, 0x88, 0xb9, 0xe7, 0xbd, 0x49, 0x62, 0xe9, 0xbd, + 0x64, 0x11, 0xf0, 0xbd, 0x85, 0xdf, 0xf2, 0xbd, 0x5c, 0x61, 0xe8, 0xbd, + 0x22, 0x46, 0xf3, 0xbd, 0x5a, 0x8e, 0xf0, 0xbd, 0x70, 0xdd, 0xf6, 0xbd, + 0x94, 0x55, 0xf3, 0xbd, 0x57, 0xba, 0xf0, 0xbd, 0x1a, 0x00, 0x00, 0x00, + 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x63, + 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x5f, 0x31, 0x2f, 0x43, 0x6f, 0x6e, 0x76, + 0x32, 0x44, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x2a, 0xfb, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, 0x94, 0x01, 0x00, 0x00, + 0x07, 0x00, 0x00, 0x00, 0x6c, 0x01, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0xb4, 0xf9, 0xff, 0xff, 0x1c, 0x01, 0x00, 0x00, 0xd4, 0x00, 0x00, 0x00, + 0x8c, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0xe6, 0x69, 0xc5, 0x3a, 0xa0, 0x8d, 0xa8, 0x3a, 0xfe, 0x5c, 0xc1, 0x3a, + 0x84, 0x01, 0xcb, 0x3a, 0xa2, 0xc2, 0xb5, 0x3a, 0x42, 0x01, 0xd1, 0x3a, + 0xd7, 0x01, 0xcc, 0x3a, 0x20, 0xd8, 0xc7, 0x3a, 0x28, 0x80, 0xa4, 0x3a, + 0xd9, 0x25, 0xbe, 0x3a, 0x39, 0x6f, 0xc4, 0x3a, 0x59, 0x6c, 0xcb, 0x3a, + 0xb8, 0x0a, 0xc2, 0x3a, 0x73, 0x3f, 0xca, 0x3a, 0xb9, 0xed, 0xc5, 0x3a, + 0xe9, 0x9f, 0xc1, 0x3a, 0x10, 0x00, 0x00, 0x00, 0x5b, 0x2e, 0x2f, 0x3e, + 0x3e, 0xd9, 0x06, 0x3e, 0x44, 0xda, 0x3f, 0x3e, 0xd3, 0x09, 0x22, 0x3e, + 0x1d, 0x57, 0x34, 0x3e, 0xa4, 0xb6, 0x44, 0x3e, 0xd3, 0x69, 0x4a, 0x3e, + 0x70, 0x48, 0x46, 0x3e, 0x28, 0x37, 0x23, 0x3e, 0xe6, 0xdb, 0x06, 0x3e, + 0x3c, 0x1d, 0x34, 0x3e, 0x36, 0xba, 0x16, 0x3e, 0x24, 0xa4, 0x34, 0x3e, + 0xf4, 0xfb, 0x37, 0x3e, 0xd6, 0x7b, 0x8a, 0x3d, 0x00, 0x85, 0xe3, 0x3d, + 0x10, 0x00, 0x00, 0x00, 0x12, 0xdf, 0x43, 0xbe, 0x85, 0x3c, 0x27, 0xbe, + 0x54, 0xcd, 0x0d, 0xbe, 0x81, 0x6b, 0x49, 0xbe, 0x33, 0xb1, 0xe7, 0xbd, + 0x3f, 0x5f, 0x4f, 0xbe, 0xa1, 0x63, 0x3e, 0xbe, 0xbb, 0xa7, 0xea, 0xbd, + 0x2d, 0x8c, 0x0e, 0xbe, 0x8d, 0xa9, 0x3c, 0xbe, 0x5b, 0xe6, 0x42, 0xbe, + 0x80, 0xd5, 0x49, 0xbe, 0xa3, 0x86, 0x40, 0xbe, 0xf4, 0xaa, 0x48, 0xbe, + 0xde, 0x61, 0x44, 0xbe, 0xa9, 0x1c, 0x40, 0xbe, 0x18, 0x00, 0x00, 0x00, + 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x63, + 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x2f, 0x43, 0x6f, 0x6e, 0x76, 0x32, 0x44, + 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0xda, 0xfc, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09, 0x64, 0x00, 0x00, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x64, 0xfb, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x07, 0x72, 0x1e, 0x3a, 0x01, 0x00, 0x00, 0x00, 0x32, 0xe2, 0x9b, 0x3d, + 0x01, 0x00, 0x00, 0x00, 0x23, 0x35, 0x9d, 0xbd, 0x17, 0x00, 0x00, 0x00, + 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x64, + 0x65, 0x6e, 0x73, 0x65, 0x2f, 0x4d, 0x61, 0x74, 0x4d, 0x75, 0x6c, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00, 0x80, 0x04, 0x00, 0x00, + 0x52, 0xfd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x02, 0x38, 0x00, 0x00, 0x00, + 0x05, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x66, + 0x6c, 0x61, 0x74, 0x74, 0x65, 0x6e, 0x2f, 0x43, 0x6f, 0x6e, 0x73, 0x74, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x9a, 0xfd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x02, 0x68, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x8c, 0xfd, 0xff, 0xff, 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xfc, 0x41, 0x4c, 0x35, 0x30, 0x00, 0x00, 0x00, + 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x64, + 0x65, 0x6e, 0x73, 0x65, 0x2f, 0x42, 0x69, 0x61, 0x73, 0x41, 0x64, 0x64, + 0x2f, 0x52, 0x65, 0x61, 0x64, 0x56, 0x61, 0x72, 0x69, 0x61, 0x62, 0x6c, + 0x65, 0x4f, 0x70, 0x2f, 0x72, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00, + 0x12, 0xfe, 0xff, 0xff, 0x00, 0x00, 0x00, 0x02, 0xdc, 0x01, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x9c, 0x01, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x04, 0xfe, 0xff, 0xff, 0x0c, 0x01, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x03, 0xf9, 0x09, 0x36, 0x3a, 0x1b, 0x0c, 0x36, 0xc6, 0xda, 0x0a, 0x36, + 0x16, 0x26, 0x0c, 0x36, 0x4b, 0x2b, 0x0a, 0x36, 0x60, 0x23, 0x0e, 0x36, + 0xd3, 0x9b, 0x0a, 0x36, 0x78, 0xe1, 0x09, 0x36, 0x78, 0xfb, 0x0c, 0x36, + 0xb6, 0x2a, 0x06, 0x36, 0x6f, 0x89, 0x08, 0x36, 0x7e, 0xe3, 0x0b, 0x36, + 0xf0, 0x9d, 0x0d, 0x36, 0xae, 0x74, 0x0a, 0x36, 0xef, 0x21, 0x0d, 0x36, + 0xe0, 0xdf, 0x0c, 0x36, 0x79, 0x8a, 0x0c, 0x36, 0x0a, 0x39, 0x0a, 0x36, + 0xbb, 0x92, 0x07, 0x36, 0x39, 0x3d, 0x09, 0x36, 0x25, 0xc9, 0x09, 0x36, + 0xd1, 0x3b, 0x0c, 0x36, 0x93, 0x54, 0x0b, 0x36, 0x9a, 0x4f, 0x06, 0x36, + 0x3c, 0xb2, 0x08, 0x36, 0x23, 0x4b, 0x0a, 0x36, 0xbe, 0x0e, 0x05, 0x36, + 0x83, 0x8b, 0x0c, 0x36, 0xc7, 0x6b, 0x0a, 0x36, 0x07, 0x91, 0x0c, 0x36, + 0x5d, 0x8e, 0x0a, 0x36, 0x7f, 0x2a, 0x09, 0x36, 0x33, 0x00, 0x00, 0x00, + 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x63, + 0x6f, 0x6e, 0x76, 0x32, 0x64, 0x5f, 0x31, 0x2f, 0x42, 0x69, 0x61, 0x73, + 0x41, 0x64, 0x64, 0x2f, 0x52, 0x65, 0x61, 0x64, 0x56, 0x61, 0x72, 0x69, + 0x61, 0x62, 0x6c, 0x65, 0x4f, 0x70, 0x2f, 0x72, 0x65, 0x73, 0x6f, 0x75, + 0x72, 0x63, 0x65, 0x00, 0x01, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x08, 0x00, 0x07, 0x00, 0x0c, 0x00, + 0x10, 0x00, 0x14, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, + 0x2c, 0x01, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0xec, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x90, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0xe1, 0x22, 0xc6, 0x36, 0x90, 0x2b, 0xa9, 0x36, 0x2d, 0x12, 0xc2, 0x36, + 0xbc, 0xbf, 0xcb, 0x36, 0xf2, 0x6c, 0xb6, 0x36, 0x19, 0xc5, 0xd1, 0x36, + 0xff, 0xc0, 0xcc, 0x36, 0x62, 0x93, 0xc8, 0x36, 0x4c, 0x1a, 0xa5, 0x36, + 0x05, 0xd8, 0xbe, 0x36, 0x49, 0x27, 0xc5, 0x36, 0xf5, 0x2a, 0xcc, 0x36, + 0x8a, 0xc0, 0xc2, 0x36, 0xf5, 0xfc, 0xca, 0x36, 0x2f, 0xa7, 0xc6, 0x36, + 0x57, 0x55, 0xc2, 0x36, 0x31, 0x00, 0x00, 0x00, 0x73, 0x65, 0x71, 0x75, + 0x65, 0x6e, 0x74, 0x69, 0x61, 0x6c, 0x2f, 0x63, 0x6f, 0x6e, 0x76, 0x32, + 0x64, 0x2f, 0x42, 0x69, 0x61, 0x73, 0x41, 0x64, 0x64, 0x2f, 0x52, 0x65, + 0x61, 0x64, 0x56, 0x61, 0x72, 0x69, 0x61, 0x62, 0x6c, 0x65, 0x4f, 0x70, + 0x2f, 0x72, 0x65, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x1c, 0x00, + 0x08, 0x00, 0x07, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x18, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x09, + 0x88, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, + 0x28, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0xff, 0xff, 0xff, 0xff, 0x10, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x14, 0x00, 0x04, 0x00, 0x08, 0x00, + 0x0c, 0x00, 0x10, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, + 0x24, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x80, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0xf0, 0x77, 0x80, 0x3b, + 0x01, 0x00, 0x00, 0x00, 0xf0, 0xee, 0x7f, 0x3f, 0x01, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, 0x63, 0x6f, 0x6e, 0x76, + 0x32, 0x64, 0x5f, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x5f, 0x69, 0x6e, 0x74, + 0x38, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x70, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, + 0x04, 0x00, 0x00, 0x00, 0xca, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x06, + 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x08, 0x00, 0x07, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x72, 0xe6, 0xff, 0xff, 0xff, + 0x00, 0x00, 0x00, 0x09, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, + 0x06, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x16, 0x0a, 0x00, + 0x0e, 0x00, 0x07, 0x00, 0x00, 0x00, 0x08, 0x00, 0x0a, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x11, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0a, 0x00, + 0x0c, 0x00, 0x07, 0x00, 0x00, 0x00, 0x08, 0x00, 0x0a, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x03, 0x03, 0x00, 0x00, 0x00}; + +const unsigned int kTestConvModelDataSize = 21344; + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.h b/src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.h new file mode 100644 index 0000000..cbdfdc4 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/testing/test_conv_model.h @@ -0,0 +1,26 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_MICRO_TESTING_TEST_CONV_MODEL_H_ +#define TENSORFLOW_LITE_MICRO_TESTING_TEST_CONV_MODEL_H_ + +// See generate_test_models.py for updating the contents of this model: +extern const unsigned char kTestConvModelData[]; +extern const unsigned int kTestConvModelDataSize; + +#endif // TENSORFLOW_LITE_MICRO_TESTING_TEST_CONV_MODEL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/Core/Include/cmsis_compiler.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/Core/Include/cmsis_compiler.h new file mode 100644 index 0000000..920cafe --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/Core/Include/cmsis_compiler.h @@ -0,0 +1,286 @@ +#if !defined(ESP32) +/**************************************************************************//** + * @file cmsis_compiler.h + * @brief CMSIS compiler generic header file + * @version V5.1.0 + * @date 09. October 2018 + ******************************************************************************/ +/* + * Copyright (c) 2009-2018 Arm Limited. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef __CMSIS_COMPILER_H +#define __CMSIS_COMPILER_H + +#include + +/* + * Arm Compiler 4/5 + */ +#if defined ( __CC_ARM ) + #include "cmsis_armcc.h" + + +/* + * Arm Compiler 6.6 LTM (armclang) + */ +#elif defined (__ARMCC_VERSION) && (__ARMCC_VERSION >= 6010050) && (__ARMCC_VERSION < 6100100) + #include "cmsis_armclang_ltm.h" + + /* + * Arm Compiler above 6.10.1 (armclang) + */ +#elif defined (__ARMCC_VERSION) && (__ARMCC_VERSION >= 6100100) + #include "cmsis_armclang.h" + + +/* + * GNU Compiler + */ +#elif defined ( __GNUC__ ) + #include "cmsis_gcc.h" + + +/* + * IAR Compiler + */ +#elif defined ( __ICCARM__ ) + #include + + +/* + * TI Arm Compiler + */ +#elif defined ( __TI_ARM__ ) + #include + + #ifndef __ASM + #define __ASM __asm + #endif + #ifndef __INLINE + #define __INLINE inline + #endif + #ifndef __STATIC_INLINE + #define __STATIC_INLINE static inline + #endif + #ifndef __STATIC_FORCEINLINE + #define __STATIC_FORCEINLINE __STATIC_INLINE + #endif + #ifndef __NO_RETURN + #define __NO_RETURN __attribute__((noreturn)) + #endif + #ifndef __USED + #define __USED __attribute__((used)) + #endif + #ifndef __WEAK + #define __WEAK __attribute__((weak)) + #endif + #ifndef __PACKED + #define __PACKED __attribute__((packed)) + #endif + #ifndef __PACKED_STRUCT + #define __PACKED_STRUCT struct __attribute__((packed)) + #endif + #ifndef __PACKED_UNION + #define __PACKED_UNION union __attribute__((packed)) + #endif + #ifndef __UNALIGNED_UINT32 /* deprecated */ + struct __attribute__((packed)) T_UINT32 { uint32_t v; }; + #define __UNALIGNED_UINT32(x) (((struct T_UINT32 *)(x))->v) + #endif + #ifndef __UNALIGNED_UINT16_WRITE + __PACKED_STRUCT T_UINT16_WRITE { uint16_t v; }; + #define __UNALIGNED_UINT16_WRITE(addr, val) (void)((((struct T_UINT16_WRITE *)(void*)(addr))->v) = (val)) + #endif + #ifndef __UNALIGNED_UINT16_READ + __PACKED_STRUCT T_UINT16_READ { uint16_t v; }; + #define __UNALIGNED_UINT16_READ(addr) (((const struct T_UINT16_READ *)(const void *)(addr))->v) + #endif + #ifndef __UNALIGNED_UINT32_WRITE + __PACKED_STRUCT T_UINT32_WRITE { uint32_t v; }; + #define __UNALIGNED_UINT32_WRITE(addr, val) (void)((((struct T_UINT32_WRITE *)(void *)(addr))->v) = (val)) + #endif + #ifndef __UNALIGNED_UINT32_READ + __PACKED_STRUCT T_UINT32_READ { uint32_t v; }; + #define __UNALIGNED_UINT32_READ(addr) (((const struct T_UINT32_READ *)(const void *)(addr))->v) + #endif + #ifndef __ALIGNED + #define __ALIGNED(x) __attribute__((aligned(x))) + #endif + #ifndef __RESTRICT + #define __RESTRICT __restrict + #endif + #ifndef __COMPILER_BARRIER + #warning No compiler specific solution for __COMPILER_BARRIER. __COMPILER_BARRIER is ignored. + #define __COMPILER_BARRIER() (void)0 + #endif + + +/* + * TASKING Compiler + */ +#elif defined ( __TASKING__ ) + /* + * The CMSIS functions have been implemented as intrinsics in the compiler. + * Please use "carm -?i" to get an up to date list of all intrinsics, + * Including the CMSIS ones. + */ + + #ifndef __ASM + #define __ASM __asm + #endif + #ifndef __INLINE + #define __INLINE inline + #endif + #ifndef __STATIC_INLINE + #define __STATIC_INLINE static inline + #endif + #ifndef __STATIC_FORCEINLINE + #define __STATIC_FORCEINLINE __STATIC_INLINE + #endif + #ifndef __NO_RETURN + #define __NO_RETURN __attribute__((noreturn)) + #endif + #ifndef __USED + #define __USED __attribute__((used)) + #endif + #ifndef __WEAK + #define __WEAK __attribute__((weak)) + #endif + #ifndef __PACKED + #define __PACKED __packed__ + #endif + #ifndef __PACKED_STRUCT + #define __PACKED_STRUCT struct __packed__ + #endif + #ifndef __PACKED_UNION + #define __PACKED_UNION union __packed__ + #endif + #ifndef __UNALIGNED_UINT32 /* deprecated */ + struct __packed__ T_UINT32 { uint32_t v; }; + #define __UNALIGNED_UINT32(x) (((struct T_UINT32 *)(x))->v) + #endif + #ifndef __UNALIGNED_UINT16_WRITE + __PACKED_STRUCT T_UINT16_WRITE { uint16_t v; }; + #define __UNALIGNED_UINT16_WRITE(addr, val) (void)((((struct T_UINT16_WRITE *)(void *)(addr))->v) = (val)) + #endif + #ifndef __UNALIGNED_UINT16_READ + __PACKED_STRUCT T_UINT16_READ { uint16_t v; }; + #define __UNALIGNED_UINT16_READ(addr) (((const struct T_UINT16_READ *)(const void *)(addr))->v) + #endif + #ifndef __UNALIGNED_UINT32_WRITE + __PACKED_STRUCT T_UINT32_WRITE { uint32_t v; }; + #define __UNALIGNED_UINT32_WRITE(addr, val) (void)((((struct T_UINT32_WRITE *)(void *)(addr))->v) = (val)) + #endif + #ifndef __UNALIGNED_UINT32_READ + __PACKED_STRUCT T_UINT32_READ { uint32_t v; }; + #define __UNALIGNED_UINT32_READ(addr) (((const struct T_UINT32_READ *)(const void *)(addr))->v) + #endif + #ifndef __ALIGNED + #define __ALIGNED(x) __align(x) + #endif + #ifndef __RESTRICT + #warning No compiler specific solution for __RESTRICT. __RESTRICT is ignored. + #define __RESTRICT + #endif + #ifndef __COMPILER_BARRIER + #warning No compiler specific solution for __COMPILER_BARRIER. __COMPILER_BARRIER is ignored. + #define __COMPILER_BARRIER() (void)0 + #endif + + +/* + * COSMIC Compiler + */ +#elif defined ( __CSMC__ ) + #include + + #ifndef __ASM + #define __ASM _asm + #endif + #ifndef __INLINE + #define __INLINE inline + #endif + #ifndef __STATIC_INLINE + #define __STATIC_INLINE static inline + #endif + #ifndef __STATIC_FORCEINLINE + #define __STATIC_FORCEINLINE __STATIC_INLINE + #endif + #ifndef __NO_RETURN + // NO RETURN is automatically detected hence no warning here + #define __NO_RETURN + #endif + #ifndef __USED + #warning No compiler specific solution for __USED. __USED is ignored. + #define __USED + #endif + #ifndef __WEAK + #define __WEAK __weak + #endif + #ifndef __PACKED + #define __PACKED @packed + #endif + #ifndef __PACKED_STRUCT + #define __PACKED_STRUCT @packed struct + #endif + #ifndef __PACKED_UNION + #define __PACKED_UNION @packed union + #endif + #ifndef __UNALIGNED_UINT32 /* deprecated */ + @packed struct T_UINT32 { uint32_t v; }; + #define __UNALIGNED_UINT32(x) (((struct T_UINT32 *)(x))->v) + #endif + #ifndef __UNALIGNED_UINT16_WRITE + __PACKED_STRUCT T_UINT16_WRITE { uint16_t v; }; + #define __UNALIGNED_UINT16_WRITE(addr, val) (void)((((struct T_UINT16_WRITE *)(void *)(addr))->v) = (val)) + #endif + #ifndef __UNALIGNED_UINT16_READ + __PACKED_STRUCT T_UINT16_READ { uint16_t v; }; + #define __UNALIGNED_UINT16_READ(addr) (((const struct T_UINT16_READ *)(const void *)(addr))->v) + #endif + #ifndef __UNALIGNED_UINT32_WRITE + __PACKED_STRUCT T_UINT32_WRITE { uint32_t v; }; + #define __UNALIGNED_UINT32_WRITE(addr, val) (void)((((struct T_UINT32_WRITE *)(void *)(addr))->v) = (val)) + #endif + #ifndef __UNALIGNED_UINT32_READ + __PACKED_STRUCT T_UINT32_READ { uint32_t v; }; + #define __UNALIGNED_UINT32_READ(addr) (((const struct T_UINT32_READ *)(const void *)(addr))->v) + #endif + #ifndef __ALIGNED + #warning No compiler specific solution for __ALIGNED. __ALIGNED is ignored. + #define __ALIGNED(x) + #endif + #ifndef __RESTRICT + #warning No compiler specific solution for __RESTRICT. __RESTRICT is ignored. + #define __RESTRICT + #endif + #ifndef __COMPILER_BARRIER + #warning No compiler specific solution for __COMPILER_BARRIER. __COMPILER_BARRIER is ignored. + #define __COMPILER_BARRIER() (void)0 + #endif + + +#else + #error Unknown compiler. +#endif + + +#endif /* __CMSIS_COMPILER_H */ + + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h new file mode 100644 index 0000000..4cc13bc --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h @@ -0,0 +1,532 @@ +#if !defined(ESP32) +/* ---------------------------------------------------------------------- + * Project: CMSIS DSP Library + * Title: arm_common_tables.h + * Description: Extern declaration for common tables + * + * $Date: 27. January 2017 + * $Revision: V.1.5.1 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ +/* + * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef _ARM_COMMON_TABLES_H +#define _ARM_COMMON_TABLES_H + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +#if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_FFT_ALLOW_TABLES) + /* Double Precision Float CFFT twiddles */ + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREV_1024) + extern const uint16_t armBitRevTable[1024]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_16) + extern const uint64_t twiddleCoefF64_16[32]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_32) + extern const uint64_t twiddleCoefF64_32[64]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_64) + extern const uint64_t twiddleCoefF64_64[128]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_128) + extern const uint64_t twiddleCoefF64_128[256]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_256) + extern const uint64_t twiddleCoefF64_256[512]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_512) + extern const uint64_t twiddleCoefF64_512[1024]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_1024) + extern const uint64_t twiddleCoefF64_1024[2048]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_2048) + extern const uint64_t twiddleCoefF64_2048[4096]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F64_4096) + extern const uint64_t twiddleCoefF64_4096[8192]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_16) + extern const float32_t twiddleCoef_16[32]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_32) + extern const float32_t twiddleCoef_32[64]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_64) + extern const float32_t twiddleCoef_64[128]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_128) + extern const float32_t twiddleCoef_128[256]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_256) + extern const float32_t twiddleCoef_256[512]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_512) + extern const float32_t twiddleCoef_512[1024]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_1024) + extern const float32_t twiddleCoef_1024[2048]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_2048) + extern const float32_t twiddleCoef_2048[4096]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_F32_4096) + extern const float32_t twiddleCoef_4096[8192]; + #define twiddleCoef twiddleCoef_4096 + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + /* Q31 */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_16) + extern const q31_t twiddleCoef_16_q31[24]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_32) + extern const q31_t twiddleCoef_32_q31[48]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_64) + extern const q31_t twiddleCoef_64_q31[96]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_128) + extern const q31_t twiddleCoef_128_q31[192]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_256) + extern const q31_t twiddleCoef_256_q31[384]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_512) + extern const q31_t twiddleCoef_512_q31[768]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_1024) + extern const q31_t twiddleCoef_1024_q31[1536]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_2048) + extern const q31_t twiddleCoef_2048_q31[3072]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q31_4096) + extern const q31_t twiddleCoef_4096_q31[6144]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_16) + extern const q15_t twiddleCoef_16_q15[24]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_32) + extern const q15_t twiddleCoef_32_q15[48]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_64) + extern const q15_t twiddleCoef_64_q15[96]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_128) + extern const q15_t twiddleCoef_128_q15[192]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_256) + extern const q15_t twiddleCoef_256_q15[384]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_512) + extern const q15_t twiddleCoef_512_q15[768]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_1024) + extern const q15_t twiddleCoef_1024_q15[1536]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_2048) + extern const q15_t twiddleCoef_2048_q15[3072]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_Q15_4096) + extern const q15_t twiddleCoef_4096_q15[6144]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + /* Double Precision Float RFFT twiddles */ + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_32) + extern const uint64_t twiddleCoefF64_rfft_32[32]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_64) + extern const uint64_t twiddleCoefF64_rfft_64[64]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_128) + extern const uint64_t twiddleCoefF64_rfft_128[128]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_256) + extern const uint64_t twiddleCoefF64_rfft_256[256]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_512) + extern const uint64_t twiddleCoefF64_rfft_512[512]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_1024) + extern const uint64_t twiddleCoefF64_rfft_1024[1024]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_2048) + extern const uint64_t twiddleCoefF64_rfft_2048[2048]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F64_4096) + extern const uint64_t twiddleCoefF64_rfft_4096[4096]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_32) + extern const float32_t twiddleCoef_rfft_32[32]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_64) + extern const float32_t twiddleCoef_rfft_64[64]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_128) + extern const float32_t twiddleCoef_rfft_128[128]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_256) + extern const float32_t twiddleCoef_rfft_256[256]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_512) + extern const float32_t twiddleCoef_rfft_512[512]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_1024) + extern const float32_t twiddleCoef_rfft_1024[1024]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_2048) + extern const float32_t twiddleCoef_rfft_2048[2048]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_TWIDDLECOEF_RFFT_F32_4096) + extern const float32_t twiddleCoef_rfft_4096[4096]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + + /* Double precision floating-point bit reversal tables */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_16) + #define ARMBITREVINDEXTABLEF64_16_TABLE_LENGTH ((uint16_t)12) + extern const uint16_t armBitRevIndexTableF64_16[ARMBITREVINDEXTABLEF64_16_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_32) + #define ARMBITREVINDEXTABLEF64_32_TABLE_LENGTH ((uint16_t)24) + extern const uint16_t armBitRevIndexTableF64_32[ARMBITREVINDEXTABLEF64_32_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_64) + #define ARMBITREVINDEXTABLEF64_64_TABLE_LENGTH ((uint16_t)56) + extern const uint16_t armBitRevIndexTableF64_64[ARMBITREVINDEXTABLEF64_64_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_128) + #define ARMBITREVINDEXTABLEF64_128_TABLE_LENGTH ((uint16_t)112) + extern const uint16_t armBitRevIndexTableF64_128[ARMBITREVINDEXTABLEF64_128_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_256) + #define ARMBITREVINDEXTABLEF64_256_TABLE_LENGTH ((uint16_t)240) + extern const uint16_t armBitRevIndexTableF64_256[ARMBITREVINDEXTABLEF64_256_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_512) + #define ARMBITREVINDEXTABLEF64_512_TABLE_LENGTH ((uint16_t)480) + extern const uint16_t armBitRevIndexTableF64_512[ARMBITREVINDEXTABLEF64_512_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_1024) + #define ARMBITREVINDEXTABLEF64_1024_TABLE_LENGTH ((uint16_t)992) + extern const uint16_t armBitRevIndexTableF64_1024[ARMBITREVINDEXTABLEF64_1024_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_2048) + #define ARMBITREVINDEXTABLEF64_2048_TABLE_LENGTH ((uint16_t)1984) + extern const uint16_t armBitRevIndexTableF64_2048[ARMBITREVINDEXTABLEF64_2048_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT64_4096) + #define ARMBITREVINDEXTABLEF64_4096_TABLE_LENGTH ((uint16_t)4032) + extern const uint16_t armBitRevIndexTableF64_4096[ARMBITREVINDEXTABLEF64_4096_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + /* floating-point bit reversal tables */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_16) + #define ARMBITREVINDEXTABLE_16_TABLE_LENGTH ((uint16_t)20) + extern const uint16_t armBitRevIndexTable16[ARMBITREVINDEXTABLE_16_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_32) + #define ARMBITREVINDEXTABLE_32_TABLE_LENGTH ((uint16_t)48) + extern const uint16_t armBitRevIndexTable32[ARMBITREVINDEXTABLE_32_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_64) + #define ARMBITREVINDEXTABLE_64_TABLE_LENGTH ((uint16_t)56) + extern const uint16_t armBitRevIndexTable64[ARMBITREVINDEXTABLE_64_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_128) + #define ARMBITREVINDEXTABLE_128_TABLE_LENGTH ((uint16_t)208) + extern const uint16_t armBitRevIndexTable128[ARMBITREVINDEXTABLE_128_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_256) + #define ARMBITREVINDEXTABLE_256_TABLE_LENGTH ((uint16_t)440) + extern const uint16_t armBitRevIndexTable256[ARMBITREVINDEXTABLE_256_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_512) + #define ARMBITREVINDEXTABLE_512_TABLE_LENGTH ((uint16_t)448) + extern const uint16_t armBitRevIndexTable512[ARMBITREVINDEXTABLE_512_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_1024) + #define ARMBITREVINDEXTABLE_1024_TABLE_LENGTH ((uint16_t)1800) + extern const uint16_t armBitRevIndexTable1024[ARMBITREVINDEXTABLE_1024_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_2048) + #define ARMBITREVINDEXTABLE_2048_TABLE_LENGTH ((uint16_t)3808) + extern const uint16_t armBitRevIndexTable2048[ARMBITREVINDEXTABLE_2048_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FLT_4096) + #define ARMBITREVINDEXTABLE_4096_TABLE_LENGTH ((uint16_t)4032) + extern const uint16_t armBitRevIndexTable4096[ARMBITREVINDEXTABLE_4096_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + + /* fixed-point bit reversal tables */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_16) + #define ARMBITREVINDEXTABLE_FIXED_16_TABLE_LENGTH ((uint16_t)12) + extern const uint16_t armBitRevIndexTable_fixed_16[ARMBITREVINDEXTABLE_FIXED_16_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_32) + #define ARMBITREVINDEXTABLE_FIXED_32_TABLE_LENGTH ((uint16_t)24) + extern const uint16_t armBitRevIndexTable_fixed_32[ARMBITREVINDEXTABLE_FIXED_32_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_64) + #define ARMBITREVINDEXTABLE_FIXED_64_TABLE_LENGTH ((uint16_t)56) + extern const uint16_t armBitRevIndexTable_fixed_64[ARMBITREVINDEXTABLE_FIXED_64_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_128) + #define ARMBITREVINDEXTABLE_FIXED_128_TABLE_LENGTH ((uint16_t)112) + extern const uint16_t armBitRevIndexTable_fixed_128[ARMBITREVINDEXTABLE_FIXED_128_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_256) + #define ARMBITREVINDEXTABLE_FIXED_256_TABLE_LENGTH ((uint16_t)240) + extern const uint16_t armBitRevIndexTable_fixed_256[ARMBITREVINDEXTABLE_FIXED_256_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_512) + #define ARMBITREVINDEXTABLE_FIXED_512_TABLE_LENGTH ((uint16_t)480) + extern const uint16_t armBitRevIndexTable_fixed_512[ARMBITREVINDEXTABLE_FIXED_512_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_1024) + #define ARMBITREVINDEXTABLE_FIXED_1024_TABLE_LENGTH ((uint16_t)992) + extern const uint16_t armBitRevIndexTable_fixed_1024[ARMBITREVINDEXTABLE_FIXED_1024_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_2048) + #define ARMBITREVINDEXTABLE_FIXED_2048_TABLE_LENGTH ((uint16_t)1984) + extern const uint16_t armBitRevIndexTable_fixed_2048[ARMBITREVINDEXTABLE_FIXED_2048_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_BITREVIDX_FXT_4096) + #define ARMBITREVINDEXTABLE_FIXED_4096_TABLE_LENGTH ((uint16_t)4032) + extern const uint16_t armBitRevIndexTable_fixed_4096[ARMBITREVINDEXTABLE_FIXED_4096_TABLE_LENGTH]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_REALCOEF_F32) + extern const float32_t realCoefA[8192]; + extern const float32_t realCoefB[8192]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_REALCOEF_Q31) + extern const q31_t realCoefAQ31[8192]; + extern const q31_t realCoefBQ31[8192]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_REALCOEF_Q15) + extern const q15_t realCoefAQ15[8192]; + extern const q15_t realCoefBQ15[8192]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_F32_128) + extern const float32_t Weights_128[256]; + extern const float32_t cos_factors_128[128]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_F32_512) + extern const float32_t Weights_512[1024]; + extern const float32_t cos_factors_512[512]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_F32_2048) + extern const float32_t Weights_2048[4096]; + extern const float32_t cos_factors_2048[2048]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_F32_8192) + extern const float32_t Weights_8192[16384]; + extern const float32_t cos_factors_8192[8192]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q15_128) + extern const q15_t WeightsQ15_128[256]; + extern const q15_t cos_factorsQ15_128[128]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q15_512) + extern const q15_t WeightsQ15_512[1024]; + extern const q15_t cos_factorsQ15_512[512]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q15_2048) + extern const q15_t WeightsQ15_2048[4096]; + extern const q15_t cos_factorsQ15_2048[2048]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q15_8192) + extern const q15_t WeightsQ15_8192[16384]; + extern const q15_t cos_factorsQ15_8192[8192]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q31_128) + extern const q31_t WeightsQ31_128[256]; + extern const q31_t cos_factorsQ31_128[128]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q31_512) + extern const q31_t WeightsQ31_512[1024]; + extern const q31_t cos_factorsQ31_512[512]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q31_2048) + extern const q31_t WeightsQ31_2048[4096]; + extern const q31_t cos_factorsQ31_2048[2048]; + #endif + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FFT_TABLES) || defined(ARM_TABLE_DCT4_Q31_8192) + extern const q31_t WeightsQ31_8192[16384]; + extern const q31_t cos_factorsQ31_8192[8192]; + #endif + +#endif /* if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_FFT_TABLES) */ + +#if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_FAST_ALLOW_TABLES) + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_RECIP_Q15) + extern const q15_t armRecipTableQ15[64]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) defined(ARM_ALL_FAST_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_RECIP_Q31) + extern const q31_t armRecipTableQ31[64]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) defined(ARM_ALL_FAST_TABLES) */ + + /* Tables for Fast Math Sine and Cosine */ + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_SIN_F32) + extern const float32_t sinTable_f32[FAST_MATH_TABLE_SIZE + 1]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) defined(ARM_ALL_FAST_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_SIN_Q31) + extern const q31_t sinTable_q31[FAST_MATH_TABLE_SIZE + 1]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) defined(ARM_ALL_FAST_TABLES) */ + + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_SIN_Q15) + extern const q15_t sinTable_q15[FAST_MATH_TABLE_SIZE + 1]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) defined(ARM_ALL_FAST_TABLES) */ + + #if defined(ARM_MATH_MVEI) + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_FAST_SQRT_Q31_MVE) + extern const q31_t sqrtTable_Q31[256]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) defined(ARM_ALL_FAST_TABLES) */ + #endif + + #if defined(ARM_MATH_MVEI) + #if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_FAST_SQRT_Q15_MVE) + extern const q15_t sqrtTable_Q15[256]; + #endif /* !defined(ARM_DSP_CONFIG_TABLES) defined(ARM_ALL_FAST_TABLES) */ + #endif + +#endif /* if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_FAST_TABLES) */ + +#if (defined(ARM_MATH_MVEF) || defined(ARM_MATH_HELIUM)) && !defined(ARM_MATH_AUTOVECTORIZE) + extern const float32_t exp_tab[8]; + extern const float32_t __logf_lut_f32[8]; +#endif /* (defined(ARM_MATH_MVEF) || defined(ARM_MATH_HELIUM)) && !defined(ARM_MATH_AUTOVECTORIZE) */ + +#if (defined(ARM_MATH_MVEI) || defined(ARM_MATH_HELIUM)) +extern const unsigned char hwLUT[256]; +#endif /* (defined(ARM_MATH_MVEI) || defined(ARM_MATH_HELIUM)) */ + +#ifdef __cplusplus +} +#endif + +#endif /* ARM_COMMON_TABLES_H */ + + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_helium_utils.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_helium_utils.h new file mode 100644 index 0000000..d80599b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_helium_utils.h @@ -0,0 +1,751 @@ +#if !defined(ESP32) +/* ---------------------------------------------------------------------- + * Project: CMSIS DSP Library + * Title: arm_helium_utils.h + * Description: Utility functions for Helium development + * + * $Date: 09. September 2019 + * $Revision: V.1.5.1 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ +/* + * Copyright (C) 2010-2019 ARM Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef _ARM_UTILS_HELIUM_H_ +#define _ARM_UTILS_HELIUM_H_ + + +#ifdef __cplusplus +extern "C" +{ +#endif +/*************************************** + +Definitions available for MVEF and MVEI + +***************************************/ +#if defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEF) || defined(ARM_MATH_MVEI) + +#define INACTIVELANE 0 /* inactive lane content */ + + +#endif /* defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEF) || defined(ARM_MATH_MVEI) */ + +/*************************************** + +Definitions available for MVEF only + +***************************************/ +#if defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEF) + +__STATIC_FORCEINLINE float32_t vecAddAcrossF32Mve(float32x4_t in) +{ + float32_t acc; + + acc = vgetq_lane(in, 0) + vgetq_lane(in, 1) + + vgetq_lane(in, 2) + vgetq_lane(in, 3); + + return acc; +} + +__STATIC_FORCEINLINE float16_t vecAddAcrossF16Mve(float16x8_t in) +{ + float16x8_t tmpVec; + _Float16 acc; + + tmpVec = (float16x8_t) vrev32q_s16((int16x8_t) in); + in = vaddq_f16(tmpVec, in); + tmpVec = (float16x8_t) vrev64q_s32((int32x4_t) in); + in = vaddq_f16(tmpVec, in); + acc = (_Float16)vgetq_lane_f16(in, 0) + (_Float16)vgetq_lane_f16(in, 4); + + return acc; +} + + +/* newton initial guess */ +#define INVSQRT_MAGIC_F32 0x5f3759df +#define INV_NEWTON_INIT_F32 0x7EF127EA + + +#define INVSQRT_NEWTON_MVE_F32(invSqrt, xHalf, xStart)\ +{ \ + float32x4_t tmp; \ + \ + /* tmp = xhalf * x * x */ \ + tmp = vmulq(xStart, xStart); \ + tmp = vmulq(tmp, xHalf); \ + /* (1.5f - xhalf * x * x) */ \ + tmp = vsubq(vdupq_n_f32(1.5f), tmp); \ + /* x = x*(1.5f-xhalf*x*x); */ \ + invSqrt = vmulq(tmp, xStart); \ +} +#endif /* defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEF) */ + + +/*************************************** + +Definitions available for f16 datatype with HW acceleration only + +***************************************/ +#if defined (ARM_MATH_MVE_FLOAT16) +__STATIC_FORCEINLINE float16x8_t __mve_cmplx_sum_intra_vec_f16( + float16x8_t vecIn) +{ + float16x8_t vecTmp, vecOut; + uint32_t tmp; + + vecTmp = (float16x8_t) vrev64q_s32((int32x4_t) vecIn); + // TO TRACK : using canonical addition leads to unefficient code generation for f16 + // vecTmp = vecTmp + vecAccCpx0; + /* + * Compute + * re0+re1 | im0+im1 | re0+re1 | im0+im1 + * re2+re3 | im2+im3 | re2+re3 | im2+im3 + */ + vecTmp = vaddq(vecTmp, vecIn); + vecOut = vecTmp; + /* + * shift left, random tmp insertion in bottom + */ + vecOut = vreinterpretq_f16_s32(vshlcq_s32(vreinterpretq_s32_f16(vecOut) , &tmp, 32)); + /* + * Compute: + * DONTCARE | DONTCARE | re0+re1+re0+re1 |im0+im1+im0+im1 + * re0+re1+re2+re3 | im0+im1+im2+im3 | re2+re3+re2+re3 |im2+im3+im2+im3 + */ + vecOut = vaddq(vecOut, vecTmp); + /* + * Cmplx sum is in 4rd & 5th f16 elt + * return full vector + */ + return vecOut; +} + + +#define mve_cmplx_sum_intra_r_i_f16(vec, Re, Im) \ +{ \ + float16x8_t vecOut = __mve_cmplx_sum_intra_vec_f16(vec); \ + Re = vgetq_lane(vecOut, 4); \ + Im = vgetq_lane(vecOut, 5); \ +} + +__STATIC_FORCEINLINE void mve_cmplx_sum_intra_vec_f16( + float16x8_t vecIn, + float16_t *pOut) +{ + float16x8_t vecOut = __mve_cmplx_sum_intra_vec_f16(vecIn); + /* + * Cmplx sum is in 4rd & 5th f16 elt + * use 32-bit extraction + */ + *(float32_t *) pOut = ((float32x4_t) vecOut)[2]; +} + + +#define INVSQRT_MAGIC_F16 0x59ba /* ( 0x1ba = 0x3759df >> 13) */ + +/* canonical version of INVSQRT_NEWTON_MVE_F16 leads to bad performance */ +#define INVSQRT_NEWTON_MVE_F16(invSqrt, xHalf, xStart) \ +{ \ + float16x8_t tmp; \ + \ + /* tmp = xhalf * x * x */ \ + tmp = vmulq(xStart, xStart); \ + tmp = vmulq(tmp, xHalf); \ + /* (1.5f - xhalf * x * x) */ \ + tmp = vsubq(vdupq_n_f16((float16_t)1.5), tmp); \ + /* x = x*(1.5f-xhalf*x*x); */ \ + invSqrt = vmulq(tmp, xStart); \ +} + +#endif + +/*************************************** + +Definitions available for MVEI and MVEF only + +***************************************/ +#if defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEF) || defined(ARM_MATH_MVEI) +/* Following functions are used to transpose matrix in f32 and q31 cases */ +__STATIC_INLINE arm_status arm_mat_trans_32bit_2x2_mve( + uint32_t * pDataSrc, + uint32_t * pDataDest) +{ + static const uint32x4_t vecOffs = { 0, 2, 1, 3 }; + /* + * + * | 0 1 | => | 0 2 | + * | 2 3 | | 1 3 | + * + */ + uint32x4_t vecIn = vldrwq_u32((uint32_t const *)pDataSrc); + vstrwq_scatter_shifted_offset_u32(pDataDest, vecOffs, vecIn); + + return (ARM_MATH_SUCCESS); +} + +__STATIC_INLINE arm_status arm_mat_trans_32bit_3x3_mve( + uint32_t * pDataSrc, + uint32_t * pDataDest) +{ + const uint32x4_t vecOffs1 = { 0, 3, 6, 1}; + const uint32x4_t vecOffs2 = { 4, 7, 2, 5}; + /* + * + * | 0 1 2 | | 0 3 6 | 4 x 32 flattened version | 0 3 6 1 | + * | 3 4 5 | => | 1 4 7 | => | 4 7 2 5 | + * | 6 7 8 | | 2 5 8 | (row major) | 8 . . . | + * + */ + uint32x4_t vecIn1 = vldrwq_u32((uint32_t const *) pDataSrc); + uint32x4_t vecIn2 = vldrwq_u32((uint32_t const *) &pDataSrc[4]); + + vstrwq_scatter_shifted_offset_u32(pDataDest, vecOffs1, vecIn1); + vstrwq_scatter_shifted_offset_u32(pDataDest, vecOffs2, vecIn2); + + pDataDest[8] = pDataSrc[8]; + + return (ARM_MATH_SUCCESS); +} + +__STATIC_INLINE arm_status arm_mat_trans_32bit_4x4_mve(uint32_t * pDataSrc, uint32_t * pDataDest) +{ + /* + * 4x4 Matrix transposition + * is 4 x de-interleave operation + * + * 0 1 2 3 0 4 8 12 + * 4 5 6 7 1 5 9 13 + * 8 9 10 11 2 6 10 14 + * 12 13 14 15 3 7 11 15 + */ + + uint32x4x4_t vecIn; + + vecIn = vld4q((uint32_t const *) pDataSrc); + vstrwq(pDataDest, vecIn.val[0]); + pDataDest += 4; + vstrwq(pDataDest, vecIn.val[1]); + pDataDest += 4; + vstrwq(pDataDest, vecIn.val[2]); + pDataDest += 4; + vstrwq(pDataDest, vecIn.val[3]); + + return (ARM_MATH_SUCCESS); +} + + +__STATIC_INLINE arm_status arm_mat_trans_32bit_generic_mve( + uint16_t srcRows, + uint16_t srcCols, + uint32_t * pDataSrc, + uint32_t * pDataDest) +{ + uint32x4_t vecOffs; + uint32_t i; + uint32_t blkCnt; + uint32_t const *pDataC; + uint32_t *pDataDestR; + uint32x4_t vecIn; + + vecOffs = vidupq_u32((uint32_t)0, 1); + vecOffs = vecOffs * srcCols; + + i = srcCols; + do + { + pDataC = (uint32_t const *) pDataSrc; + pDataDestR = pDataDest; + + blkCnt = srcRows >> 2; + while (blkCnt > 0U) + { + vecIn = vldrwq_gather_shifted_offset_u32(pDataC, vecOffs); + vstrwq(pDataDestR, vecIn); + pDataDestR += 4; + pDataC = pDataC + srcCols * 4; + /* + * Decrement the blockSize loop counter + */ + blkCnt--; + } + + /* + * tail + */ + blkCnt = srcRows & 3; + if (blkCnt > 0U) + { + mve_pred16_t p0 = vctp32q(blkCnt); + vecIn = vldrwq_gather_shifted_offset_u32(pDataC, vecOffs); + vstrwq_p(pDataDestR, vecIn, p0); + } + + pDataSrc += 1; + pDataDest += srcRows; + } + while (--i); + + return (ARM_MATH_SUCCESS); +} + +__STATIC_INLINE arm_status arm_mat_cmplx_trans_32bit( + uint16_t srcRows, + uint16_t srcCols, + uint32_t *pDataSrc, + uint16_t dstRows, + uint16_t dstCols, + uint32_t *pDataDest) +{ + uint32_t i; + uint32_t const *pDataC; + uint32_t *pDataRow; + uint32_t *pDataDestR, *pDataDestRow; + uint32x4_t vecOffsRef, vecOffsCur; + uint32_t blkCnt; + uint32x4_t vecIn; + +#ifdef ARM_MATH_MATRIX_CHECK + /* + * Check for matrix mismatch condition + */ + if ((srcRows != dstCols) || (srcCols != dstRows)) + { + /* + * Set status as ARM_MATH_SIZE_MISMATCH + */ + return = ARM_MATH_SIZE_MISMATCH; + } +#else + (void)dstRows; + (void)dstCols; +#endif + + /* 2x2, 3x3 and 4x4 specialization to be added */ + + vecOffsRef[0] = 0; + vecOffsRef[1] = 1; + vecOffsRef[2] = srcCols << 1; + vecOffsRef[3] = (srcCols << 1) + 1; + + pDataRow = pDataSrc; + pDataDestRow = pDataDest; + i = srcCols; + do + { + pDataC = (uint32_t const *) pDataRow; + pDataDestR = pDataDestRow; + vecOffsCur = vecOffsRef; + + blkCnt = (srcRows * CMPLX_DIM) >> 2; + while (blkCnt > 0U) + { + vecIn = vldrwq_gather_shifted_offset(pDataC, vecOffsCur); + vstrwq(pDataDestR, vecIn); + pDataDestR += 4; + vecOffsCur = vaddq(vecOffsCur, (srcCols << 2)); + /* + * Decrement the blockSize loop counter + */ + blkCnt--; + } + /* + * tail + * (will be merged thru tail predication) + */ + blkCnt = (srcRows * CMPLX_DIM) & 3; + if (blkCnt > 0U) + { + mve_pred16_t p0 = vctp32q(blkCnt); + vecIn = vldrwq_gather_shifted_offset(pDataC, vecOffsCur); + vstrwq_p(pDataDestR, vecIn, p0); + } + + pDataRow += CMPLX_DIM; + pDataDestRow += (srcRows * CMPLX_DIM); + } + while (--i); + + return (ARM_MATH_SUCCESS); +} + +__STATIC_INLINE arm_status arm_mat_trans_16bit_2x2(uint16_t * pDataSrc, uint16_t * pDataDest) +{ + pDataDest[0] = pDataSrc[0]; + pDataDest[3] = pDataSrc[3]; + pDataDest[2] = pDataSrc[1]; + pDataDest[1] = pDataSrc[2]; + + return (ARM_MATH_SUCCESS); +} + +__STATIC_INLINE arm_status arm_mat_trans_16bit_3x3_mve(uint16_t * pDataSrc, uint16_t * pDataDest) +{ + static const uint16_t stridesTr33[8] = { 0, 3, 6, 1, 4, 7, 2, 5 }; + uint16x8_t vecOffs1; + uint16x8_t vecIn1; + /* + * + * | 0 1 2 | | 0 3 6 | 8 x 16 flattened version | 0 3 6 1 4 7 2 5 | + * | 3 4 5 | => | 1 4 7 | => | 8 . . . . . . . | + * | 6 7 8 | | 2 5 8 | (row major) + * + */ + vecOffs1 = vldrhq_u16((uint16_t const *) stridesTr33); + vecIn1 = vldrhq_u16((uint16_t const *) pDataSrc); + + vstrhq_scatter_shifted_offset_u16(pDataDest, vecOffs1, vecIn1); + + pDataDest[8] = pDataSrc[8]; + + return (ARM_MATH_SUCCESS); +} + + +__STATIC_INLINE arm_status arm_mat_trans_16bit_4x4_mve(uint16_t * pDataSrc, uint16_t * pDataDest) +{ + static const uint16_t stridesTr44_1[8] = { 0, 4, 8, 12, 1, 5, 9, 13 }; + static const uint16_t stridesTr44_2[8] = { 2, 6, 10, 14, 3, 7, 11, 15 }; + uint16x8_t vecOffs1, vecOffs2; + uint16x8_t vecIn1, vecIn2; + uint16_t const * pDataSrcVec = (uint16_t const *) pDataSrc; + + /* + * 4x4 Matrix transposition + * + * | 0 1 2 3 | | 0 4 8 12 | 8 x 16 flattened version + * | 4 5 6 7 | => | 1 5 9 13 | => [0 4 8 12 1 5 9 13] + * | 8 9 10 11 | | 2 6 10 14 | [2 6 10 14 3 7 11 15] + * | 12 13 14 15 | | 3 7 11 15 | + */ + + vecOffs1 = vldrhq_u16((uint16_t const *) stridesTr44_1); + vecOffs2 = vldrhq_u16((uint16_t const *) stridesTr44_2); + vecIn1 = vldrhq_u16(pDataSrcVec); + pDataSrcVec += 8; + vecIn2 = vldrhq_u16(pDataSrcVec); + + vstrhq_scatter_shifted_offset_u16(pDataDest, vecOffs1, vecIn1); + vstrhq_scatter_shifted_offset_u16(pDataDest, vecOffs2, vecIn2); + + + return (ARM_MATH_SUCCESS); +} + + + +__STATIC_INLINE arm_status arm_mat_trans_16bit_generic( + uint16_t srcRows, + uint16_t srcCols, + uint16_t * pDataSrc, + uint16_t * pDataDest) +{ + uint16x8_t vecOffs; + uint32_t i; + uint32_t blkCnt; + uint16_t const *pDataC; + uint16_t *pDataDestR; + uint16x8_t vecIn; + + vecOffs = vidupq_u16((uint32_t)0, 1); + vecOffs = vecOffs * srcCols; + + i = srcCols; + while(i > 0U) + { + pDataC = (uint16_t const *) pDataSrc; + pDataDestR = pDataDest; + + blkCnt = srcRows >> 3; + while (blkCnt > 0U) + { + vecIn = vldrhq_gather_shifted_offset_u16(pDataC, vecOffs); + vstrhq_u16(pDataDestR, vecIn); + pDataDestR += 8; + pDataC = pDataC + srcCols * 8; + /* + * Decrement the blockSize loop counter + */ + blkCnt--; + } + + /* + * tail + */ + blkCnt = srcRows & 7; + if (blkCnt > 0U) + { + mve_pred16_t p0 = vctp16q(blkCnt); + vecIn = vldrhq_gather_shifted_offset_u16(pDataC, vecOffs); + vstrhq_p_u16(pDataDestR, vecIn, p0); + } + pDataSrc += 1; + pDataDest += srcRows; + i--; + } + + return (ARM_MATH_SUCCESS); +} + + +__STATIC_INLINE arm_status arm_mat_cmplx_trans_16bit( + uint16_t srcRows, + uint16_t srcCols, + uint16_t *pDataSrc, + uint16_t dstRows, + uint16_t dstCols, + uint16_t *pDataDest) +{ + static const uint16_t loadCmplxCol[8] = { 0, 0, 1, 1, 2, 2, 3, 3 }; + int i; + uint16x8_t vecOffsRef, vecOffsCur; + uint16_t const *pDataC; + uint16_t *pDataRow; + uint16_t *pDataDestR, *pDataDestRow; + uint32_t blkCnt; + uint16x8_t vecIn; + +#ifdef ARM_MATH_MATRIX_CHECK + /* + * Check for matrix mismatch condition + */ + if ((srcRows != dstCols) || (srcCols != dstRows)) + { + /* + * Set status as ARM_MATH_SIZE_MISMATCH + */ + return = ARM_MATH_SIZE_MISMATCH; + } +#else + (void)dstRows; + (void)dstCols; +#endif + + /* + * 2x2, 3x3 and 4x4 specialization to be added + */ + + + /* + * build [0, 1, 2xcol, 2xcol+1, 4xcol, 4xcol+1, 6xcol, 6xcol+1] + */ + vecOffsRef = vldrhq_u16((uint16_t const *) loadCmplxCol); + vecOffsRef = vmulq(vecOffsRef, (uint16_t) (srcCols * CMPLX_DIM)) + + viwdupq_u16((uint32_t)0, (uint16_t) 2, 1); + + pDataRow = pDataSrc; + pDataDestRow = pDataDest; + i = srcCols; + do + { + pDataC = (uint16_t const *) pDataRow; + pDataDestR = pDataDestRow; + vecOffsCur = vecOffsRef; + + blkCnt = (srcRows * CMPLX_DIM) >> 3; + while (blkCnt > 0U) + { + vecIn = vldrhq_gather_shifted_offset(pDataC, vecOffsCur); + vstrhq(pDataDestR, vecIn); + pDataDestR+= 8; // VEC_LANES_U16 + vecOffsCur = vaddq(vecOffsCur, (srcCols << 3)); + /* + * Decrement the blockSize loop counter + */ + blkCnt--; + } + /* + * tail + * (will be merged thru tail predication) + */ + blkCnt = (srcRows * CMPLX_DIM) & 0x7; + if (blkCnt > 0U) + { + mve_pred16_t p0 = vctp16q(blkCnt); + vecIn = vldrhq_gather_shifted_offset(pDataC, vecOffsCur); + vstrhq_p(pDataDestR, vecIn, p0); + } + + pDataRow += CMPLX_DIM; + pDataDestRow += (srcRows * CMPLX_DIM); + } + while (--i); + + return (ARM_MATH_SUCCESS); +} +#endif /* MVEF and MVEI */ + +/*************************************** + +Definitions available for MVEI only + +***************************************/ +#if defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEI) + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h" + +#define MVE_ASRL_SAT16(acc, shift) ((sqrshrl_sat48(acc, -(32-shift)) >> 32) & 0xffffffff) +#define MVE_ASRL_SAT32(acc, shift) ((sqrshrl(acc, -(32-shift)) >> 32) & 0xffffffff) + + +#if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_FAST_SQRT_Q31_MVE) +__STATIC_INLINE q31x4_t FAST_VSQRT_Q31(q31x4_t vecIn) +{ + q63x2_t vecTmpLL; + q31x4_t vecTmp0, vecTmp1; + q31_t scale; + q63_t tmp64; + q31x4_t vecNrm, vecDst, vecIdx, vecSignBits; + + + vecSignBits = vclsq(vecIn); + vecSignBits = vbicq(vecSignBits, 1); + /* + * in = in << no_of_sign_bits; + */ + vecNrm = vshlq(vecIn, vecSignBits); + /* + * index = in >> 24; + */ + vecIdx = vecNrm >> 24; + vecIdx = vecIdx << 1; + + vecTmp0 = vldrwq_gather_shifted_offset_s32(sqrtTable_Q31, (uint32x4_t)vecIdx); + + vecIdx = vecIdx + 1; + + vecTmp1 = vldrwq_gather_shifted_offset_s32(sqrtTable_Q31, (uint32x4_t)vecIdx); + + vecTmp1 = vqrdmulhq(vecTmp1, vecNrm); + vecTmp0 = vecTmp0 - vecTmp1; + vecTmp1 = vqrdmulhq(vecTmp0, vecTmp0); + vecTmp1 = vqrdmulhq(vecNrm, vecTmp1); + vecTmp1 = vdupq_n_s32(0x18000000) - vecTmp1; + vecTmp0 = vqrdmulhq(vecTmp0, vecTmp1); + vecTmpLL = vmullbq_int(vecNrm, vecTmp0); + + /* + * scale elements 0, 2 + */ + scale = 26 + (vecSignBits[0] >> 1); + tmp64 = asrl(vecTmpLL[0], scale); + vecDst[0] = (q31_t) tmp64; + + scale = 26 + (vecSignBits[2] >> 1); + tmp64 = asrl(vecTmpLL[1], scale); + vecDst[2] = (q31_t) tmp64; + + vecTmpLL = vmulltq_int(vecNrm, vecTmp0); + + /* + * scale elements 1, 3 + */ + scale = 26 + (vecSignBits[1] >> 1); + tmp64 = asrl(vecTmpLL[0], scale); + vecDst[1] = (q31_t) tmp64; + + scale = 26 + (vecSignBits[3] >> 1); + tmp64 = asrl(vecTmpLL[1], scale); + vecDst[3] = (q31_t) tmp64; + /* + * set negative values to 0 + */ + vecDst = vdupq_m(vecDst, 0, vcmpltq_n_s32(vecIn, 0)); + + return vecDst; +} +#endif + +#if !defined(ARM_DSP_CONFIG_TABLES) || defined(ARM_ALL_FAST_TABLES) || defined(ARM_TABLE_FAST_SQRT_Q15_MVE) +__STATIC_INLINE q15x8_t FAST_VSQRT_Q15(q15x8_t vecIn) +{ + q31x4_t vecTmpLev, vecTmpLodd, vecSignL; + q15x8_t vecTmp0, vecTmp1; + q15x8_t vecNrm, vecDst, vecIdx, vecSignBits; + + vecDst = vuninitializedq_s16(); + + vecSignBits = vclsq(vecIn); + vecSignBits = vbicq(vecSignBits, 1); + /* + * in = in << no_of_sign_bits; + */ + vecNrm = vshlq(vecIn, vecSignBits); + + vecIdx = vecNrm >> 8; + vecIdx = vecIdx << 1; + + vecTmp0 = vldrhq_gather_shifted_offset_s16(sqrtTable_Q15, (uint16x8_t)vecIdx); + + vecIdx = vecIdx + 1; + + vecTmp1 = vldrhq_gather_shifted_offset_s16(sqrtTable_Q15, (uint16x8_t)vecIdx); + + vecTmp1 = vqrdmulhq(vecTmp1, vecNrm); + vecTmp0 = vecTmp0 - vecTmp1; + vecTmp1 = vqrdmulhq(vecTmp0, vecTmp0); + vecTmp1 = vqrdmulhq(vecNrm, vecTmp1); + vecTmp1 = vdupq_n_s16(0x1800) - vecTmp1; + vecTmp0 = vqrdmulhq(vecTmp0, vecTmp1); + + vecSignBits = vecSignBits >> 1; + + vecTmpLev = vmullbq_int(vecNrm, vecTmp0); + vecTmpLodd = vmulltq_int(vecNrm, vecTmp0); + + vecTmp0 = vecSignBits + 10; + /* + * negate sign to apply register based vshl + */ + vecTmp0 = -vecTmp0; + + /* + * shift even elements + */ + vecSignL = vmovlbq(vecTmp0); + vecTmpLev = vshlq(vecTmpLev, vecSignL); + /* + * shift odd elements + */ + vecSignL = vmovltq(vecTmp0); + vecTmpLodd = vshlq(vecTmpLodd, vecSignL); + /* + * merge and narrow odd and even parts + */ + vecDst = vmovnbq_s32(vecDst, vecTmpLev); + vecDst = vmovntq_s32(vecDst, vecTmpLodd); + /* + * set negative values to 0 + */ + vecDst = vdupq_m(vecDst, 0, vcmpltq_n_s16(vecIn, 0)); + + return vecDst; +} +#endif + +#endif /* defined (ARM_MATH_HELIUM) || defined(ARM_MATH_MVEI) */ + +#ifdef __cplusplus +} +#endif + +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h new file mode 100644 index 0000000..cba217f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h @@ -0,0 +1,249 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file arm_math.h + * @brief Public header file for CMSIS DSP Library + * @version V1.7.0 + * @date 18. March 2019 + ******************************************************************************/ +/* + * Copyright (c) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/** + \mainpage CMSIS DSP Software Library + * + * \section intro Introduction + * + * This user manual describes the CMSIS DSP software library, + * a suite of common signal processing functions for use on Cortex-M and Cortex-A processor + * based devices. + * + * The library is divided into a number of functions each covering a specific category: + * - Basic math functions + * - Fast math functions + * - Complex math functions + * - Filtering functions + * - Matrix functions + * - Transform functions + * - Motor control functions + * - Statistical functions + * - Support functions + * - Interpolation functions + * - Support Vector Machine functions (SVM) + * - Bayes classifier functions + * - Distance functions + * + * The library has generally separate functions for operating on 8-bit integers, 16-bit integers, + * 32-bit integer and 32-bit floating-point values. + * + * \section using Using the Library + * + * The library installer contains prebuilt versions of the libraries in the Lib folder. + * + * Here is the list of pre-built libraries : + * - arm_cortexM7lfdp_math.lib (Cortex-M7, Little endian, Double Precision Floating Point Unit) + * - arm_cortexM7bfdp_math.lib (Cortex-M7, Big endian, Double Precision Floating Point Unit) + * - arm_cortexM7lfsp_math.lib (Cortex-M7, Little endian, Single Precision Floating Point Unit) + * - arm_cortexM7bfsp_math.lib (Cortex-M7, Big endian and Single Precision Floating Point Unit on) + * - arm_cortexM7l_math.lib (Cortex-M7, Little endian) + * - arm_cortexM7b_math.lib (Cortex-M7, Big endian) + * - arm_cortexM4lf_math.lib (Cortex-M4, Little endian, Floating Point Unit) + * - arm_cortexM4bf_math.lib (Cortex-M4, Big endian, Floating Point Unit) + * - arm_cortexM4l_math.lib (Cortex-M4, Little endian) + * - arm_cortexM4b_math.lib (Cortex-M4, Big endian) + * - arm_cortexM3l_math.lib (Cortex-M3, Little endian) + * - arm_cortexM3b_math.lib (Cortex-M3, Big endian) + * - arm_cortexM0l_math.lib (Cortex-M0 / Cortex-M0+, Little endian) + * - arm_cortexM0b_math.lib (Cortex-M0 / Cortex-M0+, Big endian) + * - arm_ARMv8MBLl_math.lib (Armv8-M Baseline, Little endian) + * - arm_ARMv8MMLl_math.lib (Armv8-M Mainline, Little endian) + * - arm_ARMv8MMLlfsp_math.lib (Armv8-M Mainline, Little endian, Single Precision Floating Point Unit) + * - arm_ARMv8MMLld_math.lib (Armv8-M Mainline, Little endian, DSP instructions) + * - arm_ARMv8MMLldfsp_math.lib (Armv8-M Mainline, Little endian, DSP instructions, Single Precision Floating Point Unit) + * + * The library functions are declared in the public file arm_math.h which is placed in the Include folder. + * Simply include this file and link the appropriate library in the application and begin calling the library functions. The Library supports single + * public header file arm_math.h for Cortex-M cores with little endian and big endian. Same header file will be used for floating point unit(FPU) variants. + * + * + * \section example Examples + * + * The library ships with a number of examples which demonstrate how to use the library functions. + * + * \section toolchain Toolchain Support + * + * The library is now tested on Fast Models building with cmake. + * Core M0, M7, A5 are tested. + * + * + * + * \section building Building the Library + * + * The library installer contains a project file to rebuild libraries on MDK toolchain in the CMSIS\\DSP\\Projects\\ARM folder. + * - arm_cortexM_math.uvprojx + * + * + * The libraries can be built by opening the arm_cortexM_math.uvprojx project in MDK-ARM, selecting a specific target, and defining the optional preprocessor macros detailed above. + * + * There is also a work in progress cmake build. The README file is giving more details. + * + * \section preprocessor Preprocessor Macros + * + * Each library project have different preprocessor macros. + * + * - ARM_MATH_BIG_ENDIAN: + * + * Define macro ARM_MATH_BIG_ENDIAN to build the library for big endian targets. By default library builds for little endian targets. + * + * - ARM_MATH_MATRIX_CHECK: + * + * Define macro ARM_MATH_MATRIX_CHECK for checking on the input and output sizes of matrices + * + * - ARM_MATH_ROUNDING: + * + * Define macro ARM_MATH_ROUNDING for rounding on support functions + * + * - ARM_MATH_LOOPUNROLL: + * + * Define macro ARM_MATH_LOOPUNROLL to enable manual loop unrolling in DSP functions + * + * - ARM_MATH_NEON: + * + * Define macro ARM_MATH_NEON to enable Neon versions of the DSP functions. + * It is not enabled by default when Neon is available because performances are + * dependent on the compiler and target architecture. + * + * - ARM_MATH_NEON_EXPERIMENTAL: + * + * Define macro ARM_MATH_NEON_EXPERIMENTAL to enable experimental Neon versions of + * of some DSP functions. Experimental Neon versions currently do not have better + * performances than the scalar versions. + * + * - ARM_MATH_HELIUM: + * + * It implies the flags ARM_MATH_MVEF and ARM_MATH_MVEI and ARM_MATH_FLOAT16. + * + * - ARM_MATH_MVEF: + * + * Select Helium versions of the f32 algorithms. + * It implies ARM_MATH_FLOAT16 and ARM_MATH_MVEI. + * + * - ARM_MATH_MVEI: + * + * Select Helium versions of the int and fixed point algorithms. + * + * - ARM_MATH_MVE_FLOAT16: + * + * MVE Float16 implementations of some algorithms (Requires MVE extension). + * + * - DISABLEFLOAT16: + * + * Disable float16 algorithms when __fp16 is not supported for a + * specific compiler / core configuration + * + *
+ * \section pack CMSIS-DSP in ARM::CMSIS Pack + * + * The following files relevant to CMSIS-DSP are present in the ARM::CMSIS Pack directories: + * |File/Folder |Content | + * |---------------------------------|------------------------------------------------------------------------| + * |\b CMSIS\\Documentation\\DSP | This documentation | + * |\b CMSIS\\DSP\\DSP_Lib_TestSuite | DSP_Lib deprecated test suite | + * |\b CMSIS\\DSP\\Examples | Example projects demonstrating the usage of the library functions | + * |\b CMSIS\\DSP\\Include | DSP_Lib include files for using and building the lib + * |\b CMSIS\\DSP\\PrivateInclude | DSP_Lib private include files for building the lib | + * |\b CMSIS\\DSP\\Lib | DSP_Lib binaries | + * |\b CMSIS\\DSP\\Projects | Projects to rebuild DSP_Lib binaries | + * |\b CMSIS\\DSP\\Source | DSP_Lib source files | + * + *
+ * \section rev Revision History of CMSIS-DSP + * Please refer to \ref ChangeLog_pg. + */ + + + + + + + + + + + +/** + * @defgroup groupExamples Examples + */ + + + + + +#ifndef _ARM_MATH_H +#define _ARM_MATH_H + + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/interpolation_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/bayes_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/matrix_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/complex_math_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/statistics_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/controller_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/support_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/distance_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/transform_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/filtering_functions.h" + + + +#ifdef __cplusplus +extern "C" +{ +#endif + + + + +//#define TABLE_SPACING_Q31 0x400000 +//#define TABLE_SPACING_Q15 0x80 + + + + + +#ifdef __cplusplus +} +#endif + + +#endif /* _ARM_MATH_H */ + +/** + * + * End of file. + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h new file mode 100644 index 0000000..ea6ec5f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h @@ -0,0 +1,243 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file arm_math_memory.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef _ARM_MATH_MEMORY_H_ + +#define _ARM_MATH_MEMORY_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" + + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + @brief definition to read/write two 16 bit values. + @deprecated + */ +#if defined ( __CC_ARM ) + #define __SIMD32_TYPE int32_t __packed +#elif defined ( __ARMCC_VERSION ) && ( __ARMCC_VERSION >= 6010050 ) + #define __SIMD32_TYPE int32_t +#elif defined ( __GNUC__ ) + #define __SIMD32_TYPE int32_t +#elif defined ( __ICCARM__ ) + #define __SIMD32_TYPE int32_t __packed +#elif defined ( __TI_ARM__ ) + #define __SIMD32_TYPE int32_t +#elif defined ( __CSMC__ ) + #define __SIMD32_TYPE int32_t +#elif defined ( __TASKING__ ) + #define __SIMD32_TYPE __un(aligned) int32_t +#elif defined(_MSC_VER ) + #define __SIMD32_TYPE int32_t +#else + #error Unknown compiler +#endif + +#define __SIMD32(addr) (*(__SIMD32_TYPE **) & (addr)) +#define __SIMD32_CONST(addr) ( (__SIMD32_TYPE * ) (addr)) +#define _SIMD32_OFFSET(addr) (*(__SIMD32_TYPE * ) (addr)) +#define __SIMD64(addr) (*( int64_t **) & (addr)) + + +/* SIMD replacement */ + + +/** + @brief Read 2 Q15 from Q15 pointer. + @param[in] pQ15 points to input value + @return Q31 value + */ +__STATIC_FORCEINLINE q31_t read_q15x2 ( + q15_t * pQ15) +{ + q31_t val; + +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (&val, pQ15, 4); +#else + val = (pQ15[1] << 16) | (pQ15[0] & 0x0FFFF) ; +#endif + + return (val); +} + +/** + @brief Read 2 Q15 from Q15 pointer and increment pointer afterwards. + @param[in] pQ15 points to input value + @return Q31 value + */ +__STATIC_FORCEINLINE q31_t read_q15x2_ia ( + q15_t ** pQ15) +{ + q31_t val; + +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (&val, *pQ15, 4); +#else + val = ((*pQ15)[1] << 16) | ((*pQ15)[0] & 0x0FFFF); +#endif + + *pQ15 += 2; + return (val); +} + +/** + @brief Read 2 Q15 from Q15 pointer and decrement pointer afterwards. + @param[in] pQ15 points to input value + @return Q31 value + */ +__STATIC_FORCEINLINE q31_t read_q15x2_da ( + q15_t ** pQ15) +{ + q31_t val; + +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (&val, *pQ15, 4); +#else + val = ((*pQ15)[1] << 16) | ((*pQ15)[0] & 0x0FFFF); +#endif + + *pQ15 -= 2; + return (val); +} + +/** + @brief Write 2 Q15 to Q15 pointer and increment pointer afterwards. + @param[in] pQ15 points to input value + @param[in] value Q31 value + @return none + */ +__STATIC_FORCEINLINE void write_q15x2_ia ( + q15_t ** pQ15, + q31_t value) +{ + q31_t val = value; +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (*pQ15, &val, 4); +#else + (*pQ15)[0] = (val & 0x0FFFF); + (*pQ15)[1] = (val >> 16) & 0x0FFFF; +#endif + + *pQ15 += 2; +} + +/** + @brief Write 2 Q15 to Q15 pointer. + @param[in] pQ15 points to input value + @param[in] value Q31 value + @return none + */ +__STATIC_FORCEINLINE void write_q15x2 ( + q15_t * pQ15, + q31_t value) +{ + q31_t val = value; + +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (pQ15, &val, 4); +#else + pQ15[0] = val & 0x0FFFF; + pQ15[1] = val >> 16; +#endif +} + + +/** + @brief Read 4 Q7 from Q7 pointer and increment pointer afterwards. + @param[in] pQ7 points to input value + @return Q31 value + */ +__STATIC_FORCEINLINE q31_t read_q7x4_ia ( + q7_t ** pQ7) +{ + q31_t val; + + +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (&val, *pQ7, 4); +#else + val =(((*pQ7)[3] & 0x0FF) << 24) | (((*pQ7)[2] & 0x0FF) << 16) | (((*pQ7)[1] & 0x0FF) << 8) | ((*pQ7)[0] & 0x0FF); +#endif + + *pQ7 += 4; + + return (val); +} + +/** + @brief Read 4 Q7 from Q7 pointer and decrement pointer afterwards. + @param[in] pQ7 points to input value + @return Q31 value + */ +__STATIC_FORCEINLINE q31_t read_q7x4_da ( + q7_t ** pQ7) +{ + q31_t val; +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (&val, *pQ7, 4); +#else + val = ((((*pQ7)[3]) & 0x0FF) << 24) | ((((*pQ7)[2]) & 0x0FF) << 16) | ((((*pQ7)[1]) & 0x0FF) << 8) | ((*pQ7)[0] & 0x0FF); +#endif + *pQ7 -= 4; + + return (val); +} + +/** + @brief Write 4 Q7 to Q7 pointer and increment pointer afterwards. + @param[in] pQ7 points to input value + @param[in] value Q31 value + @return none + */ +__STATIC_FORCEINLINE void write_q7x4_ia ( + q7_t ** pQ7, + q31_t value) +{ + q31_t val = value; +#ifdef __ARM_FEATURE_UNALIGNED + memcpy (*pQ7, &val, 4); +#else + (*pQ7)[0] = val & 0x0FF; + (*pQ7)[1] = (val >> 8) & 0x0FF; + (*pQ7)[2] = (val >> 16) & 0x0FF; + (*pQ7)[3] = (val >> 24) & 0x0FF; + +#endif + *pQ7 += 4; +} + + +#ifdef __cplusplus +} +#endif + +#endif /*ifndef _ARM_MATH_MEMORY_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h new file mode 100644 index 0000000..94946e9 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h @@ -0,0 +1,601 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file arm_math_types.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef _ARM_MATH_TYPES_H_ + +#define _ARM_MATH_TYPES_H_ + +#ifdef __cplusplus +extern "C" +{ +#endif + +/* Compiler specific diagnostic adjustment */ +#if defined ( __CC_ARM ) + +#elif defined ( __ARMCC_VERSION ) && ( __ARMCC_VERSION >= 6010050 ) + +#elif defined ( __GNUC__ ) + #pragma GCC diagnostic push + #pragma GCC diagnostic ignored "-Wsign-conversion" + #pragma GCC diagnostic ignored "-Wconversion" + #pragma GCC diagnostic ignored "-Wunused-parameter" + +#elif defined ( __ICCARM__ ) + +#elif defined ( __TI_ARM__ ) + +#elif defined ( __CSMC__ ) + +#elif defined ( __TASKING__ ) + +#elif defined ( _MSC_VER ) + +#else + #error Unknown compiler +#endif + + +/* Included for instrinsics definitions */ +#if defined (_MSC_VER ) +#include +#define __STATIC_FORCEINLINE static __forceinline +#define __STATIC_INLINE static __inline +#define __ALIGNED(x) __declspec(align(x)) + +#elif defined (__GNUC_PYTHON__) +#include +#define __ALIGNED(x) __attribute__((aligned(x))) +#define __STATIC_FORCEINLINE static __attribute__((inline)) +#define __STATIC_INLINE static __attribute__((inline)) +#pragma GCC diagnostic ignored "-Wunused-function" +#pragma GCC diagnostic ignored "-Wattributes" + +#else +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/Core/Include/cmsis_compiler.h" +#endif + + + +#include +#include +#include +#include + +/* evaluate ARM DSP feature */ +#if (defined (__ARM_FEATURE_DSP) && (__ARM_FEATURE_DSP == 1)) + #define ARM_MATH_DSP 1 +#endif + +#if defined(ARM_MATH_NEON) +#include +#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + #if !defined(ARM_MATH_NEON_FLOAT16) + #define ARM_MATH_NEON_FLOAT16 + #endif +#endif +#endif + +#if !defined(ARM_MATH_AUTOVECTORIZE) + +#if __ARM_FEATURE_MVE + #if !defined(ARM_MATH_MVEI) + #define ARM_MATH_MVEI + #endif +#endif + +#if (__ARM_FEATURE_MVE & 2) + #if !defined(ARM_MATH_MVEF) + #define ARM_MATH_MVEF + #endif + #if !defined(ARM_MATH_MVE_FLOAT16) + /* HW Float16 not yet well supported on gcc for M55 */ + #if !defined(__CMSIS_GCC_H) + #define ARM_MATH_MVE_FLOAT16 + #endif + #endif +#endif + +#endif /*!defined(ARM_MATH_AUTOVECTORIZE)*/ + + +#if defined (ARM_MATH_HELIUM) + #if !defined(ARM_MATH_MVEF) + #define ARM_MATH_MVEF + #endif + + #if !defined(ARM_MATH_MVEI) + #define ARM_MATH_MVEI + #endif + + #if !defined(ARM_MATH_MVE_FLOAT16) + /* HW Float16 not yet well supported on gcc for M55 */ + #if !defined(__CMSIS_GCC_H) + #define ARM_MATH_MVE_FLOAT16 + #endif + #endif +#endif + + + +#if defined ( __CC_ARM ) + /* Enter low optimization region - place directly above function definition */ + #if defined( __ARM_ARCH_7EM__ ) + #define LOW_OPTIMIZATION_ENTER \ + _Pragma ("push") \ + _Pragma ("O1") + #else + #define LOW_OPTIMIZATION_ENTER + #endif + + /* Exit low optimization region - place directly after end of function definition */ + #if defined ( __ARM_ARCH_7EM__ ) + #define LOW_OPTIMIZATION_EXIT \ + _Pragma ("pop") + #else + #define LOW_OPTIMIZATION_EXIT + #endif + + /* Enter low optimization region - place directly above function definition */ + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + + /* Exit low optimization region - place directly after end of function definition */ + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT + +#elif defined (__ARMCC_VERSION ) && ( __ARMCC_VERSION >= 6010050 ) + #define LOW_OPTIMIZATION_ENTER + #define LOW_OPTIMIZATION_EXIT + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT + +#elif defined ( __GNUC__ ) + #define LOW_OPTIMIZATION_ENTER \ + __attribute__(( optimize("-O1") )) + #define LOW_OPTIMIZATION_EXIT + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT + +#elif defined ( __ICCARM__ ) + /* Enter low optimization region - place directly above function definition */ + #if defined ( __ARM_ARCH_7EM__ ) + #define LOW_OPTIMIZATION_ENTER \ + _Pragma ("optimize=low") + #else + #define LOW_OPTIMIZATION_ENTER + #endif + + /* Exit low optimization region - place directly after end of function definition */ + #define LOW_OPTIMIZATION_EXIT + + /* Enter low optimization region - place directly above function definition */ + #if defined ( __ARM_ARCH_7EM__ ) + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER \ + _Pragma ("optimize=low") + #else + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + #endif + + /* Exit low optimization region - place directly after end of function definition */ + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT + +#elif defined ( __TI_ARM__ ) + #define LOW_OPTIMIZATION_ENTER + #define LOW_OPTIMIZATION_EXIT + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT + +#elif defined ( __CSMC__ ) + #define LOW_OPTIMIZATION_ENTER + #define LOW_OPTIMIZATION_EXIT + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT + +#elif defined ( __TASKING__ ) + #define LOW_OPTIMIZATION_ENTER + #define LOW_OPTIMIZATION_EXIT + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT + +#elif defined ( _MSC_VER ) || defined(__GNUC_PYTHON__) + #define LOW_OPTIMIZATION_ENTER + #define LOW_OPTIMIZATION_EXIT + #define IAR_ONLY_LOW_OPTIMIZATION_ENTER + #define IAR_ONLY_LOW_OPTIMIZATION_EXIT +#endif + + + +/* Compiler specific diagnostic adjustment */ +#if defined ( __CC_ARM ) + +#elif defined ( __ARMCC_VERSION ) && ( __ARMCC_VERSION >= 6010050 ) + +#elif defined ( __GNUC__ ) +#pragma GCC diagnostic pop + +#elif defined ( __ICCARM__ ) + +#elif defined ( __TI_ARM__ ) + +#elif defined ( __CSMC__ ) + +#elif defined ( __TASKING__ ) + +#elif defined ( _MSC_VER ) + +#else + #error Unknown compiler +#endif + +#ifdef __cplusplus +} +#endif + +#if __ARM_FEATURE_MVE +#include +#endif + +#ifdef __cplusplus +extern "C" +{ +#endif + + /** + * @brief 8-bit fractional data type in 1.7 format. + */ + typedef int8_t q7_t; + + /** + * @brief 16-bit fractional data type in 1.15 format. + */ + typedef int16_t q15_t; + + /** + * @brief 32-bit fractional data type in 1.31 format. + */ + typedef int32_t q31_t; + + /** + * @brief 64-bit fractional data type in 1.63 format. + */ + typedef int64_t q63_t; + + /** + * @brief 32-bit floating-point type definition. + */ + typedef float float32_t; + + /** + * @brief 64-bit floating-point type definition. + */ + typedef double float64_t; + + /** + * @brief vector types + */ +#if defined(ARM_MATH_NEON) || defined (ARM_MATH_MVEI) + /** + * @brief 64-bit fractional 128-bit vector data type in 1.63 format + */ + typedef int64x2_t q63x2_t; + + /** + * @brief 32-bit fractional 128-bit vector data type in 1.31 format. + */ + typedef int32x4_t q31x4_t; + + /** + * @brief 16-bit fractional 128-bit vector data type with 16-bit alignement in 1.15 format. + */ + typedef __ALIGNED(2) int16x8_t q15x8_t; + + /** + * @brief 8-bit fractional 128-bit vector data type with 8-bit alignement in 1.7 format. + */ + typedef __ALIGNED(1) int8x16_t q7x16_t; + + /** + * @brief 32-bit fractional 128-bit vector pair data type in 1.31 format. + */ + typedef int32x4x2_t q31x4x2_t; + + /** + * @brief 32-bit fractional 128-bit vector quadruplet data type in 1.31 format. + */ + typedef int32x4x4_t q31x4x4_t; + + /** + * @brief 16-bit fractional 128-bit vector pair data type in 1.15 format. + */ + typedef int16x8x2_t q15x8x2_t; + + /** + * @brief 16-bit fractional 128-bit vector quadruplet data type in 1.15 format. + */ + typedef int16x8x4_t q15x8x4_t; + + /** + * @brief 8-bit fractional 128-bit vector pair data type in 1.7 format. + */ + typedef int8x16x2_t q7x16x2_t; + + /** + * @brief 8-bit fractional 128-bit vector quadruplet data type in 1.7 format. + */ + typedef int8x16x4_t q7x16x4_t; + + /** + * @brief 32-bit fractional data type in 9.23 format. + */ + typedef int32_t q23_t; + + /** + * @brief 32-bit fractional 128-bit vector data type in 9.23 format. + */ + typedef int32x4_t q23x4_t; + + /** + * @brief 64-bit status 128-bit vector data type. + */ + typedef int64x2_t status64x2_t; + + /** + * @brief 32-bit status 128-bit vector data type. + */ + typedef int32x4_t status32x4_t; + + /** + * @brief 16-bit status 128-bit vector data type. + */ + typedef int16x8_t status16x8_t; + + /** + * @brief 8-bit status 128-bit vector data type. + */ + typedef int8x16_t status8x16_t; + + +#endif + +#if defined(ARM_MATH_NEON) || defined(ARM_MATH_MVEF) /* floating point vector*/ + /** + * @brief 32-bit floating-point 128-bit vector type + */ + typedef float32x4_t f32x4_t; + + /** + * @brief 32-bit floating-point 128-bit vector pair data type + */ + typedef float32x4x2_t f32x4x2_t; + + /** + * @brief 32-bit floating-point 128-bit vector quadruplet data type + */ + typedef float32x4x4_t f32x4x4_t; + + /** + * @brief 32-bit ubiquitous 128-bit vector data type + */ + typedef union _any32x4_t + { + float32x4_t f; + int32x4_t i; + } any32x4_t; + +#endif + +#if defined(ARM_MATH_NEON) + /** + * @brief 32-bit fractional 64-bit vector data type in 1.31 format. + */ + typedef int32x2_t q31x2_t; + + /** + * @brief 16-bit fractional 64-bit vector data type in 1.15 format. + */ + typedef __ALIGNED(2) int16x4_t q15x4_t; + + /** + * @brief 8-bit fractional 64-bit vector data type in 1.7 format. + */ + typedef __ALIGNED(1) int8x8_t q7x8_t; + + /** + * @brief 32-bit float 64-bit vector data type. + */ + typedef float32x2_t f32x2_t; + + /** + * @brief 32-bit floating-point 128-bit vector triplet data type + */ + typedef float32x4x3_t f32x4x3_t; + + + /** + * @brief 32-bit fractional 128-bit vector triplet data type in 1.31 format + */ + typedef int32x4x3_t q31x4x3_t; + + /** + * @brief 16-bit fractional 128-bit vector triplet data type in 1.15 format + */ + typedef int16x8x3_t q15x8x3_t; + + /** + * @brief 8-bit fractional 128-bit vector triplet data type in 1.7 format + */ + typedef int8x16x3_t q7x16x3_t; + + /** + * @brief 32-bit floating-point 64-bit vector pair data type + */ + typedef float32x2x2_t f32x2x2_t; + + /** + * @brief 32-bit floating-point 64-bit vector triplet data type + */ + typedef float32x2x3_t f32x2x3_t; + + /** + * @brief 32-bit floating-point 64-bit vector quadruplet data type + */ + typedef float32x2x4_t f32x2x4_t; + + + /** + * @brief 32-bit fractional 64-bit vector pair data type in 1.31 format + */ + typedef int32x2x2_t q31x2x2_t; + + /** + * @brief 32-bit fractional 64-bit vector triplet data type in 1.31 format + */ + typedef int32x2x3_t q31x2x3_t; + + /** + * @brief 32-bit fractional 64-bit vector quadruplet data type in 1.31 format + */ + typedef int32x4x3_t q31x2x4_t; + + /** + * @brief 16-bit fractional 64-bit vector pair data type in 1.15 format + */ + typedef int16x4x2_t q15x4x2_t; + + /** + * @brief 16-bit fractional 64-bit vector triplet data type in 1.15 format + */ + typedef int16x4x2_t q15x4x3_t; + + /** + * @brief 16-bit fractional 64-bit vector quadruplet data type in 1.15 format + */ + typedef int16x4x3_t q15x4x4_t; + + /** + * @brief 8-bit fractional 64-bit vector pair data type in 1.7 format + */ + typedef int8x8x2_t q7x8x2_t; + + /** + * @brief 8-bit fractional 64-bit vector triplet data type in 1.7 format + */ + typedef int8x8x3_t q7x8x3_t; + + /** + * @brief 8-bit fractional 64-bit vector quadruplet data type in 1.7 format + */ + typedef int8x8x4_t q7x8x4_t; + + /** + * @brief 32-bit ubiquitous 64-bit vector data type + */ + typedef union _any32x2_t + { + float32x2_t f; + int32x2_t i; + } any32x2_t; + + + /** + * @brief 32-bit status 64-bit vector data type. + */ + typedef int32x4_t status32x2_t; + + /** + * @brief 16-bit status 64-bit vector data type. + */ + typedef int16x8_t status16x4_t; + + /** + * @brief 8-bit status 64-bit vector data type. + */ + typedef int8x16_t status8x8_t; + +#endif + + + + + +#define F64_MAX ((float64_t)DBL_MAX) +#define F32_MAX ((float32_t)FLT_MAX) + + + +#define F64_MIN (-DBL_MAX) +#define F32_MIN (-FLT_MAX) + + + +#define F64_ABSMAX ((float64_t)DBL_MAX) +#define F32_ABSMAX ((float32_t)FLT_MAX) + + + +#define F64_ABSMIN ((float64_t)0.0) +#define F32_ABSMIN ((float32_t)0.0) + + +#define Q31_MAX ((q31_t)(0x7FFFFFFFL)) +#define Q15_MAX ((q15_t)(0x7FFF)) +#define Q7_MAX ((q7_t)(0x7F)) +#define Q31_MIN ((q31_t)(0x80000000L)) +#define Q15_MIN ((q15_t)(0x8000)) +#define Q7_MIN ((q7_t)(0x80)) + +#define Q31_ABSMAX ((q31_t)(0x7FFFFFFFL)) +#define Q15_ABSMAX ((q15_t)(0x7FFF)) +#define Q7_ABSMAX ((q7_t)(0x7F)) +#define Q31_ABSMIN ((q31_t)0) +#define Q15_ABSMIN ((q15_t)0) +#define Q7_ABSMIN ((q7_t)0) + + /* Dimension C vector space */ + #define CMPLX_DIM 2 + + /** + * @brief Error status returned by some functions in the library. + */ + + typedef enum + { + ARM_MATH_SUCCESS = 0, /**< No error */ + ARM_MATH_ARGUMENT_ERROR = -1, /**< One or more arguments are incorrect */ + ARM_MATH_LENGTH_ERROR = -2, /**< Length of data buffer is incorrect */ + ARM_MATH_SIZE_MISMATCH = -3, /**< Size of matrices is not compatible with the operation */ + ARM_MATH_NANINF = -4, /**< Not-a-number (NaN) or infinity is generated */ + ARM_MATH_SINGULAR = -5, /**< Input matrix is singular and cannot be inverted */ + ARM_MATH_TEST_FAILURE = -6 /**< Test Failed */ + } arm_status; + + +#ifdef __cplusplus +} +#endif + +#endif /*ifndef _ARM_MATH_TYPES_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h new file mode 100644 index 0000000..9d0956c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h @@ -0,0 +1,702 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file basic_math_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _BASIC_MATH_FUNCTIONS_H_ +#define _BASIC_MATH_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + * @defgroup groupMath Basic Math Functions + */ + + /** + * @brief Q7 vector multiplication. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_mult_q7( + const q7_t * pSrcA, + const q7_t * pSrcB, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q15 vector multiplication. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_mult_q15( + const q15_t * pSrcA, + const q15_t * pSrcB, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q31 vector multiplication. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_mult_q31( + const q31_t * pSrcA, + const q31_t * pSrcB, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Floating-point vector multiplication. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_mult_f32( + const float32_t * pSrcA, + const float32_t * pSrcB, + float32_t * pDst, + uint32_t blockSize); + + + + /** + * @brief Floating-point vector addition. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_add_f32( + const float32_t * pSrcA, + const float32_t * pSrcB, + float32_t * pDst, + uint32_t blockSize); + + + + /** + * @brief Q7 vector addition. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_add_q7( + const q7_t * pSrcA, + const q7_t * pSrcB, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q15 vector addition. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_add_q15( + const q15_t * pSrcA, + const q15_t * pSrcB, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q31 vector addition. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_add_q31( + const q31_t * pSrcA, + const q31_t * pSrcB, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Floating-point vector subtraction. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_sub_f32( + const float32_t * pSrcA, + const float32_t * pSrcB, + float32_t * pDst, + uint32_t blockSize); + + + + /** + * @brief Q7 vector subtraction. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_sub_q7( + const q7_t * pSrcA, + const q7_t * pSrcB, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q15 vector subtraction. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_sub_q15( + const q15_t * pSrcA, + const q15_t * pSrcB, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q31 vector subtraction. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in each vector + */ + void arm_sub_q31( + const q31_t * pSrcA, + const q31_t * pSrcB, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Multiplies a floating-point vector by a scalar. + * @param[in] pSrc points to the input vector + * @param[in] scale scale factor to be applied + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_scale_f32( + const float32_t * pSrc, + float32_t scale, + float32_t * pDst, + uint32_t blockSize); + + + + /** + * @brief Multiplies a Q7 vector by a scalar. + * @param[in] pSrc points to the input vector + * @param[in] scaleFract fractional portion of the scale value + * @param[in] shift number of bits to shift the result by + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_scale_q7( + const q7_t * pSrc, + q7_t scaleFract, + int8_t shift, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Multiplies a Q15 vector by a scalar. + * @param[in] pSrc points to the input vector + * @param[in] scaleFract fractional portion of the scale value + * @param[in] shift number of bits to shift the result by + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_scale_q15( + const q15_t * pSrc, + q15_t scaleFract, + int8_t shift, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Multiplies a Q31 vector by a scalar. + * @param[in] pSrc points to the input vector + * @param[in] scaleFract fractional portion of the scale value + * @param[in] shift number of bits to shift the result by + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_scale_q31( + const q31_t * pSrc, + q31_t scaleFract, + int8_t shift, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q7 vector absolute value. + * @param[in] pSrc points to the input buffer + * @param[out] pDst points to the output buffer + * @param[in] blockSize number of samples in each vector + */ + void arm_abs_q7( + const q7_t * pSrc, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Floating-point vector absolute value. + * @param[in] pSrc points to the input buffer + * @param[out] pDst points to the output buffer + * @param[in] blockSize number of samples in each vector + */ + void arm_abs_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + + + /** + * @brief Q15 vector absolute value. + * @param[in] pSrc points to the input buffer + * @param[out] pDst points to the output buffer + * @param[in] blockSize number of samples in each vector + */ + void arm_abs_q15( + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Q31 vector absolute value. + * @param[in] pSrc points to the input buffer + * @param[out] pDst points to the output buffer + * @param[in] blockSize number of samples in each vector + */ + void arm_abs_q31( + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Dot product of floating-point vectors. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] blockSize number of samples in each vector + * @param[out] result output result returned here + */ + void arm_dot_prod_f32( + const float32_t * pSrcA, + const float32_t * pSrcB, + uint32_t blockSize, + float32_t * result); + + + + /** + * @brief Dot product of Q7 vectors. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] blockSize number of samples in each vector + * @param[out] result output result returned here + */ + void arm_dot_prod_q7( + const q7_t * pSrcA, + const q7_t * pSrcB, + uint32_t blockSize, + q31_t * result); + + + /** + * @brief Dot product of Q15 vectors. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] blockSize number of samples in each vector + * @param[out] result output result returned here + */ + void arm_dot_prod_q15( + const q15_t * pSrcA, + const q15_t * pSrcB, + uint32_t blockSize, + q63_t * result); + + + /** + * @brief Dot product of Q31 vectors. + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] blockSize number of samples in each vector + * @param[out] result output result returned here + */ + void arm_dot_prod_q31( + const q31_t * pSrcA, + const q31_t * pSrcB, + uint32_t blockSize, + q63_t * result); + + + /** + * @brief Shifts the elements of a Q7 vector a specified number of bits. + * @param[in] pSrc points to the input vector + * @param[in] shiftBits number of bits to shift. A positive value shifts left; a negative value shifts right. + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_shift_q7( + const q7_t * pSrc, + int8_t shiftBits, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Shifts the elements of a Q15 vector a specified number of bits. + * @param[in] pSrc points to the input vector + * @param[in] shiftBits number of bits to shift. A positive value shifts left; a negative value shifts right. + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_shift_q15( + const q15_t * pSrc, + int8_t shiftBits, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Shifts the elements of a Q31 vector a specified number of bits. + * @param[in] pSrc points to the input vector + * @param[in] shiftBits number of bits to shift. A positive value shifts left; a negative value shifts right. + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_shift_q31( + const q31_t * pSrc, + int8_t shiftBits, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Adds a constant offset to a floating-point vector. + * @param[in] pSrc points to the input vector + * @param[in] offset is the offset to be added + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_offset_f32( + const float32_t * pSrc, + float32_t offset, + float32_t * pDst, + uint32_t blockSize); + + + + /** + * @brief Adds a constant offset to a Q7 vector. + * @param[in] pSrc points to the input vector + * @param[in] offset is the offset to be added + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_offset_q7( + const q7_t * pSrc, + q7_t offset, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Adds a constant offset to a Q15 vector. + * @param[in] pSrc points to the input vector + * @param[in] offset is the offset to be added + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_offset_q15( + const q15_t * pSrc, + q15_t offset, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Adds a constant offset to a Q31 vector. + * @param[in] pSrc points to the input vector + * @param[in] offset is the offset to be added + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_offset_q31( + const q31_t * pSrc, + q31_t offset, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Negates the elements of a floating-point vector. + * @param[in] pSrc points to the input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_negate_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Negates the elements of a Q7 vector. + * @param[in] pSrc points to the input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_negate_q7( + const q7_t * pSrc, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Negates the elements of a Q15 vector. + * @param[in] pSrc points to the input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_negate_q15( + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Negates the elements of a Q31 vector. + * @param[in] pSrc points to the input vector + * @param[out] pDst points to the output vector + * @param[in] blockSize number of samples in the vector + */ + void arm_negate_q31( + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + +/** + * @brief Compute the logical bitwise AND of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_and_u16( + const uint16_t * pSrcA, + const uint16_t * pSrcB, + uint16_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise AND of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_and_u32( + const uint32_t * pSrcA, + const uint32_t * pSrcB, + uint32_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise AND of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_and_u8( + const uint8_t * pSrcA, + const uint8_t * pSrcB, + uint8_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise OR of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_or_u16( + const uint16_t * pSrcA, + const uint16_t * pSrcB, + uint16_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise OR of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_or_u32( + const uint32_t * pSrcA, + const uint32_t * pSrcB, + uint32_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise OR of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_or_u8( + const uint8_t * pSrcA, + const uint8_t * pSrcB, + uint8_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise NOT of a fixed-point vector. + * @param[in] pSrc points to input vector + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_not_u16( + const uint16_t * pSrc, + uint16_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise NOT of a fixed-point vector. + * @param[in] pSrc points to input vector + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_not_u32( + const uint32_t * pSrc, + uint32_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise NOT of a fixed-point vector. + * @param[in] pSrc points to input vector + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_not_u8( + const uint8_t * pSrc, + uint8_t * pDst, + uint32_t blockSize); + +/** + * @brief Compute the logical bitwise XOR of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_xor_u16( + const uint16_t * pSrcA, + const uint16_t * pSrcB, + uint16_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise XOR of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_xor_u32( + const uint32_t * pSrcA, + const uint32_t * pSrcB, + uint32_t * pDst, + uint32_t blockSize); + + /** + * @brief Compute the logical bitwise XOR of two fixed-point vectors. + * @param[in] pSrcA points to input vector A + * @param[in] pSrcB points to input vector B + * @param[out] pDst points to output vector + * @param[in] blockSize number of samples in each vector + * @return none + */ + void arm_xor_u8( + const uint8_t * pSrcA, + const uint8_t * pSrcB, + uint8_t * pDst, + uint32_t blockSize); + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _BASIC_MATH_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/bayes_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/bayes_functions.h new file mode 100644 index 0000000..9bd90c1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/bayes_functions.h @@ -0,0 +1,89 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file bayes_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _BAYES_FUNCTIONS_H_ +#define _BAYES_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/statistics_functions.h" + +/** + * @defgroup groupBayes Bayesian estimators + * + * Implement the naive gaussian Bayes estimator. + * The training must be done from scikit-learn. + * + * The parameters can be easily + * generated from the scikit-learn object. Some examples are given in + * DSP/Testing/PatternGeneration/Bayes.py + */ + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + * @brief Instance structure for Naive Gaussian Bayesian estimator. + */ +typedef struct +{ + uint32_t vectorDimension; /**< Dimension of vector space */ + uint32_t numberOfClasses; /**< Number of different classes */ + const float32_t *theta; /**< Mean values for the Gaussians */ + const float32_t *sigma; /**< Variances for the Gaussians */ + const float32_t *classPriors; /**< Class prior probabilities */ + float32_t epsilon; /**< Additive value to variances */ +} arm_gaussian_naive_bayes_instance_f32; + +/** + * @brief Naive Gaussian Bayesian Estimator + * + * @param[in] S points to a naive bayes instance structure + * @param[in] in points to the elements of the input vector. + * @param[in] pBuffer points to a buffer of length numberOfClasses + * @return The predicted class + * + */ + + +uint32_t arm_gaussian_naive_bayes_predict_f32(const arm_gaussian_naive_bayes_instance_f32 *S, + const float32_t * in, + float32_t *pBuffer); + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _BAYES_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/complex_math_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/complex_math_functions.h new file mode 100644 index 0000000..d5d56b3 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/complex_math_functions.h @@ -0,0 +1,297 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file complex_math_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _COMPLEX_MATH_FUNCTIONS_H_ +#define _COMPLEX_MATH_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + * @defgroup groupCmplxMath Complex Math Functions + * This set of functions operates on complex data vectors. + * The data in the complex arrays is stored in an interleaved fashion + * (real, imag, real, imag, ...). + * In the API functions, the number of samples in a complex array refers + * to the number of complex values; the array contains twice this number of + * real values. + */ + + /** + * @brief Floating-point complex conjugate. + * @param[in] pSrc points to the input vector + * @param[out] pDst points to the output vector + * @param[in] numSamples number of complex samples in each vector + */ + void arm_cmplx_conj_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t numSamples); + + /** + * @brief Q31 complex conjugate. + * @param[in] pSrc points to the input vector + * @param[out] pDst points to the output vector + * @param[in] numSamples number of complex samples in each vector + */ + void arm_cmplx_conj_q31( + const q31_t * pSrc, + q31_t * pDst, + uint32_t numSamples); + + + /** + * @brief Q15 complex conjugate. + * @param[in] pSrc points to the input vector + * @param[out] pDst points to the output vector + * @param[in] numSamples number of complex samples in each vector + */ + void arm_cmplx_conj_q15( + const q15_t * pSrc, + q15_t * pDst, + uint32_t numSamples); + + + /** + * @brief Floating-point complex magnitude squared + * @param[in] pSrc points to the complex input vector + * @param[out] pDst points to the real output vector + * @param[in] numSamples number of complex samples in the input vector + */ + void arm_cmplx_mag_squared_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t numSamples); + + + /** + * @brief Q31 complex magnitude squared + * @param[in] pSrc points to the complex input vector + * @param[out] pDst points to the real output vector + * @param[in] numSamples number of complex samples in the input vector + */ + void arm_cmplx_mag_squared_q31( + const q31_t * pSrc, + q31_t * pDst, + uint32_t numSamples); + + + /** + * @brief Q15 complex magnitude squared + * @param[in] pSrc points to the complex input vector + * @param[out] pDst points to the real output vector + * @param[in] numSamples number of complex samples in the input vector + */ + void arm_cmplx_mag_squared_q15( + const q15_t * pSrc, + q15_t * pDst, + uint32_t numSamples); + + +/** + * @brief Floating-point complex magnitude + * @param[in] pSrc points to the complex input vector + * @param[out] pDst points to the real output vector + * @param[in] numSamples number of complex samples in the input vector + */ + void arm_cmplx_mag_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t numSamples); + + + /** + * @brief Q31 complex magnitude + * @param[in] pSrc points to the complex input vector + * @param[out] pDst points to the real output vector + * @param[in] numSamples number of complex samples in the input vector + */ + void arm_cmplx_mag_q31( + const q31_t * pSrc, + q31_t * pDst, + uint32_t numSamples); + + + /** + * @brief Q15 complex magnitude + * @param[in] pSrc points to the complex input vector + * @param[out] pDst points to the real output vector + * @param[in] numSamples number of complex samples in the input vector + */ + void arm_cmplx_mag_q15( + const q15_t * pSrc, + q15_t * pDst, + uint32_t numSamples); + + + /** + * @brief Q15 complex dot product + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] numSamples number of complex samples in each vector + * @param[out] realResult real part of the result returned here + * @param[out] imagResult imaginary part of the result returned here + */ + void arm_cmplx_dot_prod_q15( + const q15_t * pSrcA, + const q15_t * pSrcB, + uint32_t numSamples, + q31_t * realResult, + q31_t * imagResult); + + + /** + * @brief Q31 complex dot product + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] numSamples number of complex samples in each vector + * @param[out] realResult real part of the result returned here + * @param[out] imagResult imaginary part of the result returned here + */ + void arm_cmplx_dot_prod_q31( + const q31_t * pSrcA, + const q31_t * pSrcB, + uint32_t numSamples, + q63_t * realResult, + q63_t * imagResult); + + + /** + * @brief Floating-point complex dot product + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] numSamples number of complex samples in each vector + * @param[out] realResult real part of the result returned here + * @param[out] imagResult imaginary part of the result returned here + */ + void arm_cmplx_dot_prod_f32( + const float32_t * pSrcA, + const float32_t * pSrcB, + uint32_t numSamples, + float32_t * realResult, + float32_t * imagResult); + + + /** + * @brief Q15 complex-by-real multiplication + * @param[in] pSrcCmplx points to the complex input vector + * @param[in] pSrcReal points to the real input vector + * @param[out] pCmplxDst points to the complex output vector + * @param[in] numSamples number of samples in each vector + */ + void arm_cmplx_mult_real_q15( + const q15_t * pSrcCmplx, + const q15_t * pSrcReal, + q15_t * pCmplxDst, + uint32_t numSamples); + + + /** + * @brief Q31 complex-by-real multiplication + * @param[in] pSrcCmplx points to the complex input vector + * @param[in] pSrcReal points to the real input vector + * @param[out] pCmplxDst points to the complex output vector + * @param[in] numSamples number of samples in each vector + */ + void arm_cmplx_mult_real_q31( + const q31_t * pSrcCmplx, + const q31_t * pSrcReal, + q31_t * pCmplxDst, + uint32_t numSamples); + + + /** + * @brief Floating-point complex-by-real multiplication + * @param[in] pSrcCmplx points to the complex input vector + * @param[in] pSrcReal points to the real input vector + * @param[out] pCmplxDst points to the complex output vector + * @param[in] numSamples number of samples in each vector + */ + void arm_cmplx_mult_real_f32( + const float32_t * pSrcCmplx, + const float32_t * pSrcReal, + float32_t * pCmplxDst, + uint32_t numSamples); + + /** + * @brief Q15 complex-by-complex multiplication + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] numSamples number of complex samples in each vector + */ + void arm_cmplx_mult_cmplx_q15( + const q15_t * pSrcA, + const q15_t * pSrcB, + q15_t * pDst, + uint32_t numSamples); + + + /** + * @brief Q31 complex-by-complex multiplication + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] numSamples number of complex samples in each vector + */ + void arm_cmplx_mult_cmplx_q31( + const q31_t * pSrcA, + const q31_t * pSrcB, + q31_t * pDst, + uint32_t numSamples); + + + /** + * @brief Floating-point complex-by-complex multiplication + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[out] pDst points to the output vector + * @param[in] numSamples number of complex samples in each vector + */ + void arm_cmplx_mult_cmplx_f32( + const float32_t * pSrcA, + const float32_t * pSrcB, + float32_t * pDst, + uint32_t numSamples); + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _COMPLEX_MATH_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/controller_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/controller_functions.h new file mode 100644 index 0000000..78d29d9 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/controller_functions.h @@ -0,0 +1,793 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file controller_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _CONTROLLER_FUNCTIONS_H_ +#define _CONTROLLER_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + /** + * @brief Macros required for SINE and COSINE Controller functions + */ + +#define CONTROLLER_Q31_SHIFT (32 - 9) + /* 1.31(q31) Fixed value of 2/360 */ + /* -1 to +1 is divided into 360 values so total spacing is (2/360) */ +#define INPUT_SPACING 0xB60B61 + +/** + * @defgroup groupController Controller Functions + */ + + + /** + * @ingroup groupController + */ + + /** + * @addtogroup SinCos + * @{ + */ + +/** + * @brief Floating-point sin_cos function. + * @param[in] theta input value in degrees + * @param[out] pSinVal points to the processed sine output. + * @param[out] pCosVal points to the processed cos output. + */ + void arm_sin_cos_f32( + float32_t theta, + float32_t * pSinVal, + float32_t * pCosVal); + + + /** + * @brief Q31 sin_cos function. + * @param[in] theta scaled input value in degrees + * @param[out] pSinVal points to the processed sine output. + * @param[out] pCosVal points to the processed cosine output. + */ + void arm_sin_cos_q31( + q31_t theta, + q31_t * pSinVal, + q31_t * pCosVal); + + /** + * @} end of SinCos group + */ + + /** + * @ingroup groupController + */ + +/** + * @defgroup PID PID Motor Control + * + * A Proportional Integral Derivative (PID) controller is a generic feedback control + * loop mechanism widely used in industrial control systems. + * A PID controller is the most commonly used type of feedback controller. + * + * This set of functions implements (PID) controllers + * for Q15, Q31, and floating-point data types. The functions operate on a single sample + * of data and each call to the function returns a single processed value. + * S points to an instance of the PID control data structure. in + * is the input sample value. The functions return the output value. + * + * \par Algorithm: + * 
+   *    y[n] = y[n-1] + A0 * x[n] + A1 * x[n-1] + A2 * x[n-2]
+   *    A0 = Kp + Ki + Kd
+   *    A1 = (-Kp ) - (2 * Kd )
+   *    A2 = Kd
+   *
+ * + * \par + * where \c Kp is proportional constant, \c Ki is Integral constant and \c Kd is Derivative constant + * + * \par + * \image html PID.gif "Proportional Integral Derivative Controller" + * + * \par + * The PID controller calculates an "error" value as the difference between + * the measured output and the reference input. + * The controller attempts to minimize the error by adjusting the process control inputs. + * The proportional value determines the reaction to the current error, + * the integral value determines the reaction based on the sum of recent errors, + * and the derivative value determines the reaction based on the rate at which the error has been changing. + * + * \par Instance Structure + * The Gains A0, A1, A2 and state variables for a PID controller are stored together in an instance data structure. + * A separate instance structure must be defined for each PID Controller. + * There are separate instance structure declarations for each of the 3 supported data types. + * + * \par Reset Functions + * There is also an associated reset function for each data type which clears the state array. + * + * \par Initialization Functions + * There is also an associated initialization function for each data type. + * The initialization function performs the following operations: + * - Initializes the Gains A0, A1, A2 from Kp,Ki, Kd gains. + * - Zeros out the values in the state buffer. + * + * \par + * Instance structure cannot be placed into a const data section and it is recommended to use the initialization function. + * + * \par Fixed-Point Behavior + * Care must be taken when using the fixed-point versions of the PID Controller functions. + * In particular, the overflow and saturation behavior of the accumulator used in each function must be considered. + * Refer to the function specific documentation below for usage guidelines. + */ + + + /** + * @brief Instance structure for the Q15 PID Control. + */ + typedef struct + { + q15_t A0; /**< The derived gain, A0 = Kp + Ki + Kd . */ +#if !defined (ARM_MATH_DSP) + q15_t A1; /**< The derived gain A1 = -Kp - 2Kd */ + q15_t A2; /**< The derived gain A1 = Kd. */ +#else + q31_t A1; /**< The derived gain A1 = -Kp - 2Kd | Kd.*/ +#endif + q15_t state[3]; /**< The state array of length 3. */ + q15_t Kp; /**< The proportional gain. */ + q15_t Ki; /**< The integral gain. */ + q15_t Kd; /**< The derivative gain. */ + } arm_pid_instance_q15; + + /** + * @brief Instance structure for the Q31 PID Control. + */ + typedef struct + { + q31_t A0; /**< The derived gain, A0 = Kp + Ki + Kd . */ + q31_t A1; /**< The derived gain, A1 = -Kp - 2Kd. */ + q31_t A2; /**< The derived gain, A2 = Kd . */ + q31_t state[3]; /**< The state array of length 3. */ + q31_t Kp; /**< The proportional gain. */ + q31_t Ki; /**< The integral gain. */ + q31_t Kd; /**< The derivative gain. */ + } arm_pid_instance_q31; + + /** + * @brief Instance structure for the floating-point PID Control. + */ + typedef struct + { + float32_t A0; /**< The derived gain, A0 = Kp + Ki + Kd . */ + float32_t A1; /**< The derived gain, A1 = -Kp - 2Kd. */ + float32_t A2; /**< The derived gain, A2 = Kd . */ + float32_t state[3]; /**< The state array of length 3. */ + float32_t Kp; /**< The proportional gain. */ + float32_t Ki; /**< The integral gain. */ + float32_t Kd; /**< The derivative gain. */ + } arm_pid_instance_f32; + + + + /** + * @brief Initialization function for the floating-point PID Control. + * @param[in,out] S points to an instance of the PID structure. + * @param[in] resetStateFlag flag to reset the state. 0 = no change in state 1 = reset the state. + */ + void arm_pid_init_f32( + arm_pid_instance_f32 * S, + int32_t resetStateFlag); + + + /** + * @brief Reset function for the floating-point PID Control. + * @param[in,out] S is an instance of the floating-point PID Control structure + */ + void arm_pid_reset_f32( + arm_pid_instance_f32 * S); + + + /** + * @brief Initialization function for the Q31 PID Control. + * @param[in,out] S points to an instance of the Q15 PID structure. + * @param[in] resetStateFlag flag to reset the state. 0 = no change in state 1 = reset the state. + */ + void arm_pid_init_q31( + arm_pid_instance_q31 * S, + int32_t resetStateFlag); + + + /** + * @brief Reset function for the Q31 PID Control. + * @param[in,out] S points to an instance of the Q31 PID Control structure + */ + + void arm_pid_reset_q31( + arm_pid_instance_q31 * S); + + + /** + * @brief Initialization function for the Q15 PID Control. + * @param[in,out] S points to an instance of the Q15 PID structure. + * @param[in] resetStateFlag flag to reset the state. 0 = no change in state 1 = reset the state. + */ + void arm_pid_init_q15( + arm_pid_instance_q15 * S, + int32_t resetStateFlag); + + + /** + * @brief Reset function for the Q15 PID Control. + * @param[in,out] S points to an instance of the q15 PID Control structure + */ + void arm_pid_reset_q15( + arm_pid_instance_q15 * S); + + + + /** + * @addtogroup PID + * @{ + */ + + /** + * @brief Process function for the floating-point PID Control. + * @param[in,out] S is an instance of the floating-point PID Control structure + * @param[in] in input sample to process + * @return processed output sample. + */ + __STATIC_FORCEINLINE float32_t arm_pid_f32( + arm_pid_instance_f32 * S, + float32_t in) + { + float32_t out; + + /* y[n] = y[n-1] + A0 * x[n] + A1 * x[n-1] + A2 * x[n-2] */ + out = (S->A0 * in) + + (S->A1 * S->state[0]) + (S->A2 * S->state[1]) + (S->state[2]); + + /* Update state */ + S->state[1] = S->state[0]; + S->state[0] = in; + S->state[2] = out; + + /* return to application */ + return (out); + + } + +/** + @brief Process function for the Q31 PID Control. + @param[in,out] S points to an instance of the Q31 PID Control structure + @param[in] in input sample to process + @return processed output sample. + + \par Scaling and Overflow Behavior + The function is implemented using an internal 64-bit accumulator. + The accumulator has a 2.62 format and maintains full precision of the intermediate multiplication results but provides only a single guard bit. + Thus, if the accumulator result overflows it wraps around rather than clip. + In order to avoid overflows completely the input signal must be scaled down by 2 bits as there are four additions. + After all multiply-accumulates are performed, the 2.62 accumulator is truncated to 1.32 format and then saturated to 1.31 format. + */ +__STATIC_FORCEINLINE q31_t arm_pid_q31( + arm_pid_instance_q31 * S, + q31_t in) + { + q63_t acc; + q31_t out; + + /* acc = A0 * x[n] */ + acc = (q63_t) S->A0 * in; + + /* acc += A1 * x[n-1] */ + acc += (q63_t) S->A1 * S->state[0]; + + /* acc += A2 * x[n-2] */ + acc += (q63_t) S->A2 * S->state[1]; + + /* convert output to 1.31 format to add y[n-1] */ + out = (q31_t) (acc >> 31U); + + /* out += y[n-1] */ + out += S->state[2]; + + /* Update state */ + S->state[1] = S->state[0]; + S->state[0] = in; + S->state[2] = out; + + /* return to application */ + return (out); + } + + +/** + @brief Process function for the Q15 PID Control. + @param[in,out] S points to an instance of the Q15 PID Control structure + @param[in] in input sample to process + @return processed output sample. + + \par Scaling and Overflow Behavior + The function is implemented using a 64-bit internal accumulator. + Both Gains and state variables are represented in 1.15 format and multiplications yield a 2.30 result. + The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format. + There is no risk of internal overflow with this approach and the full precision of intermediate multiplications is preserved. + After all additions have been performed, the accumulator is truncated to 34.15 format by discarding low 15 bits. + Lastly, the accumulator is saturated to yield a result in 1.15 format. + */ +__STATIC_FORCEINLINE q15_t arm_pid_q15( + arm_pid_instance_q15 * S, + q15_t in) + { + q63_t acc; + q15_t out; + +#if defined (ARM_MATH_DSP) + /* Implementation of PID controller */ + + /* acc = A0 * x[n] */ + acc = (q31_t) __SMUAD((uint32_t)S->A0, (uint32_t)in); + + /* acc += A1 * x[n-1] + A2 * x[n-2] */ + acc = (q63_t)__SMLALD((uint32_t)S->A1, (uint32_t)read_q15x2 (S->state), (uint64_t)acc); +#else + /* acc = A0 * x[n] */ + acc = ((q31_t) S->A0) * in; + + /* acc += A1 * x[n-1] + A2 * x[n-2] */ + acc += (q31_t) S->A1 * S->state[0]; + acc += (q31_t) S->A2 * S->state[1]; +#endif + + /* acc += y[n-1] */ + acc += (q31_t) S->state[2] << 15; + + /* saturate the output */ + out = (q15_t) (__SSAT((q31_t)(acc >> 15), 16)); + + /* Update state */ + S->state[1] = S->state[0]; + S->state[0] = in; + S->state[2] = out; + + /* return to application */ + return (out); + } + + /** + * @} end of PID group + */ + + /** + * @ingroup groupController + */ + + /** + * @defgroup park Vector Park Transform + * + * Forward Park transform converts the input two-coordinate vector to flux and torque components. + * The Park transform can be used to realize the transformation of the Ialpha and the Ibeta currents + * from the stationary to the moving reference frame and control the spatial relationship between + * the stator vector current and rotor flux vector. + * If we consider the d axis aligned with the rotor flux, the diagram below shows the + * current vector and the relationship from the two reference frames: + * \image html park.gif "Stator current space vector and its component in (a,b) and in the d,q rotating reference frame" + * + * The function operates on a single sample of data and each call to the function returns the processed output. + * The library provides separate functions for Q31 and floating-point data types. + * \par Algorithm + * \image html parkFormula.gif + * where Ialpha and Ibeta are the stator vector components, + * pId and pIq are rotor vector components and cosVal and sinVal are the + * cosine and sine values of theta (rotor flux position). + * \par Fixed-Point Behavior + * Care must be taken when using the Q31 version of the Park transform. + * In particular, the overflow and saturation behavior of the accumulator used must be considered. + * Refer to the function specific documentation below for usage guidelines. + */ + + /** + * @addtogroup park + * @{ + */ + + /** + * @brief Floating-point Park transform + * @param[in] Ialpha input two-phase vector coordinate alpha + * @param[in] Ibeta input two-phase vector coordinate beta + * @param[out] pId points to output rotor reference frame d + * @param[out] pIq points to output rotor reference frame q + * @param[in] sinVal sine value of rotation angle theta + * @param[in] cosVal cosine value of rotation angle theta + * @return none + * + * The function implements the forward Park transform. + * + */ + __STATIC_FORCEINLINE void arm_park_f32( + float32_t Ialpha, + float32_t Ibeta, + float32_t * pId, + float32_t * pIq, + float32_t sinVal, + float32_t cosVal) + { + /* Calculate pId using the equation, pId = Ialpha * cosVal + Ibeta * sinVal */ + *pId = Ialpha * cosVal + Ibeta * sinVal; + + /* Calculate pIq using the equation, pIq = - Ialpha * sinVal + Ibeta * cosVal */ + *pIq = -Ialpha * sinVal + Ibeta * cosVal; + } + + +/** + @brief Park transform for Q31 version + @param[in] Ialpha input two-phase vector coordinate alpha + @param[in] Ibeta input two-phase vector coordinate beta + @param[out] pId points to output rotor reference frame d + @param[out] pIq points to output rotor reference frame q + @param[in] sinVal sine value of rotation angle theta + @param[in] cosVal cosine value of rotation angle theta + @return none + + \par Scaling and Overflow Behavior + The function is implemented using an internal 32-bit accumulator. + The accumulator maintains 1.31 format by truncating lower 31 bits of the intermediate multiplication in 2.62 format. + There is saturation on the addition and subtraction, hence there is no risk of overflow. + */ +__STATIC_FORCEINLINE void arm_park_q31( + q31_t Ialpha, + q31_t Ibeta, + q31_t * pId, + q31_t * pIq, + q31_t sinVal, + q31_t cosVal) + { + q31_t product1, product2; /* Temporary variables used to store intermediate results */ + q31_t product3, product4; /* Temporary variables used to store intermediate results */ + + /* Intermediate product is calculated by (Ialpha * cosVal) */ + product1 = (q31_t) (((q63_t) (Ialpha) * (cosVal)) >> 31); + + /* Intermediate product is calculated by (Ibeta * sinVal) */ + product2 = (q31_t) (((q63_t) (Ibeta) * (sinVal)) >> 31); + + + /* Intermediate product is calculated by (Ialpha * sinVal) */ + product3 = (q31_t) (((q63_t) (Ialpha) * (sinVal)) >> 31); + + /* Intermediate product is calculated by (Ibeta * cosVal) */ + product4 = (q31_t) (((q63_t) (Ibeta) * (cosVal)) >> 31); + + /* Calculate pId by adding the two intermediate products 1 and 2 */ + *pId = __QADD(product1, product2); + + /* Calculate pIq by subtracting the two intermediate products 3 from 4 */ + *pIq = __QSUB(product4, product3); + } + + /** + * @} end of park group + */ + + + /** + * @ingroup groupController + */ + + /** + * @defgroup inv_park Vector Inverse Park transform + * Inverse Park transform converts the input flux and torque components to two-coordinate vector. + * + * The function operates on a single sample of data and each call to the function returns the processed output. + * The library provides separate functions for Q31 and floating-point data types. + * \par Algorithm + * \image html parkInvFormula.gif + * where pIalpha and pIbeta are the stator vector components, + * Id and Iq are rotor vector components and cosVal and sinVal are the + * cosine and sine values of theta (rotor flux position). + * \par Fixed-Point Behavior + * Care must be taken when using the Q31 version of the Park transform. + * In particular, the overflow and saturation behavior of the accumulator used must be considered. + * Refer to the function specific documentation below for usage guidelines. + */ + + /** + * @addtogroup inv_park + * @{ + */ + + /** + * @brief Floating-point Inverse Park transform + * @param[in] Id input coordinate of rotor reference frame d + * @param[in] Iq input coordinate of rotor reference frame q + * @param[out] pIalpha points to output two-phase orthogonal vector axis alpha + * @param[out] pIbeta points to output two-phase orthogonal vector axis beta + * @param[in] sinVal sine value of rotation angle theta + * @param[in] cosVal cosine value of rotation angle theta + * @return none + */ + __STATIC_FORCEINLINE void arm_inv_park_f32( + float32_t Id, + float32_t Iq, + float32_t * pIalpha, + float32_t * pIbeta, + float32_t sinVal, + float32_t cosVal) + { + /* Calculate pIalpha using the equation, pIalpha = Id * cosVal - Iq * sinVal */ + *pIalpha = Id * cosVal - Iq * sinVal; + + /* Calculate pIbeta using the equation, pIbeta = Id * sinVal + Iq * cosVal */ + *pIbeta = Id * sinVal + Iq * cosVal; + } + + +/** + @brief Inverse Park transform for Q31 version + @param[in] Id input coordinate of rotor reference frame d + @param[in] Iq input coordinate of rotor reference frame q + @param[out] pIalpha points to output two-phase orthogonal vector axis alpha + @param[out] pIbeta points to output two-phase orthogonal vector axis beta + @param[in] sinVal sine value of rotation angle theta + @param[in] cosVal cosine value of rotation angle theta + @return none + + @par Scaling and Overflow Behavior + The function is implemented using an internal 32-bit accumulator. + The accumulator maintains 1.31 format by truncating lower 31 bits of the intermediate multiplication in 2.62 format. + There is saturation on the addition, hence there is no risk of overflow. + */ +__STATIC_FORCEINLINE void arm_inv_park_q31( + q31_t Id, + q31_t Iq, + q31_t * pIalpha, + q31_t * pIbeta, + q31_t sinVal, + q31_t cosVal) + { + q31_t product1, product2; /* Temporary variables used to store intermediate results */ + q31_t product3, product4; /* Temporary variables used to store intermediate results */ + + /* Intermediate product is calculated by (Id * cosVal) */ + product1 = (q31_t) (((q63_t) (Id) * (cosVal)) >> 31); + + /* Intermediate product is calculated by (Iq * sinVal) */ + product2 = (q31_t) (((q63_t) (Iq) * (sinVal)) >> 31); + + + /* Intermediate product is calculated by (Id * sinVal) */ + product3 = (q31_t) (((q63_t) (Id) * (sinVal)) >> 31); + + /* Intermediate product is calculated by (Iq * cosVal) */ + product4 = (q31_t) (((q63_t) (Iq) * (cosVal)) >> 31); + + /* Calculate pIalpha by using the two intermediate products 1 and 2 */ + *pIalpha = __QSUB(product1, product2); + + /* Calculate pIbeta by using the two intermediate products 3 and 4 */ + *pIbeta = __QADD(product4, product3); + } + + /** + * @} end of Inverse park group + */ + +/** + * @ingroup groupController + */ + + /** + * @defgroup clarke Vector Clarke Transform + * Forward Clarke transform converts the instantaneous stator phases into a two-coordinate time invariant vector. + * Generally the Clarke transform uses three-phase currents Ia, Ib and Ic to calculate currents + * in the two-phase orthogonal stator axis Ialpha and Ibeta. + * When Ialpha is superposed with Ia as shown in the figure below + * \image html clarke.gif Stator current space vector and its components in (a,b). + * and Ia + Ib + Ic = 0, in this condition Ialpha and Ibeta + * can be calculated using only Ia and Ib. + * + * The function operates on a single sample of data and each call to the function returns the processed output. + * The library provides separate functions for Q31 and floating-point data types. + * \par Algorithm + * \image html clarkeFormula.gif + * where Ia and Ib are the instantaneous stator phases and + * pIalpha and pIbeta are the two coordinates of time invariant vector. + * \par Fixed-Point Behavior + * Care must be taken when using the Q31 version of the Clarke transform. + * In particular, the overflow and saturation behavior of the accumulator used must be considered. + * Refer to the function specific documentation below for usage guidelines. + */ + + /** + * @addtogroup clarke + * @{ + */ + + /** + * + * @brief Floating-point Clarke transform + * @param[in] Ia input three-phase coordinate a + * @param[in] Ib input three-phase coordinate b + * @param[out] pIalpha points to output two-phase orthogonal vector axis alpha + * @param[out] pIbeta points to output two-phase orthogonal vector axis beta + * @return none + */ + __STATIC_FORCEINLINE void arm_clarke_f32( + float32_t Ia, + float32_t Ib, + float32_t * pIalpha, + float32_t * pIbeta) + { + /* Calculate pIalpha using the equation, pIalpha = Ia */ + *pIalpha = Ia; + + /* Calculate pIbeta using the equation, pIbeta = (1/sqrt(3)) * Ia + (2/sqrt(3)) * Ib */ + *pIbeta = (0.57735026919f * Ia + 1.15470053838f * Ib); + } + + +/** + @brief Clarke transform for Q31 version + @param[in] Ia input three-phase coordinate a + @param[in] Ib input three-phase coordinate b + @param[out] pIalpha points to output two-phase orthogonal vector axis alpha + @param[out] pIbeta points to output two-phase orthogonal vector axis beta + @return none + + \par Scaling and Overflow Behavior + The function is implemented using an internal 32-bit accumulator. + The accumulator maintains 1.31 format by truncating lower 31 bits of the intermediate multiplication in 2.62 format. + There is saturation on the addition, hence there is no risk of overflow. + */ +__STATIC_FORCEINLINE void arm_clarke_q31( + q31_t Ia, + q31_t Ib, + q31_t * pIalpha, + q31_t * pIbeta) + { + q31_t product1, product2; /* Temporary variables used to store intermediate results */ + + /* Calculating pIalpha from Ia by equation pIalpha = Ia */ + *pIalpha = Ia; + + /* Intermediate product is calculated by (1/(sqrt(3)) * Ia) */ + product1 = (q31_t) (((q63_t) Ia * 0x24F34E8B) >> 30); + + /* Intermediate product is calculated by (2/sqrt(3) * Ib) */ + product2 = (q31_t) (((q63_t) Ib * 0x49E69D16) >> 30); + + /* pIbeta is calculated by adding the intermediate products */ + *pIbeta = __QADD(product1, product2); + } + + /** + * @} end of clarke group + */ + + + /** + * @ingroup groupController + */ + + /** + * @defgroup inv_clarke Vector Inverse Clarke Transform + * Inverse Clarke transform converts the two-coordinate time invariant vector into instantaneous stator phases. + * + * The function operates on a single sample of data and each call to the function returns the processed output. + * The library provides separate functions for Q31 and floating-point data types. + * \par Algorithm + * \image html clarkeInvFormula.gif + * where pIa and pIb are the instantaneous stator phases and + * Ialpha and Ibeta are the two coordinates of time invariant vector. + * \par Fixed-Point Behavior + * Care must be taken when using the Q31 version of the Clarke transform. + * In particular, the overflow and saturation behavior of the accumulator used must be considered. + * Refer to the function specific documentation below for usage guidelines. + */ + + /** + * @addtogroup inv_clarke + * @{ + */ + + /** + * @brief Floating-point Inverse Clarke transform + * @param[in] Ialpha input two-phase orthogonal vector axis alpha + * @param[in] Ibeta input two-phase orthogonal vector axis beta + * @param[out] pIa points to output three-phase coordinate a + * @param[out] pIb points to output three-phase coordinate b + * @return none + */ + __STATIC_FORCEINLINE void arm_inv_clarke_f32( + float32_t Ialpha, + float32_t Ibeta, + float32_t * pIa, + float32_t * pIb) + { + /* Calculating pIa from Ialpha by equation pIa = Ialpha */ + *pIa = Ialpha; + + /* Calculating pIb from Ialpha and Ibeta by equation pIb = -(1/2) * Ialpha + (sqrt(3)/2) * Ibeta */ + *pIb = -0.5f * Ialpha + 0.8660254039f * Ibeta; + } + + +/** + @brief Inverse Clarke transform for Q31 version + @param[in] Ialpha input two-phase orthogonal vector axis alpha + @param[in] Ibeta input two-phase orthogonal vector axis beta + @param[out] pIa points to output three-phase coordinate a + @param[out] pIb points to output three-phase coordinate b + @return none + + \par Scaling and Overflow Behavior + The function is implemented using an internal 32-bit accumulator. + The accumulator maintains 1.31 format by truncating lower 31 bits of the intermediate multiplication in 2.62 format. + There is saturation on the subtraction, hence there is no risk of overflow. + */ +__STATIC_FORCEINLINE void arm_inv_clarke_q31( + q31_t Ialpha, + q31_t Ibeta, + q31_t * pIa, + q31_t * pIb) + { + q31_t product1, product2; /* Temporary variables used to store intermediate results */ + + /* Calculating pIa from Ialpha by equation pIa = Ialpha */ + *pIa = Ialpha; + + /* Intermediate product is calculated by (1/(2*sqrt(3)) * Ia) */ + product1 = (q31_t) (((q63_t) (Ialpha) * (0x40000000)) >> 31); + + /* Intermediate product is calculated by (1/sqrt(3) * pIb) */ + product2 = (q31_t) (((q63_t) (Ibeta) * (0x6ED9EBA1)) >> 31); + + /* pIb is calculated by subtracting the products */ + *pIb = __QSUB(product2, product1); + } + + /** + * @} end of inv_clarke group + */ + + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _CONTROLLER_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/distance_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/distance_functions.h new file mode 100644 index 0000000..da0aedc --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/distance_functions.h @@ -0,0 +1,299 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file distance_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _DISTANCE_FUNCTIONS_H_ +#define _DISTANCE_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/statistics_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + +/** + * @defgroup groupDistance Distance functions + * + * Distance functions for use with clustering algorithms. + * There are distance functions for float vectors and boolean vectors. + * + */ + +/* 6.14 bug */ +#if defined (__ARMCC_VERSION) && (__ARMCC_VERSION >= 6100100) && (__ARMCC_VERSION < 6150001) + +__attribute__((weak)) float __powisf2(float a, int b); + +#endif + +/** + * @brief Euclidean distance between two vectors + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ + +float32_t arm_euclidean_distance_f32(const float32_t *pA,const float32_t *pB, uint32_t blockSize); + +/** + * @brief Bray-Curtis distance between two vectors + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ +float32_t arm_braycurtis_distance_f32(const float32_t *pA,const float32_t *pB, uint32_t blockSize); + +/** + * @brief Canberra distance between two vectors + * + * This function may divide by zero when samples pA[i] and pB[i] are both zero. + * The result of the computation will be correct. So the division per zero may be + * ignored. + * + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ +float32_t arm_canberra_distance_f32(const float32_t *pA,const float32_t *pB, uint32_t blockSize); + + +/** + * @brief Chebyshev distance between two vectors + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ +float32_t arm_chebyshev_distance_f32(const float32_t *pA,const float32_t *pB, uint32_t blockSize); + + +/** + * @brief Cityblock (Manhattan) distance between two vectors + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ +float32_t arm_cityblock_distance_f32(const float32_t *pA,const float32_t *pB, uint32_t blockSize); + +/** + * @brief Correlation distance between two vectors + * + * The input vectors are modified in place ! + * + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ +float32_t arm_correlation_distance_f32(float32_t *pA,float32_t *pB, uint32_t blockSize); + +/** + * @brief Cosine distance between two vectors + * + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ + +float32_t arm_cosine_distance_f32(const float32_t *pA,const float32_t *pB, uint32_t blockSize); + +/** + * @brief Jensen-Shannon distance between two vectors + * + * This function is assuming that elements of second vector are > 0 + * and 0 only when the corresponding element of first vector is 0. + * Otherwise the result of the computation does not make sense + * and for speed reasons, the cases returning NaN or Infinity are not + * managed. + * + * When the function is computing x log (x / y) with x 0 and y 0, + * it will compute the right value (0) but a division per zero will occur + * and shoudl be ignored in client code. + * + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] blockSize vector length + * @return distance + * + */ + +float32_t arm_jensenshannon_distance_f32(const float32_t *pA,const float32_t *pB,uint32_t blockSize); + +/** + * @brief Minkowski distance between two vectors + * + * @param[in] pA First vector + * @param[in] pB Second vector + * @param[in] n Norm order (>= 2) + * @param[in] blockSize vector length + * @return distance + * + */ + + + +float32_t arm_minkowski_distance_f32(const float32_t *pA,const float32_t *pB, int32_t order, uint32_t blockSize); + +/** + * @brief Dice distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] order Distance order + * @param[in] blockSize Number of samples + * @return distance + * + */ + + +float32_t arm_dice_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Hamming distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_hamming_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Jaccard distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_jaccard_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Kulsinski distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_kulsinski_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Roger Stanimoto distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_rogerstanimoto_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Russell-Rao distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_russellrao_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Sokal-Michener distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_sokalmichener_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Sokal-Sneath distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_sokalsneath_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + +/** + * @brief Yule distance between two vectors + * + * @param[in] pA First vector of packed booleans + * @param[in] pB Second vector of packed booleans + * @param[in] numberOfBools Number of booleans + * @return distance + * + */ + +float32_t arm_yule_distance(const uint32_t *pA, const uint32_t *pB, uint32_t numberOfBools); + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _DISTANCE_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h new file mode 100644 index 0000000..758731a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h @@ -0,0 +1,290 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file fast_math_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _FAST_MATH_FUNCTIONS_H_ +#define _FAST_MATH_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + /** + * @brief Macros required for SINE and COSINE Fast math approximations + */ + +#define FAST_MATH_TABLE_SIZE 512 +#define FAST_MATH_Q31_SHIFT (32 - 10) +#define FAST_MATH_Q15_SHIFT (16 - 10) + +#ifndef PI + #define PI 3.14159265358979f +#endif + + +/** + * @defgroup groupFastMath Fast Math Functions + * This set of functions provides a fast approximation to sine, cosine, and square root. + * As compared to most of the other functions in the CMSIS math library, the fast math functions + * operate on individual values and not arrays. + * There are separate functions for Q15, Q31, and floating-point data. + * + */ + + /** + * @ingroup groupFastMath + */ + + +/** + @addtogroup sin + @{ + */ + +/** + * @brief Fast approximation to the trigonometric sine function for floating-point data. + * @param[in] x input value in radians. + * @return sin(x). + */ + float32_t arm_sin_f32( + float32_t x); + + + /** + * @brief Fast approximation to the trigonometric sine function for Q31 data. + * @param[in] x Scaled input value in radians. + * @return sin(x). + */ + q31_t arm_sin_q31( + q31_t x); + + + /** + * @brief Fast approximation to the trigonometric sine function for Q15 data. + * @param[in] x Scaled input value in radians. + * @return sin(x). + */ + q15_t arm_sin_q15( + q15_t x); + +/** + @} end of sin group + */ + +/** + @addtogroup cos + @{ + */ + + /** + * @brief Fast approximation to the trigonometric cosine function for floating-point data. + * @param[in] x input value in radians. + * @return cos(x). + */ + float32_t arm_cos_f32( + float32_t x); + + + /** + * @brief Fast approximation to the trigonometric cosine function for Q31 data. + * @param[in] x Scaled input value in radians. + * @return cos(x). + */ + q31_t arm_cos_q31( + q31_t x); + + + /** + * @brief Fast approximation to the trigonometric cosine function for Q15 data. + * @param[in] x Scaled input value in radians. + * @return cos(x). + */ + q15_t arm_cos_q15( + q15_t x); + +/** + @} end of cos group + */ + + +/** + @brief Floating-point vector of log values. + @param[in] pSrc points to the input vector + @param[out] pDst points to the output vector + @param[in] blockSize number of samples in each vector + @return none + */ + void arm_vlog_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + +/** + @brief Floating-point vector of exp values. + @param[in] pSrc points to the input vector + @param[out] pDst points to the output vector + @param[in] blockSize number of samples in each vector + @return none + */ + void arm_vexp_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + /** + * @defgroup SQRT Square Root + * + * Computes the square root of a number. + * There are separate functions for Q15, Q31, and floating-point data types. + * The square root function is computed using the Newton-Raphson algorithm. + * This is an iterative algorithm of the form: + * 
+   *      x1 = x0 - f(x0)/f'(x0)
+   *
+ * where x1 is the current estimate, + * x0 is the previous estimate, and + * f'(x0) is the derivative of f() evaluated at x0. + * For the square root function, the algorithm reduces to: + * 
+   *     x0 = in/2                         [initial guess]
+   *     x1 = 1/2 * ( x0 + in / x0)        [each iteration]
+   *
+ */ + + + /** + * @addtogroup SQRT + * @{ + */ + +/** + @brief Floating-point square root function. + @param[in] in input value + @param[out] pOut square root of input value + @return execution status + - \ref ARM_MATH_SUCCESS : input value is positive + - \ref ARM_MATH_ARGUMENT_ERROR : input value is negative; *pOut is set to 0 + */ +__STATIC_FORCEINLINE arm_status arm_sqrt_f32( + float32_t in, + float32_t * pOut) + { + if (in >= 0.0f) + { +#if defined ( __CC_ARM ) + #if defined __TARGET_FPU_VFP + *pOut = __sqrtf(in); + #else + *pOut = sqrtf(in); + #endif + +#elif defined ( __ICCARM__ ) + #if defined __ARMVFP__ + __ASM("VSQRT.F32 %0,%1" : "=t"(*pOut) : "t"(in)); + #else + *pOut = sqrtf(in); + #endif + +#else + *pOut = sqrtf(in); +#endif + + return (ARM_MATH_SUCCESS); + } + else + { + *pOut = 0.0f; + return (ARM_MATH_ARGUMENT_ERROR); + } + } + + +/** + @brief Q31 square root function. + @param[in] in input value. The range of the input value is [0 +1) or 0x00000000 to 0x7FFFFFFF + @param[out] pOut points to square root of input value + @return execution status + - \ref ARM_MATH_SUCCESS : input value is positive + - \ref ARM_MATH_ARGUMENT_ERROR : input value is negative; *pOut is set to 0 + */ +arm_status arm_sqrt_q31( + q31_t in, + q31_t * pOut); + + +/** + @brief Q15 square root function. + @param[in] in input value. The range of the input value is [0 +1) or 0x0000 to 0x7FFF + @param[out] pOut points to square root of input value + @return execution status + - \ref ARM_MATH_SUCCESS : input value is positive + - \ref ARM_MATH_ARGUMENT_ERROR : input value is negative; *pOut is set to 0 + */ +arm_status arm_sqrt_q15( + q15_t in, + q15_t * pOut); + + /** + * @brief Vector Floating-point square root function. + * @param[in] pIn input vector. + * @param[out] pOut vector of square roots of input elements. + * @param[in] len length of input vector. + * @return The function returns ARM_MATH_SUCCESS if input value is positive value or ARM_MATH_ARGUMENT_ERROR if + * in is negative value and returns zero output for negative values. + */ + void arm_vsqrt_f32( + float32_t * pIn, + float32_t * pOut, + uint16_t len); + + void arm_vsqrt_q31( + q31_t * pIn, + q31_t * pOut, + uint16_t len); + + void arm_vsqrt_q15( + q15_t * pIn, + q15_t * pOut, + uint16_t len); + + /** + * @} end of SQRT group + */ + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _FAST_MATH_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/filtering_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/filtering_functions.h new file mode 100644 index 0000000..00f5e0a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/filtering_functions.h @@ -0,0 +1,2442 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file filtering_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _FILTERING_FUNCTIONS_H_ +#define _FILTERING_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/support_functions.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +#define DELTA_Q31 ((q31_t)(0x100)) +#define DELTA_Q15 ((q15_t)0x5) + +/** + * @defgroup groupFilters Filtering Functions + */ + + /** + * @brief Instance structure for the Q7 FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of filter coefficients in the filter. */ + q7_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + const q7_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + } arm_fir_instance_q7; + + /** + * @brief Instance structure for the Q15 FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of filter coefficients in the filter. */ + q15_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + const q15_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + } arm_fir_instance_q15; + + /** + * @brief Instance structure for the Q31 FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of filter coefficients in the filter. */ + q31_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + const q31_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + } arm_fir_instance_q31; + + /** + * @brief Instance structure for the floating-point FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of filter coefficients in the filter. */ + float32_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + const float32_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + } arm_fir_instance_f32; + + /** + * @brief Processing function for the Q7 FIR filter. + * @param[in] S points to an instance of the Q7 FIR filter structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_fir_q7( + const arm_fir_instance_q7 * S, + const q7_t * pSrc, + q7_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the Q7 FIR filter. + * @param[in,out] S points to an instance of the Q7 FIR structure. + * @param[in] numTaps Number of filter coefficients in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of samples that are processed. + * + * For the MVE version, the coefficient length must be a multiple of 16. + * You can pad with zeros if you have less coefficients. + */ + void arm_fir_init_q7( + arm_fir_instance_q7 * S, + uint16_t numTaps, + const q7_t * pCoeffs, + q7_t * pState, + uint32_t blockSize); + + /** + * @brief Processing function for the Q15 FIR filter. + * @param[in] S points to an instance of the Q15 FIR structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_fir_q15( + const arm_fir_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + /** + * @brief Processing function for the fast Q15 FIR filter (fast version). + * @param[in] S points to an instance of the Q15 FIR filter structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_fir_fast_q15( + const arm_fir_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the Q15 FIR filter. + * @param[in,out] S points to an instance of the Q15 FIR filter structure. + * @param[in] numTaps Number of filter coefficients in the filter. Must be even and greater than or equal to 4. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of samples that are processed at a time. + * @return The function returns either + * ARM_MATH_SUCCESS if initialization was successful or + * ARM_MATH_ARGUMENT_ERROR if numTaps is not a supported value. + * + * For the MVE version, the coefficient length must be a multiple of 8. + * You can pad with zeros if you have less coefficients. + * + */ + arm_status arm_fir_init_q15( + arm_fir_instance_q15 * S, + uint16_t numTaps, + const q15_t * pCoeffs, + q15_t * pState, + uint32_t blockSize); + + /** + * @brief Processing function for the Q31 FIR filter. + * @param[in] S points to an instance of the Q31 FIR filter structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_fir_q31( + const arm_fir_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + /** + * @brief Processing function for the fast Q31 FIR filter (fast version). + * @param[in] S points to an instance of the Q31 FIR filter structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_fir_fast_q31( + const arm_fir_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the Q31 FIR filter. + * @param[in,out] S points to an instance of the Q31 FIR structure. + * @param[in] numTaps Number of filter coefficients in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of samples that are processed at a time. + * + * For the MVE version, the coefficient length must be a multiple of 4. + * You can pad with zeros if you have less coefficients. + */ + void arm_fir_init_q31( + arm_fir_instance_q31 * S, + uint16_t numTaps, + const q31_t * pCoeffs, + q31_t * pState, + uint32_t blockSize); + + /** + * @brief Processing function for the floating-point FIR filter. + * @param[in] S points to an instance of the floating-point FIR structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_fir_f32( + const arm_fir_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the floating-point FIR filter. + * @param[in,out] S points to an instance of the floating-point FIR filter structure. + * @param[in] numTaps Number of filter coefficients in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of samples that are processed at a time. + */ + void arm_fir_init_f32( + arm_fir_instance_f32 * S, + uint16_t numTaps, + const float32_t * pCoeffs, + float32_t * pState, + uint32_t blockSize); + + /** + * @brief Instance structure for the Q15 Biquad cascade filter. + */ + typedef struct + { + int8_t numStages; /**< number of 2nd order stages in the filter. Overall order is 2*numStages. */ + q15_t *pState; /**< Points to the array of state coefficients. The array is of length 4*numStages. */ + const q15_t *pCoeffs; /**< Points to the array of coefficients. The array is of length 5*numStages. */ + int8_t postShift; /**< Additional shift, in bits, applied to each output sample. */ + } arm_biquad_casd_df1_inst_q15; + + /** + * @brief Instance structure for the Q31 Biquad cascade filter. + */ + typedef struct + { + uint32_t numStages; /**< number of 2nd order stages in the filter. Overall order is 2*numStages. */ + q31_t *pState; /**< Points to the array of state coefficients. The array is of length 4*numStages. */ + const q31_t *pCoeffs; /**< Points to the array of coefficients. The array is of length 5*numStages. */ + uint8_t postShift; /**< Additional shift, in bits, applied to each output sample. */ + } arm_biquad_casd_df1_inst_q31; + + /** + * @brief Instance structure for the floating-point Biquad cascade filter. + */ + typedef struct + { + uint32_t numStages; /**< number of 2nd order stages in the filter. Overall order is 2*numStages. */ + float32_t *pState; /**< Points to the array of state coefficients. The array is of length 4*numStages. */ + const float32_t *pCoeffs; /**< Points to the array of coefficients. The array is of length 5*numStages. */ + } arm_biquad_casd_df1_inst_f32; + +#if defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE) + /** + * @brief Instance structure for the modified Biquad coefs required by vectorized code. + */ + typedef struct + { + float32_t coeffs[8][4]; /**< Points to the array of modified coefficients. The array is of length 32. There is one per stage */ + } arm_biquad_mod_coef_f32; +#endif + + /** + * @brief Processing function for the Q15 Biquad cascade filter. + * @param[in] S points to an instance of the Q15 Biquad cascade structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_df1_q15( + const arm_biquad_casd_df1_inst_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the Q15 Biquad cascade filter. + * @param[in,out] S points to an instance of the Q15 Biquad cascade structure. + * @param[in] numStages number of 2nd order stages in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] postShift Shift to be applied to the output. Varies according to the coefficients format + */ + void arm_biquad_cascade_df1_init_q15( + arm_biquad_casd_df1_inst_q15 * S, + uint8_t numStages, + const q15_t * pCoeffs, + q15_t * pState, + int8_t postShift); + + /** + * @brief Fast but less precise processing function for the Q15 Biquad cascade filter for Cortex-M3 and Cortex-M4. + * @param[in] S points to an instance of the Q15 Biquad cascade structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_df1_fast_q15( + const arm_biquad_casd_df1_inst_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + /** + * @brief Processing function for the Q31 Biquad cascade filter + * @param[in] S points to an instance of the Q31 Biquad cascade structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_df1_q31( + const arm_biquad_casd_df1_inst_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + /** + * @brief Fast but less precise processing function for the Q31 Biquad cascade filter for Cortex-M3 and Cortex-M4. + * @param[in] S points to an instance of the Q31 Biquad cascade structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_df1_fast_q31( + const arm_biquad_casd_df1_inst_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the Q31 Biquad cascade filter. + * @param[in,out] S points to an instance of the Q31 Biquad cascade structure. + * @param[in] numStages number of 2nd order stages in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] postShift Shift to be applied to the output. Varies according to the coefficients format + */ + void arm_biquad_cascade_df1_init_q31( + arm_biquad_casd_df1_inst_q31 * S, + uint8_t numStages, + const q31_t * pCoeffs, + q31_t * pState, + int8_t postShift); + + /** + * @brief Processing function for the floating-point Biquad cascade filter. + * @param[in] S points to an instance of the floating-point Biquad cascade structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_df1_f32( + const arm_biquad_casd_df1_inst_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the floating-point Biquad cascade filter. + * @param[in,out] S points to an instance of the floating-point Biquad cascade structure. + * @param[in] numStages number of 2nd order stages in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pCoeffsMod points to the modified filter coefficients (only MVE version). + * @param[in] pState points to the state buffer. + */ +#if defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE) + void arm_biquad_cascade_df1_mve_init_f32( + arm_biquad_casd_df1_inst_f32 * S, + uint8_t numStages, + const float32_t * pCoeffs, + arm_biquad_mod_coef_f32 * pCoeffsMod, + float32_t * pState); +#endif + + void arm_biquad_cascade_df1_init_f32( + arm_biquad_casd_df1_inst_f32 * S, + uint8_t numStages, + const float32_t * pCoeffs, + float32_t * pState); + + +/** + * @brief Convolution of floating-point sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the location where the output result is written. Length srcALen+srcBLen-1. + */ + void arm_conv_f32( + const float32_t * pSrcA, + uint32_t srcALen, + const float32_t * pSrcB, + uint32_t srcBLen, + float32_t * pDst); + + + /** + * @brief Convolution of Q15 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length srcALen+srcBLen-1. + * @param[in] pScratch1 points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + * @param[in] pScratch2 points to scratch buffer of size min(srcALen, srcBLen). + */ + void arm_conv_opt_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + q15_t * pScratch1, + q15_t * pScratch2); + + +/** + * @brief Convolution of Q15 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the location where the output result is written. Length srcALen+srcBLen-1. + */ + void arm_conv_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst); + + + /** + * @brief Convolution of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length srcALen+srcBLen-1. + */ + void arm_conv_fast_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst); + + + /** + * @brief Convolution of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length srcALen+srcBLen-1. + * @param[in] pScratch1 points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + * @param[in] pScratch2 points to scratch buffer of size min(srcALen, srcBLen). + */ + void arm_conv_fast_opt_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + q15_t * pScratch1, + q15_t * pScratch2); + + + /** + * @brief Convolution of Q31 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length srcALen+srcBLen-1. + */ + void arm_conv_q31( + const q31_t * pSrcA, + uint32_t srcALen, + const q31_t * pSrcB, + uint32_t srcBLen, + q31_t * pDst); + + + /** + * @brief Convolution of Q31 sequences (fast version) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length srcALen+srcBLen-1. + */ + void arm_conv_fast_q31( + const q31_t * pSrcA, + uint32_t srcALen, + const q31_t * pSrcB, + uint32_t srcBLen, + q31_t * pDst); + + + /** + * @brief Convolution of Q7 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length srcALen+srcBLen-1. + * @param[in] pScratch1 points to scratch buffer(of type q15_t) of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + * @param[in] pScratch2 points to scratch buffer (of type q15_t) of size min(srcALen, srcBLen). + */ + void arm_conv_opt_q7( + const q7_t * pSrcA, + uint32_t srcALen, + const q7_t * pSrcB, + uint32_t srcBLen, + q7_t * pDst, + q15_t * pScratch1, + q15_t * pScratch2); + + + /** + * @brief Convolution of Q7 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length srcALen+srcBLen-1. + */ + void arm_conv_q7( + const q7_t * pSrcA, + uint32_t srcALen, + const q7_t * pSrcB, + uint32_t srcBLen, + q7_t * pDst); + + + /** + * @brief Partial convolution of floating-point sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_f32( + const float32_t * pSrcA, + uint32_t srcALen, + const float32_t * pSrcB, + uint32_t srcBLen, + float32_t * pDst, + uint32_t firstIndex, + uint32_t numPoints); + + + /** + * @brief Partial convolution of Q15 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @param[in] pScratch1 points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + * @param[in] pScratch2 points to scratch buffer of size min(srcALen, srcBLen). + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_opt_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + uint32_t firstIndex, + uint32_t numPoints, + q15_t * pScratch1, + q15_t * pScratch2); + + + /** + * @brief Partial convolution of Q15 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + uint32_t firstIndex, + uint32_t numPoints); + + + /** + * @brief Partial convolution of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_fast_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + uint32_t firstIndex, + uint32_t numPoints); + + + /** + * @brief Partial convolution of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @param[in] pScratch1 points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + * @param[in] pScratch2 points to scratch buffer of size min(srcALen, srcBLen). + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_fast_opt_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + uint32_t firstIndex, + uint32_t numPoints, + q15_t * pScratch1, + q15_t * pScratch2); + + + /** + * @brief Partial convolution of Q31 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_q31( + const q31_t * pSrcA, + uint32_t srcALen, + const q31_t * pSrcB, + uint32_t srcBLen, + q31_t * pDst, + uint32_t firstIndex, + uint32_t numPoints); + + + /** + * @brief Partial convolution of Q31 sequences (fast version) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_fast_q31( + const q31_t * pSrcA, + uint32_t srcALen, + const q31_t * pSrcB, + uint32_t srcBLen, + q31_t * pDst, + uint32_t firstIndex, + uint32_t numPoints); + + + /** + * @brief Partial convolution of Q7 sequences + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @param[in] pScratch1 points to scratch buffer(of type q15_t) of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + * @param[in] pScratch2 points to scratch buffer (of type q15_t) of size min(srcALen, srcBLen). + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_opt_q7( + const q7_t * pSrcA, + uint32_t srcALen, + const q7_t * pSrcB, + uint32_t srcBLen, + q7_t * pDst, + uint32_t firstIndex, + uint32_t numPoints, + q15_t * pScratch1, + q15_t * pScratch2); + + +/** + * @brief Partial convolution of Q7 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data + * @param[in] firstIndex is the first output sample to start with. + * @param[in] numPoints is the number of output points to be computed. + * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2]. + */ + arm_status arm_conv_partial_q7( + const q7_t * pSrcA, + uint32_t srcALen, + const q7_t * pSrcB, + uint32_t srcBLen, + q7_t * pDst, + uint32_t firstIndex, + uint32_t numPoints); + + + /** + * @brief Instance structure for the Q15 FIR decimator. + */ + typedef struct + { + uint8_t M; /**< decimation factor. */ + uint16_t numTaps; /**< number of coefficients in the filter. */ + const q15_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + q15_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + } arm_fir_decimate_instance_q15; + + /** + * @brief Instance structure for the Q31 FIR decimator. + */ + typedef struct + { + uint8_t M; /**< decimation factor. */ + uint16_t numTaps; /**< number of coefficients in the filter. */ + const q31_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + q31_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + } arm_fir_decimate_instance_q31; + +/** + @brief Instance structure for floating-point FIR decimator. + */ +typedef struct + { + uint8_t M; /**< decimation factor. */ + uint16_t numTaps; /**< number of coefficients in the filter. */ + const float32_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + float32_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + } arm_fir_decimate_instance_f32; + + +/** + @brief Processing function for floating-point FIR decimator. + @param[in] S points to an instance of the floating-point FIR decimator structure + @param[in] pSrc points to the block of input data + @param[out] pDst points to the block of output data + @param[in] blockSize number of samples to process + */ +void arm_fir_decimate_f32( + const arm_fir_decimate_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + +/** + @brief Initialization function for the floating-point FIR decimator. + @param[in,out] S points to an instance of the floating-point FIR decimator structure + @param[in] numTaps number of coefficients in the filter + @param[in] M decimation factor + @param[in] pCoeffs points to the filter coefficients + @param[in] pState points to the state buffer + @param[in] blockSize number of input samples to process per call + @return execution status + - \ref ARM_MATH_SUCCESS : Operation successful + - \ref ARM_MATH_LENGTH_ERROR : blockSize is not a multiple of M + */ +arm_status arm_fir_decimate_init_f32( + arm_fir_decimate_instance_f32 * S, + uint16_t numTaps, + uint8_t M, + const float32_t * pCoeffs, + float32_t * pState, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q15 FIR decimator. + * @param[in] S points to an instance of the Q15 FIR decimator structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_decimate_q15( + const arm_fir_decimate_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q15 FIR decimator (fast variant) for Cortex-M3 and Cortex-M4. + * @param[in] S points to an instance of the Q15 FIR decimator structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_decimate_fast_q15( + const arm_fir_decimate_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q15 FIR decimator. + * @param[in,out] S points to an instance of the Q15 FIR decimator structure. + * @param[in] numTaps number of coefficients in the filter. + * @param[in] M decimation factor. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of input samples to process per call. + * @return The function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_LENGTH_ERROR if + * blockSize is not a multiple of M. + */ + arm_status arm_fir_decimate_init_q15( + arm_fir_decimate_instance_q15 * S, + uint16_t numTaps, + uint8_t M, + const q15_t * pCoeffs, + q15_t * pState, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q31 FIR decimator. + * @param[in] S points to an instance of the Q31 FIR decimator structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_decimate_q31( + const arm_fir_decimate_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + /** + * @brief Processing function for the Q31 FIR decimator (fast variant) for Cortex-M3 and Cortex-M4. + * @param[in] S points to an instance of the Q31 FIR decimator structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_decimate_fast_q31( + const arm_fir_decimate_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q31 FIR decimator. + * @param[in,out] S points to an instance of the Q31 FIR decimator structure. + * @param[in] numTaps number of coefficients in the filter. + * @param[in] M decimation factor. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of input samples to process per call. + * @return The function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_LENGTH_ERROR if + * blockSize is not a multiple of M. + */ + arm_status arm_fir_decimate_init_q31( + arm_fir_decimate_instance_q31 * S, + uint16_t numTaps, + uint8_t M, + const q31_t * pCoeffs, + q31_t * pState, + uint32_t blockSize); + + + /** + * @brief Instance structure for the Q15 FIR interpolator. + */ + typedef struct + { + uint8_t L; /**< upsample factor. */ + uint16_t phaseLength; /**< length of each polyphase filter component. */ + const q15_t *pCoeffs; /**< points to the coefficient array. The array is of length L*phaseLength. */ + q15_t *pState; /**< points to the state variable array. The array is of length blockSize+phaseLength-1. */ + } arm_fir_interpolate_instance_q15; + + /** + * @brief Instance structure for the Q31 FIR interpolator. + */ + typedef struct + { + uint8_t L; /**< upsample factor. */ + uint16_t phaseLength; /**< length of each polyphase filter component. */ + const q31_t *pCoeffs; /**< points to the coefficient array. The array is of length L*phaseLength. */ + q31_t *pState; /**< points to the state variable array. The array is of length blockSize+phaseLength-1. */ + } arm_fir_interpolate_instance_q31; + + /** + * @brief Instance structure for the floating-point FIR interpolator. + */ + typedef struct + { + uint8_t L; /**< upsample factor. */ + uint16_t phaseLength; /**< length of each polyphase filter component. */ + const float32_t *pCoeffs; /**< points to the coefficient array. The array is of length L*phaseLength. */ + float32_t *pState; /**< points to the state variable array. The array is of length phaseLength+numTaps-1. */ + } arm_fir_interpolate_instance_f32; + + + /** + * @brief Processing function for the Q15 FIR interpolator. + * @param[in] S points to an instance of the Q15 FIR interpolator structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_interpolate_q15( + const arm_fir_interpolate_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q15 FIR interpolator. + * @param[in,out] S points to an instance of the Q15 FIR interpolator structure. + * @param[in] L upsample factor. + * @param[in] numTaps number of filter coefficients in the filter. + * @param[in] pCoeffs points to the filter coefficient buffer. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of input samples to process per call. + * @return The function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_LENGTH_ERROR if + * the filter length numTaps is not a multiple of the interpolation factor L. + */ + arm_status arm_fir_interpolate_init_q15( + arm_fir_interpolate_instance_q15 * S, + uint8_t L, + uint16_t numTaps, + const q15_t * pCoeffs, + q15_t * pState, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q31 FIR interpolator. + * @param[in] S points to an instance of the Q15 FIR interpolator structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_interpolate_q31( + const arm_fir_interpolate_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q31 FIR interpolator. + * @param[in,out] S points to an instance of the Q31 FIR interpolator structure. + * @param[in] L upsample factor. + * @param[in] numTaps number of filter coefficients in the filter. + * @param[in] pCoeffs points to the filter coefficient buffer. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of input samples to process per call. + * @return The function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_LENGTH_ERROR if + * the filter length numTaps is not a multiple of the interpolation factor L. + */ + arm_status arm_fir_interpolate_init_q31( + arm_fir_interpolate_instance_q31 * S, + uint8_t L, + uint16_t numTaps, + const q31_t * pCoeffs, + q31_t * pState, + uint32_t blockSize); + + + /** + * @brief Processing function for the floating-point FIR interpolator. + * @param[in] S points to an instance of the floating-point FIR interpolator structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_interpolate_f32( + const arm_fir_interpolate_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the floating-point FIR interpolator. + * @param[in,out] S points to an instance of the floating-point FIR interpolator structure. + * @param[in] L upsample factor. + * @param[in] numTaps number of filter coefficients in the filter. + * @param[in] pCoeffs points to the filter coefficient buffer. + * @param[in] pState points to the state buffer. + * @param[in] blockSize number of input samples to process per call. + * @return The function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_LENGTH_ERROR if + * the filter length numTaps is not a multiple of the interpolation factor L. + */ + arm_status arm_fir_interpolate_init_f32( + arm_fir_interpolate_instance_f32 * S, + uint8_t L, + uint16_t numTaps, + const float32_t * pCoeffs, + float32_t * pState, + uint32_t blockSize); + + + /** + * @brief Instance structure for the high precision Q31 Biquad cascade filter. + */ + typedef struct + { + uint8_t numStages; /**< number of 2nd order stages in the filter. Overall order is 2*numStages. */ + q63_t *pState; /**< points to the array of state coefficients. The array is of length 4*numStages. */ + const q31_t *pCoeffs; /**< points to the array of coefficients. The array is of length 5*numStages. */ + uint8_t postShift; /**< additional shift, in bits, applied to each output sample. */ + } arm_biquad_cas_df1_32x64_ins_q31; + + + /** + * @param[in] S points to an instance of the high precision Q31 Biquad cascade filter structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cas_df1_32x64_q31( + const arm_biquad_cas_df1_32x64_ins_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @param[in,out] S points to an instance of the high precision Q31 Biquad cascade filter structure. + * @param[in] numStages number of 2nd order stages in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] postShift shift to be applied to the output. Varies according to the coefficients format + */ + void arm_biquad_cas_df1_32x64_init_q31( + arm_biquad_cas_df1_32x64_ins_q31 * S, + uint8_t numStages, + const q31_t * pCoeffs, + q63_t * pState, + uint8_t postShift); + + + /** + * @brief Instance structure for the floating-point transposed direct form II Biquad cascade filter. + */ + typedef struct + { + uint8_t numStages; /**< number of 2nd order stages in the filter. Overall order is 2*numStages. */ + float32_t *pState; /**< points to the array of state coefficients. The array is of length 2*numStages. */ + const float32_t *pCoeffs; /**< points to the array of coefficients. The array is of length 5*numStages. */ + } arm_biquad_cascade_df2T_instance_f32; + + /** + * @brief Instance structure for the floating-point transposed direct form II Biquad cascade filter. + */ + typedef struct + { + uint8_t numStages; /**< number of 2nd order stages in the filter. Overall order is 2*numStages. */ + float32_t *pState; /**< points to the array of state coefficients. The array is of length 4*numStages. */ + const float32_t *pCoeffs; /**< points to the array of coefficients. The array is of length 5*numStages. */ + } arm_biquad_cascade_stereo_df2T_instance_f32; + + /** + * @brief Instance structure for the floating-point transposed direct form II Biquad cascade filter. + */ + typedef struct + { + uint8_t numStages; /**< number of 2nd order stages in the filter. Overall order is 2*numStages. */ + float64_t *pState; /**< points to the array of state coefficients. The array is of length 2*numStages. */ + const float64_t *pCoeffs; /**< points to the array of coefficients. The array is of length 5*numStages. */ + } arm_biquad_cascade_df2T_instance_f64; + + + /** + * @brief Processing function for the floating-point transposed direct form II Biquad cascade filter. + * @param[in] S points to an instance of the filter data structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_df2T_f32( + const arm_biquad_cascade_df2T_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Processing function for the floating-point transposed direct form II Biquad cascade filter. 2 channels + * @param[in] S points to an instance of the filter data structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_stereo_df2T_f32( + const arm_biquad_cascade_stereo_df2T_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Processing function for the floating-point transposed direct form II Biquad cascade filter. + * @param[in] S points to an instance of the filter data structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of samples to process. + */ + void arm_biquad_cascade_df2T_f64( + const arm_biquad_cascade_df2T_instance_f64 * S, + const float64_t * pSrc, + float64_t * pDst, + uint32_t blockSize); + + +#if defined(ARM_MATH_NEON) +void arm_biquad_cascade_df2T_compute_coefs_f32( + arm_biquad_cascade_df2T_instance_f32 * S, + uint8_t numStages, + float32_t * pCoeffs); +#endif + /** + * @brief Initialization function for the floating-point transposed direct form II Biquad cascade filter. + * @param[in,out] S points to an instance of the filter data structure. + * @param[in] numStages number of 2nd order stages in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + */ + void arm_biquad_cascade_df2T_init_f32( + arm_biquad_cascade_df2T_instance_f32 * S, + uint8_t numStages, + const float32_t * pCoeffs, + float32_t * pState); + + + /** + * @brief Initialization function for the floating-point transposed direct form II Biquad cascade filter. + * @param[in,out] S points to an instance of the filter data structure. + * @param[in] numStages number of 2nd order stages in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + */ + void arm_biquad_cascade_stereo_df2T_init_f32( + arm_biquad_cascade_stereo_df2T_instance_f32 * S, + uint8_t numStages, + const float32_t * pCoeffs, + float32_t * pState); + + + /** + * @brief Initialization function for the floating-point transposed direct form II Biquad cascade filter. + * @param[in,out] S points to an instance of the filter data structure. + * @param[in] numStages number of 2nd order stages in the filter. + * @param[in] pCoeffs points to the filter coefficients. + * @param[in] pState points to the state buffer. + */ + void arm_biquad_cascade_df2T_init_f64( + arm_biquad_cascade_df2T_instance_f64 * S, + uint8_t numStages, + const float64_t * pCoeffs, + float64_t * pState); + + + /** + * @brief Instance structure for the Q15 FIR lattice filter. + */ + typedef struct + { + uint16_t numStages; /**< number of filter stages. */ + q15_t *pState; /**< points to the state variable array. The array is of length numStages. */ + const q15_t *pCoeffs; /**< points to the coefficient array. The array is of length numStages. */ + } arm_fir_lattice_instance_q15; + + /** + * @brief Instance structure for the Q31 FIR lattice filter. + */ + typedef struct + { + uint16_t numStages; /**< number of filter stages. */ + q31_t *pState; /**< points to the state variable array. The array is of length numStages. */ + const q31_t *pCoeffs; /**< points to the coefficient array. The array is of length numStages. */ + } arm_fir_lattice_instance_q31; + + /** + * @brief Instance structure for the floating-point FIR lattice filter. + */ + typedef struct + { + uint16_t numStages; /**< number of filter stages. */ + float32_t *pState; /**< points to the state variable array. The array is of length numStages. */ + const float32_t *pCoeffs; /**< points to the coefficient array. The array is of length numStages. */ + } arm_fir_lattice_instance_f32; + + + /** + * @brief Initialization function for the Q15 FIR lattice filter. + * @param[in] S points to an instance of the Q15 FIR lattice structure. + * @param[in] numStages number of filter stages. + * @param[in] pCoeffs points to the coefficient buffer. The array is of length numStages. + * @param[in] pState points to the state buffer. The array is of length numStages. + */ + void arm_fir_lattice_init_q15( + arm_fir_lattice_instance_q15 * S, + uint16_t numStages, + const q15_t * pCoeffs, + q15_t * pState); + + + /** + * @brief Processing function for the Q15 FIR lattice filter. + * @param[in] S points to an instance of the Q15 FIR lattice structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_fir_lattice_q15( + const arm_fir_lattice_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q31 FIR lattice filter. + * @param[in] S points to an instance of the Q31 FIR lattice structure. + * @param[in] numStages number of filter stages. + * @param[in] pCoeffs points to the coefficient buffer. The array is of length numStages. + * @param[in] pState points to the state buffer. The array is of length numStages. + */ + void arm_fir_lattice_init_q31( + arm_fir_lattice_instance_q31 * S, + uint16_t numStages, + const q31_t * pCoeffs, + q31_t * pState); + + + /** + * @brief Processing function for the Q31 FIR lattice filter. + * @param[in] S points to an instance of the Q31 FIR lattice structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of samples to process. + */ + void arm_fir_lattice_q31( + const arm_fir_lattice_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + +/** + * @brief Initialization function for the floating-point FIR lattice filter. + * @param[in] S points to an instance of the floating-point FIR lattice structure. + * @param[in] numStages number of filter stages. + * @param[in] pCoeffs points to the coefficient buffer. The array is of length numStages. + * @param[in] pState points to the state buffer. The array is of length numStages. + */ + void arm_fir_lattice_init_f32( + arm_fir_lattice_instance_f32 * S, + uint16_t numStages, + const float32_t * pCoeffs, + float32_t * pState); + + + /** + * @brief Processing function for the floating-point FIR lattice filter. + * @param[in] S points to an instance of the floating-point FIR lattice structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of samples to process. + */ + void arm_fir_lattice_f32( + const arm_fir_lattice_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Instance structure for the Q15 IIR lattice filter. + */ + typedef struct + { + uint16_t numStages; /**< number of stages in the filter. */ + q15_t *pState; /**< points to the state variable array. The array is of length numStages+blockSize. */ + q15_t *pkCoeffs; /**< points to the reflection coefficient array. The array is of length numStages. */ + q15_t *pvCoeffs; /**< points to the ladder coefficient array. The array is of length numStages+1. */ + } arm_iir_lattice_instance_q15; + + /** + * @brief Instance structure for the Q31 IIR lattice filter. + */ + typedef struct + { + uint16_t numStages; /**< number of stages in the filter. */ + q31_t *pState; /**< points to the state variable array. The array is of length numStages+blockSize. */ + q31_t *pkCoeffs; /**< points to the reflection coefficient array. The array is of length numStages. */ + q31_t *pvCoeffs; /**< points to the ladder coefficient array. The array is of length numStages+1. */ + } arm_iir_lattice_instance_q31; + + /** + * @brief Instance structure for the floating-point IIR lattice filter. + */ + typedef struct + { + uint16_t numStages; /**< number of stages in the filter. */ + float32_t *pState; /**< points to the state variable array. The array is of length numStages+blockSize. */ + float32_t *pkCoeffs; /**< points to the reflection coefficient array. The array is of length numStages. */ + float32_t *pvCoeffs; /**< points to the ladder coefficient array. The array is of length numStages+1. */ + } arm_iir_lattice_instance_f32; + + + /** + * @brief Processing function for the floating-point IIR lattice filter. + * @param[in] S points to an instance of the floating-point IIR lattice structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_iir_lattice_f32( + const arm_iir_lattice_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the floating-point IIR lattice filter. + * @param[in] S points to an instance of the floating-point IIR lattice structure. + * @param[in] numStages number of stages in the filter. + * @param[in] pkCoeffs points to the reflection coefficient buffer. The array is of length numStages. + * @param[in] pvCoeffs points to the ladder coefficient buffer. The array is of length numStages+1. + * @param[in] pState points to the state buffer. The array is of length numStages+blockSize-1. + * @param[in] blockSize number of samples to process. + */ + void arm_iir_lattice_init_f32( + arm_iir_lattice_instance_f32 * S, + uint16_t numStages, + float32_t * pkCoeffs, + float32_t * pvCoeffs, + float32_t * pState, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q31 IIR lattice filter. + * @param[in] S points to an instance of the Q31 IIR lattice structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_iir_lattice_q31( + const arm_iir_lattice_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q31 IIR lattice filter. + * @param[in] S points to an instance of the Q31 IIR lattice structure. + * @param[in] numStages number of stages in the filter. + * @param[in] pkCoeffs points to the reflection coefficient buffer. The array is of length numStages. + * @param[in] pvCoeffs points to the ladder coefficient buffer. The array is of length numStages+1. + * @param[in] pState points to the state buffer. The array is of length numStages+blockSize. + * @param[in] blockSize number of samples to process. + */ + void arm_iir_lattice_init_q31( + arm_iir_lattice_instance_q31 * S, + uint16_t numStages, + q31_t * pkCoeffs, + q31_t * pvCoeffs, + q31_t * pState, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q15 IIR lattice filter. + * @param[in] S points to an instance of the Q15 IIR lattice structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_iir_lattice_q15( + const arm_iir_lattice_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + +/** + * @brief Initialization function for the Q15 IIR lattice filter. + * @param[in] S points to an instance of the fixed-point Q15 IIR lattice structure. + * @param[in] numStages number of stages in the filter. + * @param[in] pkCoeffs points to reflection coefficient buffer. The array is of length numStages. + * @param[in] pvCoeffs points to ladder coefficient buffer. The array is of length numStages+1. + * @param[in] pState points to state buffer. The array is of length numStages+blockSize. + * @param[in] blockSize number of samples to process per call. + */ + void arm_iir_lattice_init_q15( + arm_iir_lattice_instance_q15 * S, + uint16_t numStages, + q15_t * pkCoeffs, + q15_t * pvCoeffs, + q15_t * pState, + uint32_t blockSize); + + + /** + * @brief Instance structure for the floating-point LMS filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + float32_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + float32_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + float32_t mu; /**< step size that controls filter coefficient updates. */ + } arm_lms_instance_f32; + + + /** + * @brief Processing function for floating-point LMS filter. + * @param[in] S points to an instance of the floating-point LMS filter structure. + * @param[in] pSrc points to the block of input data. + * @param[in] pRef points to the block of reference data. + * @param[out] pOut points to the block of output data. + * @param[out] pErr points to the block of error data. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_f32( + const arm_lms_instance_f32 * S, + const float32_t * pSrc, + float32_t * pRef, + float32_t * pOut, + float32_t * pErr, + uint32_t blockSize); + + + /** + * @brief Initialization function for floating-point LMS filter. + * @param[in] S points to an instance of the floating-point LMS filter structure. + * @param[in] numTaps number of filter coefficients. + * @param[in] pCoeffs points to the coefficient buffer. + * @param[in] pState points to state buffer. + * @param[in] mu step size that controls filter coefficient updates. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_init_f32( + arm_lms_instance_f32 * S, + uint16_t numTaps, + float32_t * pCoeffs, + float32_t * pState, + float32_t mu, + uint32_t blockSize); + + + /** + * @brief Instance structure for the Q15 LMS filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + q15_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + q15_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + q15_t mu; /**< step size that controls filter coefficient updates. */ + uint32_t postShift; /**< bit shift applied to coefficients. */ + } arm_lms_instance_q15; + + + /** + * @brief Initialization function for the Q15 LMS filter. + * @param[in] S points to an instance of the Q15 LMS filter structure. + * @param[in] numTaps number of filter coefficients. + * @param[in] pCoeffs points to the coefficient buffer. + * @param[in] pState points to the state buffer. + * @param[in] mu step size that controls filter coefficient updates. + * @param[in] blockSize number of samples to process. + * @param[in] postShift bit shift applied to coefficients. + */ + void arm_lms_init_q15( + arm_lms_instance_q15 * S, + uint16_t numTaps, + q15_t * pCoeffs, + q15_t * pState, + q15_t mu, + uint32_t blockSize, + uint32_t postShift); + + + /** + * @brief Processing function for Q15 LMS filter. + * @param[in] S points to an instance of the Q15 LMS filter structure. + * @param[in] pSrc points to the block of input data. + * @param[in] pRef points to the block of reference data. + * @param[out] pOut points to the block of output data. + * @param[out] pErr points to the block of error data. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_q15( + const arm_lms_instance_q15 * S, + const q15_t * pSrc, + q15_t * pRef, + q15_t * pOut, + q15_t * pErr, + uint32_t blockSize); + + + /** + * @brief Instance structure for the Q31 LMS filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + q31_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + q31_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + q31_t mu; /**< step size that controls filter coefficient updates. */ + uint32_t postShift; /**< bit shift applied to coefficients. */ + } arm_lms_instance_q31; + + + /** + * @brief Processing function for Q31 LMS filter. + * @param[in] S points to an instance of the Q15 LMS filter structure. + * @param[in] pSrc points to the block of input data. + * @param[in] pRef points to the block of reference data. + * @param[out] pOut points to the block of output data. + * @param[out] pErr points to the block of error data. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_q31( + const arm_lms_instance_q31 * S, + const q31_t * pSrc, + q31_t * pRef, + q31_t * pOut, + q31_t * pErr, + uint32_t blockSize); + + + /** + * @brief Initialization function for Q31 LMS filter. + * @param[in] S points to an instance of the Q31 LMS filter structure. + * @param[in] numTaps number of filter coefficients. + * @param[in] pCoeffs points to coefficient buffer. + * @param[in] pState points to state buffer. + * @param[in] mu step size that controls filter coefficient updates. + * @param[in] blockSize number of samples to process. + * @param[in] postShift bit shift applied to coefficients. + */ + void arm_lms_init_q31( + arm_lms_instance_q31 * S, + uint16_t numTaps, + q31_t * pCoeffs, + q31_t * pState, + q31_t mu, + uint32_t blockSize, + uint32_t postShift); + + + /** + * @brief Instance structure for the floating-point normalized LMS filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + float32_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + float32_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + float32_t mu; /**< step size that control filter coefficient updates. */ + float32_t energy; /**< saves previous frame energy. */ + float32_t x0; /**< saves previous input sample. */ + } arm_lms_norm_instance_f32; + + + /** + * @brief Processing function for floating-point normalized LMS filter. + * @param[in] S points to an instance of the floating-point normalized LMS filter structure. + * @param[in] pSrc points to the block of input data. + * @param[in] pRef points to the block of reference data. + * @param[out] pOut points to the block of output data. + * @param[out] pErr points to the block of error data. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_norm_f32( + arm_lms_norm_instance_f32 * S, + const float32_t * pSrc, + float32_t * pRef, + float32_t * pOut, + float32_t * pErr, + uint32_t blockSize); + + + /** + * @brief Initialization function for floating-point normalized LMS filter. + * @param[in] S points to an instance of the floating-point LMS filter structure. + * @param[in] numTaps number of filter coefficients. + * @param[in] pCoeffs points to coefficient buffer. + * @param[in] pState points to state buffer. + * @param[in] mu step size that controls filter coefficient updates. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_norm_init_f32( + arm_lms_norm_instance_f32 * S, + uint16_t numTaps, + float32_t * pCoeffs, + float32_t * pState, + float32_t mu, + uint32_t blockSize); + + + /** + * @brief Instance structure for the Q31 normalized LMS filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + q31_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + q31_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + q31_t mu; /**< step size that controls filter coefficient updates. */ + uint8_t postShift; /**< bit shift applied to coefficients. */ + const q31_t *recipTable; /**< points to the reciprocal initial value table. */ + q31_t energy; /**< saves previous frame energy. */ + q31_t x0; /**< saves previous input sample. */ + } arm_lms_norm_instance_q31; + + + /** + * @brief Processing function for Q31 normalized LMS filter. + * @param[in] S points to an instance of the Q31 normalized LMS filter structure. + * @param[in] pSrc points to the block of input data. + * @param[in] pRef points to the block of reference data. + * @param[out] pOut points to the block of output data. + * @param[out] pErr points to the block of error data. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_norm_q31( + arm_lms_norm_instance_q31 * S, + const q31_t * pSrc, + q31_t * pRef, + q31_t * pOut, + q31_t * pErr, + uint32_t blockSize); + + + /** + * @brief Initialization function for Q31 normalized LMS filter. + * @param[in] S points to an instance of the Q31 normalized LMS filter structure. + * @param[in] numTaps number of filter coefficients. + * @param[in] pCoeffs points to coefficient buffer. + * @param[in] pState points to state buffer. + * @param[in] mu step size that controls filter coefficient updates. + * @param[in] blockSize number of samples to process. + * @param[in] postShift bit shift applied to coefficients. + */ + void arm_lms_norm_init_q31( + arm_lms_norm_instance_q31 * S, + uint16_t numTaps, + q31_t * pCoeffs, + q31_t * pState, + q31_t mu, + uint32_t blockSize, + uint8_t postShift); + + + /** + * @brief Instance structure for the Q15 normalized LMS filter. + */ + typedef struct + { + uint16_t numTaps; /**< Number of coefficients in the filter. */ + q15_t *pState; /**< points to the state variable array. The array is of length numTaps+blockSize-1. */ + q15_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps. */ + q15_t mu; /**< step size that controls filter coefficient updates. */ + uint8_t postShift; /**< bit shift applied to coefficients. */ + const q15_t *recipTable; /**< Points to the reciprocal initial value table. */ + q15_t energy; /**< saves previous frame energy. */ + q15_t x0; /**< saves previous input sample. */ + } arm_lms_norm_instance_q15; + + + /** + * @brief Processing function for Q15 normalized LMS filter. + * @param[in] S points to an instance of the Q15 normalized LMS filter structure. + * @param[in] pSrc points to the block of input data. + * @param[in] pRef points to the block of reference data. + * @param[out] pOut points to the block of output data. + * @param[out] pErr points to the block of error data. + * @param[in] blockSize number of samples to process. + */ + void arm_lms_norm_q15( + arm_lms_norm_instance_q15 * S, + const q15_t * pSrc, + q15_t * pRef, + q15_t * pOut, + q15_t * pErr, + uint32_t blockSize); + + + /** + * @brief Initialization function for Q15 normalized LMS filter. + * @param[in] S points to an instance of the Q15 normalized LMS filter structure. + * @param[in] numTaps number of filter coefficients. + * @param[in] pCoeffs points to coefficient buffer. + * @param[in] pState points to state buffer. + * @param[in] mu step size that controls filter coefficient updates. + * @param[in] blockSize number of samples to process. + * @param[in] postShift bit shift applied to coefficients. + */ + void arm_lms_norm_init_q15( + arm_lms_norm_instance_q15 * S, + uint16_t numTaps, + q15_t * pCoeffs, + q15_t * pState, + q15_t mu, + uint32_t blockSize, + uint8_t postShift); + + + /** + * @brief Correlation of floating-point sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + */ + void arm_correlate_f32( + const float32_t * pSrcA, + uint32_t srcALen, + const float32_t * pSrcB, + uint32_t srcBLen, + float32_t * pDst); + + +/** + @brief Correlation of Q15 sequences + @param[in] pSrcA points to the first input sequence + @param[in] srcALen length of the first input sequence + @param[in] pSrcB points to the second input sequence + @param[in] srcBLen length of the second input sequence + @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + @param[in] pScratch points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. +*/ +void arm_correlate_opt_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + q15_t * pScratch); + + +/** + @brief Correlation of Q15 sequences. + @param[in] pSrcA points to the first input sequence + @param[in] srcALen length of the first input sequence + @param[in] pSrcB points to the second input sequence + @param[in] srcBLen length of the second input sequence + @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + */ + void arm_correlate_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst); + + +/** + @brief Correlation of Q15 sequences (fast version). + @param[in] pSrcA points to the first input sequence + @param[in] srcALen length of the first input sequence + @param[in] pSrcB points to the second input sequence + @param[in] srcBLen length of the second input sequence + @param[out] pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. + @return none + */ +void arm_correlate_fast_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst); + + +/** + @brief Correlation of Q15 sequences (fast version). + @param[in] pSrcA points to the first input sequence. + @param[in] srcALen length of the first input sequence. + @param[in] pSrcB points to the second input sequence. + @param[in] srcBLen length of the second input sequence. + @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + @param[in] pScratch points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + */ +void arm_correlate_fast_opt_q15( + const q15_t * pSrcA, + uint32_t srcALen, + const q15_t * pSrcB, + uint32_t srcBLen, + q15_t * pDst, + q15_t * pScratch); + + + /** + * @brief Correlation of Q31 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + */ + void arm_correlate_q31( + const q31_t * pSrcA, + uint32_t srcALen, + const q31_t * pSrcB, + uint32_t srcBLen, + q31_t * pDst); + + +/** + @brief Correlation of Q31 sequences (fast version). + @param[in] pSrcA points to the first input sequence + @param[in] srcALen length of the first input sequence + @param[in] pSrcB points to the second input sequence + @param[in] srcBLen length of the second input sequence + @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + */ +void arm_correlate_fast_q31( + const q31_t * pSrcA, + uint32_t srcALen, + const q31_t * pSrcB, + uint32_t srcBLen, + q31_t * pDst); + + + /** + * @brief Correlation of Q7 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + * @param[in] pScratch1 points to scratch buffer(of type q15_t) of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2. + * @param[in] pScratch2 points to scratch buffer (of type q15_t) of size min(srcALen, srcBLen). + */ + void arm_correlate_opt_q7( + const q7_t * pSrcA, + uint32_t srcALen, + const q7_t * pSrcB, + uint32_t srcBLen, + q7_t * pDst, + q15_t * pScratch1, + q15_t * pScratch2); + + + /** + * @brief Correlation of Q7 sequences. + * @param[in] pSrcA points to the first input sequence. + * @param[in] srcALen length of the first input sequence. + * @param[in] pSrcB points to the second input sequence. + * @param[in] srcBLen length of the second input sequence. + * @param[out] pDst points to the block of output data Length 2 * max(srcALen, srcBLen) - 1. + */ + void arm_correlate_q7( + const q7_t * pSrcA, + uint32_t srcALen, + const q7_t * pSrcB, + uint32_t srcBLen, + q7_t * pDst); + + + /** + * @brief Instance structure for the floating-point sparse FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + uint16_t stateIndex; /**< state buffer index. Points to the oldest sample in the state buffer. */ + float32_t *pState; /**< points to the state buffer array. The array is of length maxDelay+blockSize-1. */ + const float32_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + uint16_t maxDelay; /**< maximum offset specified by the pTapDelay array. */ + int32_t *pTapDelay; /**< points to the array of delay values. The array is of length numTaps. */ + } arm_fir_sparse_instance_f32; + + /** + * @brief Instance structure for the Q31 sparse FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + uint16_t stateIndex; /**< state buffer index. Points to the oldest sample in the state buffer. */ + q31_t *pState; /**< points to the state buffer array. The array is of length maxDelay+blockSize-1. */ + const q31_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + uint16_t maxDelay; /**< maximum offset specified by the pTapDelay array. */ + int32_t *pTapDelay; /**< points to the array of delay values. The array is of length numTaps. */ + } arm_fir_sparse_instance_q31; + + /** + * @brief Instance structure for the Q15 sparse FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + uint16_t stateIndex; /**< state buffer index. Points to the oldest sample in the state buffer. */ + q15_t *pState; /**< points to the state buffer array. The array is of length maxDelay+blockSize-1. */ + const q15_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + uint16_t maxDelay; /**< maximum offset specified by the pTapDelay array. */ + int32_t *pTapDelay; /**< points to the array of delay values. The array is of length numTaps. */ + } arm_fir_sparse_instance_q15; + + /** + * @brief Instance structure for the Q7 sparse FIR filter. + */ + typedef struct + { + uint16_t numTaps; /**< number of coefficients in the filter. */ + uint16_t stateIndex; /**< state buffer index. Points to the oldest sample in the state buffer. */ + q7_t *pState; /**< points to the state buffer array. The array is of length maxDelay+blockSize-1. */ + const q7_t *pCoeffs; /**< points to the coefficient array. The array is of length numTaps.*/ + uint16_t maxDelay; /**< maximum offset specified by the pTapDelay array. */ + int32_t *pTapDelay; /**< points to the array of delay values. The array is of length numTaps. */ + } arm_fir_sparse_instance_q7; + + + /** + * @brief Processing function for the floating-point sparse FIR filter. + * @param[in] S points to an instance of the floating-point sparse FIR structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] pScratchIn points to a temporary buffer of size blockSize. + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_sparse_f32( + arm_fir_sparse_instance_f32 * S, + const float32_t * pSrc, + float32_t * pDst, + float32_t * pScratchIn, + uint32_t blockSize); + + + /** + * @brief Initialization function for the floating-point sparse FIR filter. + * @param[in,out] S points to an instance of the floating-point sparse FIR structure. + * @param[in] numTaps number of nonzero coefficients in the filter. + * @param[in] pCoeffs points to the array of filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] pTapDelay points to the array of offset times. + * @param[in] maxDelay maximum offset time supported. + * @param[in] blockSize number of samples that will be processed per block. + */ + void arm_fir_sparse_init_f32( + arm_fir_sparse_instance_f32 * S, + uint16_t numTaps, + const float32_t * pCoeffs, + float32_t * pState, + int32_t * pTapDelay, + uint16_t maxDelay, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q31 sparse FIR filter. + * @param[in] S points to an instance of the Q31 sparse FIR structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] pScratchIn points to a temporary buffer of size blockSize. + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_sparse_q31( + arm_fir_sparse_instance_q31 * S, + const q31_t * pSrc, + q31_t * pDst, + q31_t * pScratchIn, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q31 sparse FIR filter. + * @param[in,out] S points to an instance of the Q31 sparse FIR structure. + * @param[in] numTaps number of nonzero coefficients in the filter. + * @param[in] pCoeffs points to the array of filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] pTapDelay points to the array of offset times. + * @param[in] maxDelay maximum offset time supported. + * @param[in] blockSize number of samples that will be processed per block. + */ + void arm_fir_sparse_init_q31( + arm_fir_sparse_instance_q31 * S, + uint16_t numTaps, + const q31_t * pCoeffs, + q31_t * pState, + int32_t * pTapDelay, + uint16_t maxDelay, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q15 sparse FIR filter. + * @param[in] S points to an instance of the Q15 sparse FIR structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] pScratchIn points to a temporary buffer of size blockSize. + * @param[in] pScratchOut points to a temporary buffer of size blockSize. + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_sparse_q15( + arm_fir_sparse_instance_q15 * S, + const q15_t * pSrc, + q15_t * pDst, + q15_t * pScratchIn, + q31_t * pScratchOut, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q15 sparse FIR filter. + * @param[in,out] S points to an instance of the Q15 sparse FIR structure. + * @param[in] numTaps number of nonzero coefficients in the filter. + * @param[in] pCoeffs points to the array of filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] pTapDelay points to the array of offset times. + * @param[in] maxDelay maximum offset time supported. + * @param[in] blockSize number of samples that will be processed per block. + */ + void arm_fir_sparse_init_q15( + arm_fir_sparse_instance_q15 * S, + uint16_t numTaps, + const q15_t * pCoeffs, + q15_t * pState, + int32_t * pTapDelay, + uint16_t maxDelay, + uint32_t blockSize); + + + /** + * @brief Processing function for the Q7 sparse FIR filter. + * @param[in] S points to an instance of the Q7 sparse FIR structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] pScratchIn points to a temporary buffer of size blockSize. + * @param[in] pScratchOut points to a temporary buffer of size blockSize. + * @param[in] blockSize number of input samples to process per call. + */ + void arm_fir_sparse_q7( + arm_fir_sparse_instance_q7 * S, + const q7_t * pSrc, + q7_t * pDst, + q7_t * pScratchIn, + q31_t * pScratchOut, + uint32_t blockSize); + + + /** + * @brief Initialization function for the Q7 sparse FIR filter. + * @param[in,out] S points to an instance of the Q7 sparse FIR structure. + * @param[in] numTaps number of nonzero coefficients in the filter. + * @param[in] pCoeffs points to the array of filter coefficients. + * @param[in] pState points to the state buffer. + * @param[in] pTapDelay points to the array of offset times. + * @param[in] maxDelay maximum offset time supported. + * @param[in] blockSize number of samples that will be processed per block. + */ + void arm_fir_sparse_init_q7( + arm_fir_sparse_instance_q7 * S, + uint16_t numTaps, + const q7_t * pCoeffs, + q7_t * pState, + int32_t * pTapDelay, + uint16_t maxDelay, + uint32_t blockSize); + + + + + + + /** + * @brief floating-point Circular write function. + */ + __STATIC_FORCEINLINE void arm_circularWrite_f32( + int32_t * circBuffer, + int32_t L, + uint16_t * writeOffset, + int32_t bufferInc, + const int32_t * src, + int32_t srcInc, + uint32_t blockSize) + { + uint32_t i = 0U; + int32_t wOffset; + + /* Copy the value of Index pointer that points + * to the current location where the input samples to be copied */ + wOffset = *writeOffset; + + /* Loop over the blockSize */ + i = blockSize; + + while (i > 0U) + { + /* copy the input sample to the circular buffer */ + circBuffer[wOffset] = *src; + + /* Update the input pointer */ + src += srcInc; + + /* Circularly update wOffset. Watch out for positive and negative value */ + wOffset += bufferInc; + if (wOffset >= L) + wOffset -= L; + + /* Decrement the loop counter */ + i--; + } + + /* Update the index pointer */ + *writeOffset = (uint16_t)wOffset; + } + + + + /** + * @brief floating-point Circular Read function. + */ + __STATIC_FORCEINLINE void arm_circularRead_f32( + int32_t * circBuffer, + int32_t L, + int32_t * readOffset, + int32_t bufferInc, + int32_t * dst, + int32_t * dst_base, + int32_t dst_length, + int32_t dstInc, + uint32_t blockSize) + { + uint32_t i = 0U; + int32_t rOffset; + int32_t* dst_end; + + /* Copy the value of Index pointer that points + * to the current location from where the input samples to be read */ + rOffset = *readOffset; + dst_end = dst_base + dst_length; + + /* Loop over the blockSize */ + i = blockSize; + + while (i > 0U) + { + /* copy the sample from the circular buffer to the destination buffer */ + *dst = circBuffer[rOffset]; + + /* Update the input pointer */ + dst += dstInc; + + if (dst == dst_end) + { + dst = dst_base; + } + + /* Circularly update rOffset. Watch out for positive and negative value */ + rOffset += bufferInc; + + if (rOffset >= L) + { + rOffset -= L; + } + + /* Decrement the loop counter */ + i--; + } + + /* Update the index pointer */ + *readOffset = rOffset; + } + + + /** + * @brief Q15 Circular write function. + */ + __STATIC_FORCEINLINE void arm_circularWrite_q15( + q15_t * circBuffer, + int32_t L, + uint16_t * writeOffset, + int32_t bufferInc, + const q15_t * src, + int32_t srcInc, + uint32_t blockSize) + { + uint32_t i = 0U; + int32_t wOffset; + + /* Copy the value of Index pointer that points + * to the current location where the input samples to be copied */ + wOffset = *writeOffset; + + /* Loop over the blockSize */ + i = blockSize; + + while (i > 0U) + { + /* copy the input sample to the circular buffer */ + circBuffer[wOffset] = *src; + + /* Update the input pointer */ + src += srcInc; + + /* Circularly update wOffset. Watch out for positive and negative value */ + wOffset += bufferInc; + if (wOffset >= L) + wOffset -= L; + + /* Decrement the loop counter */ + i--; + } + + /* Update the index pointer */ + *writeOffset = (uint16_t)wOffset; + } + + + /** + * @brief Q15 Circular Read function. + */ + __STATIC_FORCEINLINE void arm_circularRead_q15( + q15_t * circBuffer, + int32_t L, + int32_t * readOffset, + int32_t bufferInc, + q15_t * dst, + q15_t * dst_base, + int32_t dst_length, + int32_t dstInc, + uint32_t blockSize) + { + uint32_t i = 0; + int32_t rOffset; + q15_t* dst_end; + + /* Copy the value of Index pointer that points + * to the current location from where the input samples to be read */ + rOffset = *readOffset; + + dst_end = dst_base + dst_length; + + /* Loop over the blockSize */ + i = blockSize; + + while (i > 0U) + { + /* copy the sample from the circular buffer to the destination buffer */ + *dst = circBuffer[rOffset]; + + /* Update the input pointer */ + dst += dstInc; + + if (dst == dst_end) + { + dst = dst_base; + } + + /* Circularly update wOffset. Watch out for positive and negative value */ + rOffset += bufferInc; + + if (rOffset >= L) + { + rOffset -= L; + } + + /* Decrement the loop counter */ + i--; + } + + /* Update the index pointer */ + *readOffset = rOffset; + } + + + /** + * @brief Q7 Circular write function. + */ + __STATIC_FORCEINLINE void arm_circularWrite_q7( + q7_t * circBuffer, + int32_t L, + uint16_t * writeOffset, + int32_t bufferInc, + const q7_t * src, + int32_t srcInc, + uint32_t blockSize) + { + uint32_t i = 0U; + int32_t wOffset; + + /* Copy the value of Index pointer that points + * to the current location where the input samples to be copied */ + wOffset = *writeOffset; + + /* Loop over the blockSize */ + i = blockSize; + + while (i > 0U) + { + /* copy the input sample to the circular buffer */ + circBuffer[wOffset] = *src; + + /* Update the input pointer */ + src += srcInc; + + /* Circularly update wOffset. Watch out for positive and negative value */ + wOffset += bufferInc; + if (wOffset >= L) + wOffset -= L; + + /* Decrement the loop counter */ + i--; + } + + /* Update the index pointer */ + *writeOffset = (uint16_t)wOffset; + } + + + /** + * @brief Q7 Circular Read function. + */ + __STATIC_FORCEINLINE void arm_circularRead_q7( + q7_t * circBuffer, + int32_t L, + int32_t * readOffset, + int32_t bufferInc, + q7_t * dst, + q7_t * dst_base, + int32_t dst_length, + int32_t dstInc, + uint32_t blockSize) + { + uint32_t i = 0; + int32_t rOffset; + q7_t* dst_end; + + /* Copy the value of Index pointer that points + * to the current location from where the input samples to be read */ + rOffset = *readOffset; + + dst_end = dst_base + dst_length; + + /* Loop over the blockSize */ + i = blockSize; + + while (i > 0U) + { + /* copy the sample from the circular buffer to the destination buffer */ + *dst = circBuffer[rOffset]; + + /* Update the input pointer */ + dst += dstInc; + + if (dst == dst_end) + { + dst = dst_base; + } + + /* Circularly update rOffset. Watch out for positive and negative value */ + rOffset += bufferInc; + + if (rOffset >= L) + { + rOffset -= L; + } + + /* Decrement the loop counter */ + i--; + } + + /* Update the index pointer */ + *readOffset = rOffset; + } + + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _FILTERING_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/interpolation_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/interpolation_functions.h new file mode 100644 index 0000000..ca5a206 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/interpolation_functions.h @@ -0,0 +1,321 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file interpolation_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _INTERPOLATION_FUNCTIONS_H_ +#define _INTERPOLATION_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + +/** + * @defgroup groupInterpolation Interpolation Functions + * These functions perform 1- and 2-dimensional interpolation of data. + * Linear interpolation is used for 1-dimensional data and + * bilinear interpolation is used for 2-dimensional data. + */ + + + /** + * @brief Instance structure for the floating-point Linear Interpolate function. + */ + typedef struct + { + uint32_t nValues; /**< nValues */ + float32_t x1; /**< x1 */ + float32_t xSpacing; /**< xSpacing */ + float32_t *pYData; /**< pointer to the table of Y values */ + } arm_linear_interp_instance_f32; + + /** + * @brief Instance structure for the floating-point bilinear interpolation function. + */ + typedef struct + { + uint16_t numRows; /**< number of rows in the data table. */ + uint16_t numCols; /**< number of columns in the data table. */ + float32_t *pData; /**< points to the data table. */ + } arm_bilinear_interp_instance_f32; + + /** + * @brief Instance structure for the Q31 bilinear interpolation function. + */ + typedef struct + { + uint16_t numRows; /**< number of rows in the data table. */ + uint16_t numCols; /**< number of columns in the data table. */ + q31_t *pData; /**< points to the data table. */ + } arm_bilinear_interp_instance_q31; + + /** + * @brief Instance structure for the Q15 bilinear interpolation function. + */ + typedef struct + { + uint16_t numRows; /**< number of rows in the data table. */ + uint16_t numCols; /**< number of columns in the data table. */ + q15_t *pData; /**< points to the data table. */ + } arm_bilinear_interp_instance_q15; + + /** + * @brief Instance structure for the Q15 bilinear interpolation function. + */ + typedef struct + { + uint16_t numRows; /**< number of rows in the data table. */ + uint16_t numCols; /**< number of columns in the data table. */ + q7_t *pData; /**< points to the data table. */ + } arm_bilinear_interp_instance_q7; + + + /** + * @brief Struct for specifying cubic spline type + */ + typedef enum + { + ARM_SPLINE_NATURAL = 0, /**< Natural spline */ + ARM_SPLINE_PARABOLIC_RUNOUT = 1 /**< Parabolic runout spline */ + } arm_spline_type; + + /** + * @brief Instance structure for the floating-point cubic spline interpolation. + */ + typedef struct + { + arm_spline_type type; /**< Type (boundary conditions) */ + const float32_t * x; /**< x values */ + const float32_t * y; /**< y values */ + uint32_t n_x; /**< Number of known data points */ + float32_t * coeffs; /**< Coefficients buffer (b,c, and d) */ + } arm_spline_instance_f32; + + + + + /** + * @ingroup groupInterpolation + */ + + /** + * @addtogroup SplineInterpolate + * @{ + */ + + + /** + * @brief Processing function for the floating-point cubic spline interpolation. + * @param[in] S points to an instance of the floating-point spline structure. + * @param[in] xq points to the x values ot the interpolated data points. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples of output data. + */ + void arm_spline_f32( + arm_spline_instance_f32 * S, + const float32_t * xq, + float32_t * pDst, + uint32_t blockSize); + + /** + * @brief Initialization function for the floating-point cubic spline interpolation. + * @param[in,out] S points to an instance of the floating-point spline structure. + * @param[in] type type of cubic spline interpolation (boundary conditions) + * @param[in] x points to the x values of the known data points. + * @param[in] y points to the y values of the known data points. + * @param[in] n number of known data points. + * @param[in] coeffs coefficients array for b, c, and d + * @param[in] tempBuffer buffer array for internal computations + */ + void arm_spline_init_f32( + arm_spline_instance_f32 * S, + arm_spline_type type, + const float32_t * x, + const float32_t * y, + uint32_t n, + float32_t * coeffs, + float32_t * tempBuffer); + + + /** + * @} end of SplineInterpolate group + */ + + + + /** + * @addtogroup LinearInterpolate + * @{ + */ + + /** + * @brief Process function for the floating-point Linear Interpolation Function. + * @param[in,out] S is an instance of the floating-point Linear Interpolation structure + * @param[in] x input sample to process + * @return y processed output sample. + * + */ + float32_t arm_linear_interp_f32( + arm_linear_interp_instance_f32 * S, + float32_t x); + + /** + * + * @brief Process function for the Q31 Linear Interpolation Function. + * @param[in] pYData pointer to Q31 Linear Interpolation table + * @param[in] x input sample to process + * @param[in] nValues number of table values + * @return y processed output sample. + * + * \par + * Input sample x is in 12.20 format which contains 12 bits for table index and 20 bits for fractional part. + * This function can support maximum of table size 2^12. + * + */ + q31_t arm_linear_interp_q31( + q31_t * pYData, + q31_t x, + uint32_t nValues); + + /** + * + * @brief Process function for the Q15 Linear Interpolation Function. + * @param[in] pYData pointer to Q15 Linear Interpolation table + * @param[in] x input sample to process + * @param[in] nValues number of table values + * @return y processed output sample. + * + * \par + * Input sample x is in 12.20 format which contains 12 bits for table index and 20 bits for fractional part. + * This function can support maximum of table size 2^12. + * + */ + q15_t arm_linear_interp_q15( + q15_t * pYData, + q31_t x, + uint32_t nValues); + + /** + * + * @brief Process function for the Q7 Linear Interpolation Function. + * @param[in] pYData pointer to Q7 Linear Interpolation table + * @param[in] x input sample to process + * @param[in] nValues number of table values + * @return y processed output sample. + * + * \par + * Input sample x is in 12.20 format which contains 12 bits for table index and 20 bits for fractional part. + * This function can support maximum of table size 2^12. + */ +q7_t arm_linear_interp_q7( + q7_t * pYData, + q31_t x, + uint32_t nValues); + + /** + * @} end of LinearInterpolate group + */ + + + + + /** + * @ingroup groupInterpolation + */ + + + /** + * @addtogroup BilinearInterpolate + * @{ + */ + + /** + * @brief Floating-point bilinear interpolation. + * @param[in,out] S points to an instance of the interpolation structure. + * @param[in] X interpolation coordinate. + * @param[in] Y interpolation coordinate. + * @return out interpolated value. + */ + float32_t arm_bilinear_interp_f32( + const arm_bilinear_interp_instance_f32 * S, + float32_t X, + float32_t Y); + + /** + * @brief Q31 bilinear interpolation. + * @param[in,out] S points to an instance of the interpolation structure. + * @param[in] X interpolation coordinate in 12.20 format. + * @param[in] Y interpolation coordinate in 12.20 format. + * @return out interpolated value. + */ + q31_t arm_bilinear_interp_q31( + arm_bilinear_interp_instance_q31 * S, + q31_t X, + q31_t Y); + + + /** + * @brief Q15 bilinear interpolation. + * @param[in,out] S points to an instance of the interpolation structure. + * @param[in] X interpolation coordinate in 12.20 format. + * @param[in] Y interpolation coordinate in 12.20 format. + * @return out interpolated value. + */ + q15_t arm_bilinear_interp_q15( + arm_bilinear_interp_instance_q15 * S, + q31_t X, + q31_t Y); + + /** + * @brief Q7 bilinear interpolation. + * @param[in,out] S points to an instance of the interpolation structure. + * @param[in] X interpolation coordinate in 12.20 format. + * @param[in] Y interpolation coordinate in 12.20 format. + * @return out interpolated value. + */ + q7_t arm_bilinear_interp_q7( + arm_bilinear_interp_instance_q7 * S, + q31_t X, + q31_t Y); + /** + * @} end of BilinearInterpolate group + */ + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _INTERPOLATION_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/matrix_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/matrix_functions.h new file mode 100644 index 0000000..f51935a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/matrix_functions.h @@ -0,0 +1,600 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file matrix_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _MATRIX_FUNCTIONS_H_ +#define _MATRIX_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + * @defgroup groupMatrix Matrix Functions + * + * This set of functions provides basic matrix math operations. + * The functions operate on matrix data structures. For example, + * the type + * definition for the floating-point matrix structure is shown + * below: + * 
+ *     typedef struct
+ *     {
+ *       uint16_t numRows;     // number of rows of the matrix.
+ *       uint16_t numCols;     // number of columns of the matrix.
+ *       float32_t *pData;     // points to the data of the matrix.
+ *     } arm_matrix_instance_f32;
+ *
+ * There are similar definitions for Q15 and Q31 data types. + * + * The structure specifies the size of the matrix and then points to + * an array of data. The array is of size numRows X numCols + * and the values are arranged in row order. That is, the + * matrix element (i, j) is stored at: + * 
+ *     pData[i*numCols + j]
+ *
+ * + * \par Init Functions + * There is an associated initialization function for each type of matrix + * data structure. + * The initialization function sets the values of the internal structure fields. + * Refer to \ref arm_mat_init_f32(), \ref arm_mat_init_q31() and \ref arm_mat_init_q15() + * for floating-point, Q31 and Q15 types, respectively. + * + * \par + * Use of the initialization function is optional. However, if initialization function is used + * then the instance structure cannot be placed into a const data section. + * To place the instance structure in a const data + * section, manually initialize the data structure. For example: + * 
+ * arm_matrix_instance_f32 S = {nRows, nColumns, pData};
+ * arm_matrix_instance_q31 S = {nRows, nColumns, pData};
+ * arm_matrix_instance_q15 S = {nRows, nColumns, pData};
+ *
+ * where nRows specifies the number of rows, nColumns + * specifies the number of columns, and pData points to the + * data array. + * + * \par Size Checking + * By default all of the matrix functions perform size checking on the input and + * output matrices. For example, the matrix addition function verifies that the + * two input matrices and the output matrix all have the same number of rows and + * columns. If the size check fails the functions return: + * 
+ *     ARM_MATH_SIZE_MISMATCH
+ *
+ * Otherwise the functions return + * 
+ *     ARM_MATH_SUCCESS
+ *
+ * There is some overhead associated with this matrix size checking. + * The matrix size checking is enabled via the \#define + * 
+ *     ARM_MATH_MATRIX_CHECK
+ *
+ * within the library project settings. By default this macro is defined + * and size checking is enabled. By changing the project settings and + * undefining this macro size checking is eliminated and the functions + * run a bit faster. With size checking disabled the functions always + * return ARM_MATH_SUCCESS. + */ + + /** + * @brief Instance structure for the floating-point matrix structure. + */ + typedef struct + { + uint16_t numRows; /**< number of rows of the matrix. */ + uint16_t numCols; /**< number of columns of the matrix. */ + float32_t *pData; /**< points to the data of the matrix. */ + } arm_matrix_instance_f32; + + /** + * @brief Instance structure for the floating-point matrix structure. + */ + typedef struct + { + uint16_t numRows; /**< number of rows of the matrix. */ + uint16_t numCols; /**< number of columns of the matrix. */ + float64_t *pData; /**< points to the data of the matrix. */ + } arm_matrix_instance_f64; + + /** + * @brief Instance structure for the Q7 matrix structure. + */ + typedef struct + { + uint16_t numRows; /**< number of rows of the matrix. */ + uint16_t numCols; /**< number of columns of the matrix. */ + q7_t *pData; /**< points to the data of the matrix. */ + } arm_matrix_instance_q7; + + /** + * @brief Instance structure for the Q15 matrix structure. + */ + typedef struct + { + uint16_t numRows; /**< number of rows of the matrix. */ + uint16_t numCols; /**< number of columns of the matrix. */ + q15_t *pData; /**< points to the data of the matrix. */ + } arm_matrix_instance_q15; + + /** + * @brief Instance structure for the Q31 matrix structure. + */ + typedef struct + { + uint16_t numRows; /**< number of rows of the matrix. */ + uint16_t numCols; /**< number of columns of the matrix. */ + q31_t *pData; /**< points to the data of the matrix. */ + } arm_matrix_instance_q31; + + /** + * @brief Floating-point matrix addition. + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_add_f32( + const arm_matrix_instance_f32 * pSrcA, + const arm_matrix_instance_f32 * pSrcB, + arm_matrix_instance_f32 * pDst); + + /** + * @brief Q15 matrix addition. + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_add_q15( + const arm_matrix_instance_q15 * pSrcA, + const arm_matrix_instance_q15 * pSrcB, + arm_matrix_instance_q15 * pDst); + + /** + * @brief Q31 matrix addition. + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_add_q31( + const arm_matrix_instance_q31 * pSrcA, + const arm_matrix_instance_q31 * pSrcB, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Floating-point, complex, matrix multiplication. + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_cmplx_mult_f32( + const arm_matrix_instance_f32 * pSrcA, + const arm_matrix_instance_f32 * pSrcB, + arm_matrix_instance_f32 * pDst); + + /** + * @brief Q15, complex, matrix multiplication. + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_cmplx_mult_q15( + const arm_matrix_instance_q15 * pSrcA, + const arm_matrix_instance_q15 * pSrcB, + arm_matrix_instance_q15 * pDst, + q15_t * pScratch); + + /** + * @brief Q31, complex, matrix multiplication. + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_cmplx_mult_q31( + const arm_matrix_instance_q31 * pSrcA, + const arm_matrix_instance_q31 * pSrcB, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Floating-point matrix transpose. + * @param[in] pSrc points to the input matrix + * @param[out] pDst points to the output matrix + * @return The function returns either ARM_MATH_SIZE_MISMATCH + * or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_trans_f32( + const arm_matrix_instance_f32 * pSrc, + arm_matrix_instance_f32 * pDst); + + /** + * @brief Floating-point complex matrix transpose. + * @param[in] pSrc points to the input matrix + * @param[out] pDst points to the output matrix + * @return The function returns either ARM_MATH_SIZE_MISMATCH + * or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_cmplx_trans_f32( + const arm_matrix_instance_f32 * pSrc, + arm_matrix_instance_f32 * pDst); + + + /** + * @brief Q15 matrix transpose. + * @param[in] pSrc points to the input matrix + * @param[out] pDst points to the output matrix + * @return The function returns either ARM_MATH_SIZE_MISMATCH + * or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_trans_q15( + const arm_matrix_instance_q15 * pSrc, + arm_matrix_instance_q15 * pDst); + + /** + * @brief Q15 complex matrix transpose. + * @param[in] pSrc points to the input matrix + * @param[out] pDst points to the output matrix + * @return The function returns either ARM_MATH_SIZE_MISMATCH + * or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_cmplx_trans_q15( + const arm_matrix_instance_q15 * pSrc, + arm_matrix_instance_q15 * pDst); + + /** + * @brief Q7 matrix transpose. + * @param[in] pSrc points to the input matrix + * @param[out] pDst points to the output matrix + * @return The function returns either ARM_MATH_SIZE_MISMATCH + * or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_trans_q7( + const arm_matrix_instance_q7 * pSrc, + arm_matrix_instance_q7 * pDst); + + /** + * @brief Q31 matrix transpose. + * @param[in] pSrc points to the input matrix + * @param[out] pDst points to the output matrix + * @return The function returns either ARM_MATH_SIZE_MISMATCH + * or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_trans_q31( + const arm_matrix_instance_q31 * pSrc, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Q31 complex matrix transpose. + * @param[in] pSrc points to the input matrix + * @param[out] pDst points to the output matrix + * @return The function returns either ARM_MATH_SIZE_MISMATCH + * or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_cmplx_trans_q31( + const arm_matrix_instance_q31 * pSrc, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Floating-point matrix multiplication + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_mult_f32( + const arm_matrix_instance_f32 * pSrcA, + const arm_matrix_instance_f32 * pSrcB, + arm_matrix_instance_f32 * pDst); + + /** + * @brief Floating-point matrix and vector multiplication + * @param[in] pSrcMat points to the input matrix structure + * @param[in] pVec points to vector + * @param[out] pDst points to output vector + */ +void arm_mat_vec_mult_f32( + const arm_matrix_instance_f32 *pSrcMat, + const float32_t *pVec, + float32_t *pDst); + + /** + * @brief Q7 matrix multiplication + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @param[in] pState points to the array for storing intermediate results + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_mult_q7( + const arm_matrix_instance_q7 * pSrcA, + const arm_matrix_instance_q7 * pSrcB, + arm_matrix_instance_q7 * pDst, + q7_t * pState); + + /** + * @brief Q7 matrix and vector multiplication + * @param[in] pSrcMat points to the input matrix structure + * @param[in] pVec points to vector + * @param[out] pDst points to output vector + */ +void arm_mat_vec_mult_q7( + const arm_matrix_instance_q7 *pSrcMat, + const q7_t *pVec, + q7_t *pDst); + + /** + * @brief Q15 matrix multiplication + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @param[in] pState points to the array for storing intermediate results + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_mult_q15( + const arm_matrix_instance_q15 * pSrcA, + const arm_matrix_instance_q15 * pSrcB, + arm_matrix_instance_q15 * pDst, + q15_t * pState); + + /** + * @brief Q15 matrix and vector multiplication + * @param[in] pSrcMat points to the input matrix structure + * @param[in] pVec points to vector + * @param[out] pDst points to output vector + */ +void arm_mat_vec_mult_q15( + const arm_matrix_instance_q15 *pSrcMat, + const q15_t *pVec, + q15_t *pDst); + + /** + * @brief Q15 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @param[in] pState points to the array for storing intermediate results + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_mult_fast_q15( + const arm_matrix_instance_q15 * pSrcA, + const arm_matrix_instance_q15 * pSrcB, + arm_matrix_instance_q15 * pDst, + q15_t * pState); + + /** + * @brief Q31 matrix multiplication + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_mult_q31( + const arm_matrix_instance_q31 * pSrcA, + const arm_matrix_instance_q31 * pSrcB, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Q31 matrix and vector multiplication + * @param[in] pSrcMat points to the input matrix structure + * @param[in] pVec points to vector + * @param[out] pDst points to output vector + */ +void arm_mat_vec_mult_q31( + const arm_matrix_instance_q31 *pSrcMat, + const q31_t *pVec, + q31_t *pDst); + + /** + * @brief Q31 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4 + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_mult_fast_q31( + const arm_matrix_instance_q31 * pSrcA, + const arm_matrix_instance_q31 * pSrcB, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Floating-point matrix subtraction + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_sub_f32( + const arm_matrix_instance_f32 * pSrcA, + const arm_matrix_instance_f32 * pSrcB, + arm_matrix_instance_f32 * pDst); + + /** + * @brief Q15 matrix subtraction + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_sub_q15( + const arm_matrix_instance_q15 * pSrcA, + const arm_matrix_instance_q15 * pSrcB, + arm_matrix_instance_q15 * pDst); + + /** + * @brief Q31 matrix subtraction + * @param[in] pSrcA points to the first input matrix structure + * @param[in] pSrcB points to the second input matrix structure + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_sub_q31( + const arm_matrix_instance_q31 * pSrcA, + const arm_matrix_instance_q31 * pSrcB, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Floating-point matrix scaling. + * @param[in] pSrc points to the input matrix + * @param[in] scale scale factor + * @param[out] pDst points to the output matrix + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_scale_f32( + const arm_matrix_instance_f32 * pSrc, + float32_t scale, + arm_matrix_instance_f32 * pDst); + + /** + * @brief Q15 matrix scaling. + * @param[in] pSrc points to input matrix + * @param[in] scaleFract fractional portion of the scale factor + * @param[in] shift number of bits to shift the result by + * @param[out] pDst points to output matrix + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_scale_q15( + const arm_matrix_instance_q15 * pSrc, + q15_t scaleFract, + int32_t shift, + arm_matrix_instance_q15 * pDst); + + /** + * @brief Q31 matrix scaling. + * @param[in] pSrc points to input matrix + * @param[in] scaleFract fractional portion of the scale factor + * @param[in] shift number of bits to shift the result by + * @param[out] pDst points to output matrix structure + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + */ +arm_status arm_mat_scale_q31( + const arm_matrix_instance_q31 * pSrc, + q31_t scaleFract, + int32_t shift, + arm_matrix_instance_q31 * pDst); + + /** + * @brief Q31 matrix initialization. + * @param[in,out] S points to an instance of the floating-point matrix structure. + * @param[in] nRows number of rows in the matrix. + * @param[in] nColumns number of columns in the matrix. + * @param[in] pData points to the matrix data array. + */ +void arm_mat_init_q31( + arm_matrix_instance_q31 * S, + uint16_t nRows, + uint16_t nColumns, + q31_t * pData); + + /** + * @brief Q15 matrix initialization. + * @param[in,out] S points to an instance of the floating-point matrix structure. + * @param[in] nRows number of rows in the matrix. + * @param[in] nColumns number of columns in the matrix. + * @param[in] pData points to the matrix data array. + */ +void arm_mat_init_q15( + arm_matrix_instance_q15 * S, + uint16_t nRows, + uint16_t nColumns, + q15_t * pData); + + /** + * @brief Floating-point matrix initialization. + * @param[in,out] S points to an instance of the floating-point matrix structure. + * @param[in] nRows number of rows in the matrix. + * @param[in] nColumns number of columns in the matrix. + * @param[in] pData points to the matrix data array. + */ +void arm_mat_init_f32( + arm_matrix_instance_f32 * S, + uint16_t nRows, + uint16_t nColumns, + float32_t * pData); + + + + /** + * @brief Floating-point matrix inverse. + * @param[in] src points to the instance of the input floating-point matrix structure. + * @param[out] dst points to the instance of the output floating-point matrix structure. + * @return The function returns ARM_MATH_SIZE_MISMATCH, if the dimensions do not match. + * If the input matrix is singular (does not have an inverse), then the algorithm terminates and returns error status ARM_MATH_SINGULAR. + */ + arm_status arm_mat_inverse_f32( + const arm_matrix_instance_f32 * src, + arm_matrix_instance_f32 * dst); + + + /** + * @brief Floating-point matrix inverse. + * @param[in] src points to the instance of the input floating-point matrix structure. + * @param[out] dst points to the instance of the output floating-point matrix structure. + * @return The function returns ARM_MATH_SIZE_MISMATCH, if the dimensions do not match. + * If the input matrix is singular (does not have an inverse), then the algorithm terminates and returns error status ARM_MATH_SINGULAR. + */ + arm_status arm_mat_inverse_f64( + const arm_matrix_instance_f64 * src, + arm_matrix_instance_f64 * dst); + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _MATRIX_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h new file mode 100644 index 0000000..962088e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h @@ -0,0 +1,579 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file none.h + * @brief Intrinsincs when no DSP extension available + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + +Definitions in this file are allowing to reuse some versions of the +CMSIS-DSP to build on a core (M0 for instance) or a host where +DSP extension are not available. + +Ideally a pure C version should have been used instead. +But those are not always available or use a restricted set +of intrinsics. + +*/ + +#ifndef _NONE_H_ +#define _NONE_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + + +/* + +Normally those kind of definitions are in a compiler file +in Core or Core_A. + +But for MSVC compiler it is a bit special. The goal is very specific +to CMSIS-DSP and only to allow the use of this library from other +systems like Python or Matlab. + +MSVC is not going to be used to cross-compile to ARM. So, having a MSVC +compiler file in Core or Core_A would not make sense. + +*/ +#if defined ( _MSC_VER ) || defined(__GNUC_PYTHON__) + __STATIC_FORCEINLINE uint8_t __CLZ(uint32_t data) + { + if (data == 0U) { return 32U; } + + uint32_t count = 0U; + uint32_t mask = 0x80000000U; + + while ((data & mask) == 0U) + { + count += 1U; + mask = mask >> 1U; + } + return count; + } + + __STATIC_FORCEINLINE int32_t __SSAT(int32_t val, uint32_t sat) + { + if ((sat >= 1U) && (sat <= 32U)) + { + const int32_t max = (int32_t)((1U << (sat - 1U)) - 1U); + const int32_t min = -1 - max ; + if (val > max) + { + return max; + } + else if (val < min) + { + return min; + } + } + return val; + } + + __STATIC_FORCEINLINE uint32_t __USAT(int32_t val, uint32_t sat) + { + if (sat <= 31U) + { + const uint32_t max = ((1U << sat) - 1U); + if (val > (int32_t)max) + { + return max; + } + else if (val < 0) + { + return 0U; + } + } + return (uint32_t)val; + } + + /** + \brief Rotate Right in unsigned value (32 bit) + \details Rotate Right (immediate) provides the value of the contents of a register rotated by a variable number of bits. + \param [in] op1 Value to rotate + \param [in] op2 Number of Bits to rotate + \return Rotated value + */ +__STATIC_FORCEINLINE uint32_t __ROR(uint32_t op1, uint32_t op2) +{ + op2 %= 32U; + if (op2 == 0U) + { + return op1; + } + return (op1 >> op2) | (op1 << (32U - op2)); +} + + +#endif + +/** + * @brief Clips Q63 to Q31 values. + */ + __STATIC_FORCEINLINE q31_t clip_q63_to_q31( + q63_t x) + { + return ((q31_t) (x >> 32) != ((q31_t) x >> 31)) ? + ((0x7FFFFFFF ^ ((q31_t) (x >> 63)))) : (q31_t) x; + } + + /** + * @brief Clips Q63 to Q15 values. + */ + __STATIC_FORCEINLINE q15_t clip_q63_to_q15( + q63_t x) + { + return ((q31_t) (x >> 32) != ((q31_t) x >> 31)) ? + ((0x7FFF ^ ((q15_t) (x >> 63)))) : (q15_t) (x >> 15); + } + + /** + * @brief Clips Q31 to Q7 values. + */ + __STATIC_FORCEINLINE q7_t clip_q31_to_q7( + q31_t x) + { + return ((q31_t) (x >> 24) != ((q31_t) x >> 23)) ? + ((0x7F ^ ((q7_t) (x >> 31)))) : (q7_t) x; + } + + /** + * @brief Clips Q31 to Q15 values. + */ + __STATIC_FORCEINLINE q15_t clip_q31_to_q15( + q31_t x) + { + return ((q31_t) (x >> 16) != ((q31_t) x >> 15)) ? + ((0x7FFF ^ ((q15_t) (x >> 31)))) : (q15_t) x; + } + + /** + * @brief Multiplies 32 X 64 and returns 32 bit result in 2.30 format. + */ + __STATIC_FORCEINLINE q63_t mult32x64( + q63_t x, + q31_t y) + { + return ((((q63_t) (x & 0x00000000FFFFFFFF) * y) >> 32) + + (((q63_t) (x >> 32) * y) ) ); + } + +/* SMMLAR */ +#define multAcc_32x32_keep32_R(a, x, y) \ + a = (q31_t) (((((q63_t) a) << 32) + ((q63_t) x * y) + 0x80000000LL ) >> 32) + +/* SMMLSR */ +#define multSub_32x32_keep32_R(a, x, y) \ + a = (q31_t) (((((q63_t) a) << 32) - ((q63_t) x * y) + 0x80000000LL ) >> 32) + +/* SMMULR */ +#define mult_32x32_keep32_R(a, x, y) \ + a = (q31_t) (((q63_t) x * y + 0x80000000LL ) >> 32) + +/* SMMLA */ +#define multAcc_32x32_keep32(a, x, y) \ + a += (q31_t) (((q63_t) x * y) >> 32) + +/* SMMLS */ +#define multSub_32x32_keep32(a, x, y) \ + a -= (q31_t) (((q63_t) x * y) >> 32) + +/* SMMUL */ +#define mult_32x32_keep32(a, x, y) \ + a = (q31_t) (((q63_t) x * y ) >> 32) + +#ifndef ARM_MATH_DSP + /** + * @brief definition to pack two 16 bit values. + */ + #define __PKHBT(ARG1, ARG2, ARG3) ( (((int32_t)(ARG1) << 0) & (int32_t)0x0000FFFF) | \ + (((int32_t)(ARG2) << ARG3) & (int32_t)0xFFFF0000) ) + #define __PKHTB(ARG1, ARG2, ARG3) ( (((int32_t)(ARG1) << 0) & (int32_t)0xFFFF0000) | \ + (((int32_t)(ARG2) >> ARG3) & (int32_t)0x0000FFFF) ) +#endif + + /** + * @brief definition to pack four 8 bit values. + */ +#ifndef ARM_MATH_BIG_ENDIAN + #define __PACKq7(v0,v1,v2,v3) ( (((int32_t)(v0) << 0) & (int32_t)0x000000FF) | \ + (((int32_t)(v1) << 8) & (int32_t)0x0000FF00) | \ + (((int32_t)(v2) << 16) & (int32_t)0x00FF0000) | \ + (((int32_t)(v3) << 24) & (int32_t)0xFF000000) ) +#else + #define __PACKq7(v0,v1,v2,v3) ( (((int32_t)(v3) << 0) & (int32_t)0x000000FF) | \ + (((int32_t)(v2) << 8) & (int32_t)0x0000FF00) | \ + (((int32_t)(v1) << 16) & (int32_t)0x00FF0000) | \ + (((int32_t)(v0) << 24) & (int32_t)0xFF000000) ) +#endif + + + + +/* + * @brief C custom defined intrinsic functions + */ +#if !defined (ARM_MATH_DSP) + + + /* + * @brief C custom defined QADD8 + */ + __STATIC_FORCEINLINE uint32_t __QADD8( + uint32_t x, + uint32_t y) + { + q31_t r, s, t, u; + + r = __SSAT(((((q31_t)x << 24) >> 24) + (((q31_t)y << 24) >> 24)), 8) & (int32_t)0x000000FF; + s = __SSAT(((((q31_t)x << 16) >> 24) + (((q31_t)y << 16) >> 24)), 8) & (int32_t)0x000000FF; + t = __SSAT(((((q31_t)x << 8) >> 24) + (((q31_t)y << 8) >> 24)), 8) & (int32_t)0x000000FF; + u = __SSAT(((((q31_t)x ) >> 24) + (((q31_t)y ) >> 24)), 8) & (int32_t)0x000000FF; + + return ((uint32_t)((u << 24) | (t << 16) | (s << 8) | (r ))); + } + + + /* + * @brief C custom defined QSUB8 + */ + __STATIC_FORCEINLINE uint32_t __QSUB8( + uint32_t x, + uint32_t y) + { + q31_t r, s, t, u; + + r = __SSAT(((((q31_t)x << 24) >> 24) - (((q31_t)y << 24) >> 24)), 8) & (int32_t)0x000000FF; + s = __SSAT(((((q31_t)x << 16) >> 24) - (((q31_t)y << 16) >> 24)), 8) & (int32_t)0x000000FF; + t = __SSAT(((((q31_t)x << 8) >> 24) - (((q31_t)y << 8) >> 24)), 8) & (int32_t)0x000000FF; + u = __SSAT(((((q31_t)x ) >> 24) - (((q31_t)y ) >> 24)), 8) & (int32_t)0x000000FF; + + return ((uint32_t)((u << 24) | (t << 16) | (s << 8) | (r ))); + } + + + /* + * @brief C custom defined QADD16 + */ + __STATIC_FORCEINLINE uint32_t __QADD16( + uint32_t x, + uint32_t y) + { +/* q31_t r, s; without initialisation 'arm_offset_q15 test' fails but 'intrinsic' tests pass! for armCC */ + q31_t r = 0, s = 0; + + r = __SSAT(((((q31_t)x << 16) >> 16) + (((q31_t)y << 16) >> 16)), 16) & (int32_t)0x0000FFFF; + s = __SSAT(((((q31_t)x ) >> 16) + (((q31_t)y ) >> 16)), 16) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined SHADD16 + */ + __STATIC_FORCEINLINE uint32_t __SHADD16( + uint32_t x, + uint32_t y) + { + q31_t r, s; + + r = (((((q31_t)x << 16) >> 16) + (((q31_t)y << 16) >> 16)) >> 1) & (int32_t)0x0000FFFF; + s = (((((q31_t)x ) >> 16) + (((q31_t)y ) >> 16)) >> 1) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined QSUB16 + */ + __STATIC_FORCEINLINE uint32_t __QSUB16( + uint32_t x, + uint32_t y) + { + q31_t r, s; + + r = __SSAT(((((q31_t)x << 16) >> 16) - (((q31_t)y << 16) >> 16)), 16) & (int32_t)0x0000FFFF; + s = __SSAT(((((q31_t)x ) >> 16) - (((q31_t)y ) >> 16)), 16) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined SHSUB16 + */ + __STATIC_FORCEINLINE uint32_t __SHSUB16( + uint32_t x, + uint32_t y) + { + q31_t r, s; + + r = (((((q31_t)x << 16) >> 16) - (((q31_t)y << 16) >> 16)) >> 1) & (int32_t)0x0000FFFF; + s = (((((q31_t)x ) >> 16) - (((q31_t)y ) >> 16)) >> 1) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined QASX + */ + __STATIC_FORCEINLINE uint32_t __QASX( + uint32_t x, + uint32_t y) + { + q31_t r, s; + + r = __SSAT(((((q31_t)x << 16) >> 16) - (((q31_t)y ) >> 16)), 16) & (int32_t)0x0000FFFF; + s = __SSAT(((((q31_t)x ) >> 16) + (((q31_t)y << 16) >> 16)), 16) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined SHASX + */ + __STATIC_FORCEINLINE uint32_t __SHASX( + uint32_t x, + uint32_t y) + { + q31_t r, s; + + r = (((((q31_t)x << 16) >> 16) - (((q31_t)y ) >> 16)) >> 1) & (int32_t)0x0000FFFF; + s = (((((q31_t)x ) >> 16) + (((q31_t)y << 16) >> 16)) >> 1) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined QSAX + */ + __STATIC_FORCEINLINE uint32_t __QSAX( + uint32_t x, + uint32_t y) + { + q31_t r, s; + + r = __SSAT(((((q31_t)x << 16) >> 16) + (((q31_t)y ) >> 16)), 16) & (int32_t)0x0000FFFF; + s = __SSAT(((((q31_t)x ) >> 16) - (((q31_t)y << 16) >> 16)), 16) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined SHSAX + */ + __STATIC_FORCEINLINE uint32_t __SHSAX( + uint32_t x, + uint32_t y) + { + q31_t r, s; + + r = (((((q31_t)x << 16) >> 16) + (((q31_t)y ) >> 16)) >> 1) & (int32_t)0x0000FFFF; + s = (((((q31_t)x ) >> 16) - (((q31_t)y << 16) >> 16)) >> 1) & (int32_t)0x0000FFFF; + + return ((uint32_t)((s << 16) | (r ))); + } + + + /* + * @brief C custom defined SMUSDX + */ + __STATIC_FORCEINLINE uint32_t __SMUSDX( + uint32_t x, + uint32_t y) + { + return ((uint32_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y ) >> 16)) - + ((((q31_t)x ) >> 16) * (((q31_t)y << 16) >> 16)) )); + } + + /* + * @brief C custom defined SMUADX + */ + __STATIC_FORCEINLINE uint32_t __SMUADX( + uint32_t x, + uint32_t y) + { + return ((uint32_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y ) >> 16)) + + ((((q31_t)x ) >> 16) * (((q31_t)y << 16) >> 16)) )); + } + + + /* + * @brief C custom defined QADD + */ + __STATIC_FORCEINLINE int32_t __QADD( + int32_t x, + int32_t y) + { + return ((int32_t)(clip_q63_to_q31((q63_t)x + (q31_t)y))); + } + + + /* + * @brief C custom defined QSUB + */ + __STATIC_FORCEINLINE int32_t __QSUB( + int32_t x, + int32_t y) + { + return ((int32_t)(clip_q63_to_q31((q63_t)x - (q31_t)y))); + } + + + /* + * @brief C custom defined SMLAD + */ + __STATIC_FORCEINLINE uint32_t __SMLAD( + uint32_t x, + uint32_t y, + uint32_t sum) + { + return ((uint32_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y << 16) >> 16)) + + ((((q31_t)x ) >> 16) * (((q31_t)y ) >> 16)) + + ( ((q31_t)sum ) ) )); + } + + + /* + * @brief C custom defined SMLADX + */ + __STATIC_FORCEINLINE uint32_t __SMLADX( + uint32_t x, + uint32_t y, + uint32_t sum) + { + return ((uint32_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y ) >> 16)) + + ((((q31_t)x ) >> 16) * (((q31_t)y << 16) >> 16)) + + ( ((q31_t)sum ) ) )); + } + + + /* + * @brief C custom defined SMLSDX + */ + __STATIC_FORCEINLINE uint32_t __SMLSDX( + uint32_t x, + uint32_t y, + uint32_t sum) + { + return ((uint32_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y ) >> 16)) - + ((((q31_t)x ) >> 16) * (((q31_t)y << 16) >> 16)) + + ( ((q31_t)sum ) ) )); + } + + + /* + * @brief C custom defined SMLALD + */ + __STATIC_FORCEINLINE uint64_t __SMLALD( + uint32_t x, + uint32_t y, + uint64_t sum) + { +/* return (sum + ((q15_t) (x >> 16) * (q15_t) (y >> 16)) + ((q15_t) x * (q15_t) y)); */ + return ((uint64_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y << 16) >> 16)) + + ((((q31_t)x ) >> 16) * (((q31_t)y ) >> 16)) + + ( ((q63_t)sum ) ) )); + } + + + /* + * @brief C custom defined SMLALDX + */ + __STATIC_FORCEINLINE uint64_t __SMLALDX( + uint32_t x, + uint32_t y, + uint64_t sum) + { +/* return (sum + ((q15_t) (x >> 16) * (q15_t) y)) + ((q15_t) x * (q15_t) (y >> 16)); */ + return ((uint64_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y ) >> 16)) + + ((((q31_t)x ) >> 16) * (((q31_t)y << 16) >> 16)) + + ( ((q63_t)sum ) ) )); + } + + + /* + * @brief C custom defined SMUAD + */ + __STATIC_FORCEINLINE uint32_t __SMUAD( + uint32_t x, + uint32_t y) + { + return ((uint32_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y << 16) >> 16)) + + ((((q31_t)x ) >> 16) * (((q31_t)y ) >> 16)) )); + } + + + /* + * @brief C custom defined SMUSD + */ + __STATIC_FORCEINLINE uint32_t __SMUSD( + uint32_t x, + uint32_t y) + { + return ((uint32_t)(((((q31_t)x << 16) >> 16) * (((q31_t)y << 16) >> 16)) - + ((((q31_t)x ) >> 16) * (((q31_t)y ) >> 16)) )); + } + + + /* + * @brief C custom defined SXTB16 + */ + __STATIC_FORCEINLINE uint32_t __SXTB16( + uint32_t x) + { + return ((uint32_t)(((((q31_t)x << 24) >> 24) & (q31_t)0x0000FFFF) | + ((((q31_t)x << 8) >> 8) & (q31_t)0xFFFF0000) )); + } + + /* + * @brief C custom defined SMMLA + */ + __STATIC_FORCEINLINE int32_t __SMMLA( + int32_t x, + int32_t y, + int32_t sum) + { + return (sum + (int32_t) (((int64_t) x * y) >> 32)); + } + +#endif /* !defined (ARM_MATH_DSP) */ + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _TRANSFORM_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/statistics_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/statistics_functions.h new file mode 100644 index 0000000..944d9f0 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/statistics_functions.h @@ -0,0 +1,486 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file statistics_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _STATISTICS_FUNCTIONS_H_ +#define _STATISTICS_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/fast_math_functions.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + +/** + * @defgroup groupStats Statistics Functions + */ + +/** + * @brief Computation of the LogSumExp + * + * In probabilistic computations, the dynamic of the probability values can be very + * wide because they come from gaussian functions. + * To avoid underflow and overflow issues, the values are represented by their log. + * In this representation, multiplying the original exp values is easy : their logs are added. + * But adding the original exp values is requiring some special handling and it is the + * goal of the LogSumExp function. + * + * If the values are x1...xn, the function is computing: + * + * ln(exp(x1) + ... + exp(xn)) and the computation is done in such a way that + * rounding issues are minimised. + * + * The max xm of the values is extracted and the function is computing: + * xm + ln(exp(x1 - xm) + ... + exp(xn - xm)) + * + * @param[in] *in Pointer to an array of input values. + * @param[in] blockSize Number of samples in the input array. + * @return LogSumExp + * + */ + + +float32_t arm_logsumexp_f32(const float32_t *in, uint32_t blockSize); + +/** + * @brief Dot product with log arithmetic + * + * Vectors are containing the log of the samples + * + * @param[in] pSrcA points to the first input vector + * @param[in] pSrcB points to the second input vector + * @param[in] blockSize number of samples in each vector + * @param[in] pTmpBuffer temporary buffer of length blockSize + * @return The log of the dot product . + * + */ + + +float32_t arm_logsumexp_dot_prod_f32(const float32_t * pSrcA, + const float32_t * pSrcB, + uint32_t blockSize, + float32_t *pTmpBuffer); + +/** + * @brief Entropy + * + * @param[in] pSrcA Array of input values. + * @param[in] blockSize Number of samples in the input array. + * @return Entropy -Sum(p ln p) + * + */ + + +float32_t arm_entropy_f32(const float32_t * pSrcA,uint32_t blockSize); + + +/** + * @brief Entropy + * + * @param[in] pSrcA Array of input values. + * @param[in] blockSize Number of samples in the input array. + * @return Entropy -Sum(p ln p) + * + */ + + +float64_t arm_entropy_f64(const float64_t * pSrcA, uint32_t blockSize); + + +/** + * @brief Kullback-Leibler + * + * @param[in] pSrcA Pointer to an array of input values for probability distribution A. + * @param[in] pSrcB Pointer to an array of input values for probability distribution B. + * @param[in] blockSize Number of samples in the input array. + * @return Kullback-Leibler Divergence D(A || B) + * + */ +float32_t arm_kullback_leibler_f32(const float32_t * pSrcA + ,const float32_t * pSrcB + ,uint32_t blockSize); + + +/** + * @brief Kullback-Leibler + * + * @param[in] pSrcA Pointer to an array of input values for probability distribution A. + * @param[in] pSrcB Pointer to an array of input values for probability distribution B. + * @param[in] blockSize Number of samples in the input array. + * @return Kullback-Leibler Divergence D(A || B) + * + */ +float64_t arm_kullback_leibler_f64(const float64_t * pSrcA, + const float64_t * pSrcB, + uint32_t blockSize); + + + /** + * @brief Sum of the squares of the elements of a Q31 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_power_q31( + const q31_t * pSrc, + uint32_t blockSize, + q63_t * pResult); + + + /** + * @brief Sum of the squares of the elements of a floating-point vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_power_f32( + const float32_t * pSrc, + uint32_t blockSize, + float32_t * pResult); + + + /** + * @brief Sum of the squares of the elements of a Q15 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_power_q15( + const q15_t * pSrc, + uint32_t blockSize, + q63_t * pResult); + + + /** + * @brief Sum of the squares of the elements of a Q7 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_power_q7( + const q7_t * pSrc, + uint32_t blockSize, + q31_t * pResult); + + + /** + * @brief Mean value of a Q7 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_mean_q7( + const q7_t * pSrc, + uint32_t blockSize, + q7_t * pResult); + + + /** + * @brief Mean value of a Q15 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_mean_q15( + const q15_t * pSrc, + uint32_t blockSize, + q15_t * pResult); + + + /** + * @brief Mean value of a Q31 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_mean_q31( + const q31_t * pSrc, + uint32_t blockSize, + q31_t * pResult); + + + /** + * @brief Mean value of a floating-point vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_mean_f32( + const float32_t * pSrc, + uint32_t blockSize, + float32_t * pResult); + + + /** + * @brief Variance of the elements of a floating-point vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_var_f32( + const float32_t * pSrc, + uint32_t blockSize, + float32_t * pResult); + + + /** + * @brief Variance of the elements of a Q31 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_var_q31( + const q31_t * pSrc, + uint32_t blockSize, + q31_t * pResult); + + + /** + * @brief Variance of the elements of a Q15 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_var_q15( + const q15_t * pSrc, + uint32_t blockSize, + q15_t * pResult); + + + /** + * @brief Root Mean Square of the elements of a floating-point vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_rms_f32( + const float32_t * pSrc, + uint32_t blockSize, + float32_t * pResult); + + + /** + * @brief Root Mean Square of the elements of a Q31 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_rms_q31( + const q31_t * pSrc, + uint32_t blockSize, + q31_t * pResult); + + + /** + * @brief Root Mean Square of the elements of a Q15 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_rms_q15( + const q15_t * pSrc, + uint32_t blockSize, + q15_t * pResult); + + + /** + * @brief Standard deviation of the elements of a floating-point vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_std_f32( + const float32_t * pSrc, + uint32_t blockSize, + float32_t * pResult); + + + /** + * @brief Standard deviation of the elements of a Q31 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_std_q31( + const q31_t * pSrc, + uint32_t blockSize, + q31_t * pResult); + + + /** + * @brief Standard deviation of the elements of a Q15 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output value. + */ + void arm_std_q15( + const q15_t * pSrc, + uint32_t blockSize, + q15_t * pResult); + + + + /** + * @brief Minimum value of a Q7 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] result is output pointer + * @param[in] index is the array index of the minimum value in the input buffer. + */ + void arm_min_q7( + const q7_t * pSrc, + uint32_t blockSize, + q7_t * result, + uint32_t * index); + + + /** + * @brief Minimum value of a Q15 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output pointer + * @param[in] pIndex is the array index of the minimum value in the input buffer. + */ + void arm_min_q15( + const q15_t * pSrc, + uint32_t blockSize, + q15_t * pResult, + uint32_t * pIndex); + + + /** + * @brief Minimum value of a Q31 vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output pointer + * @param[out] pIndex is the array index of the minimum value in the input buffer. + */ + void arm_min_q31( + const q31_t * pSrc, + uint32_t blockSize, + q31_t * pResult, + uint32_t * pIndex); + + + /** + * @brief Minimum value of a floating-point vector. + * @param[in] pSrc is input pointer + * @param[in] blockSize is the number of samples to process + * @param[out] pResult is output pointer + * @param[out] pIndex is the array index of the minimum value in the input buffer. + */ + void arm_min_f32( + const float32_t * pSrc, + uint32_t blockSize, + float32_t * pResult, + uint32_t * pIndex); + + +/** + * @brief Maximum value of a Q7 vector. + * @param[in] pSrc points to the input buffer + * @param[in] blockSize length of the input vector + * @param[out] pResult maximum value returned here + * @param[out] pIndex index of maximum value returned here + */ + void arm_max_q7( + const q7_t * pSrc, + uint32_t blockSize, + q7_t * pResult, + uint32_t * pIndex); + + +/** + * @brief Maximum value of a Q15 vector. + * @param[in] pSrc points to the input buffer + * @param[in] blockSize length of the input vector + * @param[out] pResult maximum value returned here + * @param[out] pIndex index of maximum value returned here + */ + void arm_max_q15( + const q15_t * pSrc, + uint32_t blockSize, + q15_t * pResult, + uint32_t * pIndex); + + +/** + * @brief Maximum value of a Q31 vector. + * @param[in] pSrc points to the input buffer + * @param[in] blockSize length of the input vector + * @param[out] pResult maximum value returned here + * @param[out] pIndex index of maximum value returned here + */ + void arm_max_q31( + const q31_t * pSrc, + uint32_t blockSize, + q31_t * pResult, + uint32_t * pIndex); + + +/** + * @brief Maximum value of a floating-point vector. + * @param[in] pSrc points to the input buffer + * @param[in] blockSize length of the input vector + * @param[out] pResult maximum value returned here + * @param[out] pIndex index of maximum value returned here + */ + void arm_max_f32( + const float32_t * pSrc, + uint32_t blockSize, + float32_t * pResult, + uint32_t * pIndex); + + /** + @brief Maximum value of a floating-point vector. + @param[in] pSrc points to the input vector + @param[in] blockSize number of samples in input vector + @param[out] pResult maximum value returned here + @return none + */ + void arm_max_no_idx_f32( + const float32_t *pSrc, + uint32_t blockSize, + float32_t *pResult); + + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _STATISTICS_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/support_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/support_functions.h new file mode 100644 index 0000000..2a81954 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/support_functions.h @@ -0,0 +1,429 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file support_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _SUPPORT_FUNCTIONS_H_ +#define _SUPPORT_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + * @defgroup groupSupport Support Functions + */ + + +/** + * @brief Converts the elements of the floating-point vector to Q31 vector. + * @param[in] pSrc points to the floating-point input vector + * @param[out] pDst points to the Q31 output vector + * @param[in] blockSize length of the input vector + */ + void arm_float_to_q31( + const float32_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the floating-point vector to Q15 vector. + * @param[in] pSrc points to the floating-point input vector + * @param[out] pDst points to the Q15 output vector + * @param[in] blockSize length of the input vector + */ + void arm_float_to_q15( + const float32_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the floating-point vector to Q7 vector. + * @param[in] pSrc points to the floating-point input vector + * @param[out] pDst points to the Q7 output vector + * @param[in] blockSize length of the input vector + */ + void arm_float_to_q7( + const float32_t * pSrc, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q31 vector to floating-point vector. + * @param[in] pSrc is input pointer + * @param[out] pDst is output pointer + * @param[in] blockSize is the number of samples to process + */ + void arm_q31_to_float( + const q31_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q31 vector to Q15 vector. + * @param[in] pSrc is input pointer + * @param[out] pDst is output pointer + * @param[in] blockSize is the number of samples to process + */ + void arm_q31_to_q15( + const q31_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q31 vector to Q7 vector. + * @param[in] pSrc is input pointer + * @param[out] pDst is output pointer + * @param[in] blockSize is the number of samples to process + */ + void arm_q31_to_q7( + const q31_t * pSrc, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q15 vector to floating-point vector. + * @param[in] pSrc is input pointer + * @param[out] pDst is output pointer + * @param[in] blockSize is the number of samples to process + */ + void arm_q15_to_float( + const q15_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q15 vector to Q31 vector. + * @param[in] pSrc is input pointer + * @param[out] pDst is output pointer + * @param[in] blockSize is the number of samples to process + */ + void arm_q15_to_q31( + const q15_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q15 vector to Q7 vector. + * @param[in] pSrc is input pointer + * @param[out] pDst is output pointer + * @param[in] blockSize is the number of samples to process + */ + void arm_q15_to_q7( + const q15_t * pSrc, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q7 vector to floating-point vector. + * @param[in] pSrc is input pointer + * @param[out] pDst is output pointer + * @param[in] blockSize is the number of samples to process + */ + void arm_q7_to_float( + const q7_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q7 vector to Q31 vector. + * @param[in] pSrc input pointer + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_q7_to_q31( + const q7_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Converts the elements of the Q7 vector to Q15 vector. + * @param[in] pSrc input pointer + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_q7_to_q15( + const q7_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + + + + /** + * @brief Struct for specifying sorting algorithm + */ + typedef enum + { + ARM_SORT_BITONIC = 0, + /**< Bitonic sort */ + ARM_SORT_BUBBLE = 1, + /**< Bubble sort */ + ARM_SORT_HEAP = 2, + /**< Heap sort */ + ARM_SORT_INSERTION = 3, + /**< Insertion sort */ + ARM_SORT_QUICK = 4, + /**< Quick sort */ + ARM_SORT_SELECTION = 5 + /**< Selection sort */ + } arm_sort_alg; + + /** + * @brief Struct for specifying sorting algorithm + */ + typedef enum + { + ARM_SORT_DESCENDING = 0, + /**< Descending order (9 to 0) */ + ARM_SORT_ASCENDING = 1 + /**< Ascending order (0 to 9) */ + } arm_sort_dir; + + /** + * @brief Instance structure for the sorting algorithms. + */ + typedef struct + { + arm_sort_alg alg; /**< Sorting algorithm selected */ + arm_sort_dir dir; /**< Sorting order (direction) */ + } arm_sort_instance_f32; + + /** + * @param[in] S points to an instance of the sorting structure. + * @param[in] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data. + * @param[in] blockSize number of samples to process. + */ + void arm_sort_f32( + const arm_sort_instance_f32 * S, + float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + /** + * @param[in,out] S points to an instance of the sorting structure. + * @param[in] alg Selected algorithm. + * @param[in] dir Sorting order. + */ + void arm_sort_init_f32( + arm_sort_instance_f32 * S, + arm_sort_alg alg, + arm_sort_dir dir); + + /** + * @brief Instance structure for the sorting algorithms. + */ + typedef struct + { + arm_sort_dir dir; /**< Sorting order (direction) */ + float32_t * buffer; /**< Working buffer */ + } arm_merge_sort_instance_f32; + + /** + * @param[in] S points to an instance of the sorting structure. + * @param[in,out] pSrc points to the block of input data. + * @param[out] pDst points to the block of output data + * @param[in] blockSize number of samples to process. + */ + void arm_merge_sort_f32( + const arm_merge_sort_instance_f32 * S, + float32_t *pSrc, + float32_t *pDst, + uint32_t blockSize); + + /** + * @param[in,out] S points to an instance of the sorting structure. + * @param[in] dir Sorting order. + * @param[in] buffer Working buffer. + */ + void arm_merge_sort_init_f32( + arm_merge_sort_instance_f32 * S, + arm_sort_dir dir, + float32_t * buffer); + + + + /** + * @brief Copies the elements of a floating-point vector. + * @param[in] pSrc input pointer + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_copy_f32( + const float32_t * pSrc, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Copies the elements of a Q7 vector. + * @param[in] pSrc input pointer + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_copy_q7( + const q7_t * pSrc, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Copies the elements of a Q15 vector. + * @param[in] pSrc input pointer + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_copy_q15( + const q15_t * pSrc, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Copies the elements of a Q31 vector. + * @param[in] pSrc input pointer + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_copy_q31( + const q31_t * pSrc, + q31_t * pDst, + uint32_t blockSize); + + + /** + * @brief Fills a constant value into a floating-point vector. + * @param[in] value input value to be filled + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_fill_f32( + float32_t value, + float32_t * pDst, + uint32_t blockSize); + + + /** + * @brief Fills a constant value into a Q7 vector. + * @param[in] value input value to be filled + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_fill_q7( + q7_t value, + q7_t * pDst, + uint32_t blockSize); + + + /** + * @brief Fills a constant value into a Q15 vector. + * @param[in] value input value to be filled + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_fill_q15( + q15_t value, + q15_t * pDst, + uint32_t blockSize); + + + /** + * @brief Fills a constant value into a Q31 vector. + * @param[in] value input value to be filled + * @param[out] pDst output pointer + * @param[in] blockSize number of samples to process + */ + void arm_fill_q31( + q31_t value, + q31_t * pDst, + uint32_t blockSize); + + + + + + + +/** + * @brief Weighted sum + * + * + * @param[in] *in Array of input values. + * @param[in] *weigths Weights + * @param[in] blockSize Number of samples in the input array. + * @return Weighted sum + * + */ +float32_t arm_weighted_sum_f32(const float32_t *in + , const float32_t *weigths + , uint32_t blockSize); + + +/** + * @brief Barycenter + * + * + * @param[in] in List of vectors + * @param[in] weights Weights of the vectors + * @param[out] out Barycenter + * @param[in] nbVectors Number of vectors + * @param[in] vecDim Dimension of space (vector dimension) + * @return None + * + */ +void arm_barycenter_f32(const float32_t *in + , const float32_t *weights + , float32_t *out + , uint32_t nbVectors + , uint32_t vecDim); + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _SUPPORT_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_defines.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_defines.h new file mode 100644 index 0000000..3374bcd --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_defines.h @@ -0,0 +1,45 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file svm_defines.h + * @brief Public header file for CMSIS DSP Library + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _SVM_DEFINES_H_ +#define _SVM_DEFINES_H_ + +/** + * @brief Struct for specifying SVM Kernel + */ +typedef enum +{ + ARM_ML_KERNEL_LINEAR = 0, + /**< Linear kernel */ + ARM_ML_KERNEL_POLYNOMIAL = 1, + /**< Polynomial kernel */ + ARM_ML_KERNEL_RBF = 2, + /**< Radial Basis Function kernel */ + ARM_ML_KERNEL_SIGMOID = 3 + /**< Sigmoid kernel */ +} arm_ml_kernel_type; + +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_functions.h new file mode 100644 index 0000000..2f9f4e7 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_functions.h @@ -0,0 +1,301 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file svm_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _SVM_FUNCTIONS_H_ +#define _SVM_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/svm_defines.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +#define STEP(x) (x) <= 0 ? 0 : 1 + +/** + * @defgroup groupSVM SVM Functions + * This set of functions is implementing SVM classification on 2 classes. + * The training must be done from scikit-learn. The parameters can be easily + * generated from the scikit-learn object. Some examples are given in + * DSP/Testing/PatternGeneration/SVM.py + * + * If more than 2 classes are needed, the functions in this folder + * will have to be used, as building blocks, to do multi-class classification. + * + * No multi-class classification is provided in this SVM folder. + * + */ + +/** + * @brief Integer exponentiation + * @param[in] x value + * @param[in] nb integer exponent >= 1 + * @return x^nb + * + */ +__STATIC_INLINE float32_t arm_exponent_f32(float32_t x, int32_t nb) +{ + float32_t r = x; + nb --; + while(nb > 0) + { + r = r * x; + nb--; + } + return(r); +} + + + + + +/** + * @brief Instance structure for linear SVM prediction function. + */ +typedef struct +{ + uint32_t nbOfSupportVectors; /**< Number of support vectors */ + uint32_t vectorDimension; /**< Dimension of vector space */ + float32_t intercept; /**< Intercept */ + const float32_t *dualCoefficients; /**< Dual coefficients */ + const float32_t *supportVectors; /**< Support vectors */ + const int32_t *classes; /**< The two SVM classes */ +} arm_svm_linear_instance_f32; + + +/** + * @brief Instance structure for polynomial SVM prediction function. + */ +typedef struct +{ + uint32_t nbOfSupportVectors; /**< Number of support vectors */ + uint32_t vectorDimension; /**< Dimension of vector space */ + float32_t intercept; /**< Intercept */ + const float32_t *dualCoefficients; /**< Dual coefficients */ + const float32_t *supportVectors; /**< Support vectors */ + const int32_t *classes; /**< The two SVM classes */ + int32_t degree; /**< Polynomial degree */ + float32_t coef0; /**< Polynomial constant */ + float32_t gamma; /**< Gamma factor */ +} arm_svm_polynomial_instance_f32; + +/** + * @brief Instance structure for rbf SVM prediction function. + */ +typedef struct +{ + uint32_t nbOfSupportVectors; /**< Number of support vectors */ + uint32_t vectorDimension; /**< Dimension of vector space */ + float32_t intercept; /**< Intercept */ + const float32_t *dualCoefficients; /**< Dual coefficients */ + const float32_t *supportVectors; /**< Support vectors */ + const int32_t *classes; /**< The two SVM classes */ + float32_t gamma; /**< Gamma factor */ +} arm_svm_rbf_instance_f32; + +/** + * @brief Instance structure for sigmoid SVM prediction function. + */ +typedef struct +{ + uint32_t nbOfSupportVectors; /**< Number of support vectors */ + uint32_t vectorDimension; /**< Dimension of vector space */ + float32_t intercept; /**< Intercept */ + const float32_t *dualCoefficients; /**< Dual coefficients */ + const float32_t *supportVectors; /**< Support vectors */ + const int32_t *classes; /**< The two SVM classes */ + float32_t coef0; /**< Independant constant */ + float32_t gamma; /**< Gamma factor */ +} arm_svm_sigmoid_instance_f32; + +/** + * @brief SVM linear instance init function + * @param[in] S Parameters for SVM functions + * @param[in] nbOfSupportVectors Number of support vectors + * @param[in] vectorDimension Dimension of vector space + * @param[in] intercept Intercept + * @param[in] dualCoefficients Array of dual coefficients + * @param[in] supportVectors Array of support vectors + * @param[in] classes Array of 2 classes ID + * @return none. + * + */ + + +void arm_svm_linear_init_f32(arm_svm_linear_instance_f32 *S, + uint32_t nbOfSupportVectors, + uint32_t vectorDimension, + float32_t intercept, + const float32_t *dualCoefficients, + const float32_t *supportVectors, + const int32_t *classes); + +/** + * @brief SVM linear prediction + * @param[in] S Pointer to an instance of the linear SVM structure. + * @param[in] in Pointer to input vector + * @param[out] pResult Decision value + * @return none. + * + */ + +void arm_svm_linear_predict_f32(const arm_svm_linear_instance_f32 *S, + const float32_t * in, + int32_t * pResult); + + +/** + * @brief SVM polynomial instance init function + * @param[in] S points to an instance of the polynomial SVM structure. + * @param[in] nbOfSupportVectors Number of support vectors + * @param[in] vectorDimension Dimension of vector space + * @param[in] intercept Intercept + * @param[in] dualCoefficients Array of dual coefficients + * @param[in] supportVectors Array of support vectors + * @param[in] classes Array of 2 classes ID + * @param[in] degree Polynomial degree + * @param[in] coef0 coeff0 (scikit-learn terminology) + * @param[in] gamma gamma (scikit-learn terminology) + * @return none. + * + */ + + +void arm_svm_polynomial_init_f32(arm_svm_polynomial_instance_f32 *S, + uint32_t nbOfSupportVectors, + uint32_t vectorDimension, + float32_t intercept, + const float32_t *dualCoefficients, + const float32_t *supportVectors, + const int32_t *classes, + int32_t degree, + float32_t coef0, + float32_t gamma + ); + +/** + * @brief SVM polynomial prediction + * @param[in] S Pointer to an instance of the polynomial SVM structure. + * @param[in] in Pointer to input vector + * @param[out] pResult Decision value + * @return none. + * + */ +void arm_svm_polynomial_predict_f32(const arm_svm_polynomial_instance_f32 *S, + const float32_t * in, + int32_t * pResult); + + +/** + * @brief SVM radial basis function instance init function + * @param[in] S points to an instance of the polynomial SVM structure. + * @param[in] nbOfSupportVectors Number of support vectors + * @param[in] vectorDimension Dimension of vector space + * @param[in] intercept Intercept + * @param[in] dualCoefficients Array of dual coefficients + * @param[in] supportVectors Array of support vectors + * @param[in] classes Array of 2 classes ID + * @param[in] gamma gamma (scikit-learn terminology) + * @return none. + * + */ + +void arm_svm_rbf_init_f32(arm_svm_rbf_instance_f32 *S, + uint32_t nbOfSupportVectors, + uint32_t vectorDimension, + float32_t intercept, + const float32_t *dualCoefficients, + const float32_t *supportVectors, + const int32_t *classes, + float32_t gamma + ); + +/** + * @brief SVM rbf prediction + * @param[in] S Pointer to an instance of the rbf SVM structure. + * @param[in] in Pointer to input vector + * @param[out] pResult decision value + * @return none. + * + */ +void arm_svm_rbf_predict_f32(const arm_svm_rbf_instance_f32 *S, + const float32_t * in, + int32_t * pResult); + +/** + * @brief SVM sigmoid instance init function + * @param[in] S points to an instance of the rbf SVM structure. + * @param[in] nbOfSupportVectors Number of support vectors + * @param[in] vectorDimension Dimension of vector space + * @param[in] intercept Intercept + * @param[in] dualCoefficients Array of dual coefficients + * @param[in] supportVectors Array of support vectors + * @param[in] classes Array of 2 classes ID + * @param[in] coef0 coeff0 (scikit-learn terminology) + * @param[in] gamma gamma (scikit-learn terminology) + * @return none. + * + */ + +void arm_svm_sigmoid_init_f32(arm_svm_sigmoid_instance_f32 *S, + uint32_t nbOfSupportVectors, + uint32_t vectorDimension, + float32_t intercept, + const float32_t *dualCoefficients, + const float32_t *supportVectors, + const int32_t *classes, + float32_t coef0, + float32_t gamma + ); + +/** + * @brief SVM sigmoid prediction + * @param[in] S Pointer to an instance of the rbf SVM structure. + * @param[in] in Pointer to input vector + * @param[out] pResult Decision value + * @return none. + * + */ +void arm_svm_sigmoid_predict_f32(const arm_svm_sigmoid_instance_f32 *S, + const float32_t * in, + int32_t * pResult); + + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _SVM_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/transform_functions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/transform_functions.h new file mode 100644 index 0000000..b9c5644 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/transform_functions.h @@ -0,0 +1,594 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file transform_functions.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +#ifndef _TRANSFORM_FUNCTIONS_H_ +#define _TRANSFORM_FUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_memory.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/none.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h" + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/basic_math_functions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/complex_math_functions.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + +/** + * @defgroup groupTransforms Transform Functions + */ + + + /** + * @brief Instance structure for the Q15 CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + uint8_t ifftFlag; /**< flag that selects forward (ifftFlag=0) or inverse (ifftFlag=1) transform. */ + uint8_t bitReverseFlag; /**< flag that enables (bitReverseFlag=1) or disables (bitReverseFlag=0) bit reversal of output. */ + const q15_t *pTwiddle; /**< points to the Sin twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t twidCoefModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + uint16_t bitRevFactor; /**< bit reversal modifier that supports different size FFTs with the same bit reversal table. */ + } arm_cfft_radix2_instance_q15; + +/* Deprecated */ + arm_status arm_cfft_radix2_init_q15( + arm_cfft_radix2_instance_q15 * S, + uint16_t fftLen, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + +/* Deprecated */ + void arm_cfft_radix2_q15( + const arm_cfft_radix2_instance_q15 * S, + q15_t * pSrc); + + + /** + * @brief Instance structure for the Q15 CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + uint8_t ifftFlag; /**< flag that selects forward (ifftFlag=0) or inverse (ifftFlag=1) transform. */ + uint8_t bitReverseFlag; /**< flag that enables (bitReverseFlag=1) or disables (bitReverseFlag=0) bit reversal of output. */ + const q15_t *pTwiddle; /**< points to the twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t twidCoefModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + uint16_t bitRevFactor; /**< bit reversal modifier that supports different size FFTs with the same bit reversal table. */ + } arm_cfft_radix4_instance_q15; + +/* Deprecated */ + arm_status arm_cfft_radix4_init_q15( + arm_cfft_radix4_instance_q15 * S, + uint16_t fftLen, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + +/* Deprecated */ + void arm_cfft_radix4_q15( + const arm_cfft_radix4_instance_q15 * S, + q15_t * pSrc); + + /** + * @brief Instance structure for the Radix-2 Q31 CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + uint8_t ifftFlag; /**< flag that selects forward (ifftFlag=0) or inverse (ifftFlag=1) transform. */ + uint8_t bitReverseFlag; /**< flag that enables (bitReverseFlag=1) or disables (bitReverseFlag=0) bit reversal of output. */ + const q31_t *pTwiddle; /**< points to the Twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t twidCoefModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + uint16_t bitRevFactor; /**< bit reversal modifier that supports different size FFTs with the same bit reversal table. */ + } arm_cfft_radix2_instance_q31; + +/* Deprecated */ + arm_status arm_cfft_radix2_init_q31( + arm_cfft_radix2_instance_q31 * S, + uint16_t fftLen, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + +/* Deprecated */ + void arm_cfft_radix2_q31( + const arm_cfft_radix2_instance_q31 * S, + q31_t * pSrc); + + /** + * @brief Instance structure for the Q31 CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + uint8_t ifftFlag; /**< flag that selects forward (ifftFlag=0) or inverse (ifftFlag=1) transform. */ + uint8_t bitReverseFlag; /**< flag that enables (bitReverseFlag=1) or disables (bitReverseFlag=0) bit reversal of output. */ + const q31_t *pTwiddle; /**< points to the twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t twidCoefModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + uint16_t bitRevFactor; /**< bit reversal modifier that supports different size FFTs with the same bit reversal table. */ + } arm_cfft_radix4_instance_q31; + +/* Deprecated */ + void arm_cfft_radix4_q31( + const arm_cfft_radix4_instance_q31 * S, + q31_t * pSrc); + +/* Deprecated */ + arm_status arm_cfft_radix4_init_q31( + arm_cfft_radix4_instance_q31 * S, + uint16_t fftLen, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + + /** + * @brief Instance structure for the floating-point CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + uint8_t ifftFlag; /**< flag that selects forward (ifftFlag=0) or inverse (ifftFlag=1) transform. */ + uint8_t bitReverseFlag; /**< flag that enables (bitReverseFlag=1) or disables (bitReverseFlag=0) bit reversal of output. */ + const float32_t *pTwiddle; /**< points to the Twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t twidCoefModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + uint16_t bitRevFactor; /**< bit reversal modifier that supports different size FFTs with the same bit reversal table. */ + float32_t onebyfftLen; /**< value of 1/fftLen. */ + } arm_cfft_radix2_instance_f32; + + +/* Deprecated */ + arm_status arm_cfft_radix2_init_f32( + arm_cfft_radix2_instance_f32 * S, + uint16_t fftLen, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + +/* Deprecated */ + void arm_cfft_radix2_f32( + const arm_cfft_radix2_instance_f32 * S, + float32_t * pSrc); + + /** + * @brief Instance structure for the floating-point CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + uint8_t ifftFlag; /**< flag that selects forward (ifftFlag=0) or inverse (ifftFlag=1) transform. */ + uint8_t bitReverseFlag; /**< flag that enables (bitReverseFlag=1) or disables (bitReverseFlag=0) bit reversal of output. */ + const float32_t *pTwiddle; /**< points to the Twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t twidCoefModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + uint16_t bitRevFactor; /**< bit reversal modifier that supports different size FFTs with the same bit reversal table. */ + float32_t onebyfftLen; /**< value of 1/fftLen. */ + } arm_cfft_radix4_instance_f32; + + + +/* Deprecated */ + arm_status arm_cfft_radix4_init_f32( + arm_cfft_radix4_instance_f32 * S, + uint16_t fftLen, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + +/* Deprecated */ + void arm_cfft_radix4_f32( + const arm_cfft_radix4_instance_f32 * S, + float32_t * pSrc); + + /** + * @brief Instance structure for the fixed-point CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + const q15_t *pTwiddle; /**< points to the Twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t bitRevLength; /**< bit reversal table length. */ +#if defined(ARM_MATH_MVEI) + const uint32_t *rearranged_twiddle_tab_stride1_arr; /**< Per stage reordered twiddle pointer (offset 1) */ \ + const uint32_t *rearranged_twiddle_tab_stride2_arr; /**< Per stage reordered twiddle pointer (offset 2) */ \ + const uint32_t *rearranged_twiddle_tab_stride3_arr; /**< Per stage reordered twiddle pointer (offset 3) */ \ + const q15_t *rearranged_twiddle_stride1; /**< reordered twiddle offset 1 storage */ \ + const q15_t *rearranged_twiddle_stride2; /**< reordered twiddle offset 2 storage */ \ + const q15_t *rearranged_twiddle_stride3; +#endif + } arm_cfft_instance_q15; + +arm_status arm_cfft_init_q15( + arm_cfft_instance_q15 * S, + uint16_t fftLen); + +void arm_cfft_q15( + const arm_cfft_instance_q15 * S, + q15_t * p1, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + + /** + * @brief Instance structure for the fixed-point CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + const q31_t *pTwiddle; /**< points to the Twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t bitRevLength; /**< bit reversal table length. */ +#if defined(ARM_MATH_MVEI) + const uint32_t *rearranged_twiddle_tab_stride1_arr; /**< Per stage reordered twiddle pointer (offset 1) */ \ + const uint32_t *rearranged_twiddle_tab_stride2_arr; /**< Per stage reordered twiddle pointer (offset 2) */ \ + const uint32_t *rearranged_twiddle_tab_stride3_arr; /**< Per stage reordered twiddle pointer (offset 3) */ \ + const q31_t *rearranged_twiddle_stride1; /**< reordered twiddle offset 1 storage */ \ + const q31_t *rearranged_twiddle_stride2; /**< reordered twiddle offset 2 storage */ \ + const q31_t *rearranged_twiddle_stride3; +#endif + } arm_cfft_instance_q31; + +arm_status arm_cfft_init_q31( + arm_cfft_instance_q31 * S, + uint16_t fftLen); + +void arm_cfft_q31( + const arm_cfft_instance_q31 * S, + q31_t * p1, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + + /** + * @brief Instance structure for the floating-point CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + const float32_t *pTwiddle; /**< points to the Twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t bitRevLength; /**< bit reversal table length. */ +#if defined(ARM_MATH_MVEF) && !defined(ARM_MATH_AUTOVECTORIZE) + const uint32_t *rearranged_twiddle_tab_stride1_arr; /**< Per stage reordered twiddle pointer (offset 1) */ \ + const uint32_t *rearranged_twiddle_tab_stride2_arr; /**< Per stage reordered twiddle pointer (offset 2) */ \ + const uint32_t *rearranged_twiddle_tab_stride3_arr; /**< Per stage reordered twiddle pointer (offset 3) */ \ + const float32_t *rearranged_twiddle_stride1; /**< reordered twiddle offset 1 storage */ \ + const float32_t *rearranged_twiddle_stride2; /**< reordered twiddle offset 2 storage */ \ + const float32_t *rearranged_twiddle_stride3; +#endif + } arm_cfft_instance_f32; + + + + arm_status arm_cfft_init_f32( + arm_cfft_instance_f32 * S, + uint16_t fftLen); + + void arm_cfft_f32( + const arm_cfft_instance_f32 * S, + float32_t * p1, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + + + /** + * @brief Instance structure for the Double Precision Floating-point CFFT/CIFFT function. + */ + typedef struct + { + uint16_t fftLen; /**< length of the FFT. */ + const float64_t *pTwiddle; /**< points to the Twiddle factor table. */ + const uint16_t *pBitRevTable; /**< points to the bit reversal table. */ + uint16_t bitRevLength; /**< bit reversal table length. */ + } arm_cfft_instance_f64; + + arm_status arm_cfft_init_f64( + arm_cfft_instance_f64 * S, + uint16_t fftLen); + + void arm_cfft_f64( + const arm_cfft_instance_f64 * S, + float64_t * p1, + uint8_t ifftFlag, + uint8_t bitReverseFlag); + + /** + * @brief Instance structure for the Q15 RFFT/RIFFT function. + */ + typedef struct + { + uint32_t fftLenReal; /**< length of the real FFT. */ + uint8_t ifftFlagR; /**< flag that selects forward (ifftFlagR=0) or inverse (ifftFlagR=1) transform. */ + uint8_t bitReverseFlagR; /**< flag that enables (bitReverseFlagR=1) or disables (bitReverseFlagR=0) bit reversal of output. */ + uint32_t twidCoefRModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + const q15_t *pTwiddleAReal; /**< points to the real twiddle factor table. */ + const q15_t *pTwiddleBReal; /**< points to the imag twiddle factor table. */ +#if defined(ARM_MATH_MVEI) + arm_cfft_instance_q15 cfftInst; +#else + const arm_cfft_instance_q15 *pCfft; /**< points to the complex FFT instance. */ +#endif + } arm_rfft_instance_q15; + + arm_status arm_rfft_init_q15( + arm_rfft_instance_q15 * S, + uint32_t fftLenReal, + uint32_t ifftFlagR, + uint32_t bitReverseFlag); + + void arm_rfft_q15( + const arm_rfft_instance_q15 * S, + q15_t * pSrc, + q15_t * pDst); + + /** + * @brief Instance structure for the Q31 RFFT/RIFFT function. + */ + typedef struct + { + uint32_t fftLenReal; /**< length of the real FFT. */ + uint8_t ifftFlagR; /**< flag that selects forward (ifftFlagR=0) or inverse (ifftFlagR=1) transform. */ + uint8_t bitReverseFlagR; /**< flag that enables (bitReverseFlagR=1) or disables (bitReverseFlagR=0) bit reversal of output. */ + uint32_t twidCoefRModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + const q31_t *pTwiddleAReal; /**< points to the real twiddle factor table. */ + const q31_t *pTwiddleBReal; /**< points to the imag twiddle factor table. */ +#if defined(ARM_MATH_MVEI) + arm_cfft_instance_q31 cfftInst; +#else + const arm_cfft_instance_q31 *pCfft; /**< points to the complex FFT instance. */ +#endif + } arm_rfft_instance_q31; + + arm_status arm_rfft_init_q31( + arm_rfft_instance_q31 * S, + uint32_t fftLenReal, + uint32_t ifftFlagR, + uint32_t bitReverseFlag); + + void arm_rfft_q31( + const arm_rfft_instance_q31 * S, + q31_t * pSrc, + q31_t * pDst); + + /** + * @brief Instance structure for the floating-point RFFT/RIFFT function. + */ + typedef struct + { + uint32_t fftLenReal; /**< length of the real FFT. */ + uint16_t fftLenBy2; /**< length of the complex FFT. */ + uint8_t ifftFlagR; /**< flag that selects forward (ifftFlagR=0) or inverse (ifftFlagR=1) transform. */ + uint8_t bitReverseFlagR; /**< flag that enables (bitReverseFlagR=1) or disables (bitReverseFlagR=0) bit reversal of output. */ + uint32_t twidCoefRModifier; /**< twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table. */ + const float32_t *pTwiddleAReal; /**< points to the real twiddle factor table. */ + const float32_t *pTwiddleBReal; /**< points to the imag twiddle factor table. */ + arm_cfft_radix4_instance_f32 *pCfft; /**< points to the complex FFT instance. */ + } arm_rfft_instance_f32; + + arm_status arm_rfft_init_f32( + arm_rfft_instance_f32 * S, + arm_cfft_radix4_instance_f32 * S_CFFT, + uint32_t fftLenReal, + uint32_t ifftFlagR, + uint32_t bitReverseFlag); + + void arm_rfft_f32( + const arm_rfft_instance_f32 * S, + float32_t * pSrc, + float32_t * pDst); + + /** + * @brief Instance structure for the Double Precision Floating-point RFFT/RIFFT function. + */ +typedef struct + { + arm_cfft_instance_f64 Sint; /**< Internal CFFT structure. */ + uint16_t fftLenRFFT; /**< length of the real sequence */ + const float64_t * pTwiddleRFFT; /**< Twiddle factors real stage */ + } arm_rfft_fast_instance_f64 ; + +arm_status arm_rfft_fast_init_f64 ( + arm_rfft_fast_instance_f64 * S, + uint16_t fftLen); + + +void arm_rfft_fast_f64( + arm_rfft_fast_instance_f64 * S, + float64_t * p, float64_t * pOut, + uint8_t ifftFlag); + + + /** + * @brief Instance structure for the floating-point RFFT/RIFFT function. + */ +typedef struct + { + arm_cfft_instance_f32 Sint; /**< Internal CFFT structure. */ + uint16_t fftLenRFFT; /**< length of the real sequence */ + const float32_t * pTwiddleRFFT; /**< Twiddle factors real stage */ + } arm_rfft_fast_instance_f32 ; + +arm_status arm_rfft_fast_init_f32 ( + arm_rfft_fast_instance_f32 * S, + uint16_t fftLen); + + + void arm_rfft_fast_f32( + const arm_rfft_fast_instance_f32 * S, + float32_t * p, float32_t * pOut, + uint8_t ifftFlag); + + /** + * @brief Instance structure for the floating-point DCT4/IDCT4 function. + */ + typedef struct + { + uint16_t N; /**< length of the DCT4. */ + uint16_t Nby2; /**< half of the length of the DCT4. */ + float32_t normalize; /**< normalizing factor. */ + const float32_t *pTwiddle; /**< points to the twiddle factor table. */ + const float32_t *pCosFactor; /**< points to the cosFactor table. */ + arm_rfft_instance_f32 *pRfft; /**< points to the real FFT instance. */ + arm_cfft_radix4_instance_f32 *pCfft; /**< points to the complex FFT instance. */ + } arm_dct4_instance_f32; + + + /** + * @brief Initialization function for the floating-point DCT4/IDCT4. + * @param[in,out] S points to an instance of floating-point DCT4/IDCT4 structure. + * @param[in] S_RFFT points to an instance of floating-point RFFT/RIFFT structure. + * @param[in] S_CFFT points to an instance of floating-point CFFT/CIFFT structure. + * @param[in] N length of the DCT4. + * @param[in] Nby2 half of the length of the DCT4. + * @param[in] normalize normalizing factor. + * @return arm_status function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_ARGUMENT_ERROR if fftLenReal is not a supported transform length. + */ + arm_status arm_dct4_init_f32( + arm_dct4_instance_f32 * S, + arm_rfft_instance_f32 * S_RFFT, + arm_cfft_radix4_instance_f32 * S_CFFT, + uint16_t N, + uint16_t Nby2, + float32_t normalize); + + + /** + * @brief Processing function for the floating-point DCT4/IDCT4. + * @param[in] S points to an instance of the floating-point DCT4/IDCT4 structure. + * @param[in] pState points to state buffer. + * @param[in,out] pInlineBuffer points to the in-place input and output buffer. + */ + void arm_dct4_f32( + const arm_dct4_instance_f32 * S, + float32_t * pState, + float32_t * pInlineBuffer); + + + /** + * @brief Instance structure for the Q31 DCT4/IDCT4 function. + */ + typedef struct + { + uint16_t N; /**< length of the DCT4. */ + uint16_t Nby2; /**< half of the length of the DCT4. */ + q31_t normalize; /**< normalizing factor. */ + const q31_t *pTwiddle; /**< points to the twiddle factor table. */ + const q31_t *pCosFactor; /**< points to the cosFactor table. */ + arm_rfft_instance_q31 *pRfft; /**< points to the real FFT instance. */ + arm_cfft_radix4_instance_q31 *pCfft; /**< points to the complex FFT instance. */ + } arm_dct4_instance_q31; + + + /** + * @brief Initialization function for the Q31 DCT4/IDCT4. + * @param[in,out] S points to an instance of Q31 DCT4/IDCT4 structure. + * @param[in] S_RFFT points to an instance of Q31 RFFT/RIFFT structure + * @param[in] S_CFFT points to an instance of Q31 CFFT/CIFFT structure + * @param[in] N length of the DCT4. + * @param[in] Nby2 half of the length of the DCT4. + * @param[in] normalize normalizing factor. + * @return arm_status function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_ARGUMENT_ERROR if N is not a supported transform length. + */ + arm_status arm_dct4_init_q31( + arm_dct4_instance_q31 * S, + arm_rfft_instance_q31 * S_RFFT, + arm_cfft_radix4_instance_q31 * S_CFFT, + uint16_t N, + uint16_t Nby2, + q31_t normalize); + + + /** + * @brief Processing function for the Q31 DCT4/IDCT4. + * @param[in] S points to an instance of the Q31 DCT4 structure. + * @param[in] pState points to state buffer. + * @param[in,out] pInlineBuffer points to the in-place input and output buffer. + */ + void arm_dct4_q31( + const arm_dct4_instance_q31 * S, + q31_t * pState, + q31_t * pInlineBuffer); + + + /** + * @brief Instance structure for the Q15 DCT4/IDCT4 function. + */ + typedef struct + { + uint16_t N; /**< length of the DCT4. */ + uint16_t Nby2; /**< half of the length of the DCT4. */ + q15_t normalize; /**< normalizing factor. */ + const q15_t *pTwiddle; /**< points to the twiddle factor table. */ + const q15_t *pCosFactor; /**< points to the cosFactor table. */ + arm_rfft_instance_q15 *pRfft; /**< points to the real FFT instance. */ + arm_cfft_radix4_instance_q15 *pCfft; /**< points to the complex FFT instance. */ + } arm_dct4_instance_q15; + + + /** + * @brief Initialization function for the Q15 DCT4/IDCT4. + * @param[in,out] S points to an instance of Q15 DCT4/IDCT4 structure. + * @param[in] S_RFFT points to an instance of Q15 RFFT/RIFFT structure. + * @param[in] S_CFFT points to an instance of Q15 CFFT/CIFFT structure. + * @param[in] N length of the DCT4. + * @param[in] Nby2 half of the length of the DCT4. + * @param[in] normalize normalizing factor. + * @return arm_status function returns ARM_MATH_SUCCESS if initialization is successful or ARM_MATH_ARGUMENT_ERROR if N is not a supported transform length. + */ + arm_status arm_dct4_init_q15( + arm_dct4_instance_q15 * S, + arm_rfft_instance_q15 * S_RFFT, + arm_cfft_radix4_instance_q15 * S_CFFT, + uint16_t N, + uint16_t Nby2, + q15_t normalize); + + + /** + * @brief Processing function for the Q15 DCT4/IDCT4. + * @param[in] S points to an instance of the Q15 DCT4 structure. + * @param[in] pState points to state buffer. + * @param[in,out] pInlineBuffer points to the in-place input and output buffer. + */ + void arm_dct4_q15( + const arm_dct4_instance_q15 * S, + q15_t * pState, + q15_t * pInlineBuffer); + + + +#ifdef __cplusplus +} +#endif + +#endif /* ifndef _TRANSFORM_FUNCTIONS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h new file mode 100644 index 0000000..86b00a1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/dsp/utils.h @@ -0,0 +1,242 @@ +#if !defined(ESP32) +/****************************************************************************** + * @file arm_math_utils.h + * @brief Public header file for CMSIS DSP Library + * @version V1.9.0 + * @date 20. July 2020 + ******************************************************************************/ +/* + * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef _ARM_MATH_UTILS_H_ + +#define _ARM_MATH_UTILS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + + /** + * @brief Macros required for reciprocal calculation in Normalized LMS + */ + +#define INDEX_MASK 0x0000003F + + +#define SQ(x) ((x) * (x)) + + + + /** + * @brief Function to Calculates 1/in (reciprocal) value of Q31 Data type. + */ + __STATIC_FORCEINLINE uint32_t arm_recip_q31( + q31_t in, + q31_t * dst, + const q31_t * pRecipTable) + { + q31_t out; + uint32_t tempVal; + uint32_t index, i; + uint32_t signBits; + + if (in > 0) + { + signBits = ((uint32_t) (__CLZ( in) - 1)); + } + else + { + signBits = ((uint32_t) (__CLZ(-in) - 1)); + } + + /* Convert input sample to 1.31 format */ + in = (in << signBits); + + /* calculation of index for initial approximated Val */ + index = (uint32_t)(in >> 24); + index = (index & INDEX_MASK); + + /* 1.31 with exp 1 */ + out = pRecipTable[index]; + + /* calculation of reciprocal value */ + /* running approximation for two iterations */ + for (i = 0U; i < 2U; i++) + { + tempVal = (uint32_t) (((q63_t) in * out) >> 31); + tempVal = 0x7FFFFFFFu - tempVal; + /* 1.31 with exp 1 */ + /* out = (q31_t) (((q63_t) out * tempVal) >> 30); */ + out = clip_q63_to_q31(((q63_t) out * tempVal) >> 30); + } + + /* write output */ + *dst = out; + + /* return num of signbits of out = 1/in value */ + return (signBits + 1U); + } + + + /** + * @brief Function to Calculates 1/in (reciprocal) value of Q15 Data type. + */ + __STATIC_FORCEINLINE uint32_t arm_recip_q15( + q15_t in, + q15_t * dst, + const q15_t * pRecipTable) + { + q15_t out = 0; + uint32_t tempVal = 0; + uint32_t index = 0, i = 0; + uint32_t signBits = 0; + + if (in > 0) + { + signBits = ((uint32_t)(__CLZ( in) - 17)); + } + else + { + signBits = ((uint32_t)(__CLZ(-in) - 17)); + } + + /* Convert input sample to 1.15 format */ + in = (in << signBits); + + /* calculation of index for initial approximated Val */ + index = (uint32_t)(in >> 8); + index = (index & INDEX_MASK); + + /* 1.15 with exp 1 */ + out = pRecipTable[index]; + + /* calculation of reciprocal value */ + /* running approximation for two iterations */ + for (i = 0U; i < 2U; i++) + { + tempVal = (uint32_t) (((q31_t) in * out) >> 15); + tempVal = 0x7FFFu - tempVal; + /* 1.15 with exp 1 */ + out = (q15_t) (((q31_t) out * tempVal) >> 14); + /* out = clip_q31_to_q15(((q31_t) out * tempVal) >> 14); */ + } + + /* write output */ + *dst = out; + + /* return num of signbits of out = 1/in value */ + return (signBits + 1); + } + + +/** + * @brief 64-bit to 32-bit unsigned normalization + * @param[in] in is input unsigned long long value + * @param[out] normalized is the 32-bit normalized value + * @param[out] norm is norm scale + */ +__STATIC_INLINE void arm_norm_64_to_32u(uint64_t in, int32_t * normalized, int32_t *norm) +{ + int32_t n1; + int32_t hi = (int32_t) (in >> 32); + int32_t lo = (int32_t) ((in << 32) >> 32); + + n1 = __CLZ(hi) - 32; + if (!n1) + { + /* + * input fits in 32-bit + */ + n1 = __CLZ(lo); + if (!n1) + { + /* + * MSB set, need to scale down by 1 + */ + *norm = -1; + *normalized = (((uint32_t) lo) >> 1); + } else + { + if (n1 == 32) + { + /* + * input is zero + */ + *norm = 0; + *normalized = 0; + } else + { + /* + * 32-bit normalization + */ + *norm = n1 - 1; + *normalized = lo << *norm; + } + } + } else + { + /* + * input fits in 64-bit + */ + n1 = 1 - n1; + *norm = -n1; + /* + * 64 bit normalization + */ + *normalized = (((uint32_t) lo) >> n1) | (hi << (32 - n1)); + } +} + +__STATIC_INLINE q31_t arm_div_q63_to_q31(q63_t num, q31_t den) +{ + q31_t result; + uint64_t absNum; + int32_t normalized; + int32_t norm; + + /* + * if sum fits in 32bits + * avoid costly 64-bit division + */ + absNum = num > 0 ? num : -num; + arm_norm_64_to_32u(absNum, &normalized, &norm); + if (norm > 0) + /* + * 32-bit division + */ + result = (q31_t) num / den; + else + /* + * 64-bit division + */ + result = (q31_t) (num / den); + + return result; +} + + +#ifdef __cplusplus +} +#endif + +#endif /*ifndef _ARM_MATH_UTILS_H_ */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_tables.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_tables.h new file mode 100644 index 0000000..9e9d114 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_tables.h @@ -0,0 +1,59 @@ +#if !defined(ESP32) +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_tables.h + * Description: Extern declaration for NN tables + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef _ARM_NN_TABLES_H +#define _ARM_NN_TABLES_H + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" + +/** +* @brief tables for various activation functions +* +*/ + +extern const q15_t sigmoidTable_q15[256]; +extern const q7_t sigmoidTable_q7[256]; + +extern const q7_t tanhTable_q7[256]; +extern const q15_t tanhTable_q15[256]; + + /** + * @brief 2-way tables for various activation functions + * + * 2-way table, H table for value larger than 1/4 + * L table for value smaller than 1/4, H table for remaining + * We have this only for the q15_t version. It does not make + * sense to have it for q7_t type + */ +extern const q15_t sigmoidHTable_q15[192]; +extern const q15_t sigmoidLTable_q15[128]; + +#endif /* ARM_NN_TABLES_H */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h new file mode 100644 index 0000000..9798022 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h @@ -0,0 +1,133 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_types.h + * Description: Public header file to contain the CMSIS-NN structs for the + * TensorFlowLite micro compliant functions + * + * $Date: April 23, 2020 + * $Revision: V.0.5.0 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + + +#ifndef _ARM_NN_TYPES_H +#define _ARM_NN_TYPES_H + +/** CMSIS-NN object to contain the width and height of a tile */ +typedef struct +{ + int32_t w; /**< Width */ + int32_t h; /**< Height */ +} cmsis_nn_tile; + +/** CMSIS-NN object used for the function context. */ +typedef struct +{ + void *buf; /**< Pointer to a buffer needed for the optimization */ + int32_t size; /**< Buffer size */ +} cmsis_nn_context; + +/** CMSIS-NN object to contain the dimensions of the tensors */ +typedef struct +{ + int32_t n; /**< Generic dimension to contain either the batch size or output channels. Please refer to the function documentation for more information */ + int32_t h; /**< Height */ + int32_t w; /**< Width */ + int32_t c; /**< Input channels */ +} cmsis_nn_dims; + +/** CMSIS-NN object for the per-channel quantization parameters */ +typedef struct +{ + int32_t *multiplier; /**< Multiplier values */ + int32_t *shift; /**< Shift values */ +} cmsis_nn_per_channel_quant_params; + +/** CMSIS-NN object for the per-tensor quantization parameters */ +typedef struct +{ + int32_t multiplier; /**< Multiplier value */ + int32_t shift; /**< Shift value */ +} cmsis_nn_per_tensor_quant_params; + +/** CMSIS-NN object for the quantized Relu activation */ +typedef struct +{ + int32_t min; /**< Min value used to clamp the result */ + int32_t max; /**< Max value used to clamp the result */ +} cmsis_nn_activation; + +/** CMSIS-NN object for the convolution layer parameters */ +typedef struct +{ + int32_t input_offset; /**< Zero value for the input tensor */ + int32_t output_offset; /**< Zero value for the output tensor */ + cmsis_nn_tile stride; + cmsis_nn_tile padding; + cmsis_nn_tile dilation; + cmsis_nn_activation activation; +} cmsis_nn_conv_params; + +/** CMSIS-NN object for Depthwise convolution layer parameters */ +typedef struct +{ + int32_t input_offset; /**< Zero value for the input tensor */ + int32_t output_offset; /**< Zero value for the output tensor */ + int32_t ch_mult; /**< Channel Multiplier. ch_mult * in_ch = out_ch */ + cmsis_nn_tile stride; + cmsis_nn_tile padding; + cmsis_nn_tile dilation; + cmsis_nn_activation activation; +} cmsis_nn_dw_conv_params; +/** CMSIS-NN object for pooling layer parameters */ +typedef struct +{ + cmsis_nn_tile stride; + cmsis_nn_tile padding; + cmsis_nn_activation activation; +} cmsis_nn_pool_params; + +/** CMSIS-NN object for Fully Connected layer parameters */ +typedef struct +{ + int32_t input_offset; /**< Zero value for the input tensor */ + int32_t filter_offset; /**< Zero value for the filter tensor */ + int32_t output_offset; /**< Zero value for the output tensor */ + cmsis_nn_activation activation; +} cmsis_nn_fc_params; + +/** CMSIS-NN object for SVDF layer parameters */ +typedef struct +{ + int32_t rank; + int32_t input_offset; /**< Zero value for the input tensor */ + int32_t output_offset; /**< Zero value for the output tensor */ + cmsis_nn_activation input_activation; + cmsis_nn_activation output_activation; +} cmsis_nn_svdf_params; + +#endif // _ARM_NN_TYPES_H + + + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h new file mode 100644 index 0000000..a760aa7 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h @@ -0,0 +1,2127 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nnfunctions.h + * Description: Public header file for CMSIS NN Library + * + * $Date: August 21, 2020 + * $Revision: V.6.5.2 + * + * Target Processor: Cortex-M CPUs + * -------------------------------------------------------------------- */ + +/** + \mainpage CMSIS NN Software Library + * + * Introduction + * ------------ + * + * This user manual describes the CMSIS NN software library, + * a collection of efficient neural network kernels developed to maximize the + * performance and minimize the memory footprint of neural networks on Cortex-M processor cores. + * + * The library is divided into a number of functions each covering a specific category: + * - Convolution Functions + * - Activation Functions + * - Fully-connected Layer Functions + * - SVDF Layer Functions + * - Pooling Functions + * - Softmax Functions + * - Basic math Functions + * + * The library has separate functions for operating on different weight and activation data + * types including 8-bit integers (q7_t) and 16-bit integers (q15_t). The descrition of the + * kernels are included in the function description. The implementation details are also + * described in this paper [1]. + * + * Function Classification + * -------- + * The functions can be classified into two segments + * - Legacy functions supporting ARM's internal symmetric quantization(8 bits). + * - Functions that support TensorFlow Lite framework with symmetric quantization(8 bits). + * + * The legacy functions can be identified with their suffix of _q7 or _q15 and are no new development is done there. The article in [2] describes in detail + * how to run a network using the legacy functions. + * + * The functions supporting TensorFlow Lite framework is identified by the _s8 suffix and can be invoked from TFL micro. The functions are bit exact to + * TensorFlow Lite. Refer to the TensorFlow's documentation in [3] on how to run a TensorFlow Lite model using optimized CMSIS-NN kernels. + * + * Block Diagram + * -------- + * \image html CMSIS-NN-OVERVIEW.PNG + * + * Examples + * -------- + * + * The library ships with a number of examples which demonstrate how to use the library functions. + * + * Pre-processor Macros + * ------------ + * + * Each library project have different pre-processor macros. + * + * - ARM_MATH_DSP: + * + * Define macro ARM_MATH_DSP, If the silicon supports DSP instructions(DSP extension). + * + * - ARM_MATH_MVEI: + * + * Define macro ARM_MATH_MVEI, If the silicon supports M-Profile Vector Extension. + + * - ARM_MATH_AUTOVECTORIZE + * Used in conjucture with ARM_MATH_MVEI to let the compiler auto vectorize for the functions that uses inline assembly. + * It does not affect functions that use C or intrinsics. + * - ARM_MATH_BIG_ENDIAN: + * + * Define macro ARM_MATH_BIG_ENDIAN to build the library for big endian targets. This is supported only for the legacy functions i.e, functions targetted at + * TensorFlow Lite do not support big endianness. By default library builds for little endian targets. + * + * - ARM_NN_TRUNCATE: + * + * Define macro ARM_NN_TRUNCATE to use floor instead of round-to-the-nearest-int for the computation. + * + * Upcoming Interface Change + * -------- + * Starting from the 1.4.0 next release, CMSIS-NN will gradually switch to a new API interface to: + * + * -# have a stable API + * -# avoid passing many variables by value + * -# improve security + * -# improve validation + * -# improve code readability + * + * The upcoming API interface change will be based on "struct" and only affect the TensorFlowLite micro compliant APIs [4] (functions with _s8 suffix) + * + * Below you can find a snapshot of how the new API interface will look like (names can change) + * + * i.e. arm_convolve_1x1_s8_fast + * + * Current API interface | New API interface proposal + * ------------- | ------------- + * const q7_t *input | const cmsis_nn_context &ctx + * const uint16_t input_x | const cmsis_nn_conv_params ¶ms + * const uint16_t input_y | const cmsis_nn_dims &input_dims + * const uint16_t input_ch | const q7_t *input_data + * const uint16_t input_batches | const cmsis_nn_dims &filter_dims + * const q7_t *kernel | const q7_t *filter_data + * const uint16_t output_ch | const cmsis_nn_dims &bias_dims + * const uint16_t pad_x | const q31_t *bias_data + * const uint16_t pad_y | const cmsis_nn_dims &output_dims + * const uint16_t stride_x | q7_t *output_data + * const uint16_t stride_y |
+ * const int32_t *bias |
+ * q7_t *output |
+ * const int32_t *output_shift |
+ * const int32_t *output_mult |
+ * const int32_t out_offset |
+ * const int32_t input_offset |
+ * const int32_t out_activation_min |
+ * const int32_t out_activation_max |
+ * const uint16_t output_x |
+ * const uint16_t output_y |
+ * q15_t *buffer_a |
+ * + * Copyright Notice + * ------------ + * + * Copyright (C) 2010-2019 Arm Limited. All rights reserved. + * + * [1] CMSIS-NN: Efficient Neural Network Kernels for Arm Cortex-M CPUs https://arxiv.org/abs/1801.06601 + * + * [2] Converting a Neural Network for Arm Cortex-M with CMSIS-NN + * https://developer.arm.com/solutions/machine-learning-on-arm/developer-material/how-to-guides/converting-a-neural-network-for-arm-cortex-m-with-cmsis-nn/single-page + * [3] https://www.tensorflow.org/lite/microcontrollers/library + * + * [4] https://github.com/ARM-software/CMSIS_5/tree/develop/CMSIS/NN#legacy-vs-tfl-micro-compliant-apis + */ + +/** + * @defgroup groupNN Neural Network Functions + * A collection of functions to perform basic operations for neural network layers. Functions with a _s8 suffix support + * TensorFlow Lite framework. + */ + +#ifndef _ARM_NNFUNCTIONS_H +#define _ARM_NNFUNCTIONS_H + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_tables.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" + +#define USE_INTRINSIC + +//#define ARM_NN_TRUNCATE /* This config the rounding model to floor or round to the nearest int */ + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + * @defgroup NNConv Convolution Functions + * + * Collection of convolution, depthwise convolution functions and their variants. + * + * The convolution is implemented in 2 steps: im2col and GEMM + * + * im2col is a process of converting each patch of image data into + * a column. After im2col, the convolution is computed as matrix-matrix + * multiplication. + * + * To reduce the memory footprint, the im2col is performed partially. + * Each iteration, only a few column (i.e., patches) are generated and + * computed with GEMM kernels similar to CMSIS-DSP arm_mat_mult functions. + * + */ + + /** + * @brief s8 convolution layer wrapper function with the main purpose to call the optimal kernel available in cmsis-nn to perform the convolution. + * + * @param[in, out] ctx Function context that contains the additional buffer if required by the implementation. + arm_convolve_wrapper_s8_get_buffer_size will return the buffer_size if required + * @param[in] conv_params Convolution parameters (e.g. strides, dilations, pads,...). + * Range of conv_params->input_offset : [-127, 128] + * Range of conv_params->output_offset : [-128, 127] + * @param[in] quant_params Per-channel quantization info. + * It contains the multiplier and shift values to be applied to each output channel + * @param[in] input_dims Input (activation) tensor dimensions. Format: [N, H, W, C_IN] + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions + * @param[in] filter_data Filter data pointer. Data type: int8 + * @param[in] bias_dims Bias tensor dimensions. Format: [C_OUT] + * @param[in] bias_data Bias data pointer. Data type: int32 + * @param[in] output_dims Output tensor dimensions. Format: [N, H, W, C_OUT] + * @param[out] output_data Output data pointer. Data type: int8 + * + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH if argument constraints fail. or, + * ARM_MATH_SUCCESS on successful completion. + * + */ + arm_status arm_convolve_wrapper_s8(const cmsis_nn_context* ctx, + const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, + const q7_t *input_data, + const cmsis_nn_dims* filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims* bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims* output_dims, + q7_t *output_data); + + /** + * @brief Get the required buffer size for arm_convolve_wrapper_s8 + * + * @param[in] conv_params Convolution parameters (e.g. strides, dilations, pads,...). + * Range of conv_params->input_offset : [-127, 128] + * Range of conv_params->output_offset : [-128, 127] + * @param[in] input_dims Input (activation) dimensions. Format: [N, H, W, C_IN] + * @param[in] filter_dims Filter dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions + * @param[in] output_dims Output tensor dimensions. Format: [N, H, W, C_OUT] + * + * @return The function returns required buffer size(bytes) + * + */ + int32_t arm_convolve_wrapper_s8_get_buffer_size(const cmsis_nn_conv_params* conv_params, + const cmsis_nn_dims* input_dims, + const cmsis_nn_dims* filter_dims, + const cmsis_nn_dims* output_dims); + + /** + * @brief Basic s8 convolution function + * @param[in, out] ctx Function context that contains the additional buffer if required by the implementation. + arm_convolve_s8_get_buffer_size will return the buffer_size if required + * @param[in] conv_params Convolution parameters (e.g. strides, dilations, pads,...). + * Range of conv_params->input_offset : [-127, 128] + * Range of conv_params->output_offset : [-128, 127] + * @param[in] quant_params Per-channel quantization info. + * It contains the multiplier and shift values to be applied to each output channel + * @param[in] input_dims Input (activation) tensor dimensions. Format: [N, H, W, C_IN] + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions + * @param[in] filter_data Filter data pointer. Data type: int8 + * @param[in] bias_dims Bias tensor dimensions. Format: [C_OUT] + * @param[in] bias_data Optional bias data pointer. Data type: int32 + * @param[in] output_dims Output tensor dimensions. Format: [N, H, W, C_OUT] + * @param[out] output_data Output data pointer. Data type: int8 + + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * 1. Supported framework: TensorFlow Lite micro + * 2. q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs. + * 3. Additional memory is required for optimization. Refer to argument 'ctx' for details. + * + */ + arm_status arm_convolve_s8(const cmsis_nn_context* ctx, + const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, + const q7_t *input_data, + const cmsis_nn_dims* filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims* bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims* output_dims, + q7_t *output_data); + + /** + * @brief Get the required buffer size for s8 convolution function + * + * @param[in] input_dims Input (activation) tensor dimensions. Format: [N, H, W, C_IN] + * @param[in] filter_dims Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions + * @return The function returns required buffer size(bytes) + * + */ + int32_t arm_convolve_s8_get_buffer_size(const cmsis_nn_dims* input_dims, + const cmsis_nn_dims* filter_dims); + + /** + * @brief Basic Q7 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns ARM_MATH_SUCCESS + * + */ + arm_status arm_convolve_HWC_q7_basic(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Basic Q7 convolution function (non-square shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimension x + * @param[in] dim_im_in_y input tensor dimension y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns ARM_MATH_SUCCESS + */ + arm_status arm_convolve_HWC_q7_basic_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Basic Q15 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns ARM_MATH_SUCCESS + * + */ + arm_status arm_convolve_HWC_q15_basic(const q15_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q15_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q15_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q15_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Fast Q7 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 4 + * ch_im_out is multiple of 2 + */ + arm_status arm_convolve_HWC_q7_fast(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Fast Q7 convolution function (non-sqaure shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimension x + * @param[in] dim_im_in_y input tensor dimension y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 4 + * ch_im_out is multiple of 2 + */ + + arm_status arm_convolve_HWC_q7_fast_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Fast Q7 version of 1x1 convolution (non-sqaure shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimension x + * @param[in] dim_im_in_y input tensor dimension y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH if argument constraints fail. or, + * ARM_MATH_SUCCESS on successful completion. + * + * This function implement convolution with 1x1 kernel size (i.e., dim_kernel_x=1 + * and dim_kernel_y=1). It can be used for + * second half of MobileNets after depthwise separable convolution. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 4 + * ch_im_out is multiple of 2 + */ + arm_status arm_convolve_1x1_HWC_q7_fast_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Fast s8 version for 1x1 convolution (non-square shape) + * + * @param[in, out] ctx Function context that contains the additional buffer if required by the implementation. + arm_convolve_1x1_s8_fast_get_buffer_size will return the buffer_size if required + * @param[in] conv_params Convolution parameters (e.g. strides, dilations, pads,...). + * Range of conv_params->input_offset : [-127, 128] + * Range of conv_params->output_offset : [-128, 127] + * @param[in] quant_params Per-channel quantization info. + * It contains the multiplier and shift values to be applied to each output channel + * @param[in] input_dims Input (activation) tensor dimensions. Format: [N, H, W, C_IN] + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN] + * @param[in] filter_data Filter data pointer. Data type: int8 + * @param[in] bias_dims Bias tensor dimensions. Format: [C_OUT] + * @param[in] bias_data Optional bias data pointer. Data type: int32 + * @param[in] output_dims Output tensor dimensions. Format: [N, H, W, C_OUT] + * @param[out] output_data Output data pointer. Data type: int8 + * + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH if argument constraints fail. or, + * ARM_MATH_SUCCESS on successful completion. + * + * @details + * - Supported framework : TensorFlow Lite Micro + * - The following constrains on the arguments apply + * -# input_dims->c is a multiple of 4 + * -# conv_params->padding.w = conv_params->padding.h = 0 + * -# conv_params->stride.w = conv_params->stride.h = 1 + * + */ + arm_status arm_convolve_1x1_s8_fast(const cmsis_nn_context* ctx, + const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, + const q7_t *input_data, + const cmsis_nn_dims* filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims* bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims* output_dims, + q7_t *output_data); + + /** + * @brief Get the required buffer size for arm_convolve_1x1_s8_fast + * + * @param[in] input_dims Input (activation) dimensions + * @return The function returns the required buffer size in bytes + * + */ + int32_t arm_convolve_1x1_s8_fast_get_buffer_size(const cmsis_nn_dims* input_dims); + + /** + * @brief 1xn convolution + * + * @param[in, out] ctx Function context that contains the additional buffer if required by the implementation. + arm_convolve_1_x_n_s8_get_buffer_size will return the buffer_size if required + * @param[in] conv_params Convolution parameters (e.g. strides, dilations, pads,...). + * Range of conv_params->input_offset : [-127, 128] + * Range of conv_params->output_offset : [-128, 127] + * @param[in] quant_params Per-channel quantization info. + * It contains the multiplier and shift values to be applied to each output channel + * @param[in] input_dims Input (activation) tensor dimensions. Format: [N, H, W, C_IN] + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [C_OUT, 1, WK, C_IN] where WK is the horizontal spatial filter dimension + * @param[in] filter_data Filter data pointer. Data type: int8 + * @param[in] bias_dims Bias tensor dimensions. Format: [C_OUT] + * @param[in] bias_data Optional bias data pointer. Data type: int32 + * @param[in] output_dims Output tensor dimensions. Format: [N, H, W, C_OUT] + * @param[out] output_data Output data pointer. Data type: int8 + * + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH if argument constraints fail. or, + * ARM_MATH_SUCCESS on successful completion. + * + * @details + * - Supported framework : TensorFlow Lite Micro + * - The following constrains on the arguments apply + * -# input_dims->n equals 1 + * -# ouput_dims->w is a multiple of 4 + * -# Explicit constraints(since it is for 1xN convolution) + * -## input_dims->h equals 1 + * -## output_dims->h equals 1 + * -## filter_dims->h equals 1 + *@todo Remove constraint on output_dims->w to make the function generic. + * + */ + arm_status arm_convolve_1_x_n_s8(const cmsis_nn_context* ctx, + const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, + const q7_t *input_data, + const cmsis_nn_dims* filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims* bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims* output_dims, + q7_t *output_data); + + /** + * @brief Get the required additional buffer size for 1xn convolution + * + * @param[in] input_dims Input (activation) tensor dimensions. Format: [N, H, W, C_IN] + * @param[in] filter_dims Filter tensor dimensions. Format: [C_OUT, 1, WK, C_IN] where WK is the horizontal spatial filter dimension + * @return The function returns required buffer size(bytes) + * + */ + int32_t arm_convolve_1_x_n_s8_get_buffer_size(const cmsis_nn_dims* input_dims, + const cmsis_nn_dims* filter_dims); + + /** + * @brief Q7 version of convolution for RGB image + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This kernel is written exclusively for convolution with ch_im_in + * equals 3. This applies on the first layer of CNNs which has input + * image with RGB format. + */ + + arm_status arm_convolve_HWC_q7_RGB(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Fast Q15 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 2 + * ch_im_out is multiple of 2 + */ + + arm_status arm_convolve_HWC_q15_fast(const q15_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q15_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q15_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q15_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Fast Q15 convolution function (non-sqaure shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimension x + * @param[in] dim_im_in_y input tensor dimension y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * @details + * + * Buffer size: + * + * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * Input dimension constraints: + * + * ch_im_in is multiple of 2 + * + * ch_im_out is multipe of 2 + * + */ + + arm_status + arm_convolve_HWC_q15_fast_nonsquare(const q15_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q15_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q15_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q15_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Q7 depthwise separable convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 2 + * ch_im_out is multiple of 2 + */ + + arm_status arm_depthwise_separable_conv_HWC_q7(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Q7 depthwise separable convolution function (non-square shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimension x + * @param[in] dim_im_in_y input tensor dimension y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding sizes x + * @param[in] padding_y padding sizes y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 2 + * ch_im_out is multiple of 2 + */ + arm_status arm_depthwise_separable_conv_HWC_q7_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB); + + /** + * @brief Wrapper function to pick the right optimized s8 depthwise convolution function + * + * @param[in, out] ctx Function context (e.g. temporary buffer). Check the function + * definition file to see if an additional buffer is required. + * Optional function {API}_get_buffer_size() provides the buffer + * size if required. + * @param[in] dw_conv_params Depthwise convolution parameters (e.g. strides, dilations, pads,...) + * dw_conv_params->dilation is not used. + * Range of dw_conv_params->input_offset : [-127, 128] + * Range of dw_conv_params->output_offset : [-128, 127] + * @param[in] quant_params Per-channel quantization info. + * It contains the multiplier and shift values to be applied to each + * output channel + * @param[in] input_dims Input (activation) tensor dimensions. Format: [H, W, C_IN] + * Batch argument N is not used and assumed to be 1. + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [1, H, W, C_OUT] + * @param[in] filter_data Filter data pointer. Data type: int8 + * @param[in] bias_dims Bias tensor dimensions. Format: [C_OUT] + * @param[in] bias_data Bias data pointer. Data type: int32 + * @param[in] output_dims Output tensor dimensions. Format: [1, H, W, C_OUT] + * @param[in, out] output_data Output data pointer. Data type: int8 + * @return The function returns + * ARM_MATH_SUCCESS - Successful completion. + * + * @details + * - Supported framework: TensorFlow Lite + * - Picks one of the the following functions + * -# arm_depthwise_conv_s8() + * -# arm_depthwise_conv_3x3_s8() - Cortex-M CPUs with DSP extension only + * -# arm_depthwise_conv_s8_opt() + * - q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs. + * - Check details of arm_depthwise_conv_s8_opt() for potential data that can be accessed outside of the boundary. + */ + arm_status arm_depthwise_conv_wrapper_s8(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims *bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data); + + /** + * @brief Get size of additional buffer required by arm_depthwise_conv_wrapper_s8() + * + * @param[in] dw_conv_params Depthwise convolution parameters (e.g. strides, dilations, pads,...) + * dw_conv_params->dilation is not used. + * Range of dw_conv_params->input_offset : [-127, 128] + * Range of dw_conv_params->input_offset : [-128, 127] + * @param[in] input_dims Input (activation) tensor dimensions. Format: [H, W, C_IN] + * Batch argument N is not used and assumed to be 1. + * @param[in] filter_dims Filter tensor dimensions. Format: [1, H, W, C_OUT] + * @param[in] output_dims Output tensor dimensions. Format: [1, H, W, C_OUT] + * @return Size of additional memory required for optimizations in bytes. + * + */ + int32_t arm_depthwise_conv_wrapper_s8_get_buffer_size(const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_dims *input_dims, + const cmsis_nn_dims *filter_dims, + const cmsis_nn_dims *output_dims); + + /** + * @brief Basic s8 depthwise convolution function that doesn't have any constraints on the input dimensions. + * + * @param[in, out] ctx Function context (e.g. temporary buffer). Check the function + * definition file to see if an additional buffer is required. + * Optional function {API}_get_buffer_size() provides the buffer + * size if an additional buffer is required. + * exists if additional memory is. + * @param[in] dw_conv_params Depthwise convolution parameters (e.g. strides, dilations, pads,...) + * dw_conv_params->dilation is not used. + * Range of dw_conv_params->input_offset : [-127, 128] + * Range of dw_conv_params->input_offset : [-128, 127] + * @param[in] quant_params Per-channel quantization info. + * It contains the multiplier and shift values to be applied to each + * output channel + * @param[in] input_dims Input (activation) tensor dimensions. Format: [1, H, W, C_IN] + * Batch argument N is not used. + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [1, H, W, C_OUT] + * @param[in] filter_data Filter data pointer. Data type: int8 + * @param[in] bias_dims Bias tensor dimensions. Format: [C_OUT] + * @param[in] bias_data Bias data pointer. Data type: int32 + * @param[in] output_dims Output tensor dimensions. Format: [1, H, W, C_OUT] + * @param[in, out] output_data Output data pointer. Data type: int8 + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * - Supported framework: TensorFlow Lite + * - q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs. + */ + arm_status arm_depthwise_conv_s8(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims *bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data); + + /** + * @brief Optimized s8 depthwise convolution function for 3x3 kernel size with some constraints on + * the input arguments(documented below). Refer arm_depthwise_conv_s8() for function + * argument details. + * + * @return The function returns one of the following + * ARM_MATH_SIZE_MISMATCH - Unsupported dimension of tensors + * ARM_MATH_ARGUMENT_ERROR - Unsupported pad size along the x axis + * ARM_MATH_SUCCESS - Successful operation + * + * @details + * - Supported framework : TensorFlow Lite Micro + * - The following constrains on the arguments apply + * -# Number of input channel equals number of output channels + * -# Filter height and width equals 3 + * -# Padding along x is either 0 or 1. + * + */ + arm_status arm_depthwise_conv_3x3_s8(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims *bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data); + + /** + * @brief Optimized s8 depthwise convolution function with constraint that in_channel equals out_channel. + * Refer arm_depthwise_conv_s8() for function argument details. + * + * @return The function returns one of the following + * ARM_MATH_SIZE_MISMATCH - input channel != output channel or + * ch_mult != 1 + * ARM_MATH_SUCCESS - Successful operation + * + * @note If number of channels is not a multiple of 4, upto 3 elements outside the boundary will be read out + * for the following if MVE optimizations(Arm Helium Technology) are used. + * - Output shift + * - Output multiplier + * - Output bias + * - kernel + * @details + * - Supported framework: TensorFlow Lite + * - The following constrains on the arguments apply + * -# Number of input channel equals number of output channels or ch_mult equals 1 + * - q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs. + * - Reccomended when number of channels is 4 or greater. + * + */ + arm_status arm_depthwise_conv_s8_opt(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims *bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data); + + /** + * @brief Get the required buffer size for optimized s8 depthwise convolution + * function with constraint that in_channel equals out_channel. + * @param[in] input_dims Input (activation) tensor dimensions. Format: [1, H, W, C_IN] + * Batch argument N is not used. + * @param[in] filter_dims Filter tensor dimensions. Format: [1, H, W, C_OUT] + * @return The function returns required buffer size in bytes + * + */ + int32_t arm_depthwise_conv_s8_opt_get_buffer_size(const cmsis_nn_dims* input_dims, + const cmsis_nn_dims* filter_dims); + + /** + * @defgroup FC Fully-connected Layer Functions + * + * Collection of fully-connected and matrix multiplication functions. + * + * Fully-connected layer is basically a matrix-vector multiplication + * with bias. The matrix is the weights and the input/output vectors + * are the activation values. Supported {weight, activation} precisions + * include {8-bit, 8-bit}, {16-bit, 16-bit}, and {8-bit, 16-bit}. + * + * Here we have two types of kernel functions. The basic function + * implements the function using regular GEMV approach. The opt functions + * operates with weights in interleaved formats. + * + */ + + /** + * @brief Q7 basic fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + */ + + arm_status arm_fully_connected_q7(const q7_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q7_t * pOut, + q15_t * vec_buffer); + + /** + * @brief Basic s8 Fully Connected function. + * + * @param[in, out] ctx Function context (e.g. temporary buffer). Check the function + * definition file to see if an additional buffer is required. + * Optional function {API}_get_buffer_size() provides the buffer + * size if an additional buffer is required. + * @param[in] fc_params Fully Connected layer parameters (e.g. strides, dilations, pads,...) + * Range of fc_params->input_offset : [-127, 128] + * Range of fc_params->filter_offset : [-127, 128] + * Range of fc_params->output_offset : [-128, 127] + * @param[in] quant_params Per-tensor quantization info. + * It contains the multiplier and shift values to be applied to the output tensor. + * @param[in] input_dims Input (activation) tensor dimensions. Format: [N, H, W, C_IN] + * Input dimension is taken as Nx(H * W * C_IN) + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Two dimensional filter dimensions. Format: [N, C] + * N : accumulation depth and equals (H * W * C_IN) from input_dims + * C : output depth and equals C_OUT in output_dims + * H & W : Not used + * @param[in] filter_data Filter data pointer. Data type: int8 + * @param[in] bias_dims Bias tensor dimensions. Format: [C_OUT] + * N, H, W : Not used + * @param[in] bias_data Bias data pointer. Data type: int32 + * @param[in] output_dims Output tensor dimensions. Format: [N, C_OUT] + * N : Batches + * C_OUT : Output depth + * H & W : Not used. + * @param[in, out] output_data Output data pointer. Data type: int8 + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * - Supported framework: TensorFlow Lite + * - q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs. + */ + arm_status + arm_fully_connected_s8(const cmsis_nn_context *ctx, + const cmsis_nn_fc_params *fc_params, + const cmsis_nn_per_tensor_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims *bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data); + + /** + * @brief Get the required buffer size for S8 basic fully-connected and + * matrix multiplication layer function for TF Lite + * @param[in] filter_dims dimension of filter + * @return The function returns required buffer size in bytes + * + */ + int32_t arm_fully_connected_s8_get_buffer_size(const cmsis_nn_dims *filter_dims); + + /** + * @brief Q7 opt fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + */ + + arm_status arm_fully_connected_q7_opt(const q7_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q7_t * pOut, + q15_t * vec_buffer); + + /** + * @brief Q15 basic fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + */ + + arm_status arm_fully_connected_q15(const q15_t * pV, + const q15_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q15_t * bias, + q15_t * pOut, + q15_t * vec_buffer); + + /** + * @brief Q15 opt fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + */ + + arm_status arm_fully_connected_q15_opt(const q15_t * pV, + const q15_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q15_t * bias, + q15_t * pOut, + q15_t * vec_buffer); + + /** + * @brief Mixed Q15-Q7 fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + */ + + arm_status arm_fully_connected_mat_q7_vec_q15(const q15_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q15_t * pOut, + q15_t * vec_buffer); + + /** + * @brief Mixed Q15-Q7 opt fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + */ + + arm_status arm_fully_connected_mat_q7_vec_q15_opt(const q15_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q15_t * pOut, + q15_t * vec_buffer); + +/** + * @brief Matrix-Multiplication Kernels for Convolution + * + * These functions are used within convolution layer functions for + * matrix multiplication. + * + * The implementation is similar to CMSIS-DSP arm_mat_mult functions + * with one Q7 and one Q15 operands. The Q15 operand is the im2col + * output which is always with 2 columns. + * + */ + + /** + * @brief Matrix-multiplication function for convolution + * @param[in] pA pointer to operand A + * @param[in] pInBuffer pointer to operand B, always conssists of 2 vectors + * @param[in] ch_im_out numRow of A + * @param[in] numCol_A numCol of A + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias the bias + * @param[in,out] pOut pointer to output + * @return The function returns the incremented output pointer + */ + + q7_t *arm_nn_mat_mult_kernel_q7_q15(const q7_t * pA, + const q15_t * pInBuffer, + const uint16_t ch_im_out, + const uint16_t numCol_A, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q7_t * pOut); + /** + * @brief Matrix-multiplication function for convolution with per-channel requantization. + * @param[in] input_a pointer to operand A + * @param[in] input_b pointer to operand B, always consists of 2 vectors. + * @param[in] output_ch number of rows of A + * @param[in] out_shift pointer to per output channel requantization shift parameter. + * @param[in] out_mult pointer to per output channel requantization multiplier parameter. + * @param[in] out_offset output tensor offset. + * @param[in] activation_min minimum value to clamp the output to. Range : int8 + * @param[in] activation_max maximum value to clamp the output to. Range : int8 + * @param[in] num_col_a number of columns of A + * @param[in] output_bias per output channel bias. Range : int32 + * @param[in,out] out_0 pointer to output + * @return The function returns one of the two + * 1. The incremented output pointer for a successful operation or + * 2. NULL if implementation is not available. + * + * @details This function does the matrix multiplication of weight matrix for all output channels + * with 2 columns from im2col and produces two elements/output_channel. The outputs are + * clamped in the range provided by activation min and max. + * Supported framework: TensorFlow Lite micro. + */ + q7_t *arm_nn_mat_mult_kernel_s8_s16(const q7_t *input_a, + const q15_t *input_b, + const uint16_t output_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int16_t activation_min, + const int16_t activation_max, + const uint16_t num_col_a, + const int32_t *const output_bias, + q7_t *out_0); + + /** + * @brief Matrix-multiplication of re-ordered input B with A. + * + * @details For arguments, refer arm_nn_mat_mult_kernel_s8_s16. The re-ordering is a consequence + * of sign extension done by the SXTB16 command on input_b. The outputs are clamped in the range + * provided by activation min and max. + * * @details + * - Supported framework : TensorFlow Lite Micro + * - The following constrains on the arguments apply + * -# num_col_a is a multiple of 4 + * -# output_ch is a multiple of 2 + * + */ + q7_t *arm_nn_mat_mult_kernel_s8_s16_reordered(const q7_t *input_a, + const q15_t *input_b, + const uint16_t output_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int16_t activation_min, + const int16_t activation_max, + const uint16_t num_col_a, + const int32_t *const output_bias, + q7_t *out_0); + + /** + * @brief Matrix-multiplication function for convolution with reordered columns + * @param[in] pA pointer to operand A + * @param[in] pInBuffer pointer to operand B, always conssists of 2 vectors + * @param[in] ch_im_out numRow of A + * @param[in] numCol_A numCol of A + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias the bias + * @param[in,out] pOut pointer to output + * @return The function returns the incremented output pointer + * + * @details This function assumes that data in pInBuffer are reordered + */ + q7_t *arm_nn_mat_mult_kernel_q7_q15_reordered(const q7_t * pA, + const q15_t * pInBuffer, + const uint16_t ch_im_out, + const uint16_t numCol_A, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q7_t * pOut); + +#ifdef __cplusplus +} +#endif + +/* + * Other functions + * These layers are typically not timing critical + * Basic implementation is supported here + */ + +#ifdef __cplusplus +extern "C" +{ +#endif + +/** + * @defgroup BasicMath Basic math functions + * + * Element wise add and multiplication functions. + * + */ + +/** + * @brief s8 element wise add of two vectors + * @param[in] input_1_vect pointer to input vector 1 + * @param[in] input_2_vect pointer to input vector 2 + * @param[in] input_1_offset offset for input 1. Range: Range: -127 to 128 + * @param[in] input_1_mult multiplier for input 1 + * @param[in] input_1_shift shift for input 1 + * @param[in] input_2_offset offset for input 2. Range: Range: -127 to 128 + * @param[in] input_2_mult multiplier for input 2 + * @param[in] input_2_shift shift for input 2 + * @param[in] left_shift input left shift + * @param[in,out] output pointer to output vector + * @param[in] out_offset output offset + * @param[in] out_mult output multiplier + * @param[in] out_shift output shift + * @param[in] out_activation_min minimum value to clamp output to + * @param[in] out_activation_max maximum value to clamp output to + * @param[in] block_size number of samples + * @return The function returns ARM_MATH_SUCCESS + */ + arm_status arm_elementwise_add_s8(const int8_t *input_1_vect, + const int8_t *input_2_vect, + const int32_t input_1_offset, + const int32_t input_1_mult, + const int32_t input_1_shift, + const int32_t input_2_offset, + const int32_t input_2_mult, + const int32_t input_2_shift, + const int32_t left_shift, + int8_t *output, + const int32_t out_offset, + const int32_t out_mult, + const int32_t out_shift, + const int32_t out_activation_min, + const int32_t out_activation_max, + const uint32_t block_size); + +/** + * @brief s8 element wise multiplication + * @param[in] input_1_vect pointer to input vector 1 + * @param[in] input_2_vect pointer to input vector 2 + * @param[in] input_1_offset offset for input 1. Range: Range: -127 to 128 + * @param[in] input_2_offset offset for input 2. Range: Range: -127 to 128 + * @param[in,out] output pointer to output vector + * @param[in] out_offset output offset + * @param[in] out_mult output multiplier + * @param[in] out_shift output shift + * @param[in] out_activation_min minimum value to clamp output to + * @param[in] out_activation_max maximum value to clamp output to + * @param[in] block_size number of samples + * @return The function returns ARM_MATH_SUCCESS + * + * @details Supported framework: TensorFlow Lite micro + */ + arm_status arm_elementwise_mul_s8(const int8_t *input_1_vect, + const int8_t *input_2_vect, + const int32_t input_1_offset, + const int32_t input_2_offset, + int8_t *output, + const int32_t out_offset, + const int32_t out_mult, + const int32_t out_shift, + const int32_t out_activation_min, + const int32_t out_activation_max, + const uint32_t block_size); +/** + * @defgroup Acti Activation Functions + * + * Perform activation layers, including ReLU (Rectified Linear Unit), + * sigmoid and tanh + * + */ + + /** + * @brief Q7 RELU function + * @param[in,out] data pointer to input + * @param[in] size number of elements + * @return none. + */ + + void arm_relu_q7(q7_t *data, uint16_t size); + + /** + * @brief s8 ReLU6 function + * @param[in,out] data pointer to input + * @param[in] size number of elements + */ + + void arm_relu6_s8(q7_t *data, uint16_t size); + + /** + * @brief Q15 RELU function + * @param[in,out] data pointer to input + * @param[in] size number of elements + * @return none. + */ + + void arm_relu_q15(q15_t *data, uint16_t size); + + /** + * @brief Q7 neural network activation function using direct table look-up + * @param[in,out] data pointer to input + * @param[in] size number of elements + * @param[in] int_width bit-width of the integer part, assume to be smaller than 3 + * @param[in] type type of activation functions + * @return none. + */ + + void arm_nn_activations_direct_q7(q7_t * data, uint16_t size, uint16_t int_width, + arm_nn_activation_type type); + + /** + * @brief Q15 neural network activation function using direct table look-up + * @param[in,out] data pointer to input + * @param[in] size number of elements + * @param[in] int_width bit-width of the integer part, assume to be smaller than 3 + * @param[in] type type of activation functions + * @return none. + * + * @details + * + * This is the direct table look-up approach. + * + * Assume here the integer part of the fixed-point is <= 3. + * More than 3 just not making much sense, makes no difference with + * saturation followed by any of these activation functions. + */ + + void arm_nn_activations_direct_q15(q15_t * data, uint16_t size, uint16_t int_width, + arm_nn_activation_type type); + +/** + * @defgroup Pooling Pooling Functions + * + * Perform pooling functions, including max pooling and average pooling + * + */ + + /** + * @brief Q7 max pooling function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] Im_out pointer to output tensor + * @return none. + * + */ + + void arm_maxpool_q7_HWC(q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const uint16_t dim_im_out, + q7_t * bufferA, + q7_t * Im_out); + + /** + * @brief Q7 average pooling function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimension + * @param[in] ch_im_in number of input tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] Im_out pointer to output tensor + * @return none. + * + */ + + void arm_avepool_q7_HWC(q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const uint16_t dim_im_out, + q7_t * bufferA, + q7_t * Im_out); + + /** + * @brief s8 average pooling function. + * + * @param[in, out] ctx Function context (e.g. temporary buffer). Check the function + * definition file to see if an additional buffer is required. + * Optional function {API}_get_buffer_size() provides the buffer + * size if an additional buffer is required. + * @param[in] pool_params Pooling parameters + * @param[in] input_dims Input (activation) tensor dimensions. Format: [H, W, C_IN] + * Argument 'N' is not used. + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [H, W] + * Argument N and C are not used. + * @param[in] output_dims Output tensor dimensions. Format: [H, W, C_OUT] + * Argument N is not used. + * C_OUT equals C_IN. + * @param[in, out] output_data Output data pointer. Data type: int8 + * @return The function returns + * ARM_MATH_SUCCESS - Successful operation + * + * @details + * - Supported Framework: TensorFlow Lite + * + */ + arm_status arm_avgpool_s8(const cmsis_nn_context *ctx, + const cmsis_nn_pool_params *pool_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const cmsis_nn_dims *output_dims, + q7_t *output_data); + + /** + * @brief Get the required buffer size for S8 average pooling function + * @param[in] dim_dst_width output tensor dimension + * @param[in] ch_src number of input tensor channels + * @return The function returns required buffer size in bytes + * + */ + int32_t arm_avgpool_s8_get_buffer_size(const int dim_dst_width, + const int ch_src); + + /** + * @brief s8 max pooling function. + * + * @param[in, out] ctx Function context (e.g. temporary buffer). Check the function + * definition file to see if an additional buffer is required. + * Optional function {API}_get_buffer_size() provides the buffer + * size if an additional buffer is required. + * @param[in] pool_params Pooling parameters + * @param[in] input_dims Input (activation) tensor dimensions. Format: [H, W, C_IN] + * Argument 'N' is not used. + * @param[in] input_data Input (activation) data pointer. Data type: int8 + * @param[in] filter_dims Filter tensor dimensions. Format: [H, W] + * Argument N and C are not used. + * @param[in] output_dims Output tensor dimensions. Format: [H, W, C_OUT] + * Argument N is not used. + * C_OUT equals C_IN. + * @param[in, out] output_data Output data pointer. Data type: int8 + * @return The function returns + * ARM_MATH_SUCCESS - Successful operation + * + * @details + * - Supported Framework: TensorFlow Lite + * + */ + arm_status arm_max_pool_s8(const cmsis_nn_context *ctx, + const cmsis_nn_pool_params *pool_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const cmsis_nn_dims *output_dims, + q7_t *output_data); +/** + * @defgroup Softmax Softmax Functions + * + * EXP(2) based softmax functions. + * + */ + + /** + * @brief Q7 softmax function + * @param[in] vec_in pointer to input vector + * @param[in] dim_vec input vector dimension + * @param[out] p_out pointer to output vector + * + * @note This function is an optimized version which is not bit-accurate with + * TensorFlow Lite's kernel + * + */ + +void arm_softmax_q7(const q7_t * vec_in, const uint16_t dim_vec, q7_t * p_out); + + /** + * @brief Q7 softmax function with batch parameter + * @param[in] vec_in pointer to input vector + * @param[in] nb_batches number of batches + * @param[in] dim_vec input vector dimension + * @param[out] p_out pointer to output vector + * @return none. + * + * @note This function is an optimized version which is not bit-accurate with + * TensorFlow Lite's kernel + * + */ + +void arm_softmax_with_batch_q7(const q7_t * vec_in, const uint16_t nb_batches,const uint16_t dim_vec, q7_t * p_out ); + /** + * @brief Q15 softmax function + * @param[in] vec_in pointer to input vector + * @param[in] dim_vec input vector dimension + * @param[out] p_out pointer to output vector + * @return none. + * + * @note This function is an optimized version which is not bit-accurate with + * TensorFlow Lite's kernel + * + */ + +void arm_softmax_q15(const q15_t * vec_in, const uint16_t dim_vec, q15_t * p_out); + + /** + * @brief S8 softmax function + * @param[in] input Pointer to the input tensor + * @param[in] num_rows Number of rows in the input tensor + * @param[in] row_size Number of elements in each input row + * @param[in] mult Input quantization multiplier + * @param[in] shift Input quantization shift within the range [0, 31] + * @param[in] diff_min Minimum difference with max in row. Used to check if + * the quantized exponential operation can be performed + * @param[out] output Pointer to the output tensor + * + * @note Supported framework: TensorFlow Lite micro (bit-accurate) + * + */ + +void arm_softmax_s8(const int8_t *input, + const int32_t num_rows, + const int32_t row_size, + const int32_t mult, + const int32_t shift, + const int32_t diff_min, + int8_t *output); + + /** + * @brief U8 softmax function + * @param[in] input Pointer to the input tensor + * @param[in] num_rows Number of rows in the input tensor + * @param[in] row_size Number of elements in each input row + * @param[in] mult Input quantization multiplier + * @param[in] shift Input quantization shift within the range [0, 31] + * @param[in] diff_min Minimum difference with max in row. Used to check if + * the quantized exponential operation can be performed + * @param[out] output Pointer to the output tensor + * + * @note Supported framework: TensorFlow Lite micro (bit-accurate) + * + */ + +void arm_softmax_u8(const uint8_t *input, + const int32_t num_rows, + const int32_t row_size, + const int32_t mult, + const int32_t shift, + const int32_t diff_min, + uint8_t *output); + + /** + * @brief uint8 depthwise convolution function with asymmetric quantization + * Unless specified otherwise, arguments are mandatory. + * + * @param[in] input Pointer to input tensor + * @param[in] input_x Width of input tensor + * @param[in] input_y Height of input tensor + * @param[in] input_ch Channels in input tensor + * @param[in] kernel Pointer to kernel weights + * @param[in] kernel_x Width of kernel + * @param[in] kernel_y Height of kernel + * @param[in] ch_mult Number of channel multiplier + * @param[in] pad_x Padding sizes x + * @param[in] pad_y Padding sizes y + * @param[in] stride_x stride along the width + * @param[in] stride_y stride along the height + * @param[in] dilation_x Dilation along width. Not used and intended for future enhancement. + * @param[in] dilation_y Dilation along height. Not used and intended for future enhancement. + * @param[in] bias Pointer to optional bias values. If no bias is + * availble, NULL is expected + * @param[in] input_offset Input tensor zero offset + * @param[in] filter_offset Kernel tensor zero offset + * @param[in] output_offset Output tensor zero offset + * @param[in,out] output Pointer to output tensor + * @param[in] output_x Width of output tensor + * @param[in] output_y Height of output tensor + * @param[in] output_activation_min Minimum value to clamp the output to. Range : {0, 255} + * @param[in] output_activation_max Minimum value to clamp the output to. Range : {0, 255} + * @param[in] out_shift Amount of right-shift for output + * @param[in] out_mult Output multiplier for requantization + * @return The function returns the following + * ARM_MATH_SUCCESS - Successful operation + * + */ + arm_status arm_depthwise_conv_u8_basic_ver1(const uint8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_ch, + const uint8_t *kernel, + const uint16_t kernel_x, + const uint16_t kernel_y, + const int16_t ch_mult, + const int16_t pad_x, + const int16_t pad_y, + const int16_t stride_x, + const int16_t stride_y, + const int16_t dilation_x, + const int16_t dilation_y, + const int32_t *bias, + const int32_t input_offset, + const int32_t filter_offset, + const int32_t output_offset, + uint8_t *output, + const uint16_t output_x, + const uint16_t output_y, + const int32_t output_activation_min, + const int32_t output_activation_max, + const int32_t out_shift, + const int32_t out_mult); + +/** + * @defgroup Reshape Reshape Functions + * + */ + + /** + * @brief Reshape a s8 vector into another with different shape + * @param[in] input points to the s8 input vector + * @param[out] output points to the s8 output vector + * @param[in] total_size total size of the input and output vectors in bytes + * + * @note The output is expected to be in a memory area that does not overlap with the input's + * + */ + void arm_reshape_s8(const int8_t *input, + int8_t *output, + const uint32_t total_size); + +/** + * @defgroup Concatenation Concatenation Functions + * + */ + + /** + * @brief int8/uint8 concatenation function to be used for concatenating N-tensors along the X axis + * This function should be called for each input tensor to concatenate. The argument offset_x + * will be used to store the input tensor in the correct position in the output tensor + * + * i.e. offset_x = 0 + * for(i = 0 i < num_input_tensors; ++i) + * { + * arm_concatenation_s8_x(&input[i], ..., &output, ..., ..., offset_x) + * offset_x += input_x[i] + * } + * + * This function assumes that the output tensor has: + * -# The same height of the input tensor + * -# The same number of channels of the input tensor + * -# The same batch size of the input tensor + * + * Unless specified otherwise, arguments are mandatory. + * + * @note This function, data layout independent, can be used to concatenate either int8 or uint8 tensors because does not involve any arithmetic operation + * + * @param[in] input Pointer to input tensor + * @param[in] input_x Width of input tensor + * @param[in] input_y Height of input tensor + * @param[in] input_z Channels in input tensor + * @param[in] input_w Batch size in input tensor + * @param[out] output Pointer to output tensor + * @param[in] output_x Width of output tensor + * @param[in] offset_x The offset (in number of elements) on the X axis to start concatenating the input tensor + * It is user responsibility to provide the correct value + * + * Input constraints + * offset_x is less than output_x + * + */ + void arm_concatenation_s8_x(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint16_t output_x, + const uint32_t offset_x); + + /** + * @brief int8/uint8 concatenation function to be used for concatenating N-tensors along the Y axis + * This function should be called for each input tensor to concatenate. The argument offset_y + * will be used to store the input tensor in the correct position in the output tensor + * + * i.e. offset_y = 0 + * for(i = 0 i < num_input_tensors; ++i) + * { + * arm_concatenation_s8_y(&input[i], ..., &output, ..., ..., offset_y) + * offset_y += input_y[i] + * } + * + * This function assumes that the output tensor has: + * -# The same width of the input tensor + * -# The same number of channels of the input tensor + * -# The same batch size of the input tensor + * + * Unless specified otherwise, arguments are mandatory. + * + * @note This function, data layout independent, can be used to concatenate either int8 or uint8 tensors because does not involve any arithmetic operation + * + * @param[in] input Pointer to input tensor + * @param[in] input_x Width of input tensor + * @param[in] input_y Height of input tensor + * @param[in] input_z Channels in input tensor + * @param[in] input_w Batch size in input tensor + * @param[out] output Pointer to output tensor + * @param[in] output_y Height of output tensor + * @param[in] offset_y The offset on the Y axis to start concatenating the input tensor + * It is user responsibility to provide the correct value + * + * Input constraints + * offset_y is less than output_y + * + */ + void arm_concatenation_s8_y(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint16_t output_y, + const uint32_t offset_y); + + /** + * @brief int8/uint8 concatenation function to be used for concatenating N-tensors along the Z axis + * This function should be called for each input tensor to concatenate. The argument offset_z + * will be used to store the input tensor in the correct position in the output tensor + * + * i.e. offset_z = 0 + * for(i = 0 i < num_input_tensors; ++i) + * { + * arm_concatenation_s8_z(&input[i], ..., &output, ..., ..., offset_z) + * offset_z += input_z[i] + * } + * + * This function assumes that the output tensor has: + * -# The same width of the input tensor + * -# The same height of the input tensor + * -# The same batch size of the input tensor + * + * Unless specified otherwise, arguments are mandatory. + * + * @note This function, data layout independent, can be used to concatenate either int8 or uint8 tensors because does not involve any arithmetic operation + * + * @param[in] input Pointer to input tensor + * @param[in] input_x Width of input tensor + * @param[in] input_y Height of input tensor + * @param[in] input_z Channels in input tensor + * @param[in] input_w Batch size in input tensor + * @param[out] output Pointer to output tensor + * @param[in] output_z Channels in output tensor + * @param[in] offset_z The offset on the Z axis to start concatenating the input tensor + * It is user responsibility to provide the correct value + * + * Input constraints + * offset_z is less than output_z + * + */ + void arm_concatenation_s8_z(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint16_t output_z, + const uint32_t offset_z); + + /** + * @brief int8/uint8 concatenation function to be used for concatenating N-tensors along the W axis (Batch size) + * This function should be called for each input tensor to concatenate. The argument offset_w + * will be used to store the input tensor in the correct position in the output tensor + * + * i.e. offset_w = 0 + * for(i = 0 i < num_input_tensors; ++i) + * { + * arm_concatenation_s8_w(&input[i], ..., &output, ..., ..., offset_w) + * offset_w += input_w[i] + * } + * + * This function assumes that the output tensor has: + * -# The same width of the input tensor + * -# The same height of the input tensor + * -# The same number o channels of the input tensor + * + * Unless specified otherwise, arguments are mandatory. + * + * @note This function, data layout independent, can be used to concatenate either int8 or uint8 tensors because does not involve any arithmetic operation + * + * @param[in] input Pointer to input tensor + * @param[in] input_x Width of input tensor + * @param[in] input_y Height of input tensor + * @param[in] input_z Channels in input tensor + * @param[in] input_w Batch size in input tensor + * @param[out] output Pointer to output tensor + * @param[in] offset_w The offset on the W axis to start concatenating the input tensor + * It is user responsibility to provide the correct value + * + */ + void arm_concatenation_s8_w(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint32_t offset_w); +/** + * @defgroup SVDF SVDF Layer Functions + * + */ + + /** + * @brief s8 SVDF function + * + * @param[in] input_ctx Temporary scratch buffer + * @param[in] output_ctx Temporary output scratch buffer + * @param[in] svdf_params SVDF Parameters + * Range of svdf_params->input_offset : [-128, 127] + * Range of svdf_params->output_offset : [-128, 127] + * @param[in] input_quant_params Input quantization parameters + * @param[in] output_quant_params Output quantization parameters + * @param[in] input_dims Input tensor dimensions + * @param[in] input_data Pointer to input tensor + * @param[in] state_dims State tensor dimensions + * @param[in] state_data Pointer to state tensor + * @param[in] weights_feature_dims Weights (feature) tensor dimensions + * @param[in] weights_feature_data Pointer to the weights (feature) tensor + * @param[in] weights_time_dims Weights (time) tensor dimensions + * @param[in] weights_time_data Pointer to the weights (time) tensor + * @param[in] bias_dims Bias tensor dimensions + * @param[in] bias_data Pointer to bias tensor + * @param[in] output_dims Output tensor dimensions + * @param[out] output_data Pointer to the output tensor + * + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * 1. Supported framework: TensorFlow Lite micro + * 2. q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs. + * + */ + arm_status + arm_svdf_s8(const cmsis_nn_context *input_ctx, + const cmsis_nn_context *output_ctx, + const cmsis_nn_svdf_params *svdf_params, + const cmsis_nn_per_tensor_quant_params *input_quant_params, + const cmsis_nn_per_tensor_quant_params *output_quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *state_dims, + q15_t *state_data, + const cmsis_nn_dims *weights_feature_dims, + const q7_t *weights_feature_data, + const cmsis_nn_dims *weights_time_dims, + const q15_t *weights_time_data, + const cmsis_nn_dims *bias_dims, + const q31_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data); + + +#ifdef __cplusplus +} +#endif + +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h new file mode 100644 index 0000000..543c69e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h @@ -0,0 +1,957 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nnsupportfunctions.h + * Description: Public header file of support functions for CMSIS NN Library + * + * $Date: July 31, 2020 + * $Revision: V.4.5.4 + * + * Target Processor: Cortex-M CPUs + * -------------------------------------------------------------------- */ + +#ifndef _ARM_NNSUPPORTFUNCTIONS_H_ +#define _ARM_NNSUPPORTFUNCTIONS_H_ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h" + +#ifdef __cplusplus +extern "C" +{ +#endif + +#define LEFT_SHIFT(_shift) (_shift > 0 ? _shift : 0) +#define RIGHT_SHIFT(_shift) (_shift > 0 ? 0 : -_shift) +#define MASK_IF_ZERO(x) (x) == 0 ? ~0 : 0 +#define MASK_IF_NON_ZERO(x) (x) != 0 ? ~0 : 0 +#define SELECT_USING_MASK(mask, a, b) ((mask) & (a)) ^ (~(mask) & (b)) + +#define MAX(A,B) ((A) > (B) ? (A) : (B)) +#define MIN(A,B) ((A) < (B) ? (A) : (B)) +#define CLAMP(x, h, l) MAX(MIN((x), (h)), (l)) + +/** + * @brief Union for SIMD access of q31/q15/q7 types + */ +union arm_nnword +{ + q31_t word; + /**< q31 type */ + q15_t half_words[2]; + /**< q15 type */ + q7_t bytes[4]; + /**< q7 type */ +}; + +/** + * @brief Union for data type long long + */ +struct arm_nn_double +{ + uint32_t low; + int32_t high; +}; + +union arm_nn_long_long +{ + int64_t long_long; + struct arm_nn_double word; +}; + +/** + * @brief Struct for specifying activation function types + * + */ +typedef enum +{ + ARM_SIGMOID = 0, + /**< Sigmoid activation function */ + ARM_TANH = 1, + /**< Tanh activation function */ +} arm_nn_activation_type; + +/** + * @defgroup nndata_convert Neural Network Data Conversion Functions + * + * Perform data type conversion in-between neural network operations + * + */ + +/** + * @brief Converts the elements of the q7 vector to q15 vector without left-shift + * @param[in] *pSrc points to the q7 input vector + * @param[out] *pDst points to the q15 output vector + * @param[in] blockSize length of the input vector + * + */ +void arm_q7_to_q15_no_shift(const q7_t *pSrc, q15_t *pDst, uint32_t blockSize); + +/** + * @brief Non-saturating addition of elements of a q7 vector + * @param[in] *input Pointer to the q7 input vector + * @param[out] *output Pointer to the q31 output variable. + * @param[in] block_size length of the input vector + * \par Description: + * + * 2^24 samples can be added without saturating the result. + * + * The equation used for the conversion process is: + * + * 
+ *  sum = input[0] + input[1] + .. + input[block_size -1]
+ *
+ * + * */ +void arm_nn_add_q7(const q7_t *input, q31_t *output, uint32_t block_size); + +/** + * @brief Converts the elements of the q7 vector to reordered q15 vector without left-shift + * @param[in] *pSrc points to the q7 input vector + * @param[out] *pDst points to the q15 output vector + * @param[in] blockSize length of the input vector + * @return none. + * + */ +void arm_q7_to_q15_reordered_no_shift(const q7_t * pSrc, q15_t * pDst, uint32_t blockSize); + +/** + * @brief Converts the elements from a q7 vector to a q15 vector with an added offset + * @param[in] src pointer to the q7 input vector + * @param[out] dst pointer to the q15 output vector + * @param[in] block_size length of the input vector + * @param[in] offset q7 offset to be added to each input vector element. + * + * \par Description: + * + * The equation used for the conversion process is: + * + * 
+ *  dst[n] = (q15_t) src[n] + offset;   0 <= n < block_size.
+ *
+ * + */ +void arm_q7_to_q15_with_offset(const q7_t *src, q15_t *dst, uint32_t block_size, q15_t offset); + +/** + * @brief Converts the elements of the q7 vector to reordered q15 vector with an added offset + * @param[in] src pointer to the q7 input vector + * @param[out] dst pointer to the q15 output vector + * @param[in] block_size length of the input vector + * @param[in] offset offset to be added to each input vector element. + * @return none. + * + * @details This function does the q7 to q15 expansion with re-ordering of bytes. Re-ordering is a consequence of + * the sign extension intrinsic(DSP extension). The tail (i.e., last (N % 4) elements) retains its original + * order. + * + */ +void arm_q7_to_q15_reordered_with_offset(const q7_t *src, q15_t *dst, uint32_t block_size, q15_t offset); + +/** + * @brief Converts the elements from a q7 vector and accumulate to a q15 vector + * @param[in] *src points to the q7 input vector + * @param[out] *dst points to the q15 output vector + * @param[in] block_size length of the input vector + * + * \par Description: + * + * The equation used for the conversion process is: + * + * 
+ *  dst[n] += (q15_t) src[n] ;   0 <= n < block_size.
+ *
+ * + */ +void arm_nn_accumulate_q7_to_q15(q15_t *dst, const q7_t *src, uint32_t block_size); + +/** + * @brief Depthwise conv on an im2col buffer where the input channel equals output channel. + * @param[in] row pointer to row + * @param[in] col pointer to im2col buffer, always consists of 2 columns. + * @param[in] num_ch number of channels + * @param[in] out_shift pointer to per output channel requantization shift parameter. + * @param[in] out_mult pointer to per output channel requantization multiplier parameter. + * @param[in] out_offset output tensor offset. + * @param[in] activation_min minimum value to clamp the output to. Range : int8 + * @param[in] activation_max maximum value to clamp the output to. Range : int8 + * @param[in] kernel_size number of elements in one column. + * @param[in] output_bias per output channel bias. Range : int32 + * @param[out] out pointer to output + * @return The function returns one of the two + * 1. The incremented output pointer for a successful operation or + * 2. NULL if implementation is not available. + * + * @details Supported framework: TensorFlow Lite micro. + */ +q7_t *arm_nn_depthwise_conv_s8_core(const q7_t *row, + const q15_t *col, + const uint16_t num_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int32_t activation_min, + const int32_t activation_max, + const uint16_t kernel_size, + const int32_t *const output_bias, + q7_t *out); + +/** + * @brief General Matrix-multiplication function with per-channel requantization. + * @param[in] input_row pointer to row operand + * @param[in] input_col pointer to col operand + * @param[in] output_ch number of rows of input_row + * @param[in] col_batches number of column batches. Range: 1 to 4 + * @param[in] output_shift pointer to per output channel requantization shift parameter. + * @param[in] output_mult pointer to per output channel requantization multiplier parameter. + * @param[in] out_offset output tensor offset. + * @param[in] col_offset input tensor(col) offset. + * @param[in] row_offset kernel offset(row). Not used. + * @param[in] out_activation_min minimum value to clamp the output to. Range : int8 + * @param[in] out_activation_max maximum value to clamp the output to. Range : int8 + * @param[in] row_len number of elements in each row + * @param[in] bias per output channel bias. Range : int32 + * @param[in,out] out pointer to output + * @return The function returns one of the two + * 1. The incremented output pointer for a successful operation or + * 2. NULL if implementation is not available. + * + * @details Supported framework: TensorFlow Lite +*/ +q7_t *arm_nn_mat_mult_s8(const q7_t *input_row, + const q7_t *input_col, + const uint16_t output_ch, + const uint16_t col_batches, + const int32_t *output_shift, + const int32_t *output_mult, + const int32_t out_offset, + const int32_t col_offset, + const int32_t row_offset, + const int16_t out_activation_min, + const int16_t out_activation_max, + const uint16_t row_len, + const int32_t *const bias, + q7_t *out); + +/** + * @brief General Matrix-multiplication without requantization for one row & one column + * @param[in] row_elements number of row elements + * @param[in] row_base pointer to row operand + * @param[in] col_base pointer to col operand + * @param[out] sum_col pointer to store sum of column elements + * @param[out] output pointer to store result of multiply-accumulate + * @return The function returns the multiply-accumulated result of the row by column. + * + * @details Pseudo-code + * *output = 0 + * sum_col = 0 + * for (i = 0; i < row_elements; i++) + * *output += row_base[i] * col_base[i] + * sum_col += col_base[i] + * +*/ +arm_status arm_nn_mat_mul_core_1x_s8(int32_t row_elements, + const int8_t *row_base, + const int8_t *col_base, + int32_t *const sum_col, + int32_t *const output); + +/** + * @brief General Matrix-multiplication without requantization for four rows and one column + * @param[in] row_elements number of row elements + * @param[in] offset offset between rows. Can be the same as row_elements. + * For e.g, in a 1x1 conv scenario with stride as 1. + * @param[in] row_base pointer to row operand + * @param[in] col_base pointer to col operand + * @param[out] sum_col pointer to store sum of column elements + * @param[out] output pointer to store result(4 int32's) of multiply-accumulate + * @return The function returns the multiply-accumulated result of the row by column + * + * @details Pseudo-code + * output[0] = 0 + * .. + * output[3] = 0 + * sum_col = 0 + * for (i = 0; i < row_elements; i++) + * output[0] += row_base[i] * col_base[i] + * .. + * output[3] += row_base[i + (row_elements * 3)] * col_base[i] + * sum_col += col_base[i] +*/ +arm_status arm_nn_mat_mul_core_4x_s8(const int32_t row_elements, + const int32_t offset, + const int8_t *row_base, + const int8_t *col_base, + int32_t *const sum_col, + int32_t *const output); + +/** +* @brief General Matrix-multiplication function with per-channel requantization. +* This function assumes: +* - LHS input matrix NOT transposed (nt) +* - RHS input matrix transposed (t) +* +* @note This operation also performs the broadcast bias addition before the requantization +* +* @param[in] lhs Pointer to the LHS input matrix +* @param[in] rhs Pointer to the RHS input matrix +* @param[in] bias Pointer to the bias vector. The length of this vector is equal to the number of output columns (or RHS input rows) +* @param[out] dst Pointer to the output matrix with "m" rows and "n" columns +* @param[in] dst_multipliers Pointer to the multipliers vector needed for the per-channel requantization. The length of this vector is equal to +* the number of output columns (or RHS input rows) +* @param[in] dst_shifts Pointer to the shifts vector needed for the per-channel requantization. The length of this vector is equal to +* the number of output columns (or RHS input rows) +* @param[in] lhs_rows Number of LHS input rows +* @param[in] rhs_rows Number of RHS input rows +* @param[in] rhs_cols Number of LHS/RHS input columns +* @param[in] lhs_offset Offset to be applied to the LHS input value +* @param[in] dst_offset Offset to be applied the output result +* @param[in] activation_min Minimum value to clamp down the output. Range : int8 +* @param[in] activation_max Maximum value to clamp up the output. Range : int8 +* +* @return The function returns ARM_MATH_SUCCESS +* +*/ +arm_status arm_nn_mat_mult_nt_t_s8(const q7_t *lhs, + const q7_t *rhs, + const q31_t *bias, + q7_t *dst, + const int32_t *dst_multipliers, + const int32_t *dst_shifts, + const int32_t lhs_rows, + const int32_t rhs_rows, + const int32_t rhs_cols, + const int32_t lhs_offset, + const int32_t dst_offset, + const int32_t activation_min, + const int32_t activation_max); + +/** + * @brief s8 Vector by Matrix (transposed) multiplication + * + * @param[in] lhs Input left-hand side vector + * @param[in] rhs Input right-hand side matrix (transposed) + * @param[in] bias Input bias + * @param[out] dst Output vector + * @param[in] lhs_offset Offset to be added to the input values of the left-hand side vector. Range: -127 to 128 + * @param[in] rhs_offset Offset to be added to the input values of the right-hand side matrix. Range: -127 to 128 + * @param[in] dst_offset Offset to be added to the output values. Range: -127 to 128 + * @param[in] dst_multiplier Output multiplier + * @param[in] dst_shift Output shift + * @param[in] rhs_cols Number of columns in the right-hand side input matrix + * @param[in] rhs_rows Number of rows in the right-hand side input matrix + * @param[in] activation_min Minimum value to clamp the output to. Range: int8 + * @param[in] activation_max Maximum value to clamp the output to. Range: int8 + * + * @return The function returns ARM_MATH_SUCCESS + * + */ +arm_status arm_nn_vec_mat_mult_t_s8(const q7_t *lhs, + const q7_t *rhs, + const q31_t *bias, + q7_t *dst, + const int32_t lhs_offset, + const int32_t rhs_offset, + const int32_t dst_offset, + const int32_t dst_multiplier, + const int32_t dst_shift, + const int32_t rhs_cols, + const int32_t rhs_rows, + const int32_t activation_min, + const int32_t activation_max); + +/** + * @brief Depthwise convolution of transposed rhs matrix with 4 lhs matrices. To be used in padded cases where + * the padding is -lhs_offset(Range: int8). Dimensions are the same for lhs and rhs. + * + * @param[in] lhs Input left-hand side matrix + * @param[in] rhs Input right-hand side matrix (transposed) + * @param[in] lhs_offset LHS matrix offset(input offset). Range: -127 to 128 + * @param[in] num_ch Number of channels in LHS/RHS + * @param[in] out_shift Per channel output shift. Length of vector is equal to number of channels + * @param[in] out_mult Per channel output multiplier. Length of vector is equal to number of channels + * @param[in] out_offset Offset to be added to the output values. Range: -127 to 128 + * @param[in] activation_min Minimum value to clamp the output to. Range: int8 + * @param[in] activation_max Maximum value to clamp the output to. Range: int8 + * @param[in] row_x_col (row_dimension * col_dimension) of LHS/RHS matrix + * @param[in] output_bias Per channel output bias. Length of vector is equal to number of channels + * @param[in] out Output pointer + * + * @return The function returns one of the two + * - Updated output pointer if an implementaiton is available + * - NULL if no implementation is available. + * + * @note If number of channels is not a multiple of 4, upto 3 elements outside the boundary will be read out + * for the following. + * - Output shift + * - Output multiplier + * - Output bias + * - rhs + */ +q7_t *arm_nn_depthwise_conv_nt_t_padded_s8(const q7_t *lhs, + const q7_t *rhs, + const int32_t lhs_offset, + const uint16_t num_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int32_t activation_min, + const int32_t activation_max, + const uint16_t row_x_col, + const int32_t *const output_bias, + q7_t *out); + +/** + * @brief Depthwise convolution of transposed rhs matrix with 4 lhs matrices. To be used in non-padded cases. + * Dimensions are the same for lhs and rhs. + * + * @param[in] lhs Input left-hand side matrix + * @param[in] rhs Input right-hand side matrix (transposed) + * @param[in] lhs_offset LHS matrix offset(input offset). Range: -127 to 128 + * @param[in] num_ch Number of channels in LHS/RHS + * @param[in] out_shift Per channel output shift. Length of vector is equal to number of channels. + * @param[in] out_mult Per channel output multiplier. Length of vector is equal to number of channels. + * @param[in] out_offset Offset to be added to the output values. Range: -127 to 128 + * @param[in] activation_min Minimum value to clamp the output to. Range: int8 + * @param[in] activation_max Maximum value to clamp the output to. Range: int8 + * @param[in] row_x_col (row_dimension * col_dimension) of LHS/RHS matrix + * @param[in] output_bias Per channel output bias. Length of vector is equal to number of channels. + * @param[in] out Output pointer + * + * @return The function returns one of the two + * - Updated output pointer if an implementaiton is available + * - NULL if no implementation is available. + * + * @note If number of channels is not a multiple of 4, upto 3 elements outside the boundary will be read out + * for the following. + * - Output shift + * - Output multiplier + * - Output bias + * - rhs + */ +q7_t *arm_nn_depthwise_conv_nt_t_s8(const q7_t *lhs, + const q7_t *rhs, + const int32_t lhs_offset, + const uint16_t num_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int32_t activation_min, + const int32_t activation_max, + const uint16_t row_x_col, + const int32_t *const output_bias, + q7_t *out); + +/** + @brief Read 2 q15 elements and post increment pointer. + @param[in] in_q15 Pointer to pointer that holds address of input. + @return q31 value + */ +__STATIC_FORCEINLINE q31_t arm_nn_read_q15x2_ia(const q15_t **in_q15) +{ + q31_t val; + + memcpy(&val, *in_q15, 4); + *in_q15 += 2; + + return (val); +} + +/** + @brief Read 4 q7 from q7 pointer and post increment pointer. + @param[in] in_q7 Pointer to pointer that holds address of input. + @return q31 value + */ +__STATIC_FORCEINLINE q31_t arm_nn_read_q7x4_ia(const q7_t **in_q7) +{ + q31_t val; + memcpy(&val, *in_q7, 4); + *in_q7 += 4; + + return (val); +} + +/** + @brief Read 2 q15 from q15 pointer. + @param[in] in_q15 pointer to address of input. + @return q31 value + */ +__STATIC_FORCEINLINE q31_t arm_nn_read_q15x2(const q15_t *in_q15) +{ + q31_t val; + memcpy(&val, in_q15, 4); + + return (val); +} + +/** + @brief Read 4 q7 values. + @param[in] in_q7 pointer to address of input. + @return q31 value + */ +__STATIC_FORCEINLINE q31_t arm_nn_read_q7x4(const q7_t *in_q7) +{ + q31_t val; + memcpy(&val, in_q7, 4); + + return (val); +} + +/** + * @brief memset optimized for MVE + * @param[in, out] dst Destination pointer + * @param[in] val Value to set + * @param[in] block_size Number of bytes to copy. + * + */ +__STATIC_FORCEINLINE void arm_memset_q7(q7_t *dst, + const q7_t val, + uint32_t block_size) +{ +#if defined(ARM_MATH_MVEI) + __asm volatile ( + " vdup.8 q0, %[set_val] \n" + " wlstp.8 lr, %[cnt], 1f \n" + "2: \n" + " vstrb.8 q0, [%[in]], 16 \n" + " letp lr, 2b \n" + "1: \n" + :[in] "+r"(dst) + :[cnt] "r"(block_size), [set_val] "r"(val) + :"q0", "memory", "r14"); +#else + memset(dst, val, block_size); +#endif +} + +#if defined (ARM_MATH_DSP) + +/** + * @brief read and expand one q7 word into two q15 words + */ + +__STATIC_FORCEINLINE const q7_t *read_and_pad(const q7_t *source, q31_t * out1, q31_t * out2) +{ + q31_t inA = arm_nn_read_q7x4_ia(&source); + q31_t inAbuf1 = __SXTB16(__ROR((uint32_t)inA, 8)); + q31_t inAbuf2 = __SXTB16(inA); + +#ifndef ARM_MATH_BIG_ENDIAN + *out2 = (int32_t) (__PKHTB (inAbuf1, inAbuf2, 16)); + *out1 = (int32_t) (__PKHBT (inAbuf2, inAbuf1, 16)); +#else + *out1 = (int32_t) (__PKHTB(inAbuf1, inAbuf2, 16)); + *out2 = (int32_t) (__PKHBT(inAbuf2, inAbuf1, 16)); +#endif + + return source; +} + +/** + * @brief read and expand one q7 word into two q15 words with reordering + */ + +__STATIC_FORCEINLINE const q7_t *read_and_pad_reordered(const q7_t *source, q31_t * out1, q31_t * out2) +{ + q31_t inA = arm_nn_read_q7x4_ia(&source); +#ifndef ARM_MATH_BIG_ENDIAN + *out2 = __SXTB16(__ROR((uint32_t)inA, 8)); + *out1 = __SXTB16(inA); +#else + *out1 = __SXTB16(__ROR((uint32_t)inA, 8)); + *out2 = __SXTB16(inA); +#endif + + return source; +} + +/** + * @brief read and expand one q7 word into two q15 words with reordering and add an offset + */ +__STATIC_FORCEINLINE const q7_t *read_and_pad_reordered_with_offset(const q7_t *source, q31_t * out1, q31_t * out2, q31_t offset) +{ + q31_t inA = arm_nn_read_q7x4_ia(&source); + +#ifndef ARM_MATH_BIG_ENDIAN + *out2 = __SXTB16(__ROR((uint32_t)inA, 8)); + *out1 = __SXTB16(inA); +#else + *out1 = __SXTB16(__ROR((uint32_t)inA, 8)); + *out2 = __SXTB16(inA); +#endif + *out1 = __QADD16(*out1,offset); + *out2 = __QADD16(*out2,offset); + + return source; +} + +#endif + + + +/** + * @defgroup NNBasicMath Basic Math Functions for Neural Network Computation + * + * Basic Math Functions for Neural Network Computation + * + */ + +/** + * @brief q7 vector multiplication with variable output shifts + * @param[in] *pSrcA pointer to the first input vector + * @param[in] *pSrcB pointer to the second input vector + * @param[out] *pDst pointer to the output vector + * @param[in] out_shift amount of right-shift for output + * @param[in] blockSize number of samples in each vector + * @return none. + * + * Scaling and Overflow Behavior: + * \par + * The function uses saturating arithmetic. + * Results outside of the allowable q15 range [0x8000 0x7FFF] will be saturated. + */ + +void arm_nn_mult_q15( + q15_t * pSrcA, + q15_t * pSrcB, + q15_t * pDst, + const uint16_t out_shift, + uint32_t blockSize); + +/** + * @brief q7 vector multiplication with variable output shifts + * @param[in] *pSrcA pointer to the first input vector + * @param[in] *pSrcB pointer to the second input vector + * @param[out] *pDst pointer to the output vector + * @param[in] out_shift amount of right-shift for output + * @param[in] blockSize number of samples in each vector + * @return none. + * + * Scaling and Overflow Behavior: + * \par + * The function uses saturating arithmetic. + * Results outside of the allowable q7 range [0x80 0x7F] will be saturated. + */ + +void arm_nn_mult_q7( + q7_t * pSrcA, + q7_t * pSrcB, + q7_t * pDst, + const uint16_t out_shift, + uint32_t blockSize); + +/** + * @brief macro for adding rounding offset + */ +#ifndef ARM_NN_TRUNCATE + #define NN_ROUND(out_shift) ( (0x1u << out_shift) >> 1 ) +#else + #define NN_ROUND(out_shift) 0 +#endif + +// Macros for shortening quantization functions' names and avoid long lines +#define MUL_SAT(a, b) arm_nn_doubling_high_mult((a), (b)) +#define MUL_SAT_MVE(a, b) arm_doubling_high_mult_mve_32x4((a), (b)) +#define MUL_POW2(a, b) arm_nn_mult_by_power_of_two((a), (b)) + + +#define DIV_POW2(a, b) arm_nn_divide_by_power_of_two((a), (b)) +#define DIV_POW2_MVE(a, b) arm_divide_by_power_of_two_mve((a), (b)) + + +#define EXP_ON_NEG(x) arm_nn_exp_on_negative_values((x)) +#define ONE_OVER1(x) arm_nn_one_over_one_plus_x_for_x_in_0_1((x)) + +/** + * @brief Saturating doubling high multiply. Result matches + * NEON instruction VQRDMULH. + * @param[in] m1 Multiplicand. Range: {Q31_MIN, Q31_MAX} + * @param[in] m2 Multiplier. Range: {Q31_MIN, Q31_MAX} + * @return Result of multiplication. + * + */ +__STATIC_FORCEINLINE q31_t arm_nn_doubling_high_mult(const q31_t m1, const q31_t m2) +{ + q31_t result = 0; + // Rounding offset to add for a right shift of 31 + q63_t mult = 1 << 30; + + if ((m1 < 0) ^ (m2 < 0)) + { + mult = 1 - mult; + } + // Gets resolved as a SMLAL instruction + mult = mult + (q63_t)m1 * m2; + + // Utilize all of the upper 32 bits. This is the doubling step + // as well. + result = (int32_t) (mult / (1ll << 31)); + + if ((m1 == m2) && (m1 == (int32_t)Q31_MIN)) + { + result = Q31_MAX; + } + return result; +} + +/** + * @brief Doubling high multiply without saturation. This is intended + * for requantization where the scale is a positive integer + * + * @param[in] m1 Multiplicand. Range: {Q31_MIN, Q31_MAX} + * @param[in] m2 Multiplier Range: {Q31_MIN, Q31_MAX} + * @return Result of multiplication. + * @note The result of this matches that of neon instruction + * VQRDMULH for m1 in range {Q31_MIN, Q31_MAX} and m2 in + * range {Q31_MIN + 1, Q31_MAX}. Saturation occurs when + * m1 equals m2 equals Q31_MIN and that is not handled by + * this function. + * + */ +__STATIC_FORCEINLINE q31_t arm_nn_doubling_high_mult_no_sat(const q31_t m1, const q31_t m2) +{ + q31_t result = 0; + union arm_nn_long_long mult; + + // Rounding offset to add for a right shift of 31 + mult.word.low = 1 << 30; + mult.word.high = 0; + + // Gets resolved as a SMLAL instruction + mult.long_long = mult.long_long + (q63_t)m1 * m2; + + // Utilize all of the upper 32 bits. This is the doubling step + // as well. + result = (int32_t)(mult.long_long >> 31); + + return result; +} + +/** + * @brief Rounding divide by power of two. + * @param[in] dividend - Dividend + * @param[in] exponent - Divisor = power(2, exponent) + * Range: [0, 31] + * @return Rounded result of division. Midpoint is rounded away from zero. + * + */ +__STATIC_FORCEINLINE q31_t arm_nn_divide_by_power_of_two(const q31_t dividend, const q31_t exponent) +{ + q31_t result = 0; + const q31_t remainder_mask = (1 << exponent) - 1; + int32_t remainder = remainder_mask & dividend; + + // Basic division + result = dividend >> exponent; + + // Adjust 'result' for rounding (mid point away from zero) + q31_t threshold = remainder_mask >> 1; + if (result < 0) + { + threshold++; + } + if (remainder > threshold) + { + result++; + } + + return result; +} + +/** + * @brief Requantize a given value. + * @param[in] val Value to be requantized + * @param[in] multiplier multiplier. Range {Q31_MIN + 1, Q32_MAX} + * @param[in] shift left or right shift for 'val * multiplier' + * + * @return Returns (val * multiplier)/(2 ^ shift) + * + */ +__STATIC_FORCEINLINE q31_t arm_nn_requantize(const q31_t val, const q31_t multiplier, const q31_t shift) +{ + return arm_nn_divide_by_power_of_two( + arm_nn_doubling_high_mult_no_sat(val * (1 << LEFT_SHIFT(shift)), multiplier), + RIGHT_SHIFT(shift)); +} + +/** + * @brief memcpy optimized for MVE + * @param[in, out] dst Destination pointer + * @param[in] src Source pointer. + * @param[in] block_size Number of bytes to copy. + * + */ +__STATIC_FORCEINLINE void arm_memcpy_q7(q7_t *__RESTRICT dst, + const q7_t *__RESTRICT src, + uint32_t block_size) +{ +#if defined(ARM_MATH_MVEI) + __asm volatile ( + " wlstp.8 lr, %[cnt], 1f \n" + "2: \n" + " vldrb.8 q0, [%[in]], 16 \n" + " vstrb.8 q0, [%[out]], 16 \n" + " letp lr, 2b \n" + "1: \n" + :[in] "+r"(src) + ,[out] "+r"(dst) + :[cnt] "r"(block_size) + :"q0", "memory", "r14"); +#else + memcpy(dst, src, block_size); +#endif +} + +#if defined(ARM_MATH_MVEI) +/** + * @brief Vector saturating doubling high multiply returning high half. + * @param[in] m1 Multiplicand + * @param[in] m2 Multiplier + * @return Result of multiplication. + * + */ +__STATIC_FORCEINLINE int32x4_t arm_doubling_high_mult_mve(const int32x4_t m1, const q31_t m2) +{ + return vqrdmulhq_n_s32(m1, m2); +} + +/** + * @brief Vector rounding divide by power of two. + * @param[in] dividend - Dividend vector + * @param[in] exponent - Divisor = power(2, exponent) + * Range: [0, 31] + * @return Rounded result of division. Midpoint is rounded away from zero. + * + */ +__STATIC_FORCEINLINE int32x4_t arm_divide_by_power_of_two_mve(const int32x4_t dividend, const q31_t exponent) +{ + const int32x4_t shift = vdupq_n_s32(-exponent); + const int32x4_t fixup = vshrq_n_s32(vandq_s32(dividend, shift), 31); + const int32x4_t fixed_up_dividend = vqaddq_s32(dividend, fixup); + return vrshlq_s32(fixed_up_dividend, shift); +} + +/** + * @brief Requantize a given vector. + * @param[in] val Vector to be requantized + * @param[in] multiplier multiplier + * @param[in] shift shift + * + * @return Returns (val * multiplier)/(2 ^ shift) + * + */ +__STATIC_FORCEINLINE int32x4_t arm_requantize_mve(const int32x4_t val, const q31_t multiplier, const q31_t shift) +{ + return arm_divide_by_power_of_two_mve( + arm_doubling_high_mult_mve(vshlq_s32(val, vdupq_n_s32(LEFT_SHIFT(shift))), multiplier), + RIGHT_SHIFT(shift)); +} + +__STATIC_FORCEINLINE int32x4_t arm_doubling_high_mult_mve_32x4(const int32x4_t m1, const int32x4_t m2) +{ + return vqrdmulhq_s32(m1, m2); +} + +__STATIC_FORCEINLINE int32x4_t arm_divide_by_power_of_two_mve_32x4(const int32x4_t dividend, const int32x4_t exponent) +{ + const int32x4_t shift = -exponent; + const int32x4_t fixup = vshrq_n_s32(vandq_s32(dividend, shift), 31); + const int32x4_t fixed_up_dividend = vqaddq_s32(dividend, fixup); + return vrshlq_s32(fixed_up_dividend, shift); +} + +__STATIC_FORCEINLINE int32x4_t arm_requantize_mve_32x4(const int32x4_t val, const int32x4_t multiplier, const int32x4_t shift) +{ + const int32x4_t zz = vdupq_n_s32(0); + const mve_pred16_t p = vcmpgtq_n_s32(shift, 0); + + const int32x4_t left_shift = vpselq_s32(shift, zz, p); + const int32x4_t right_shift = -vpselq_s32(zz, shift, p); + + return arm_divide_by_power_of_two_mve_32x4(arm_doubling_high_mult_mve_32x4(vshlq_s32(val, left_shift), multiplier), right_shift); +} +#endif + +// @note The following functions are used only for softmax layer, scaled bits = 5 assumed + +__STATIC_FORCEINLINE int32_t arm_nn_exp_on_negative_values(int32_t val) +{ + int32_t mask = 0; + int32_t shift = 24; + + const int32_t val_mod_minus_quarter = (val & ((1 << shift) - 1)) - (1 << shift); + const int32_t remainder = val_mod_minus_quarter - val; + const int32_t x = (val_mod_minus_quarter << 5) + (1 << 28); + const int32_t x2 = MUL_SAT(x, x); + + int32_t result = 1895147668 + MUL_SAT(1895147668, x + + DIV_POW2(MUL_SAT(DIV_POW2(MUL_SAT(x2, x2), 2) + MUL_SAT(x2, x), 715827883) + x2, 1)); + +#define SELECT_IF_NON_ZERO(x) \ +{ \ + mask = MASK_IF_NON_ZERO(remainder & (1 << shift++)); \ + result = SELECT_USING_MASK(mask, MUL_SAT(result, x), result); \ +} + + SELECT_IF_NON_ZERO(1672461947) + SELECT_IF_NON_ZERO(1302514674) + SELECT_IF_NON_ZERO(790015084) + SELECT_IF_NON_ZERO(290630308) + SELECT_IF_NON_ZERO(39332535) + SELECT_IF_NON_ZERO(720401) + SELECT_IF_NON_ZERO(242) + +#undef SELECT_IF_NON_ZERO + + mask = MASK_IF_ZERO(val); + return SELECT_USING_MASK(mask, Q31_MAX, result); +} + +__STATIC_FORCEINLINE q31_t arm_nn_mult_by_power_of_two(const int32_t val, const int32_t exp) +{ + const int32_t thresh = ((1 << (31 - exp)) - 1); + int32_t result = val << exp; + result = SELECT_USING_MASK(MASK_IF_NON_ZERO(val > thresh), Q31_MAX, result); + result = SELECT_USING_MASK(MASK_IF_NON_ZERO(val < -thresh), Q31_MIN, result); + return result; +} + +__STATIC_FORCEINLINE int32_t arm_nn_one_over_one_plus_x_for_x_in_0_1(int32_t val) +{ + const int64_t sum = (int64_t)val + (int64_t)Q31_MAX; + const int32_t half_denominator = (int32_t)((sum + (sum >= 0 ? 1 : -1)) / 2L); + int32_t x = 1515870810 + MUL_SAT(half_denominator, -1010580540); + + const int32_t shift = (1 << 29); + x += MUL_POW2(MUL_SAT(x, shift - MUL_SAT(half_denominator, x)), 2); + x += MUL_POW2(MUL_SAT(x, shift - MUL_SAT(half_denominator, x)), 2); + x += MUL_POW2(MUL_SAT(x, shift - MUL_SAT(half_denominator, x)), 2); + + return MUL_POW2(x, 1); +} + +#ifdef __cplusplus +} +#endif + +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q15.c new file mode 100644 index 0000000..9829557 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q15.c @@ -0,0 +1,100 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_activations_q15.c + * Description: Q15 neural network activation function using direct table look-up + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Acti + * @{ + */ + +/** + * @brief neural network activation function using direct table look-up + * + * @note Refer header file for details. + * + */ + +void arm_nn_activations_direct_q15(q15_t * data, uint16_t size, uint16_t int_width, arm_nn_activation_type type) +{ + uint16_t i = size; + q15_t *pIn = data; + q15_t *pOut = data; + uint16_t shift_size = 8 + 3 - int_width; + uint32_t bit_mask = 0x7FF >> int_width; + uint32_t full_frac = bit_mask + 1; + const q15_t *lookup_table; + + switch (type) + { + case ARM_SIGMOID: + lookup_table = sigmoidTable_q15; + break; + case ARM_TANH: + default: + lookup_table = tanhTable_q15; + break; + } + + while (i) + { + q15_t out; + q15_t in = *pIn++; + q15_t frac = (uint32_t) in & bit_mask; + q15_t value = lookup_table[(uint8_t)(in >> shift_size)]; + if ((in >> shift_size) != 0x7f) + { + q15_t value2 = lookup_table[(uint8_t)(1 + ((uint8_t)(in >> shift_size)))]; + /* doing the interpolation here for better accuracy */ + out = ((q31_t) (full_frac - frac) * value + (q31_t) value2 * frac) >> shift_size; + } else + { + /* the largest positive value does not have a right side for linear interpolation */ + out = value; + } + + *pOut++ = out; + i--; + } + +} + +/** + * @} end of Acti group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q7.c new file mode 100644 index 0000000..e246340 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q7.c @@ -0,0 +1,93 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_activations_q7.c + * Description: Q7 neural network activation function using direct table look-up + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_common_tables.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Acti + * @{ + */ + + /** + * @brief Q7 neural network activation function using direct table look-up + * @param[in,out] data pointer to input + * @param[in] size number of elements + * @param[in] int_width bit-width of the integer part, assume to be smaller than 3 + * @param[in] type type of activation functions + * + * @details + * + * This is the direct table look-up approach. + * + * Assume here the integer part of the fixed-point is <= 3. + * More than 3 just not making much sense, makes no difference with + * saturation followed by any of these activation functions. + */ + +void arm_nn_activations_direct_q7(q7_t * data, uint16_t size, uint16_t int_width, arm_nn_activation_type type) +{ + uint16_t i = size; + q7_t *pIn = data; + q7_t *pOut = data; + q7_t in; + q7_t out; + uint16_t shift_size = 3 - int_width; + const q7_t *lookup_table; + switch (type) + { + case ARM_SIGMOID: + lookup_table = sigmoidTable_q7; + break; + case ARM_TANH: + default: + lookup_table = tanhTable_q7; + break; + } + while (i) + { + in = *pIn++; + out = lookup_table[(uint8_t) (in >> shift_size)]; + *pOut++ = out; + i--; + } +} + +/** + * @} end of Acti group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu6_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu6_s8.c new file mode 100644 index 0000000..5a16e99 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu6_s8.c @@ -0,0 +1,68 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_relu6_s8.c + * Description: Basic s8 version of ReLU6 + * + * $Date: Spetember 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Acti + * @{ + */ + + /* + * Basic ReLU6 function + * + * Refer to header file for details. + * + */ + +void arm_relu6_s8(q7_t *data, uint16_t size) +{ + int32_t i; + + for (i = 0; i < size; i++) + { + int32_t ip = data[i]; + + ip = MAX(ip, 0); + data[i] = MIN(ip, 6); + } +} + +/** + * @} end of Acti group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q15.c new file mode 100644 index 0000000..b457314 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q15.c @@ -0,0 +1,107 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_relu_q15.c + * Description: Q15 version of ReLU + * + * $Date: February 27, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Acti + * @{ + */ + +/** + * @brief Q15 RELU function + * @param[in,out] data pointer to input + * @param[in] size number of elements + * + * @details + * + * Optimized relu with QSUB instructions. + * + */ + +void arm_relu_q15(q15_t *data, uint16_t size) +{ + +#if defined(ARM_MATH_DSP) + /* Run the following code for M cores with DSP extension */ + + uint16_t i = size >> 1; + q15_t *input = data; + q15_t *output = data; + q31_t in; + q31_t buf; + q31_t mask; + + while (i) + { + in = read_q15x2_ia(&input); + + /* extract the first bit */ + buf = __ROR(in & 0x80008000, 15); + + /* if MSB=1, mask will be 0xFF, 0x0 otherwise */ + mask = __QSUB16(0x00000000, buf); + + write_q15x2_ia(&output, in & (~mask)); + i--; + } + + if (size & 0x1) + { + if (*input < 0) + { + *input = 0; + } + input++; + } +#else + /* Run the following code as reference implementation for M cores without DSP extension */ + uint16_t i; + + for (i = 0; i < size; i++) + { + if (data[i] < 0) + data[i] = 0; + } + +#endif /* ARM_MATH_DSP */ +} + +/** + * @} end of Acti group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q7.c new file mode 100644 index 0000000..02a2533 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q7.c @@ -0,0 +1,112 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_relu_q7.c + * Description: Q7 version of ReLU + * + * $Date: May 29, 2020 + * $Revision: V.1.0.2 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Acti + * @{ + */ + + /** + * @brief Q7 RELU function + * @param[in,out] data pointer to input + * @param[in] size number of elements + * + * @details + * + * Optimized relu with QSUB instructions. + * + */ + +void arm_relu_q7(q7_t *data, uint16_t size) +{ + +#if defined(ARM_MATH_DSP) + /* Run the following code for M cores with DSP extension */ + + uint16_t i = size >> 2; + q7_t *input = data; + q7_t *output = data; + q31_t in; + q31_t buf; + q31_t mask; + + while (i) + { + in = read_q7x4_ia(&input); + + /* extract the first bit */ + buf = (int32_t)__ROR((uint32_t)in & 0x80808080, 7); + + /* if MSB=1, mask will be 0xFF, 0x0 otherwise */ + mask = __QSUB8(0x00000000, buf); + + write_q7x4_ia(&output, in & (~mask)); + + i--; + } + + i = size & 0x3; + while (i) + { + if (*input < 0) + { + *input = 0; + } + input++; + i--; + } + +#else + /* Run the following code as reference implementation for cores without DSP extension */ + + uint16_t i; + + for (i = 0; i < size; i++) + { + if (data[i] < 0) + data[i] = 0; + } + +#endif +} + +/** + * @} end of Acti group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_add_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_add_s8.c new file mode 100644 index 0000000..3683ef1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_add_s8.c @@ -0,0 +1,261 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_elementwise_add_s8 + * Description: Element wise add + * + * $Date: July 31, 2020 + * $Revision: V.2.5.1 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" +#if defined(ARM_MATH_MVEI) +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_helium_utils.h" +#endif + +#if defined(ARM_MATH_MVEI) +#define SAT_INPUT_VECT(__INPUT_V, __MULT, __SHIFT) \ + __INPUT_V = arm_doubling_high_mult_mve(__INPUT_V, __MULT); \ + __INPUT_V = arm_divide_by_power_of_two_mve(__INPUT_V, -__SHIFT); +#endif + + +/** + * @note The *_no_sat API does not mean that the input not saturated, Since + * __MULT is a positive integer, it is saturated. The API definition + * has more info about it. + */ +#define SAT_INPUT(__INPUT, __MULT, __SHIFT) \ + __INPUT = arm_nn_doubling_high_mult_no_sat(__INPUT, __MULT); \ + __INPUT = arm_nn_divide_by_power_of_two(__INPUT, -__SHIFT); + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup BasicMath + * @{ + */ + +/* + * s8 element wise add + * + * Refer header file for details. + * + */ + +/* Note: __SHIFT is expected to be <=0 */ + + +arm_status +arm_elementwise_add_s8(const int8_t *input_1_vect, + const int8_t *input_2_vect, + const int32_t input_1_offset, + const int32_t input_1_mult, + const int32_t input_1_shift, + const int32_t input_2_offset, + const int32_t input_2_mult, + const int32_t input_2_shift, + const int32_t left_shift, + int8_t *output, + const int32_t out_offset, + const int32_t out_mult, + const int32_t out_shift, + const int32_t out_activation_min, + const int32_t out_activation_max, + const uint32_t block_size) +{ +#if defined(ARM_MATH_MVEI) + int32_t count = (int32_t)block_size; + + while (count > 0) + { + int32x4_t vect_1; + int32x4_t vect_2; + + mve_pred16_t p = vctp32q((uint32_t)count); + + vect_1 = vldrbq_z_s32(input_1_vect, p); + vect_2 = vldrbq_z_s32(input_2_vect, p); + + vect_1 = vaddq_s32(vect_1, vdupq_n_s32(input_1_offset)); + vect_2 = vaddq_s32(vect_2, vdupq_n_s32(input_2_offset)); + + vect_1 = vshlq_r_s32(vect_1, left_shift); + vect_2 = vshlq_r_s32(vect_2, left_shift); + + SAT_INPUT_VECT(vect_1, input_1_mult, input_1_shift); + SAT_INPUT_VECT(vect_2, input_2_mult, input_2_shift); + + vect_1 = vaddq_s32(vect_1, vect_2); + SAT_INPUT_VECT(vect_1, out_mult, out_shift); + + vect_1 = vaddq_n_s32(vect_1, out_offset); + + vect_1 = vmaxq_s32(vect_1, vdupq_n_s32(out_activation_min)); + vect_1 = vminq_s32(vect_1, vdupq_n_s32(out_activation_max)); + + input_1_vect += 4; + input_2_vect += 4; + vstrbq_p_s32(output, vect_1, p); + + output += 4; + count -= 4; + } +#else + uint32_t loop_count; + int32_t input_1; + int32_t input_2; + int32_t sum; + +#if defined(ARM_MATH_DSP) + int32_t a_1, b_1, a_2, b_2; + + int32_t offset_1_packed, offset_2_packed; + + int8_t r1, r2, r3, r4; + + offset_1_packed = (input_1_offset << 16U) | (input_1_offset & 0x0FFFFL); + offset_2_packed = (input_2_offset << 16U) | (input_2_offset & 0x0FFFFL); + + loop_count = block_size >> 2; + + while (loop_count > 0U) + { + /* 4 outputs are calculated in one loop. The order of calculation is follows the order of output sign extension + intrinsic */ + input_1_vect = read_and_pad_reordered(input_1_vect, &b_1, &a_1); + input_2_vect = read_and_pad_reordered(input_2_vect, &b_2, &a_2); + + a_1 = __SADD16(a_1, offset_1_packed); + b_1 = __SADD16(b_1, offset_1_packed); + + a_2 = __SADD16(a_2, offset_2_packed); + b_2 = __SADD16(b_2, offset_2_packed); + + /* Sum 1 */ + input_1 = (int16_t)(b_1 & 0x0FFFFL) << left_shift; + SAT_INPUT(input_1, input_1_mult, input_1_shift); + + input_2 = (int16_t)(b_2 & 0x0FFFFL) << left_shift; + SAT_INPUT(input_2, input_2_mult, input_2_shift); + + sum = input_1 + input_2; + SAT_INPUT(sum, out_mult, out_shift); + sum += out_offset; + sum = MAX(sum, out_activation_min); + sum = MIN(sum, out_activation_max); + r1 = (q7_t)sum; + + /* Sum 3 */ + input_1 = (int16_t)((b_1 >> 16) & 0x0FFFFL) << left_shift; + SAT_INPUT(input_1, input_1_mult, input_1_shift); + + input_2 = (int16_t)((b_2 >> 16) & 0x0FFFFL) << left_shift; + SAT_INPUT(input_2, input_2_mult, input_2_shift); + + sum = input_1 + input_2; + SAT_INPUT(sum, out_mult, out_shift); + sum += out_offset; + sum = MAX(sum, out_activation_min); + sum = MIN(sum, out_activation_max); + r3 = (q7_t)sum; + + /* Sum 2 */ + input_1 = (int16_t)(a_1 & 0x0FFFFL) << left_shift; + SAT_INPUT(input_1, input_1_mult, input_1_shift); + + input_2 = (int16_t)(a_2 & 0x0FFFFL) << left_shift; + SAT_INPUT(input_2, input_2_mult, input_2_shift); + + sum = input_1 + input_2; + SAT_INPUT(sum, out_mult, out_shift); + sum += out_offset; + sum = MAX(sum, out_activation_min); + sum = MIN(sum, out_activation_max); + r2 = (q7_t)sum; + + /* Sum 4 */ + input_1 = (int16_t)((a_1 >> 16) & 0x0FFFFL) << left_shift; + SAT_INPUT(input_1, input_1_mult, input_1_shift); + + input_2 = (int16_t)((a_2 >> 16) & 0x0FFFFL) << left_shift; + SAT_INPUT(input_2, input_2_mult, input_2_shift); + + sum = input_1 + input_2; + SAT_INPUT(sum, out_mult, out_shift); + sum += out_offset; + sum = MAX(sum, out_activation_min); + sum = MIN(sum, out_activation_max); + r4 = (q7_t)sum; + + write_q7x4_ia(&output, __PACKq7(r1, r2, r3, r4)); + + loop_count--; + } + + loop_count = block_size & 0x3; +#else + loop_count = block_size; +#endif + + while (loop_count > 0U) + { + /* C = A + B */ + + input_1 = (*input_1_vect++ + input_1_offset) << left_shift; + input_2 = (*input_2_vect++ + input_2_offset) << left_shift; + + input_1 = arm_nn_doubling_high_mult(input_1, input_1_mult); + input_1 = arm_nn_divide_by_power_of_two(input_1, -input_1_shift); + + input_2 = arm_nn_doubling_high_mult(input_2, input_2_mult); + input_2 = arm_nn_divide_by_power_of_two(input_2, -input_2_shift); + + sum = input_1 + input_2; + SAT_INPUT(sum, out_mult, out_shift); + sum += out_offset; + + sum = MAX(sum, out_activation_min); + sum = MIN(sum, out_activation_max); + + *output++ = (q7_t)sum; + + /* Decrement loop counter */ + loop_count--; + } + +#endif /* ARM_MATH_MVEI */ + + return (ARM_MATH_SUCCESS); +} + +/** + * @} end of BasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_mul_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_mul_s8.c new file mode 100644 index 0000000..13f65a4 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_mul_s8.c @@ -0,0 +1,205 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_elementwise_mul_s8 + * Description: Element wise multiplication + * + * $Date: May 29, 2020 + * $Revision: V.1.0.3 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup BasicMath + * @{ + */ + +/** + * @brief s8 element wise multiplication of two vectors + * + * @note Refer header file for details. + * + */ + +arm_status +arm_elementwise_mul_s8(const int8_t *input_1_vect, + const int8_t *input_2_vect, + const int32_t input_1_offset, + const int32_t input_2_offset, + int8_t *output, + const int32_t out_offset, + const int32_t out_mult, + const int32_t out_shift, + const int32_t out_activation_min, + const int32_t out_activation_max, + const uint32_t block_size) +{ + + int32_t loop_count; +#if defined(ARM_MATH_MVEI) + + loop_count = (block_size + 3) / 4; + uint32_t num_elements = block_size; + + for (int i = 0; i < loop_count; i++) + { + mve_pred16_t p = vctp32q(num_elements); + + int32x4_t input_1 = vldrbq_z_s32(input_1_vect, p); + input_1 = vaddq_n_s32(input_1, input_1_offset); + + int32x4_t input_2 = vldrbq_z_s32(input_2_vect, p); + input_2 = vaddq_n_s32(input_2, input_2_offset); + + int32x4_t res_0 = vmulq_s32(input_1, input_2); + + res_0 = arm_requantize_mve_32x4(res_0, vdupq_n_s32(out_mult), vdupq_n_s32(out_shift)); + + res_0 += vdupq_n_s32(out_offset); + + res_0 = vmaxq_s32(res_0, vdupq_n_s32(out_activation_min)); + res_0 = vminq_s32(res_0, vdupq_n_s32(out_activation_max)); + + vstrbq_p_s32(output, res_0, p); + input_1_vect += 4; + input_2_vect += 4; + output += 4; + num_elements -= 4; + } + +#else + int32_t input_1; + int32_t input_2; + int32_t mul_res; + +#if defined(ARM_MATH_DSP) + int32_t a_1, b_1, a_2, b_2; + + int32_t offset_1_packed, offset_2_packed; + + int8_t r1, r2, r3, r4; + + offset_1_packed = (input_1_offset << 16U) | (input_1_offset & 0x0FFFFL); + offset_2_packed = (input_2_offset << 16U) | (input_2_offset & 0x0FFFFL); + + loop_count = block_size >> 2; + + while (loop_count > 0U) + { + /* 4 outputs are calculated in one loop. The order of calculation is follows the order of output sign extension + intrinsic */ + input_1_vect = read_and_pad_reordered(input_1_vect, &b_1, &a_1); + input_2_vect = read_and_pad_reordered(input_2_vect, &b_2, &a_2); + + a_1 = __SADD16(a_1, offset_1_packed); + b_1 = __SADD16(b_1, offset_1_packed); + + a_2 = __SADD16(a_2, offset_2_packed); + b_2 = __SADD16(b_2, offset_2_packed); + + /* Mul 1 */ + input_1 = (int16_t)(b_1 & 0x0FFFFL); + input_2 = (int16_t)(b_2 & 0x0FFFFL); + + mul_res = input_1 * input_2; + mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; + + mul_res = MAX(mul_res, out_activation_min); + mul_res = MIN(mul_res, out_activation_max); + r1 = (q7_t)mul_res; + + /* Mul 3 */ + input_1 = (int16_t)((b_1 >> 16U) & 0x0FFFFL); + input_2 = (int16_t)((b_2 >> 16U) & 0x0FFFFL); + + mul_res = input_1 * input_2; + mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; + mul_res = MAX(mul_res, out_activation_min); + mul_res = MIN(mul_res, out_activation_max); + r3 = (q7_t)mul_res; + + /* Mul 2 */ + input_1 = (int16_t)(a_1 & 0x0FFFFL); + input_2 = (int16_t)(a_2 & 0x0FFFFL); + + mul_res = input_1 * input_2; + mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; + mul_res = MAX(mul_res, out_activation_min); + mul_res = MIN(mul_res, out_activation_max); + r2 = (q7_t)mul_res; + + /* Mul 4 */ + input_1 = (int16_t)((a_1 >> 16U) & 0x0FFFFL); + input_2 = (int16_t)((a_2 >> 16U) & 0x0FFFFL); + + mul_res = input_1 * input_2; + mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; + mul_res = MAX(mul_res, out_activation_min); + mul_res = MIN(mul_res, out_activation_max); + r4 = (q7_t)mul_res; + + write_q7x4_ia(&output, __PACKq7(r1, r2, r3, r4)); + + loop_count--; + } + + loop_count = block_size & 0x3; +#else + loop_count = block_size; +#endif + + while (loop_count > 0U) + { + /* C = A * B */ + + input_1 = *input_1_vect++ + input_1_offset; + input_2 = *input_2_vect++ + input_2_offset; + + mul_res = input_1 * input_2; + mul_res = arm_nn_requantize(mul_res, out_mult, out_shift) + out_offset; + + mul_res = MAX(mul_res, out_activation_min); + mul_res = MIN(mul_res, out_activation_max); + + *output++ = (q7_t)mul_res; + + /* Decrement loop counter */ + loop_count--; + } +#endif + return ARM_MATH_SUCCESS; +} + +/** + * @} end of BasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_w.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_w.c new file mode 100644 index 0000000..23be5cb --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_w.c @@ -0,0 +1,68 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_concatenation_s8_w.c + * Description: s8 version of concatenation along the W axis + * + * $Date: October 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Concatenation + * @{ + */ + + /* + * s8 version of concatenation along the W axis + * + * Refer to header file for details. + * + */ +void arm_concatenation_s8_w(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint32_t offset_w) +{ + const uint32_t input_copy_size = input_x * input_y * input_z * input_w; + + output += offset_w * (input_x * input_y * input_z); + + memcpy(output, input, input_copy_size); +} + +/** + * @} end of Concatenation group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_x.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_x.c new file mode 100644 index 0000000..e40b353 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_x.c @@ -0,0 +1,77 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_concatenation_s8_x.c + * Description: s8 version of concatenation along the X axis + * + * $Date: October 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Concatenation + * @{ + */ + + /* + * s8 version of concatenation along the X axis + * + * Refer to header file for details. + * + */ +void arm_concatenation_s8_x(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint16_t output_x, + const uint32_t offset_x) +{ + const uint32_t num_iterations = input_y * input_z * input_w; + + output += offset_x; + + uint32_t i; + + // Copy per row + for (i = 0; i < num_iterations; ++i) + { + memcpy(output, input, input_x); + input += input_x; + output += output_x; + } +} + +/** + * @} end of Concatenation group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_y.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_y.c new file mode 100644 index 0000000..4376945 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_y.c @@ -0,0 +1,78 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_concatenation_s8_y.c + * Description: s8 version of concatenation along the Y axis + * + * $Date: October 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Concatenation + * @{ + */ + + /* + * s8 version of concatenation along the Y axis + * + * Refer to header file for details. + * + */ +void arm_concatenation_s8_y(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint16_t output_y, + const uint32_t offset_y) +{ + const uint32_t num_iterations = input_z * input_w; + const uint32_t input_copy_size = input_x * input_y; + const uint32_t output_stride = input_x * output_y; + + output += offset_y * input_x; + uint32_t i; + + // Copy per tile + for (i = 0; i < num_iterations; ++i) + { + memcpy(output, input, input_copy_size); + input += input_copy_size; + output += output_stride; + } +} + +/** + * @} end of Concatenation group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_z.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_z.c new file mode 100644 index 0000000..527bc84 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_z.c @@ -0,0 +1,77 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_concatenation_s8_z.c + * Description: s8 version of concatenation along the Z axis + * + * $Date: October 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Concatenation + * @{ + */ + + /* + * s8 version of concatenation along the Z axis + * + * Refer to header file for details. + * + */ +void arm_concatenation_s8_z(const int8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_z, + const uint16_t input_w, + int8_t *output, + const uint16_t output_z, + const uint32_t offset_z) +{ + const uint32_t input_copy_size = input_x * input_y * input_z; + const uint32_t output_stride = input_x * input_y * output_z; + + output += offset_z * (input_x * input_y); + + uint32_t i; + + for (i = 0; i < input_w; ++i) + { + memcpy(output, input, input_copy_size); + input += input_copy_size; + output += output_stride; + } +} + +/** + * @} end of Concatenation group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1_x_n_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1_x_n_s8.c new file mode 100644 index 0000000..79ad61c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1_x_n_s8.c @@ -0,0 +1,204 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_1_x_n_s8.c + * Description: s8 version of 1xN convolution using symmetric quantization. + * + * $Date: July 27, 2020 + * $Revision: V.2.0.1 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/* + * 1xN s8 convolution function. + * + * Refer header file for details. + * + */ + +arm_status arm_convolve_1_x_n_s8(const cmsis_nn_context* ctx, + const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, + const q7_t *input_data, + const cmsis_nn_dims* filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims* bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims* output_dims, + q7_t *output_data) +{ + (void)bias_dims; + arm_status status = ARM_MATH_SUCCESS; + if (output_dims->w % 4 != 0) + { + status = ARM_MATH_SIZE_MISMATCH; + goto out; + } + +#if defined(ARM_MATH_MVEI) + q15_t *buffer_a = (q15_t *)ctx->buf; + + const uint16_t input_x = input_dims->w; + const uint16_t kernel_x = filter_dims->w; + const uint16_t output_x = output_dims->w; + const uint16_t output_ch = output_dims->c; + const uint16_t input_ch = input_dims->c; + const uint16_t pad_x = conv_params->padding.w; + const uint16_t stride_x = conv_params->stride.w; + + const int32_t input_offset = conv_params->input_offset; + const int32_t out_offset = conv_params->output_offset; + const int32_t out_activation_min = conv_params->activation.min; + const int32_t out_activation_max = conv_params->activation.max; + int32_t *output_mult = quant_params->multiplier; + int32_t *output_shift = quant_params->shift; + + for (int i_out_x = 0; i_out_x <= (output_x - 4); i_out_x += 4) + { + int32_t input_begin_idx[4]; + int32_t ker_begin_idx[4]; + int32_t ker_end_idx[4]; + + for (int i = 0; i < 4; i++) + { + const int32_t est_input_x_idx = stride_x * (i_out_x + i) - pad_x; + input_begin_idx[i] = MAX(0, est_input_x_idx); + ker_begin_idx[i] = MAX(0, -est_input_x_idx); + ker_end_idx[i] = MIN(kernel_x, input_x - est_input_x_idx); + } + + for (int i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) + { + int32x4_t s_offset; + int32_t acc[4]; + if ((ker_begin_idx[0] != 0) || (ker_end_idx[3] != kernel_x)) + { + int32_t sum_row[4]; + + (void)arm_nn_mat_mul_core_1x_s8((ker_end_idx[0] - ker_begin_idx[0]) * input_ch, + input_data + input_begin_idx[0] * input_ch, + filter_data + (input_ch * kernel_x * i_out_ch) + (ker_begin_idx[0] * input_ch), + &sum_row[0], + &acc[0]); + (void)arm_nn_mat_mul_core_1x_s8((ker_end_idx[1] - ker_begin_idx[1]) * input_ch, + input_data + input_begin_idx[1] * input_ch, + filter_data + (input_ch * kernel_x * i_out_ch) + (ker_begin_idx[1] * input_ch), + &sum_row[1], + &acc[1]); + + (void)arm_nn_mat_mul_core_1x_s8((ker_end_idx[2] - ker_begin_idx[2]) * input_ch, + input_data + input_begin_idx[2] * input_ch, + filter_data + (input_ch * kernel_x * i_out_ch) + (ker_begin_idx[2] * input_ch), + &sum_row[2], + &acc[2]); + + (void)arm_nn_mat_mul_core_1x_s8((ker_end_idx[3] - ker_begin_idx[3]) * input_ch, + input_data + input_begin_idx[3] * input_ch, + filter_data + (input_ch * kernel_x * i_out_ch) + (ker_begin_idx[3] * input_ch), + &sum_row[3], + &acc[3]); + + s_offset = vldrwq_s32(sum_row); + } + else + { + int32_t sum_row; + (void)arm_nn_mat_mul_core_4x_s8(kernel_x * input_ch, + stride_x * input_ch, + input_data + input_begin_idx[0] * input_ch, + filter_data + (input_ch * kernel_x * i_out_ch), + &sum_row, + acc); + + s_offset = vdupq_n_s32(sum_row); + } + int32x4_t res = vldrwq_s32(acc); + s_offset = vmulq_n_s32(s_offset, input_offset); + res = vaddq_s32(res, s_offset); + if (bias_data) + { + res = vaddq_n_s32(res, bias_data[i_out_ch]); + } + res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]); + res = vaddq_n_s32(res, out_offset); + + res = vmaxq_s32(res, vdupq_n_s32(out_activation_min)); + res = vminq_s32(res, vdupq_n_s32(out_activation_max)); + + const uint32x4_t scatter_offset = {0, output_ch, output_ch * 2, output_ch * 3}; + vstrbq_scatter_offset_s32(output_data, scatter_offset, res); + output_data++; + } + output_data += (3 * output_ch); + } + +#else + status = arm_convolve_s8(ctx, + conv_params, + quant_params, + input_dims, + input_data, + filter_dims, + filter_data, + bias_dims, + bias_data, + output_dims, + output_data); +#endif + +out: + /* Return to application */ + return status; +} + +int32_t arm_convolve_1_x_n_s8_get_buffer_size(const cmsis_nn_dims* input_dims, + const cmsis_nn_dims* filter_dims) +{ +#if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI) + return (2 * input_dims->c * filter_dims->w * filter_dims->h) * sizeof(int16_t); +#else + (void)input_dims; + (void)filter_dims; + return 0; +#endif +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c new file mode 100644 index 0000000..c529f2e --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c @@ -0,0 +1,239 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_1x1_HWC_q7_fast_nonsquare.c + * Description: Fast Q7 version of 1x1 convolution (non-square shape) + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/** + * @brief Fast Q7 version of 1x1 convolution (non-sqaure shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimention x + * @param[in] dim_im_in_y input tensor dimention y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is optimized for convolution with 1x1 kernel size (i.e., dim_kernel_x=1 + * and dim_kernel_y=1). It can be used for the second half of MobileNets [1] after depthwise + * separable convolution. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 4 + * ch_im_out is multiple of 2 + * + * [1] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications + * https://arxiv.org/abs/1704.04861 + */ + +arm_status arm_convolve_1x1_HWC_q7_fast_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + (void)dim_im_in_y; + int16_t i_out_y, i_out_x; + int16_t i_ch_out; + + /* ----------------------- + * Here we use bufferA as q15_t internally as computation are done with q15_t level + * im2col are done to output in q15_t format from q7_t input + */ + + q15_t *pBuffer = bufferA; + q7_t *pOut = Im_out; + + if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0 || dim_kernel_x != 1 || dim_kernel_y != 1 + || padding_x != 0 || padding_y != 0 || stride_x != 1 || stride_y != 1) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) + { + /* This part implements the im2col function */ + arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_out_y * dim_im_in_x + i_out_x) * ch_im_in, pBuffer, + ch_im_in); + pBuffer += ch_im_in; + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in, bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* check if there is left-over for compute */ + if (pBuffer != bufferA) + { + const q7_t *pA = wt; + for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++) + { + q31_t sum = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift); + const q15_t *pB = bufferA; + /* basically each time it process 4 entries */ + uint16_t colCnt = ch_im_in * dim_kernel_x * dim_kernel_y >> 2; + + while (colCnt) + { + + q31_t inA1, inA2; + q31_t inB1, inB2; + + pA = read_and_pad_reordered(pA, &inA1, &inA2); + + inB1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inA1, inB1, sum); + inB2 = arm_nn_read_q15x2_ia(&pB); + + sum = __SMLAD(inA2, inB2, sum); + + colCnt--; + } + colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + sum += inA1 * inB1; + colCnt--; + } + *pOut = (q7_t) __SSAT((sum >> out_shift), 8); + pOut++; + + } + + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0 || dim_kernel_x != 1 || dim_kernel_y != 1 + || padding_x != 0 || padding_y != 0 || stride_x != 1 || stride_y != 1) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out_y; j++) + { + for (k = 0; k < dim_im_out_x; k++) + { + conv_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel_y; m++) + { + for (n = 0; n < dim_kernel_x; n++) + { + // if-for implementation + in_row = stride_y * j + m - padding_y; + in_col = stride_x * k + n - padding_x; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] * + wt[i * ch_im_in * dim_kernel_y * dim_kernel_x + (m * dim_kernel_y + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c new file mode 100644 index 0000000..0a53f16 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c @@ -0,0 +1,189 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_1x1_s8_fast.c + * Description: Fast q7 version of 1x1 convolution (non-square shape) + * + * $Date: July 27, 2020 + * $Revision: V.2.0.2 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" + +#define DIM_KER_X (1U) +#define DIM_KER_Y (1U) + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/* + * Fast s8 version for 1x1 convolution (non-square shape) + * + * Refer header file for details. + * + */ + +arm_status arm_convolve_1x1_s8_fast(const cmsis_nn_context *ctx, + const cmsis_nn_conv_params *conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims *bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data) +{ + if (input_dims->c % 4 != 0 || + conv_params->padding.w != 0 || conv_params->padding.h != 0 || + conv_params->stride.w != 1 || conv_params->stride.h != 1) + { + return ARM_MATH_SIZE_MISMATCH; + } + + (void)ctx; + (void)filter_dims; + (void)bias_dims; + +#if defined(ARM_MATH_MVEI) + + const int32_t col_len = input_dims->w * input_dims->h * input_dims->n; + const int32_t output_ch = output_dims->c; + const int32_t input_ch = input_dims->c; + const int32_t input_offset = conv_params->input_offset; + const int32_t out_offset = conv_params->output_offset; + const int32_t out_activation_min = conv_params->activation.min; + const int32_t out_activation_max = conv_params->activation.max; + int32_t *output_mult = quant_params->multiplier; + int32_t *output_shift = quant_params->shift; + + for (int i_items = 0; i_items <= (col_len - 4); i_items += 4) + { + for (int i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) + { + int32_t sum_row = 0; + int32_t temp_out[4]; + + (void)arm_nn_mat_mul_core_4x_s8(input_ch, + input_ch, + input_data + i_items * input_ch, + filter_data + i_out_ch * input_ch, + &sum_row, + temp_out); + int32x4_t res = vldrwq_s32(temp_out); + if (bias_data) + { + res = vaddq_n_s32(res, bias_data[i_out_ch]); + } + sum_row = sum_row * input_offset; + res = vaddq_n_s32(res, sum_row); + res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]); + res = vaddq_n_s32(res, out_offset); + + res = vmaxq_s32(res, vdupq_n_s32(out_activation_min)); + res = vminq_s32(res, vdupq_n_s32(out_activation_max)); + + const uint32x4_t scatter_offset = {0, (uint32_t)output_ch, + (uint32_t)output_ch * 2, + (uint32_t)output_ch * 3}; + vstrbq_scatter_offset_s32(output_data, scatter_offset, res); + output_data++; + } + output_data += (3 * output_ch); + } + + /* Handle left over elements */ + for (int i_items = (col_len & ~0x3); i_items < col_len; i_items++) + { + for (int i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) + { + int32_t sum_row = 0; + + int32_t acc; + (void)arm_nn_mat_mul_core_1x_s8(input_ch, + input_data + i_items * input_ch, + filter_data + i_out_ch * input_ch, + &sum_row, + &acc); + if (bias_data) + { + acc += bias_data[i_out_ch]; + } + sum_row = (sum_row * input_offset); + acc += sum_row; + acc = arm_nn_requantize(acc, output_mult[i_out_ch], output_shift[i_out_ch]); + acc += out_offset; + + acc = MAX(acc, out_activation_min); + acc = MIN(acc, out_activation_max); + *output_data++ = acc; + } + } + +#else + /* Run the following code as reference implementation for Cortex-M processors with or without DSP extension */ + + const int32_t lhs_rows = input_dims->w * input_dims->h * input_dims->n; + const int32_t rhs_rows = output_dims->c; + const int32_t rhs_cols = input_dims->c; + + arm_nn_mat_mult_nt_t_s8(input_data, + filter_data, + bias_data, + output_data, + quant_params->multiplier, + quant_params->shift, + lhs_rows, + rhs_rows, + rhs_cols, + conv_params->input_offset, + conv_params->output_offset, + conv_params->activation.min, + conv_params->activation.max); + +#endif + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +int32_t arm_convolve_1x1_s8_fast_get_buffer_size(const cmsis_nn_dims *input_dims) +{ + (void)input_dims; + return 0; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c new file mode 100644 index 0000000..72201f5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c @@ -0,0 +1,210 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q15_basic.c + * Description: Q15 version of convolution + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + + /** + * @brief Basic Q15 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * + * Buffer size: + * + * bufferA size: ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * This basic version is designed to work for any input tensor and weight + * dimension. + */ + +arm_status +arm_convolve_HWC_q15_basic(const q15_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q15_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q15_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q15_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + uint16_t im2col_out_pixel_index = 0; + q15_t *pBuffer = bufferA; + q15_t *pOut = Im_out; + q15_t *im_buffer = bufferA; + const q15_t *pA; + int i; + + /* This part implements the im2col function */ + for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* Filling 0 for out-of-bound paddings */ + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */ + memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in); + } + pBuffer += ch_im_in; + } + } + + pA = wt; + for (i = 0; i < ch_im_out; i++) + { + q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + const q15_t *pB = im_buffer; + uint16_t colCnt = ch_im_in * dim_kernel * dim_kernel >> 2; + while (colCnt) + { + q31_t inA1 = arm_nn_read_q15x2_ia(&pA); + q31_t inB1 = arm_nn_read_q15x2_ia(&pB); + q31_t inA2 = arm_nn_read_q15x2_ia(&pA); + q31_t inB2 = arm_nn_read_q15x2_ia(&pB); + + sum = __SMLAD(inA1, inB1, sum); + sum = __SMLAD(inA2, inB2, sum); + + colCnt--; + } + colCnt = ch_im_in * dim_kernel * dim_kernel & 0x3; + while (colCnt) + { + q15_t inA1 = *pA++; + q15_t inB1 = *pB++; + sum += inA1 * inB1; + colCnt--; + } + *pOut = (q15_t) __SSAT((sum >> out_shift), 16); + pOut++; + } + + /* counter reset */ + pBuffer = im_buffer; + im2col_out_pixel_index++; + } + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out; j++) + { + for (k = 0; k < dim_im_out; k++) + { + conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel; m++) + { + for (n = 0; n < dim_kernel; n++) + { + in_row = stride * j + m - padding; + in_col = stride * k + n - padding; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += + Im_in[(in_row * dim_im_in + in_col) * ch_im_in + + l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel + + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c new file mode 100644 index 0000000..6545129 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c @@ -0,0 +1,258 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q15_fast.c + * Description: Fast Q15 version of convolution + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + + /** + * @brief Fast Q15 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * @details + * + * Buffer size: + * + * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * Input dimension constraints: + * + * ch_im_in is multiple of 2 + * + * ch_im_out is multipe of 2 + * + */ + +arm_status +arm_convolve_HWC_q15_fast(const q15_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q15_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q15_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q15_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + q15_t *pBuffer = bufferA; + q15_t *im_buffer = bufferA; + q15_t *pOut = Im_out; + + if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + /* This part implements the im2col function */ + for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */ + memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (i_out_x & 0x1) + { + int i; + /* initialize the matrix pointers for A */ + const q15_t *pA = wt; + + /* set up the second output pointers */ + q15_t *pOut2 = pOut + ch_im_out; + + /* this loop over rows in A */ + for (i = 0; i < ch_im_out; i += 2) + { + /* setup pointers for B */ + const q15_t *pB = im_buffer; + const q15_t *pB2 = pB + ch_im_in * dim_kernel * dim_kernel; + + /* aling the second pointer for A */ + const q15_t *pA2 = pA + ch_im_in * dim_kernel * dim_kernel; + + /* init the sum with bias */ + q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = ch_im_in * dim_kernel * dim_kernel >> 1; + /* accumulate over the vector */ + while (colCnt) + { + q31_t inA1 = arm_nn_read_q15x2_ia(&pA); + q31_t inB1 = arm_nn_read_q15x2_ia(&pB); + q31_t inA2 = arm_nn_read_q15x2_ia(&pA2); + q31_t inB2 = arm_nn_read_q15x2_ia(&pB2); + + sum = __SMLAD(inA1, inB1, sum); + sum2 = __SMLAD(inA1, inB2, sum2); + sum3 = __SMLAD(inA2, inB1, sum3); + sum4 = __SMLAD(inA2, inB2, sum4); + + colCnt--; + } /* while over colCnt */ + colCnt = ch_im_in * dim_kernel * dim_kernel & 0x1; + while (colCnt) + { + q15_t inA1 = *pA++; + q15_t inB1 = *pB++; + q15_t inA2 = *pA2++; + q15_t inB2 = *pB2++; + + sum += inA1 * inB1; + sum2 += inA1 * inB2; + sum3 += inA2 * inB1; + sum4 += inA2 * inB2; + colCnt--; + } /* while over colCnt */ + *pOut++ = (q15_t) __SSAT(sum >> out_shift, 16); + *pOut++ = (q15_t) __SSAT(sum3 >> out_shift, 16); + *pOut2++ = (q15_t) __SSAT(sum2 >> out_shift, 16); + *pOut2++ = (q15_t) __SSAT(sum4 >> out_shift, 16); + + /* skip the row computed with A2 */ + pA += ch_im_in * dim_kernel * dim_kernel; + } /* for over ch_im_out */ + + pOut += ch_im_out; + /* counter reset */ + pBuffer = im_buffer; + } + } + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out; j++) + { + for (k = 0; k < dim_im_out; k++) + { + conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel; m++) + { + for (n = 0; n < dim_kernel; n++) + { + in_row = stride * j + m - padding; + in_col = stride * k + n - padding; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += + Im_in[(in_row * dim_im_in + in_col) * ch_im_in + + l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel + + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c new file mode 100644 index 0000000..8befc55 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c @@ -0,0 +1,268 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q15_fast.c + * Description: Fast Q15 version of convolution + * + * $Date: 24. May 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + + /** + * @brief Fast Q15 convolution function (non-sqaure shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimention x + * @param[in] dim_im_in_y input tensor dimention y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * @details + * + * Buffer size: + * + * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * Input dimension constraints: + * + * ch_im_in is multiple of 2 + * + * ch_im_out is multipe of 2 + * + */ + +arm_status +arm_convolve_HWC_q15_fast_nonsquare(const q15_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q15_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q15_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q15_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + q15_t *pBuffer = bufferA; + q15_t *im_buffer = bufferA; + q15_t *pOut = Im_out; + + if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + /* This part implements the im2col function */ + for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) + { + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */ + memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (i_out_x & 0x1) + { + int i; + /* initialize the matrix pointers for A */ + const q15_t *pA = wt; + + /* set up the second output pointers */ + q15_t *pOut2 = pOut + ch_im_out; + + /* this loop over rows in A */ + for (i = 0; i < ch_im_out; i += 2) + { + /* setup pointers for B */ + const q15_t *pB = im_buffer; + const q15_t *pB2 = pB + ch_im_in * dim_kernel_y * dim_kernel_x; + + /* aling the second pointer for A */ + const q15_t *pA2 = pA + ch_im_in * dim_kernel_y * dim_kernel_x; + + /* init the sum with bias */ + q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = ch_im_in * dim_kernel_y * dim_kernel_x >> 1; + /* accumulate over the vector */ + while (colCnt) + { + q31_t inA1 = arm_nn_read_q15x2_ia(&pA); + q31_t inB1 = arm_nn_read_q15x2_ia(&pB); + q31_t inA2 = arm_nn_read_q15x2_ia(&pA2); + q31_t inB2 = arm_nn_read_q15x2_ia(&pB2); + + sum = __SMLAD(inA1, inB1, sum); + sum2 = __SMLAD(inA1, inB2, sum2); + sum3 = __SMLAD(inA2, inB1, sum3); + sum4 = __SMLAD(inA2, inB2, sum4); + + colCnt--; + } /* while over colCnt */ + colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x1; + while (colCnt) + { + q15_t inA1 = *pA++; + q15_t inB1 = *pB++; + q15_t inA2 = *pA2++; + q15_t inB2 = *pB2++; + + sum += inA1 * inB1; + sum2 += inA1 * inB2; + sum3 += inA2 * inB1; + sum4 += inA2 * inB2; + colCnt--; + } /* while over colCnt */ + *pOut++ = (q15_t) __SSAT(sum >> out_shift, 16); + *pOut++ = (q15_t) __SSAT(sum3 >> out_shift, 16); + *pOut2++ = (q15_t) __SSAT(sum2 >> out_shift, 16); + *pOut2++ = (q15_t) __SSAT(sum4 >> out_shift, 16); + + /* skip the row computed with A2 */ + pA += ch_im_in * dim_kernel_y * dim_kernel_x; + } /* for over ch_im_out */ + + pOut += ch_im_out; + /* counter reset */ + pBuffer = im_buffer; + } + } + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out_y; j++) + { + for (k = 0; k < dim_im_out_x; k++) + { + conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel_y; m++) + { + for (n = 0; n < dim_kernel_x; n++) + { + in_row = stride_y * j + m - padding_y; + in_col = stride_x * k + n - padding_x; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += + Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + + l] * wt[i * ch_im_in * dim_kernel_x * dim_kernel_y + (m * dim_kernel_x + + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c new file mode 100644 index 0000000..86bd70d --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c @@ -0,0 +1,282 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q7_RGB.c + * Description: Q7 version of convolution for RGB image + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + + /** + * @brief Q7 convolution function for RGB image + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * @details + * + * Buffer size: + * + * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * Input dimension constraints: + * + * ch_im_in equals 3 + * + * This kernel is written exclusively for convolution with ch_im_in + * equals 3. This applies on the first layer of CNNs which has input + * image with RGB format. + */ + +arm_status +arm_convolve_HWC_q7_RGB(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, const uint16_t dim_im_out, q15_t * bufferA, q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + /* + * Here we use bufferA as q15_t internally as computation are done with q15_t level + * im2col are done to output in q15_t format from q7_t input + */ + q15_t *pBuffer = bufferA; + q7_t *pOut = Im_out; + + // check if number of input channels is 3 + if (ch_im_in != 3) + { + return ARM_MATH_SIZE_MISMATCH; + } + // This part implements the im2col function + for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* Equivalent to arm_fill_q15(0, pBuffer, ch_im_in) with assumption: ch_im_in = 3 */ + *__SIMD32(pBuffer) = 0x0; + *(pBuffer + 2) = 0; + pBuffer += 3; + } else + { + /* + * Equivalent to: + * arm_q7_to_q15_no_shift( (q7_t*)Im_in+(i_ker_y*dim_im_in+i_ker_x)*3, pBuffer, 3); + */ + + const q7_t *pPixel = Im_in + (i_ker_y * dim_im_in + i_ker_x) * 3; + q31_t buf = arm_nn_read_q7x4(pPixel); + + union arm_nnword top; + union arm_nnword bottom; + + top.word = __SXTB16(buf); + bottom.word = __SXTB16(__ROR(buf, 8)); + +#ifndef ARM_MATH_BIG_ENDIAN + /* + * little-endian, | omit | 3rd | 2nd | 1st | + * MSB LSB + * top | 3rd | 1st |; bottom | omit | 2nd | + * + * version 1, need to swap 2nd and 3rd weight + * *__SIMD32(pBuffer) = top.word; + * *(pBuffer+2) = bottom.half_words[0]; + * + * version 2, no weight shuffling required + */ + *pBuffer++ = top.half_words[0]; + *__SIMD32(pBuffer) = __PKHBT(bottom.word, top.word, 0); +#else + /* + * big-endian, | 1st | 2nd | 3rd | omit | + * MSB LSB + * top | 2nd | omit |; bottom | 1st | 3rd | + * + * version 1, need to swap 2nd and 3rd weight + * *__SIMD32(pBuffer) = bottom.word; + * *(pBuffer+2) = top.half_words[1]; + * + * version 2, no weight shuffling required + */ + *pBuffer++ = bottom.half_words[0]; + *__SIMD32(pBuffer) = __PKHTB(top.word, bottom.word, 0); +#endif + pBuffer += 2; + } + } + } + + if (pBuffer == bufferA + 2 * 3 * dim_kernel * dim_kernel) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15(wt, bufferA, + ch_im_out, + 3 * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); + + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* left-over because odd number of output pixels */ + if (pBuffer != bufferA) + { + const q7_t *pA = wt; + int i; + + for (i = 0; i < ch_im_out; i++) + { + q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + q15_t *pB = bufferA; + /* basically each time it process 4 entries */ + uint16_t colCnt = 3 * dim_kernel * dim_kernel >> 2; + + while (colCnt) + { + + q31_t inA1, inA2; + q31_t inB1, inB2; + + pA = read_and_pad(pA, &inA1, &inA2); + + inB1 = arm_nn_read_q15x2_ia((const q15_t **)&pB); + sum = __SMLAD(inA1, inB1, sum); + inB2 = arm_nn_read_q15x2_ia((const q15_t **)&pB); + sum = __SMLAD(inA2, inB2, sum); + + colCnt--; + } + colCnt = 3 * dim_kernel * dim_kernel & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + sum += inA1 * inB1; + colCnt--; + } + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + } + } +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + // check if number of input channels is 3 + if (ch_im_in != 3) + { + return ARM_MATH_SIZE_MISMATCH; + } + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out; j++) + { + for (k = 0; k < dim_im_out; k++) + { + conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel; m++) + { + for (n = 0; n < dim_kernel; n++) + { + /* if-for implementation */ + in_row = stride * j + m - padding; + in_col = stride * k + n - padding; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += + Im_in[(in_row * dim_im_in + in_col) * ch_im_in + + l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel + + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return (ARM_MATH_SUCCESS); +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c new file mode 100644 index 0000000..68477dd --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c @@ -0,0 +1,234 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q7_basic.c + * Description: Q7 version of convolution + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + + /** + * @brief Basic Q7 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * + * Buffer size: + * + * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * This basic version is designed to work for any input tensor and weight + * dimension. + */ + +arm_status +arm_convolve_HWC_q7_basic(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + /* + * Here we use bufferA as q15_t internally as computation are done with q15_t level + * im2col are done to output in q15_t format from q7_t input + */ + q15_t *pBuffer = bufferA; + q7_t *pOut = Im_out; + + /* This part implements the im2col function */ + for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* Filling 0 for out-of-bound paddings */ + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + /* Copying the pixel data to column */ + arm_q7_to_q15_no_shift((q7_t *) + Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + /* Computation is filed for every 2 columns */ + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15(wt, bufferA, + ch_im_out, + ch_im_in * + dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); + + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* left-over because odd number of output pixels */ + if (pBuffer != bufferA) + { + const q7_t *pA = wt; + int i; + + for (i = 0; i < ch_im_out; i++) + { + /* Load the accumulator with bias first */ + q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + + /* Point to the beging of the im2col buffer */ + const q15_t *pB = bufferA; + + /* Each time it process 4 entries */ + uint16_t colCnt = ch_im_in * dim_kernel * dim_kernel >> 2; + + while (colCnt) + { + q31_t inA1, inA2; + q31_t inB1, inB2; + + pA = read_and_pad(pA, &inA1, &inA2); + + inB1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inA1, inB1, sum); + inB2 = arm_nn_read_q15x2_ia(&pB); + + sum = __SMLAD(inA2, inB2, sum); + + colCnt--; + } + colCnt = ch_im_in * dim_kernel * dim_kernel & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + sum += inA1 * inB1; + colCnt--; + } + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + } + } +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out; j++) + { + for (k = 0; k < dim_im_out; k++) + { + conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel; m++) + { + for (n = 0; n < dim_kernel; n++) + { + // if-for implementation + in_row = stride * j + m - padding; + in_col = stride * k + n - padding; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += + Im_in[(in_row * dim_im_in + in_col) * ch_im_in + + l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel + + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c new file mode 100644 index 0000000..51bd134 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c @@ -0,0 +1,232 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q7_basic.c + * Description: Q7 version of convolution + * + * $Date: 13. July 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + + /** + * @brief Basic Q7 convolution function (non-sqaure shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimention x + * @param[in] dim_im_in_y input tensor dimention y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns ARM_MATH_SUCCESS + */ + +arm_status arm_convolve_HWC_q7_basic_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + /* + * Here we use bufferA as q15_t internally as computation are done with q15_t level + * im2col are done to output in q15_t format from q7_t input + */ + q15_t *pBuffer = bufferA; + q7_t *pOut = Im_out; + + /* This part implements the im2col function */ + for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) + { + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x) + { + /* Filling 0 for out-of-bound paddings */ + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + /* Copying the pixel data to column */ + arm_q7_to_q15_no_shift((q7_t *) + Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + /* Computation is filed for every 2 columns */ + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_y * dim_kernel_x) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15(wt, bufferA, + ch_im_out, + ch_im_in * + dim_kernel_y * dim_kernel_x, bias_shift, out_shift, bias, pOut); + + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* left-over because odd number of output pixels */ + if (pBuffer != bufferA) + { + const q7_t *pA = wt; + int i; + + for (i = 0; i < ch_im_out; i++) + { + /* Load the accumulator with bias first */ + q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + + /* Point to the beging of the im2col buffer */ + const q15_t *pB = bufferA; + + /* Each time it process 4 entries */ + uint16_t colCnt = ch_im_in * dim_kernel_y * dim_kernel_x >> 2; + + while (colCnt) + { + q31_t inA1, inA2; + q31_t inB1, inB2; + + pA = read_and_pad(pA, &inA1, &inA2); + + inB1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inA1, inB1, sum); + inB2 = arm_nn_read_q15x2_ia(&pB); + + sum = __SMLAD(inA2, inB2, sum); + + colCnt--; + } + colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + sum += inA1 * inB1; + colCnt--; + } + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + } + } +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out_y; j++) + { + for (k = 0; k < dim_im_out_x; k++) + { + conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel_y; m++) + { + for (n = 0; n < dim_kernel_x; n++) + { + // if-for implementation + in_row = stride_y * j + m - padding_y; + in_col = stride_x * k + n - padding_x; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += + Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] * + wt[i * ch_im_in * dim_kernel_y * dim_kernel_x + + (m * dim_kernel_x + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c new file mode 100644 index 0000000..3941aa0 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c @@ -0,0 +1,411 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q7_fast.c + * Description: Fast Q7 version of convolution + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + + /** + * @brief Fast Q7 convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * @details + * + * Buffer size: + * + * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * Input dimension constraints: + * + * ch_im_in is multiple of 4 ( because of the SIMD32 read and swap ) + * + * ch_im_out is multipe of 2 ( bacause 2x2 mat_mult kernel ) + * + * The im2col converts the Q7 tensor input into Q15 column, which is stored in + * bufferA. There is reordering happenning during this im2col process with + * arm_q7_to_q15_reordered_no_shift. For every four elements, the second and + * third elements are swapped. + * + * The computation kernel arm_nn_mat_mult_kernel_q7_q15_reordered does the + * GEMM computation with the reordered columns. + * + * To speed-up the determination of the padding condition, we split the + * computation into 3x3 parts, i.e., {top, mid, bottom} X {left, mid, right}. + * This reduces the total number of boundary condition checks and improves + * the data copying performance. + */ + +arm_status +arm_convolve_HWC_q7_fast(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + /* + * Here we use bufferA as q15_t internally as computation are done with q15_t level + * im2col are done to output in q15_t format from q7_t input + */ + + q15_t *pBuffer = bufferA; + q7_t *pOut = Im_out; + + if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + /* + * Here we split the entire matrix into three regions depending on the padding situation + * Top: i_out_y from 0 to padding - 1 + * Middle: i_out_y from padding to dim_im_out-padding-1 + * Bottom: i_out_y from dim_im_out-padding to dim_im_out-1 + */ + + /* top part */ + for (i_out_y = 0; i_out_y < padding; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift + ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, + bufferA, + ch_im_out, + ch_im_in + * + dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* middle part, here we also divide the x into left, mid and right */ + for (; i_out_y < dim_im_out - padding; i_out_y++) + { + + /* left part */ + for (i_out_x = 0; i_out_x < padding; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift + ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, + bufferA, + ch_im_out, + ch_im_in + * + dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + + /* mid part */ + for (; i_out_x < dim_im_out - padding; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + + + (i_ker_y * + dim_im_in + + i_out_x * + stride - padding) * ch_im_in, pBuffer, ch_im_in * dim_kernel); + pBuffer += ch_im_in * dim_kernel; + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, + bufferA, + ch_im_out, + ch_im_in + * + dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + + /* right part */ + for (; i_out_x < dim_im_out; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift + ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, + bufferA, + ch_im_out, + ch_im_in + * + dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + } + + for (; i_out_y < dim_im_out; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift + ((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel * dim_kernel) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, + bufferA, + ch_im_out, + ch_im_in + * + dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* check if there is left-over for compute */ + if (pBuffer != bufferA) + { + const q7_t *pA = wt; + int i; + + for (i = 0; i < ch_im_out; i++) + { + q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift); + const q15_t *pB = bufferA; + /* each time it process 4 entries */ + uint16_t colCnt = ch_im_in * dim_kernel * dim_kernel >> 2; + + while (colCnt) + { + + q31_t inA1, inA2; + q31_t inB1, inB2; + + pA = read_and_pad_reordered(pA, &inA1, &inA2); + + inB1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inA1, inB1, sum); + inB2 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inA2, inB2, sum); + + colCnt--; + } + colCnt = ch_im_in * dim_kernel * dim_kernel & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + sum += inA1 * inB1; + colCnt--; + } + *pOut = (q7_t) __SSAT((sum >> out_shift), 8); + pOut++; + + } + + } +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out; j++) + { + for (k = 0; k < dim_im_out; k++) + { + conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel; m++) + { + for (n = 0; n < dim_kernel; n++) + { + // if-for implementation + in_row = stride * j + m - padding; + in_col = stride * k + n - padding; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += + Im_in[(in_row * dim_im_in + in_col) * ch_im_in + + l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel + + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c new file mode 100644 index 0000000..ef96d2f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c @@ -0,0 +1,382 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_HWC_q7_fast_nonsquare.c + * Description: Fast Q7 version of convolution (non-sqaure shape) + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/** + * @brief Fast Q7 convolution function (non-sqaure shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimention x + * @param[in] dim_im_in_y input tensor dimention y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding size x + * @param[in] padding_y padding size y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is multiple of 4 + * ch_im_out is multiple of 2 + */ + +arm_status arm_convolve_HWC_q7_fast_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + int16_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + /* ----------------------- + * Here we use bufferA as q15_t internally as computation are done with q15_t level + * im2col are done to output in q15_t format from q7_t input + */ + + q15_t *pBuffer = bufferA; + q7_t *pOut = Im_out; + + if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + /* + * Here we split the entire matrix into three regions depending on the padding situation + * Top: i_out_y from 0 to padding - 1 + * Middle: i_out_y from padding to dim_im_out-padding-1 + * Bottom: i_out_y from dim_im_out-padding to dim_im_out-1 + */ + + /* top part */ + for (i_out_y = 0; i_out_y < padding_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; + i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; + i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, + pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, + bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* middle part, here we also divide the x into left, mid and right */ + for (; i_out_y < dim_im_out_y - padding_y; i_out_y++) + { + + /* left part */ + for (i_out_x = 0; i_out_x < padding_x; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; + i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; + i_ker_x++) + { + if (i_ker_x < 0 || i_ker_x >= dim_im_in_x) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, + pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, + bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + + /* mid part */ + for (; i_out_x < dim_im_out_x - padding_x; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; + i_ker_y++) + { + arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + + (i_ker_y * dim_im_in_x + i_out_x * stride_x - padding_x) * ch_im_in, + pBuffer, ch_im_in * dim_kernel_x); + pBuffer += ch_im_in * dim_kernel_x; + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, + bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + + /* right part */ + for (; i_out_x < dim_im_out_x; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; + i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; + i_ker_x++) + { + if (i_ker_x < 0 || i_ker_x >= dim_im_in_x) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, + pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, + bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + } + + for (; i_out_y < dim_im_out_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) + { + /* This part implements the im2col function */ + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; + i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; + i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x) + { + /* arm_fill_q15(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, sizeof(q15_t)*ch_im_in); + } else + { + arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, + pBuffer, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y) + { + pOut = + arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in * dim_kernel_x * dim_kernel_y, + bias_shift, out_shift, bias, pOut); + /* counter reset */ + pBuffer = bufferA; + } + } + } + + /* check if there is left-over for compute */ + if (pBuffer != bufferA) + { + const q7_t *pA = wt; + int i; + for (i = 0; i < ch_im_out; i++) + { + q31_t sum = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + const q15_t *pB = bufferA; + /* basically each time it process 4 entries */ + uint16_t colCnt = ch_im_in * dim_kernel_x * dim_kernel_y >> 2; + + while (colCnt) + { + + q31_t inA1, inA2; + q31_t inB1, inB2; + + pA = read_and_pad_reordered(pA, &inA1, &inA2); + + inB1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inA1, inB1, sum); + inB2 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inA2, inB2, sum); + + colCnt--; + } + colCnt = (ch_im_in * dim_kernel_y * dim_kernel_x) & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + sum += inA1 * inB1; + colCnt--; + } + *pOut = (q7_t) __SSAT((sum >> out_shift), 8); + pOut++; + + } + + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + int i, j, k, l, m, n; + int conv_out; + int in_row, in_col; + + if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0) + { + /* check if the input dimension meets the constraints */ + return ARM_MATH_SIZE_MISMATCH; + } + + for (i = 0; i < ch_im_out; i++) + { + for (j = 0; j < dim_im_out_y; j++) + { + for (k = 0; k < dim_im_out_x; k++) + { + conv_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + for (m = 0; m < dim_kernel_y; m++) + { + for (n = 0; n < dim_kernel_x; n++) + { + /* if-for implementation */ + in_row = stride_y * j + m - padding_y; + in_col = stride_x * k + n - padding_x; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x) + { + for (l = 0; l < ch_im_in; l++) + { + conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] * + wt[i * ch_im_in * dim_kernel_y * dim_kernel_x + (m * dim_kernel_x + n) * ch_im_in + l]; + } + } + } + } + Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c new file mode 100644 index 0000000..38ee4d3 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c @@ -0,0 +1,385 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_s8.c + * Description: s8 version of convolution using symmetric quantization. + * + * $Date: July 27, 2020 + * $Revision: V.2.0.2 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/* + * Basic s8 convolution function. + * + * Refer header file for details. Optimal use case for the DSP/MVE implementation is when input and output channels + * are multiples of 4 or atleast greater than 4. + * + */ + +arm_status arm_convolve_s8(const cmsis_nn_context* ctx, + const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, + const q7_t *input_data, + const cmsis_nn_dims* filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims* bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims* output_dims, + q7_t *output_data) +{ + q15_t *buffer_a = (q15_t *)ctx->buf; + + const uint16_t input_batches = input_dims->n; + const uint16_t input_x = input_dims->w; + const uint16_t input_y = input_dims->h; + const uint16_t input_ch = input_dims->c; + const uint16_t kernel_x = filter_dims->w; + const uint16_t kernel_y = filter_dims->h; + const uint16_t output_x = output_dims->w; + const uint16_t output_y = output_dims->h; + const uint16_t output_ch = output_dims->c; + + const uint16_t pad_x = conv_params->padding.w; + const uint16_t pad_y = conv_params->padding.h; + const uint16_t stride_x = conv_params->stride.w; + const uint16_t stride_y = conv_params->stride.h; + + const int32_t input_offset = conv_params->input_offset; + const int32_t out_offset = conv_params->output_offset; + const int32_t out_activation_min = conv_params->activation.min; + const int32_t out_activation_max = conv_params->activation.max; + int32_t *output_mult = quant_params->multiplier; + int32_t *output_shift = quant_params->shift; + + int i_batch; + for (i_batch = 0; i_batch < input_batches; i_batch++) + { +#if defined(ARM_MATH_MVEI) + (void)bias_dims; + /* Generate upto four columns from the input tensor a GEMM computation */ + q7_t *im2col_buf = (q7_t *)buffer_a; + q7_t *out = output_data; + int32_t buffer_fill_cnt = 0; + int32_t padded = 0; + const int32_t num_elem = kernel_x * kernel_y * input_ch; + + /* This part implements the im2col function */ + for (int i_out_y = 0; i_out_y < output_y; i_out_y++) + { + for (int i_out_x = 0; i_out_x < output_x; i_out_x++) + { + for (int i_ker_y = i_out_y * stride_y - pad_y; i_ker_y < i_out_y * stride_y - pad_y + kernel_y; i_ker_y++) + { + for (int i_ker_x = i_out_x * stride_x - pad_x; i_ker_x < i_out_x * stride_x - pad_x + kernel_x; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= input_y || i_ker_x < 0 || i_ker_x >= input_x) + { + memset(im2col_buf, (int8_t)-input_offset, sizeof(q7_t) * input_ch); + padded = 1; + } + else + { + arm_memcpy_q7(im2col_buf, input_data + (i_ker_y * input_x + i_ker_x) * input_ch, input_ch); + } + im2col_buf += input_ch; + } + } + + buffer_fill_cnt++; + + /* Computation is filed for every 4 columns */ + if (buffer_fill_cnt == 4 && (padded == 0)) + { + buffer_fill_cnt = 0; + for (int i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) + { + int32_t sum_row; + int32_t acc[4]; + + (void)arm_nn_mat_mul_core_4x_s8(num_elem, + num_elem, + (q7_t *)buffer_a, + filter_data + num_elem * i_out_ch, + &sum_row, + acc); + int32x4_t s_offset = vdupq_n_s32(sum_row); + + int32x4_t res = vldrwq_s32(acc); + s_offset = vmulq_n_s32(s_offset, input_offset); + if (bias_data) + { + res = vaddq_n_s32(res, bias_data[i_out_ch]); + } + res = vaddq_s32(res, s_offset); + res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]); + res = vaddq_n_s32(res, out_offset); + + res = vmaxq_s32(res, vdupq_n_s32(out_activation_min)); + res = vminq_s32(res, vdupq_n_s32(out_activation_max)); + + const uint32x4_t scatter_offset = {0, output_ch, output_ch * 2, output_ch * 3}; + vstrbq_scatter_offset_s32(out, scatter_offset, res); + out++; + } + out += (3 * output_ch); + im2col_buf = (q7_t *)buffer_a; + } + else if (buffer_fill_cnt == 4 && (padded != 0)) + { + buffer_fill_cnt = 0; + out = arm_nn_mat_mult_s8(filter_data, + (q7_t *)buffer_a, + output_ch, + 4, + output_shift, + output_mult, + out_offset, + input_offset, + 0, + out_activation_min, + out_activation_max, + num_elem, + bias_data, + out); + + im2col_buf = (q7_t *)buffer_a; + padded = 0; + } + } + } + /* Handle left over columns */ + if (buffer_fill_cnt != 0) + { + out = arm_nn_mat_mult_s8(filter_data, + (q7_t *)buffer_a, + output_ch, + buffer_fill_cnt, + output_shift, + output_mult, + out_offset, + input_offset, + 0, + out_activation_min, + out_activation_max, + num_elem, + bias_data, + out); + } + +#elif defined(ARM_MATH_DSP) + (void)bias_dims; + int32_t i_out_y, i_out_x, i_ker_y, i_ker_x; + + /* Generate two columns from the input tensor a GEMM computation */ + q15_t *two_column_buf = buffer_a; + q7_t *out = output_data; + + /* This part implements the im2col function */ + for (i_out_y = 0; i_out_y < output_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < output_x; i_out_x++) + { + for (i_ker_y = i_out_y * stride_y - pad_y; i_ker_y < i_out_y * stride_y - pad_y + kernel_y; i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - pad_x; i_ker_x < i_out_x * stride_x - pad_x + kernel_x; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= input_y || i_ker_x < 0 || i_ker_x >= input_x) + { + /* Filling 0 for out-of-bound paddings */ + memset(two_column_buf, 0, sizeof(q15_t) * input_ch); + } + else + { + /* Copying the pixel data to column */ + arm_q7_to_q15_with_offset(input_data + (i_ker_y * input_x + i_ker_x) * input_ch, two_column_buf, input_ch, input_offset); + } + two_column_buf += input_ch; + } + } + + /* Computation is filed for every 2 columns */ + if (two_column_buf == buffer_a + 2 * input_ch * kernel_y * kernel_x) + { + out = + arm_nn_mat_mult_kernel_s8_s16(filter_data, + buffer_a, + output_ch, + output_shift, + output_mult, + out_offset, + out_activation_min, + out_activation_max, + input_ch * kernel_y * kernel_x, + bias_data, + out); + + /* counter reset */ + two_column_buf = buffer_a; + } + } + } + + /* left-over because odd number of output pixels */ + if (two_column_buf != buffer_a) + { + const q7_t *ker_a = filter_data; + int i; + + for (i = 0; i < output_ch; i++) + { + /* Load the accumulator with bias first */ + q31_t sum = 0; + if (bias_data) + { + sum = bias_data[i]; + } + + /* Point to the beginning of the im2col buffer where the input is available as a rearranged column */ + const q15_t *ip_as_col = buffer_a; + + /* 4 multiply and accumulates are done in one loop. */ + uint16_t col_count = (input_ch * kernel_y * kernel_x) >> 2; + + while (col_count) + { + q31_t ker_a1, ker_a2; + q31_t ip_b1, ip_b2; + + ker_a = read_and_pad(ker_a, &ker_a1, &ker_a2); + + ip_b1 = arm_nn_read_q15x2_ia(&ip_as_col); + sum = __SMLAD(ker_a1, ip_b1, sum); + ip_b2 = arm_nn_read_q15x2_ia(&ip_as_col); + sum = __SMLAD(ker_a2, ip_b2, sum); + + col_count--; + } + /* Handle left over mac */ + col_count = input_ch * kernel_y * kernel_x & 0x3; + while (col_count) + { + q7_t ker_a1 = *ker_a++; + q15_t ip_b1 = *ip_as_col++; + sum += ker_a1 * ip_b1; + col_count--; + } + + sum = arm_nn_requantize(sum, output_mult[i], output_shift[i]); + sum += out_offset; + sum = MAX(sum, out_activation_min); + sum = MIN(sum, out_activation_max); + *out++ = (q7_t)sum; + } + } +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + (void)buffer_a; + int32_t i_out_ch, i_out_y, i_out_x, i_input_ch, i_ker_y, i_ker_x; + int32_t conv_out; + + for (i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) + { + for (i_out_y = 0; i_out_y < output_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < output_x; i_out_x++) + { + conv_out = 0; + + const int32_t base_idx_y = stride_y * i_out_y - pad_y; + const int32_t base_idx_x = stride_x * i_out_x - pad_x; + + const int32_t ker_y_start = MAX(0, -base_idx_y); + const int32_t ker_x_start = MAX(0, -base_idx_x); + + const int32_t ker_y_end = MIN(kernel_y, input_y - base_idx_y); + const int32_t ker_x_end = MIN(kernel_x, input_x - base_idx_x); + + for (i_ker_y = ker_y_start; i_ker_y < ker_y_end; i_ker_y++) + { + for (i_ker_x = ker_x_start; i_ker_x < ker_x_end; i_ker_x++) + { + const int32_t in_row = base_idx_y + i_ker_y; + const int32_t in_col = base_idx_x + i_ker_x; + for (i_input_ch = 0; i_input_ch < input_ch; i_input_ch++) + { + conv_out += + (input_data[(in_row * input_x + in_col) * input_ch + i_input_ch] + input_offset) * + filter_data[i_out_ch * input_ch * kernel_y * kernel_x + + (i_ker_y * kernel_x + i_ker_x) * input_ch + i_input_ch]; + } + } + } + if (bias_data) + { + conv_out += bias_data[i_out_ch]; + } + conv_out = arm_nn_requantize(conv_out, output_mult[i_out_ch], output_shift[i_out_ch]); + conv_out += out_offset; + conv_out = MAX(conv_out, out_activation_min); + conv_out = MIN(conv_out, out_activation_max); + output_data[i_out_ch + (i_out_y * output_x + i_out_x) * output_ch] = (int8_t)conv_out; + } + } + } +#endif + /* Advance to the next batch */ + input_data += (input_x * input_y * input_ch); + output_data += (output_x * output_y * output_ch); + } + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +int32_t arm_convolve_s8_get_buffer_size(const cmsis_nn_dims* input_dims, + const cmsis_nn_dims* filter_dims) +{ +#if defined(ARM_MATH_DSP) + return (2 * input_dims->c * filter_dims->w * filter_dims->h) * (int32_t)sizeof(int16_t); +#else + (void)input_dims; + (void)filter_dims; + return 0; +#endif +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_wrapper_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_wrapper_s8.c new file mode 100644 index 0000000..dc9fd3c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_wrapper_s8.c @@ -0,0 +1,151 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_convolve_wrapper_s8.c + * Description: s8 convolution layer wrapper function with the main purpose to call the optimal kernel available in cmsis-nn to perform the convolution. + * + * $Date: May 18, 2020 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/* + * Convolution layer + * + * Refer header file for details. + * + */ + +arm_status arm_convolve_wrapper_s8(const cmsis_nn_context* ctx, + const cmsis_nn_conv_params* conv_params, + const cmsis_nn_per_channel_quant_params* quant_params, + const cmsis_nn_dims* input_dims, + const q7_t *input_data, + const cmsis_nn_dims* filter_dims, + const q7_t *filter_data, + const cmsis_nn_dims* bias_dims, + const int32_t *bias_data, + const cmsis_nn_dims* output_dims, + q7_t *output_data) +{ + if ((conv_params->padding.w == 0) && + (conv_params->padding.h == 0) && + (input_dims->c % 4 == 0) && + (conv_params->stride.w == 1) && + (conv_params->stride.h == 1) && + (filter_dims->w == 1) && + (filter_dims->h == 1)) + { + return arm_convolve_1x1_s8_fast(ctx, + conv_params, + quant_params, + input_dims, + input_data, + filter_dims, + filter_data, + bias_dims, + bias_data, + output_dims, + output_data); + } + else if ((output_dims->h == 1) && + (input_dims->h == 1) && + (filter_dims->h == 1) && + (output_dims->w % 4 == 0) && + (input_dims->n == 1)) + { + return arm_convolve_1_x_n_s8(ctx, + conv_params, + quant_params, + input_dims, + input_data, + filter_dims, + filter_data, + bias_dims, + bias_data, + output_dims, + output_data); + } + else + { + return arm_convolve_s8(ctx, + conv_params, + quant_params, + input_dims, + input_data, + filter_dims, + filter_data, + bias_dims, + bias_data, + output_dims, + output_data); + } +} + +int32_t arm_convolve_wrapper_s8_get_buffer_size(const cmsis_nn_conv_params* conv_params, + const cmsis_nn_dims* input_dims, + const cmsis_nn_dims* filter_dims, + const cmsis_nn_dims* output_dims) +{ + if ((conv_params->padding.w == 0) && + (conv_params->padding.h == 0) && + (input_dims->c % 4 == 0) && + (conv_params->stride.w == 1) && + (conv_params->stride.h == 1) && + (filter_dims->w == 1) && + (filter_dims->h == 1)) + { + return arm_convolve_1x1_s8_fast_get_buffer_size(input_dims); + } + else if ((output_dims->h == 1) && + (input_dims->h == 1) && + (filter_dims->h == 1) && + (output_dims->w % 4 == 0) && + (input_dims->n == 1)) + { + return arm_convolve_1_x_n_s8_get_buffer_size(input_dims, filter_dims); + } + else + { + return arm_convolve_s8_get_buffer_size(input_dims, filter_dims); + } +} + +/** + * @} end of NNConv group + */ + + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_3x3_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_3x3_s8.c new file mode 100644 index 0000000..b6303d0 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_3x3_s8.c @@ -0,0 +1,216 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_depthwise_conv_3x3_s8.c + * Description: Optimized s8 depthwise convolution function for channel + * multiplier of 1 and 3x3 kernel size. + * + * $Date: May 14, 2020 + * $Revision: V.2.0.0 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/* + * Optimized s8 depthwise convolution function with constraint that + * in_channel == out_channel and kernel_x == kernel_y == 3 with pads at most 1 + * + * Refer prototype header file for details. + * + */ + +arm_status arm_depthwise_conv_3x3_s8(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input, + const cmsis_nn_dims *filter_dims, + const q7_t *kernel, + const cmsis_nn_dims *bias_dims, + const int32_t *bias, + const cmsis_nn_dims *output_dims, + q7_t *output) +{ + (void)ctx; + (void)bias_dims; + + const int32_t input_x = input_dims->w; + const int32_t input_y = input_dims->h; + const int32_t input_ch = input_dims->c; + const int32_t output_ch = output_dims->c; + const int32_t pad_x = dw_conv_params->padding.w; + const int32_t pad_y = dw_conv_params->padding.h; + const int32_t stride_x = dw_conv_params->stride.w; + const int32_t stride_y = dw_conv_params->stride.h; + const int32_t *output_shift = quant_params->shift; + const int32_t *output_mult = quant_params->multiplier; + const int32_t output_x = output_dims->w; + const int32_t output_y = output_dims->h; + const int32_t output_offset = dw_conv_params->output_offset; + const int32_t input_offset = dw_conv_params->input_offset; + const int32_t output_activation_min = dw_conv_params->activation.min; + const int32_t output_activation_max = dw_conv_params->activation.max; + + /* Check input constraints input_ch == output_ch */ + if (input_ch != output_ch) + { + return ARM_MATH_SIZE_MISMATCH; + } + /* Check input constraints pad_x <= 1 */ + if (pad_x > 1 || filter_dims->w != 3 || filter_dims->h != 3) + { + return ARM_MATH_ARGUMENT_ERROR; + } + + for (int32_t in_h = -pad_y, out_h = 0, out_idx = 0; out_h < output_y; in_h += stride_y, ++out_h) + { + for (int32_t in_w = -pad_x, out_w = 0, ker_h_start = MAX(0, -in_h); out_w < output_x; in_w += stride_x, ++out_w) + { + int32_t in_ch = 0; + int32_t ker_w_start = MAX(0, -in_w); + + for (; in_ch <= (input_ch - 4); in_ch += 4) + { + int32_t out_buff0 = bias[in_ch + 0]; + int32_t out_buff1 = bias[in_ch + 1]; + int32_t out_buff2 = bias[in_ch + 2]; + int32_t out_buff3 = bias[in_ch + 3]; + + const int8_t *input_ptr = input + (in_h + ker_h_start) * (input_ch * input_x) + in_w * input_ch + in_ch; + const int8_t *kernel_ptr = kernel + ker_h_start * (input_ch * 3) + in_ch; + + for (int32_t ker_h = ker_h_start; ker_h < MIN(3, input_y - in_h); ++ker_h) + { + int32_t in_val = 0; + int32_t ker_val = 0; + + if (ker_w_start == 0) + { + in_val = arm_nn_read_q7x4(input_ptr); + ker_val = arm_nn_read_q7x4(kernel_ptr); + + out_buff0 += ((int8_t)in_val + input_offset) * (int8_t)ker_val; + out_buff1 += ((int8_t)(in_val >> 8) + input_offset) * (int8_t)(ker_val >> 8); + out_buff2 += ((int8_t)(in_val >> 16) + input_offset) * (int8_t)(ker_val >> 16); + out_buff3 += ((int8_t)(in_val >> 24) + input_offset) * (int8_t)(ker_val >> 24); + } + + in_val = arm_nn_read_q7x4(input_ptr + input_ch); + ker_val = arm_nn_read_q7x4(kernel_ptr + input_ch); + + out_buff0 += ((int8_t)in_val + input_offset) * (int8_t)ker_val; + out_buff1 += ((int8_t)(in_val >> 8) + input_offset) * (int8_t)(ker_val >> 8); + out_buff2 += ((int8_t)(in_val >> 16) + input_offset) * (int8_t)(ker_val >> 16); + out_buff3 += ((int8_t)(in_val >> 24) + input_offset) * (int8_t)(ker_val >> 24); + + if ((input_x - in_w) >= 3) + { + in_val = arm_nn_read_q7x4(input_ptr + (input_ch << 1)); + ker_val = arm_nn_read_q7x4(kernel_ptr + (input_ch << 1)); + + out_buff0 += ((int8_t)in_val + input_offset) * (int8_t)ker_val; + out_buff1 += ((int8_t)(in_val >> 8) + input_offset) * (int8_t)(ker_val >> 8); + out_buff2 += ((int8_t)(in_val >> 16) + input_offset) * (int8_t)(ker_val >> 16); + out_buff3 += ((int8_t)(in_val >> 24) + input_offset) * (int8_t)(ker_val >> 24); + } + + input_ptr += (input_ch * input_x); + kernel_ptr += (input_ch * 3); + } + + out_buff0 = arm_nn_requantize(out_buff0, output_mult[in_ch + 0], output_shift[in_ch + 0]); + out_buff1 = arm_nn_requantize(out_buff1, output_mult[in_ch + 1], output_shift[in_ch + 1]); + out_buff2 = arm_nn_requantize(out_buff2, output_mult[in_ch + 2], output_shift[in_ch + 2]); + out_buff3 = arm_nn_requantize(out_buff3, output_mult[in_ch + 3], output_shift[in_ch + 3]); + + out_buff0 += output_offset; + out_buff1 += output_offset; + out_buff2 += output_offset; + out_buff3 += output_offset; + + out_buff0 = MIN(MAX(out_buff0, output_activation_min), output_activation_max); + out_buff1 = MIN(MAX(out_buff1, output_activation_min), output_activation_max); + out_buff2 = MIN(MAX(out_buff2, output_activation_min), output_activation_max); + out_buff3 = MIN(MAX(out_buff3, output_activation_min), output_activation_max); + + output[out_idx++] = (int8_t)out_buff0; + output[out_idx++] = (int8_t)out_buff1; + output[out_idx++] = (int8_t)out_buff2; + output[out_idx++] = (int8_t)out_buff3; + } + + // Leftover + for (; in_ch < input_ch; ++in_ch) + { + int32_t out_buff = bias[in_ch]; + + const int8_t *input_ptr = input + (in_h + ker_h_start) * (input_ch * input_x) + in_w * input_ch + in_ch; + const int8_t *kernel_ptr = kernel + ker_h_start * (input_ch * 3) + in_ch; + + for (int32_t ker_h = ker_h_start; ker_h < MIN(3, input_y - in_h); ++ker_h) + { + if (ker_w_start == 0) + { + out_buff += (*(input_ptr) + input_offset) * *(kernel_ptr); + } + + out_buff += (*(input_ptr + input_ch) + input_offset) * *(kernel_ptr + input_ch); + + if ((input_x - in_w) >= 3) + { + out_buff += (*(input_ptr + (input_ch << 1)) + input_offset) * *(kernel_ptr + (input_ch << 1)); + } + + input_ptr += (input_ch * input_x); + kernel_ptr += (input_ch * 3); + } + + out_buff = arm_nn_requantize(out_buff, output_mult[in_ch], output_shift[in_ch]); + out_buff += output_offset; + out_buff = MIN(MAX(out_buff, output_activation_min), output_activation_max); + output[out_idx++] = (int8_t)out_buff; + } + } + } + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c new file mode 100644 index 0000000..7664c60 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c @@ -0,0 +1,252 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_depthwise_conv_s8.c + * Description: s8 version of depthwise convolution. + * + * $Date: May 14, 2020 + * $Revision: V.2.0.0 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +static void depthwise_conv_s8_mult_4(const int8_t *input, + const int32_t input_x, + const int32_t input_y, + const int32_t input_ch, + const int8_t *kernel, + const int32_t output_ch, + const int32_t ch_mult, + const int32_t kernel_x, + const int32_t kernel_y, + const int32_t pad_x, + const int32_t pad_y, + const int32_t stride_x, + const int32_t stride_y, + const int32_t *bias, + int8_t *output, + const int32_t *output_shift, + const int32_t *output_mult, + const int32_t output_x, + const int32_t output_y, + const int32_t output_offset, + const int32_t input_offset, + const int32_t output_activation_min, + const int32_t output_activation_max) +{ + for (int32_t in_h = -pad_y, out_h = 0, out_idx = 0; out_h < output_y; in_h += stride_y, ++out_h) + { + for (int32_t in_w = -pad_x, out_w = 0, ker_h_start = MAX(0, -in_h); out_w < output_x; in_w += stride_x, ++out_w) + { + for (int32_t in_ch = 0, out_ch = 0, ker_w_start = MAX(0, -in_w); out_ch < output_ch; ++in_ch, out_ch += ch_mult) + { + for (int mult_tile = 0; mult_tile < ch_mult; mult_tile += 4) + { + int32_t out_buff[4]; + + out_buff[0] = bias[out_ch + 0 + mult_tile]; + out_buff[1] = bias[out_ch + 1 + mult_tile]; + out_buff[2] = bias[out_ch + 2 + mult_tile]; + out_buff[3] = bias[out_ch + 3 + mult_tile]; + + for (int32_t ker_h = ker_h_start; ker_h < MIN(kernel_y, input_y - in_h); ++ker_h) + { + int32_t ker_idx = ker_h * (output_ch * kernel_x) + ker_w_start * output_ch + out_ch; + int32_t in_idx = (in_h + ker_h) * (input_ch * input_x) + in_w * input_ch + in_ch; + + for (int32_t ker_w = ker_w_start; ker_w < MIN(kernel_x, input_x - in_w); ++ker_w, ker_idx += output_ch) + { + int32_t in_val = input[in_idx + ker_w * input_ch] + input_offset; + out_buff[0] += in_val * kernel[ker_idx + 0 + mult_tile]; + out_buff[1] += in_val * kernel[ker_idx + 1 + mult_tile]; + out_buff[2] += in_val * kernel[ker_idx + 2 + mult_tile]; + out_buff[3] += in_val * kernel[ker_idx + 3 + mult_tile]; + } + } +#if defined(ARM_MATH_MVEI) + (void)out_idx; + int32x4_t res = vldrwq_s32(out_buff); + res = arm_requantize_mve_32x4(res, vldrwq_s32(&output_mult[out_ch + mult_tile]), vldrwq_s32(&output_shift[out_ch + mult_tile])); + res = vaddq_n_s32(res, output_offset); + + res = vmaxq_s32(res, vdupq_n_s32(output_activation_min)); + res = vminq_s32(res, vdupq_n_s32(output_activation_max)); + vstrbq_s32(output, res); + output += 4; +#else + out_buff[0] = arm_nn_requantize(out_buff[0], output_mult[out_ch + 0 + mult_tile], output_shift[out_ch + 0 + mult_tile]); + out_buff[1] = arm_nn_requantize(out_buff[1], output_mult[out_ch + 1 + mult_tile], output_shift[out_ch + 1 + mult_tile]); + out_buff[2] = arm_nn_requantize(out_buff[2], output_mult[out_ch + 2 + mult_tile], output_shift[out_ch + 2 + mult_tile]); + out_buff[3] = arm_nn_requantize(out_buff[3], output_mult[out_ch + 3 + mult_tile], output_shift[out_ch + 3 + mult_tile]); + + out_buff[0] += output_offset; + out_buff[1] += output_offset; + out_buff[2] += output_offset; + out_buff[3] += output_offset; + + out_buff[0] = MIN(MAX(out_buff[0], output_activation_min), output_activation_max); + out_buff[1] = MIN(MAX(out_buff[1], output_activation_min), output_activation_max); + out_buff[2] = MIN(MAX(out_buff[2], output_activation_min), output_activation_max); + out_buff[3] = MIN(MAX(out_buff[3], output_activation_min), output_activation_max); + + output[out_idx++] = (int8_t)out_buff[0]; + output[out_idx++] = (int8_t)out_buff[1]; + output[out_idx++] = (int8_t)out_buff[2]; + output[out_idx++] = (int8_t)out_buff[3]; + +#endif + } + } + } + } +} + +static void depthwise_conv_s8_generic(const q7_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_ch, + const q7_t *kernel, + const uint16_t output_ch, + const uint16_t ch_mult, + const uint16_t kernel_x, + const uint16_t kernel_y, + const uint16_t pad_x, + const uint16_t pad_y, + const uint16_t stride_x, + const uint16_t stride_y, + const int32_t *bias, + q7_t *output, + const int32_t *output_shift, + const int32_t *output_mult, + const uint16_t output_x, + const uint16_t output_y, + const int32_t output_offset, + const int32_t input_offset, + const int32_t output_activation_min, + const int32_t output_activation_max) +{ + (void)output_ch; + int i_out = 0; + for (int i_out_y = 0; i_out_y < output_y; i_out_y++) + { + const int16_t base_idx_y = (i_out_y * stride_y) - pad_y; + for (int i_out_x = 0; i_out_x < output_x; i_out_x++) + { + const int16_t base_idx_x = (i_out_x * stride_x) - pad_x; + for (int i_input_ch = 0; i_input_ch < input_ch; i_input_ch++) + { + for (int i_ch_mult = 0; i_ch_mult < ch_mult; i_ch_mult++) + { + const int idx_out_ch = i_ch_mult + i_input_ch * ch_mult; + int32_t acc_0; + /* Condition for kernel start dimension: (base_idx_ + ker__start) >= 0 */ + const int ker_y_start = MAX(0, -base_idx_y); + const int ker_x_start = MAX(0, -base_idx_x); + /* Condition for kernel end dimension: (base_idx_ + ker__end) < input_ */ + const int ker_y_end = MIN(kernel_y, input_y - base_idx_y); + const int ker_x_end = MIN(kernel_x, input_x - base_idx_x); + acc_0 = bias[idx_out_ch]; + + for (int i_ker_y = ker_y_start; i_ker_y < ker_y_end; i_ker_y++) + { + const int32_t idx_y = base_idx_y + i_ker_y; + for (int i_ker_x = ker_x_start; i_ker_x < ker_x_end; i_ker_x++) + { + const int32_t idx_x = base_idx_x + i_ker_x; + int32_t idx_0 = (idx_y * input_x + idx_x) * input_ch + i_input_ch; + int32_t ker_idx_0 = (i_ker_y * kernel_x + i_ker_x) * (input_ch * ch_mult) + idx_out_ch; + + acc_0 += (input[idx_0] + input_offset) * kernel[ker_idx_0]; + } + } + + /* Requantize and clamp output to provided range */ + acc_0 = arm_nn_requantize(acc_0, output_mult[idx_out_ch], output_shift[idx_out_ch]); + acc_0 += output_offset; + acc_0 = MAX(acc_0, output_activation_min); + acc_0 = MIN(acc_0, output_activation_max); + + output[i_out++] = acc_0; + } + } + } + } +} + +/* + * Basic s8 depthwise convolution function. + * + * Refer header file for details. + * Optimization using DSP extension is not available for the generic case where channel multiplier is > 1. + * + */ +arm_status arm_depthwise_conv_s8(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input, + const cmsis_nn_dims *filter_dims, + const q7_t *kernel, + const cmsis_nn_dims *bias_dims, + const int32_t *bias, + const cmsis_nn_dims *output_dims, + q7_t *output) +{ + (void)dw_conv_params->dilation; + (void)ctx; + + if (dw_conv_params->ch_mult % 4 == 0) + { + depthwise_conv_s8_mult_4(input, input_dims->w, input_dims->h, input_dims->c, kernel, output_dims->c, dw_conv_params->ch_mult, filter_dims->w, filter_dims->h, + dw_conv_params->padding.w, dw_conv_params->padding.h, dw_conv_params->stride.w, dw_conv_params->stride.h, bias, output, + quant_params->shift, quant_params->multiplier, output_dims->w, output_dims->h, dw_conv_params->output_offset, + dw_conv_params->input_offset, dw_conv_params->activation.min, dw_conv_params->activation.max); + } + else + { + depthwise_conv_s8_generic(input, input_dims->w, input_dims->h, input_dims->c, kernel, output_dims->c, dw_conv_params->ch_mult, filter_dims->w, filter_dims->h, + dw_conv_params->padding.w, dw_conv_params->padding.h, dw_conv_params->stride.w, dw_conv_params->stride.h, bias, output, + quant_params->shift, quant_params->multiplier, output_dims->w, output_dims->h, dw_conv_params->output_offset, + dw_conv_params->input_offset, dw_conv_params->activation.min, dw_conv_params->activation.max); + } + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8_opt.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8_opt.c new file mode 100644 index 0000000..267a384 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8_opt.c @@ -0,0 +1,428 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_depthwise_conv_s8_opt.c + * Description: Optimized s8 depthwise separable convolution function for + * channel multiplier of 1. + * + * $Date: May 29, 2020 + * $Revision: V.2.0.1 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/* + * Optimized s8 depthwise convolution function with constraint that in_channel equals out_channel + * + * Refer prototype header file for details. + * + */ + +arm_status arm_depthwise_conv_s8_opt(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input, + const cmsis_nn_dims *filter_dims, + const q7_t *kernel, + const cmsis_nn_dims *bias_dims, + const int32_t *bias, + const cmsis_nn_dims *output_dims, + q7_t *output) +{ + const int32_t input_x = input_dims->w; + const int32_t input_y = input_dims->h; + const int32_t input_ch = input_dims->c; + const int32_t output_ch = output_dims->c; + const int32_t kernel_x = filter_dims->w; + const int32_t kernel_y = filter_dims->h; + const int32_t pad_x = dw_conv_params->padding.w; + const int32_t pad_y = dw_conv_params->padding.h; + const int32_t stride_x = dw_conv_params->stride.w; + const int32_t stride_y = dw_conv_params->stride.h; + const int32_t *output_shift = quant_params->shift; + const int32_t *output_mult = quant_params->multiplier; + const int32_t output_x = output_dims->w; + const int32_t output_y = output_dims->h; + const int32_t output_offset = dw_conv_params->output_offset; + const int32_t input_offset = dw_conv_params->input_offset; + const int32_t output_activation_min = dw_conv_params->activation.min; + const int32_t output_activation_max = dw_conv_params->activation.max; + q15_t *buffer_a = (q15_t *)ctx->buf; + + /* Check input constraints input_ch == output_ch */ + if (input_ch != output_ch) + { + return ARM_MATH_SIZE_MISMATCH; + } +#ifdef ARM_MATH_MVEI + (void)bias_dims; + /* Generate two columns from the input tensor */ + q7_t *lhs_buffer = (q7_t *)buffer_a; + q7_t *out = output; + int padded = 0; + int buffer_count = 0; + const int32_t kernel_size = kernel_x * kernel_y; + + /* This part implements the im2col function */ + for (int i_out_y = 0, base_idx_y = -pad_y; i_out_y < output_y; base_idx_y += stride_y, i_out_y++) + { + for (int i_out_x = 0, base_idx_x = -pad_x; i_out_x < output_x; base_idx_x += stride_x, i_out_x++) + { + for (int i_ker_y = base_idx_y; i_ker_y < base_idx_y + kernel_y; i_ker_y++) + { + for (int i_ker_x = base_idx_x; i_ker_x < base_idx_x + kernel_x; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= input_y || i_ker_x < 0 || i_ker_x >= input_x) + { + arm_memset_q7(lhs_buffer, (int8_t)-input_offset, (uint32_t)input_ch); + padded = 1; + } + else + { + arm_memcpy_q7(lhs_buffer, input + (i_ker_y * input_x + i_ker_x) * input_ch, (uint32_t)input_ch); + } + lhs_buffer += input_ch; + } + } + buffer_count++; + + if (buffer_count == 4) + { + lhs_buffer = (q7_t *)buffer_a; + if (padded == 0) + { + out = arm_nn_depthwise_conv_nt_t_s8(lhs_buffer, + kernel, + input_offset, + input_ch, + output_shift, + output_mult, + output_offset, + output_activation_min, + output_activation_max, + kernel_size, + bias, + out); + } + else + { + out = arm_nn_depthwise_conv_nt_t_padded_s8(lhs_buffer, + kernel, + input_offset, + input_ch, + output_shift, + output_mult, + output_offset, + output_activation_min, + output_activation_max, + kernel_size, + bias, + out); + padded = 0; + } + buffer_count = 0; + } + } + } + + /* Handle left over buffers */ + lhs_buffer = (q7_t *)buffer_a; + + for (int i_buf = 0; i_buf < buffer_count; i_buf++) + { + int32_t loop_count = (input_ch + 3) / 4; + + int32_t num_ch_to_process = input_ch; + for (int i_loop_cnt = 0, offset = 0; i_loop_cnt < loop_count; + num_ch_to_process -= 4, offset += 4, i_loop_cnt++) + { + const int8_t *col_0 = lhs_buffer + (kernel_size * input_ch * i_buf) + offset; + const int8_t *row_0 = kernel + offset; + int32x4_t out_0 = vldrwq_s32(&bias[offset]); + + for (int i_ker = 0; i_ker < kernel_size; i_ker++) + { + const int32x4_t ker_0 = vldrbq_s32(row_0); + + int32x4_t ip_0 = vldrbq_s32(col_0); + ip_0 = vaddq_n_s32(ip_0, input_offset); + out_0 += vmulq_s32(ip_0, ker_0); + + col_0 += input_ch; + row_0 += input_ch; + } + + const int32x4_t mult = vldrwq_s32(&output_mult[offset]); + const int32x4_t shift = vldrwq_s32(&output_shift[offset]); + + out_0 = arm_requantize_mve_32x4(out_0, mult, shift); + out_0 = vaddq_n_s32(out_0, output_offset); + out_0 = vmaxq_s32(out_0, vdupq_n_s32(output_activation_min)); + out_0 = vminq_s32(out_0, vdupq_n_s32(output_activation_max)); + mve_pred16_t p = vctp32q((uint32_t)num_ch_to_process); + vstrbq_p_s32(out, out_0, p); + + out += 4; + } + + const int tail_ch = input_ch & 0x3; + if (tail_ch != 0) + { + out -= (4 - tail_ch); + } + } + +#elif defined(ARM_MATH_DSP) + (void)bias_dims; + /* Run the following code in cores using DSP extension */ + q15_t *const col_buffer_start = buffer_a; + q15_t *col_buffer = col_buffer_start; + const int32_t *const bias_start_pos = bias; + const q31_t *const out_mult_start_pos = output_mult; + const q31_t *const out_shift_start_pos = output_shift; + uint16_t row_count; + uint16_t row_shift; + + for (int i_out_y = 0; i_out_y < output_y; i_out_y++) + { + const int16_t base_idx_y = (i_out_y * stride_y) - pad_y; + for (int i_out_x = 0; i_out_x < output_x; i_out_x++) + { + const int16_t base_idx_x = (i_out_x * stride_x) - pad_x; + + /* Out of bounds is only considered for the y axis as it provides a contiguous zero'ing opportunity than along + the x axis */ + const int ker_y_start = MAX(0, -base_idx_y); + /* Condition for kernel end dimension: (base_idx_y + ker_y_end) < input_y */ + const int ker_y_end = MIN(kernel_y, input_y - base_idx_y); + + int32_t index = 0; + if (ker_y_start != 0) + { + memset(&col_buffer[index], 0, (kernel_x * input_ch) * ker_y_start * sizeof(q15_t)); + index += (kernel_x * input_ch) * ker_y_start; + } + + for (int i_ker_y = ker_y_start; i_ker_y < ker_y_end; i_ker_y++) + { + const int32_t idx_y = base_idx_y + i_ker_y; + + for (int i_ker_x = 0; i_ker_x < kernel_x; i_ker_x++) + { + const int32_t idx_x = base_idx_x + i_ker_x; + if (idx_x < 0 || idx_x >= input_x) + { + memset(&col_buffer[index], 0, input_ch * sizeof(q15_t)); + } + else + { + arm_q7_to_q15_with_offset((q7_t *)input + (idx_y * input_x + idx_x) * input_ch, &col_buffer[index], input_ch, input_offset); + } + index += input_ch; + } + } + + const int diff = kernel_y - ker_y_end; + if (diff != 0) + { + memset(&col_buffer[index], 0, (kernel_x * input_ch) * diff * sizeof(q15_t)); + } + + row_count = output_ch / 4; + row_shift = 0; + bias = bias_start_pos; + output_mult = out_mult_start_pos; + output_shift = out_shift_start_pos; + + while (row_count) + { + q31_t sum = *bias++; + q31_t sum_2 = *bias++; + q31_t sum_3 = *bias++; + q31_t sum_4 = *bias++; + + uint16_t col_count = (kernel_x * kernel_y) / 2; + q15_t *col_pos = col_buffer_start + row_shift; + const q7_t *row_pos = kernel + row_shift; + row_shift += 4; + + while (col_count) + { + /* General idea is to read 4 + 4 (input, kernel) pair and re-arrange them in the right order to + use in a SMLAD instruction . One run of this loop produces 4 partial outputs with 8 MACs. */ + /* Note: variable names can be improved here to align with rows and columns. */ + q31_t ip_a1, ip_a2, ip_b1, ip_b2, op_a, op_b, op_c; + /* Read 4 weights */ + ip_b1 = arm_nn_read_q7x4(row_pos); + ip_a1 = arm_nn_read_q7x4(row_pos + input_ch); + op_a = arm_nn_read_q15x2(col_pos); + op_b = arm_nn_read_q15x2(col_pos + input_ch); + + ip_a2 = __SXTB16(ip_b1); + ip_b1 = __SXTB16(__ROR(ip_b1, 8)); + + ip_b2 = __SXTB16(ip_a1); + ip_a1 = __SXTB16(__ROR(ip_a1, 8)); + + op_c = __PKHBT(op_b, op_a, 16); + op_a = __PKHTB(op_b, op_a, 16); + op_b = __PKHBT(ip_b2, ip_a2, 16); + sum = __SMLAD(op_c, op_b, sum); + + op_b = __PKHBT(ip_b1, ip_a1, 16); + sum_2 = __SMLAD(op_a, op_b, sum_2); + + op_a = arm_nn_read_q15x2(col_pos + 2); + op_b = arm_nn_read_q15x2(col_pos + input_ch + 2); + + op_c = __PKHBT(op_b, op_a, 16); + op_a = __PKHTB(op_b, op_a, 16); + op_b = __PKHTB(ip_a2, ip_b2, 16); + sum_3 = __SMLAD(op_c, op_b, sum_3); + + op_b = __PKHTB(ip_a1, ip_b1, 16); + sum_4 = __SMLAD(op_a, op_b, sum_4); + + row_pos += input_ch << 1; + col_pos += input_ch << 1; + col_count--; + } + + col_count = (kernel_x * kernel_y) & 0x1; + while (col_count) + { + sum += row_pos[0] * col_pos[0]; + sum_2 += row_pos[1] * col_pos[1]; + sum_3 += row_pos[2] * col_pos[2]; + sum_4 += row_pos[3] * col_pos[3]; + + row_pos += input_ch; + col_pos += input_ch; + + col_count--; + } + sum = arm_nn_requantize(sum, *output_mult++, *output_shift++); + sum += output_offset; + sum = MAX(sum, output_activation_min); + sum = MIN(sum, output_activation_max); + *output++ = (q7_t)sum; + + sum_2 = arm_nn_requantize(sum_2, *output_mult++, *output_shift++); + sum_2 += output_offset; + sum_2 = MAX(sum_2, output_activation_min); + sum_2 = MIN(sum_2, output_activation_max); + *output++ = (q7_t)sum_2; + sum_3 = arm_nn_requantize(sum_3, *output_mult++, *output_shift++); + sum_3 += output_offset; + sum_3 = MAX(sum_3, output_activation_min); + sum_3 = MIN(sum_3, output_activation_max); + *output++ = (q7_t)sum_3; + + sum_4 = arm_nn_requantize(sum_4, *output_mult++, *output_shift++); + sum_4 += output_offset; + sum_4 = MAX(sum_4, output_activation_min); + sum_4 = MIN(sum_4, output_activation_max); + *output++ = (q7_t)sum_4; + + row_count--; + } + + row_count = output_ch & 0x3; + while (row_count) + { + q15_t *col_pos = col_buffer_start + row_shift; + const q7_t *row_pos = kernel + row_shift; + q31_t sum = *bias++; + const uint16_t col_count = (kernel_x * kernel_y); + row_shift += 1; + + for (int i = 0; i < col_count; i++) + { + sum += row_pos[i * input_ch] * col_pos[i * input_ch]; + } + sum = arm_nn_requantize(sum, *output_mult++, *output_shift++); + sum += output_offset; + sum = MAX(sum, output_activation_min); + sum = MIN(sum, output_activation_max); + *output++ = (q7_t)sum; + + row_count--; + } + + // clear counter and pointers + col_buffer = col_buffer_start; + } + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + return arm_depthwise_conv_s8(ctx, + dw_conv_params, + quant_params, + input_dims, + input, + filter_dims, + kernel, + bias_dims, + bias, + output_dims, + output); +#endif /* ARM_MATH_MVEI | ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +int32_t arm_depthwise_conv_s8_opt_get_buffer_size(const cmsis_nn_dims *input_dims, + const cmsis_nn_dims *filter_dims) +{ +#if defined(ARM_MATH_MVEI) + /* The + 4 accounts for out of bounds read of the lhs buffers in the *_nt_t_* functions. */ + return (2 * input_dims->c * filter_dims->w * filter_dims->h) * (int32_t)sizeof(int16_t) + 4; +#elif defined(ARM_MATH_DSP) + return (input_dims->c * filter_dims->w * filter_dims->h) * sizeof(int16_t); +#else + (void)input_dims; + (void)filter_dims; + return 0; +#endif +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_u8_basic_ver1.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_u8_basic_ver1.c new file mode 100644 index 0000000..53a422d --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_u8_basic_ver1.c @@ -0,0 +1,297 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_depthwise_conv_u8_basic_ver1.c + * Description: u8 depthwise convolution function + * + * $Date: May 29, 2020 + * $Revision: V.1.1.0 + * + * Target : Cortex-M CPUs + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +static void depthwise_conv_u8_mult_4(const uint8_t *input, + const int32_t input_x, + const int32_t input_y, + const int32_t input_ch, + const uint8_t *kernel, + const int32_t output_ch, + const int32_t ch_mult, + const int32_t kernel_x, + const int32_t kernel_y, + const int32_t pad_x, + const int32_t pad_y, + const int32_t stride_x, + const int32_t stride_y, + const int32_t *bias, + uint8_t *output, + const int32_t output_shift, + const int32_t output_mult, + const int32_t output_x, + const int32_t output_y, + const int32_t output_offset, + const int32_t input_offset, + const int32_t filter_offset, + const int32_t output_activation_min, + const int32_t output_activation_max) +{ + for (int32_t in_h = -pad_y, out_h = 0, out_idx = 0; out_h < output_y; in_h += stride_y, ++out_h) + { + for (int32_t in_w = -pad_x, out_w = 0, ker_h_start = MAX(0, -in_h); out_w < output_x; in_w += stride_x, ++out_w) + { + for (int32_t in_ch = 0, out_ch = 0, ker_w_start = MAX(0, -in_w); out_ch < output_ch; ++in_ch, out_ch += ch_mult) + { + for (int mult_tile = 0; mult_tile < ch_mult; mult_tile += 4) + { + int32_t out_buff[4]; + + out_buff[0] = 0; + out_buff[1] = 0; + out_buff[2] = 0; + out_buff[3] = 0; + + for (int32_t ker_h = ker_h_start; ker_h < MIN(kernel_y, input_y - in_h); ++ker_h) + { + int32_t ker_idx = ker_h * (output_ch * kernel_x) + ker_w_start * output_ch + out_ch; + int32_t in_idx = (in_h + ker_h) * (input_ch * input_x) + in_w * input_ch + in_ch; + + for (int32_t ker_w = ker_w_start; ker_w < MIN(kernel_x, input_x - in_w); ++ker_w, ker_idx += output_ch) + { + int32_t in_val = input[in_idx + ker_w * input_ch] + input_offset; + out_buff[0] += in_val * (kernel[ker_idx + 0 + mult_tile] + filter_offset); + out_buff[1] += in_val * (kernel[ker_idx + 1 + mult_tile] + filter_offset); + out_buff[2] += in_val * (kernel[ker_idx + 2 + mult_tile] + filter_offset); + out_buff[3] += in_val * (kernel[ker_idx + 3 + mult_tile] + filter_offset); + } + } + + if (bias != NULL) + { + out_buff[0] += bias[out_ch + 0 + mult_tile]; + out_buff[1] += bias[out_ch + 1 + mult_tile]; + out_buff[2] += bias[out_ch + 2 + mult_tile]; + out_buff[3] += bias[out_ch + 3 + mult_tile]; + } + out_buff[0] = arm_nn_requantize(out_buff[0], output_mult, output_shift); + out_buff[1] = arm_nn_requantize(out_buff[1], output_mult, output_shift); + out_buff[2] = arm_nn_requantize(out_buff[2], output_mult, output_shift); + out_buff[3] = arm_nn_requantize(out_buff[3], output_mult, output_shift); + + out_buff[0] += output_offset; + out_buff[1] += output_offset; + out_buff[2] += output_offset; + out_buff[3] += output_offset; + + out_buff[0] = MIN(MAX(out_buff[0], output_activation_min), output_activation_max); + out_buff[1] = MIN(MAX(out_buff[1], output_activation_min), output_activation_max); + out_buff[2] = MIN(MAX(out_buff[2], output_activation_min), output_activation_max); + out_buff[3] = MIN(MAX(out_buff[3], output_activation_min), output_activation_max); + + output[out_idx++] = (uint8_t)out_buff[0]; + output[out_idx++] = (uint8_t)out_buff[1]; + output[out_idx++] = (uint8_t)out_buff[2]; + output[out_idx++] = (uint8_t)out_buff[3]; + } + } + } + } +} + +static void depthwise_conv_u8_generic(const uint8_t *input, + const int32_t input_x, + const int32_t input_y, + const int32_t input_ch, + const uint8_t *kernel, + const int32_t output_ch, + const int32_t ch_mult, + const int32_t kernel_x, + const int32_t kernel_y, + const int32_t pad_x, + const int32_t pad_y, + const int32_t stride_x, + const int32_t stride_y, + const int32_t *bias, + uint8_t *output, + const int32_t output_shift, + const int32_t output_mult, + const int32_t output_x, + const int32_t output_y, + const int32_t output_offset, + const int32_t input_offset, + const int32_t filter_offset, + const int32_t output_activation_min, + const int32_t output_activation_max) +{ + (void)output_ch; + int i_out = 0; + for (int i_out_y = 0; i_out_y < output_y; i_out_y++) + { + const int16_t base_idx_y = (i_out_y * stride_y) - pad_y; + for (int i_out_x = 0; i_out_x < output_x; i_out_x++) + { + const int16_t base_idx_x = (i_out_x * stride_x) - pad_x; + for (int i_input_ch = 0; i_input_ch < input_ch; i_input_ch++) + { + for (int i_ch_mult = 0; i_ch_mult < ch_mult; i_ch_mult++) + { + const int idx_out_ch = i_ch_mult + i_input_ch * ch_mult; + int32_t acc_0; + /* Condition for kernel start dimension: (base_idx_ + ker__start) >= 0 */ + const int ker_y_start = MAX(0, -base_idx_y); + const int ker_x_start = MAX(0, -base_idx_x); + /* Condition for kernel end dimension: (base_idx_ + ker__end) < input_ */ + const int ker_y_end = MIN(kernel_y, input_y - base_idx_y); + const int ker_x_end = MIN(kernel_x, input_x - base_idx_x); + acc_0 = 0; + + for (int i_ker_y = ker_y_start; i_ker_y < ker_y_end; i_ker_y++) + { + const int32_t idx_y = base_idx_y + i_ker_y; + for (int i_ker_x = ker_x_start; i_ker_x < ker_x_end; i_ker_x++) + { + const int32_t idx_x = base_idx_x + i_ker_x; + int32_t idx_0 = (idx_y * input_x + idx_x) * input_ch + i_input_ch; + int32_t ker_idx_0 = (i_ker_y * kernel_x + i_ker_x) * (input_ch * ch_mult) + idx_out_ch; + + acc_0 += (input[idx_0] + input_offset) * (kernel[ker_idx_0] + filter_offset); + } + } + if (bias != NULL) + { + acc_0 += bias[idx_out_ch]; + } + + /* Requantize and clamp output to provided range */ + acc_0 = arm_nn_requantize(acc_0, output_mult, output_shift); + acc_0 += output_offset; + acc_0 = MAX(acc_0, output_activation_min); + acc_0 = MIN(acc_0, output_activation_max); + + output[i_out++] = acc_0; + } + } + } + } +} + +/** + * @brief uint8 depthwise convolution function with asymmetric quantization + * + * @param[in] input Pointer to input tensor + * @param[in] input_x Width of input tensor + * @param[in] input_y Height of input tensor + * @param[in] input_ch Channels in input tensor + * @param[in] kernel Pointer to kernel weights + * @param[in] kernel_x Width of kernel + * @param[in] kernel_y Height of kernel + * @param[in] ch_mult Number of channel multiplier + * @param[in] pad_x Padding sizes x + * @param[in] pad_y Padding sizes y + * @param[in] stride_x Convolution stride along the width + * @param[in] stride_y Convolution stride along the height + * @param[in] dilation_x Dilation along width. Not used and intended for future enhancement. + * @param[in] dilation_y Dilation along height. Not used and intended for future enhancement. + * @param[in] bias Pointer to optional bias values. If no bias is + * availble, NULL is expected + * @param[in] input_offset Input tensor zero offset + * @param[in] filter_offset Kernel tensor zero offset + * @param[in] output_offset Output tensor zero offset + * @param[in,out] output Pointer to output tensor + * @param[in] output_x Width of output tensor + * @param[in] output_y Height of output tensor + * @param[in] output_activation_min Minimum value to clamp the output to. Range : {0, 255} + * @param[in] output_activation_max Minimum value to clamp the output to. Range : {0, 255} + * @param[in] output_shift Amount of right-shift for output + * @param[in] output_mult Output multiplier for requantization + * @return The function returns one of the following + * ARM_MATH_SIZE_MISMATCH - Not supported dimension of tensors + * ARM_MATH_SUCCESS - Successful operation + * ARM_MATH_ARGUMENT_ERROR - Implementation not available + * + * + */ + +arm_status arm_depthwise_conv_u8_basic_ver1(const uint8_t *input, + const uint16_t input_x, + const uint16_t input_y, + const uint16_t input_ch, + const uint8_t *kernel, + const uint16_t kernel_x, + const uint16_t kernel_y, + const int16_t ch_mult, + const int16_t pad_x, + const int16_t pad_y, + const int16_t stride_x, + const int16_t stride_y, + const int16_t dilation_x, + const int16_t dilation_y, + const int32_t *bias, + const int32_t input_offset, + const int32_t filter_offset, + const int32_t output_offset, + uint8_t *output, + const uint16_t output_x, + const uint16_t output_y, + const int32_t output_activation_min, + const int32_t output_activation_max, + const int32_t output_shift, + const int32_t output_mult) +{ + (void)dilation_x; + (void)dilation_y; + + if (ch_mult % 4 == 0) + { + depthwise_conv_u8_mult_4(input, input_x, input_y, input_ch, kernel, ch_mult * input_ch, ch_mult, + kernel_x, kernel_y, pad_x, pad_y, stride_x, stride_y, bias, output, + output_shift, output_mult, output_x, output_y, output_offset, input_offset, + filter_offset, output_activation_min, output_activation_max); + } + else + { + depthwise_conv_u8_generic(input, input_x, input_y, input_ch, kernel, ch_mult * input_ch, ch_mult, + kernel_x, kernel_y, pad_x, pad_y, stride_x, stride_y, bias, + output, output_shift, output_mult, output_x, output_y, output_offset, + input_offset, filter_offset, output_activation_min, output_activation_max); + } + + /* Return to application */ + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_wrapper_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_wrapper_s8.c new file mode 100644 index 0000000..74cc4e1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_wrapper_s8.c @@ -0,0 +1,136 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_depthwise_conv_wrapper_s8.c + * Description: Wrapper API to select appropriate depthwise conv API based + * on dimensions. + * + * $Date: May 29, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/* + * s8 Depthwise conv wrapper function + * + * Refer header file for details. + * + */ +arm_status arm_depthwise_conv_wrapper_s8(const cmsis_nn_context *ctx, + const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_per_channel_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input, + const cmsis_nn_dims *filter_dims, + const q7_t *filter, + const cmsis_nn_dims *bias_dims, + const int32_t *bias, + const cmsis_nn_dims *output_dims, + q7_t *output) +{ + arm_status status = ARM_MATH_SUCCESS; + if (1 == dw_conv_params->ch_mult) + { +#if !defined(ARM_MATH_MVEI) + if ((filter_dims->w == 3) && (filter_dims->h == 3) && (dw_conv_params->padding.h <= 1)) + { + status = arm_depthwise_conv_3x3_s8(ctx, + dw_conv_params, + quant_params, + input_dims, + input, + filter_dims, + filter, + bias_dims, + bias, + output_dims, + output); + } + else +#endif + { + status = arm_depthwise_conv_s8_opt(ctx, + dw_conv_params, + quant_params, + input_dims, + input, + filter_dims, + filter, + bias_dims, + bias, + output_dims, + output); + } + } + else + { + status = arm_depthwise_conv_s8(ctx, + dw_conv_params, + quant_params, + input_dims, + input, + filter_dims, + filter, + bias_dims, + bias, + output_dims, + output); + } + + /* Return to application */ + return status; +} + +int32_t arm_depthwise_conv_wrapper_s8_get_buffer_size(const cmsis_nn_dw_conv_params *dw_conv_params, + const cmsis_nn_dims *input_dims, + const cmsis_nn_dims *filter_dims, + const cmsis_nn_dims *output_dims) +{ + (void)dw_conv_params; + int32_t size = 0; + + if (input_dims->c == output_dims->c) + { + size = arm_depthwise_conv_s8_opt_get_buffer_size(input_dims, filter_dims); + } + + return size; +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c new file mode 100644 index 0000000..8e5748c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c @@ -0,0 +1,421 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_depthwise_separable_conv_HWC_q7.c + * Description: Q7 depthwise separable convolution function + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/** + * @brief Q7 depthwise separable convolution function + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * @details + * + * Buffer size: + * + * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel + * + * bufferB size: 0 + * + * Input dimension constraints: + * + * ch_im_in equals ch_im_out + * + * Implementation: + * There are 3 nested loop here: + * Inner loop: calculate each output value with MAC instruction over an accumulator + * Mid loop: loop over different output channel + * Outer loop: loop over different output (x, y) + */ + +arm_status arm_depthwise_separable_conv_HWC_q7(const q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out, + q15_t * bufferA, + q7_t * bufferB) +{ + (void)bufferB; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + int16_t i_out_y, i_out_x; + int16_t i_ker_y, i_ker_x; + q7_t *colBuffer = (q7_t *) bufferA; + q7_t *pBuffer = colBuffer; + const q7_t *pBias = bias; + q7_t *pOut = Im_out; + uint16_t rowCnt; + uint16_t row_shift; + + /* do some checking here, basically ch_im_in == ch_im_out */ + if (ch_im_in != ch_im_out) + { + return ARM_MATH_SIZE_MISMATCH; + } + + for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + /* we first do im2col here */ + for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++) + { + for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in) + { + /* arm_fill_q7(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, ch_im_in); + } else + { + /* arm_copy_q7((q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */ + memcpy(pBuffer, (q7_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + /* we will do the computation here for each channel */ + rowCnt = ch_im_out >> 2; + row_shift = 0; + pBias = bias; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = (dim_kernel * dim_kernel) >> 1; + q7_t *pB = colBuffer + row_shift; + const q7_t *pA = wt + row_shift; + row_shift += 4; + +#ifdef USE_INTRINSIC + +#ifndef ARM_MATH_BIG_ENDIAN + + while (colCnt) + { + q31_t inA1, inA2, inB1, inB2, opA, opB; + + inB1 = arm_nn_read_q7x4(pB); + pB += ch_im_in; + opB = arm_nn_read_q7x4(pB); + pB += ch_im_in; + inB2 = __PKHTB(opB, inB1, 16); + inB1 = __PKHBT(inB1, opB, 16); + inA1 = arm_nn_read_q7x4(pA); + pA += ch_im_in; + opB = arm_nn_read_q7x4(pA); + pA += ch_im_in; + inA2 = __PKHTB(opB, inA1, 16); + inA1 = __PKHBT(inA1, opB, 16); + opA = __SXTB16(inA1); + opB = __SXTB16(inB1); + sum = __SMLAD(opA, opB, sum); + opA = __SXTB16(__ROR(inA1, 8)); + opB = __SXTB16(__ROR(inB1, 8)); + sum2 = __SMLAD(opA, opB, sum2); + opA = __SXTB16(inA2); + opB = __SXTB16(inB2); + sum3 = __SMLAD(opA, opB, sum3); + opA = __SXTB16(__ROR(inA2, 8)); + opB = __SXTB16(__ROR(inB2, 8)); + sum4 = __SMLAD(opA, opB, sum4); + colCnt--; + } +#else + + while (colCnt) + { + q31_t inA1, inA2, inB1, inB2, opA, opB; + + inB1 = arm_nn_read_q7x4(pB); + pB += ch_im_in; + opB = arm_nn_read_q7x4(pB); + pB += ch_im_in; + inB2 = __PKHBT(opB, inB1, 16); + inB1 = __PKHTB(inB1, opB, 16); + inA1 = arm_nn_read_q7x4(pA); + pA += ch_im_in; + opB = arm_nn_read_q7x4(pA); + pA += ch_im_in; + inA2 = __PKHBT(opB, inA1, 16); + inA1 = __PKHTB(inA1, opB, 16); + opA = __SXTB16(inA1); + opB = __SXTB16(inB1); + sum2 = __SMLAD(opA, opB, sum2); + opA = __SXTB16(__ROR(inA1, 8)); + opB = __SXTB16(__ROR(inB1, 8)); + sum = __SMLAD(opA, opB, sum); + opA = __SXTB16(inA2); + opB = __SXTB16(inB2); + sum4 = __SMLAD(opA, opB, sum4); + opA = __SXTB16(__ROR(inA2, 8)); + opB = __SXTB16(__ROR(inB2, 8)); + sum3 = __SMLAD(opA, opB, sum3); + colCnt--; + } + +#endif /* ARM_MATH_BIG_ENDIAN */ + +#else + +#ifndef ARM_MATH_BIG_ENDIAN + /* + * r0 r1 r2 r3 r4 r5 + * inA1, inA2, inB1, inB2, opA, opB + */ + + asm volatile ("COL_LOOP_%=:\n" + "ldr.w r2, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "ldr.w r5, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "pkhtb r3, r5, r2, ASR #16\n" + "pkhbt r2, r2, r5, LSL #16\n" + "ldr.w r0, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "ldr.w r5, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "pkhtb r1, r5, r0, ASR #16\n" + "pkhbt r0, r0, r5, LSL #16\n" + "sxtb16 r4, r0\n" + "sxtb16 r5, r2\n" + "smlad %[sum], r4, r5, %[sum]\n" + "mov.w r4, r0, ror #8\n" + "mov.w r5, r2, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum2], r4, r5, %[sum2]\n" + "sxtb16 r4, r1\n" + "sxtb16 r5, r3\n" + "smlad %[sum3], r4, r5, %[sum3]\n" + "mov.w r4, r1, ror #8\n" + "mov.w r5, r3, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum4], r4, r5, %[sum4]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP_%=\n":[sum] + "+r"(sum),[sum2] "+r"(sum2), + [sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB), + [pA] "+r"(pA):[colCnt] + "r"(colCnt),[ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5"); +#else + /* + * r0 r1 r2 r3 r4 r5 + * inA1, inA2, inB1, inB2, opA, opB + */ + asm volatile ("COL_LOOP_%=:\n" + "ldr.w r2, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "ldr.w r5, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "pkhbt r3, r5, r2, LSL #16\n" + "pkhtb r2, r2, r5, ASR #16\n" + "ldr.w r0, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "ldr.w r5, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "pkhbt r1, r5, r0, LSL #16\n" + "pkhtb r0, r0, r5, ASR #16\n" + "sxtb16 r4, r0\n" + "sxtb16 r5, r2\n" + "smlad %[sum2], r4, r5, %[sum2]\n" + "mov.w r4, r0, ror #8\n" + "mov.w r5, r2, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum], r4, r5, %[sum]\n" + "sxtb16 r4, r1\n" + "sxtb16 r5, r3\n" + "smlad %[sum4], r4, r5, %[sum4]\n" + "mov.w r4, r1, ror #8\n" + "mov.w r5, r3, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum3], r4, r5, %[sum3]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP_%=\n":[sum] + "+r"(sum),[sum2] "+r"(sum2), + [sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB), + [pA] "+r"(pA):[colCnt] + "r"(colCnt),[ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5"); + +#endif /* ARM_MATH_BIG_ENDIAN */ + +#endif /* USE_INTRINSIC */ + + colCnt = (dim_kernel * dim_kernel) & 0x1; + while (colCnt) + { + union arm_nnword inA, inB; + inA.word = arm_nn_read_q7x4(pA); + pA += ch_im_in; + inB.word = arm_nn_read_q7x4(pB); + pB += ch_im_in; + sum += inA.bytes[0] * inB.bytes[0]; + sum2 += inA.bytes[1] * inB.bytes[1]; + sum3 += inA.bytes[2] * inB.bytes[2]; + sum4 += inA.bytes[3] * inB.bytes[3]; + colCnt--; + } + + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum2 >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum3 >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum4 >> out_shift), 8); + + rowCnt--; + } + + rowCnt = ch_im_out & 0x3; + while (rowCnt) + { + q7_t *pB = colBuffer + row_shift; + const q7_t *pA = wt + row_shift; + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + uint16_t colCnt = (dim_kernel * dim_kernel); + + row_shift += 1; + + while (colCnt) + { + q7_t A1 = *pA; + q7_t B1 = *pB; + pA += ch_im_in; + pB += ch_im_in; + sum += A1 * B1; + + colCnt--; + } + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + rowCnt--; + } + + /* clear counter and pointers */ + pBuffer = colBuffer; + } + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + int i_out_y, i_out_x, i_ch_out, i_ker_x, i_ker_y; + int conv_out; + + /* do some checking here, basically ch_im_in == ch_im_out */ + if (ch_im_in != ch_im_out) + { + return ARM_MATH_SIZE_MISMATCH; + } + + for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++) + { + for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++) + { + // for each output + conv_out = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift); + for (i_ker_y = 0; i_ker_y < dim_kernel; i_ker_y++) + { + for (i_ker_x = 0; i_ker_x < dim_kernel; i_ker_x++) + { + int in_row = stride * i_out_y + i_ker_y - padding; + int in_col = stride * i_out_x + i_ker_x - padding; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in) + { + conv_out += + Im_in[(in_row * + dim_im_in + + in_col) * + ch_im_in + + i_ch_out] * wt[(i_ker_y * dim_kernel + i_ker_x) * ch_im_out + i_ch_out]; + } + } + } + Im_out[(i_out_y * dim_im_out + + i_out_x) * ch_im_out + i_ch_out] = (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return ARM_MATH_SUCCESS; + +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c new file mode 100644 index 0000000..534748a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c @@ -0,0 +1,416 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_depthwise_separable_conv_HWC_q7_nonsquare.c + * Description: Q7 depthwise separable convolution function (non-square shape) + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup NNConv + * @{ + */ + +/** + * @brief Q7 depthwise separable convolution function (non-square shape) + * @param[in] Im_in pointer to input tensor + * @param[in] dim_im_in_x input tensor dimention x + * @param[in] dim_im_in_y input tensor dimention y + * @param[in] ch_im_in number of input tensor channels + * @param[in] wt pointer to kernel weights + * @param[in] ch_im_out number of filters, i.e., output tensor channels + * @param[in] dim_kernel_x filter kernel size x + * @param[in] dim_kernel_y filter kernel size y + * @param[in] padding_x padding sizes x + * @param[in] padding_y padding sizes y + * @param[in] stride_x convolution stride x + * @param[in] stride_y convolution stride y + * @param[in] bias pointer to bias + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in,out] Im_out pointer to output tensor + * @param[in] dim_im_out_x output tensor dimension x + * @param[in] dim_im_out_y output tensor dimension y + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] bufferB pointer to buffer space for output + * @return The function returns either + * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking. + * + * This function is the version with full list of optimization tricks, but with + * some contraints: + * ch_im_in is equal to ch_im_out + * + */ + +arm_status arm_depthwise_separable_conv_HWC_q7_nonsquare(const q7_t * Im_in, + const uint16_t dim_im_in_x, + const uint16_t dim_im_in_y, + const uint16_t ch_im_in, + const q7_t * wt, + const uint16_t ch_im_out, + const uint16_t dim_kernel_x, + const uint16_t dim_kernel_y, + const uint16_t padding_x, + const uint16_t padding_y, + const uint16_t stride_x, + const uint16_t stride_y, + const q7_t * bias, + const uint16_t bias_shift, + const uint16_t out_shift, + q7_t * Im_out, + const uint16_t dim_im_out_x, + const uint16_t dim_im_out_y, + q15_t * bufferA, + q7_t * bufferB) +{ + + (void)bufferB; + +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + +/* + * Implementation: + * There are 3 nested loop here: + * Inner loop: calculate each output value with MAC instruction over an accumulator + * Mid loop: loop over different output channel + * Outer loop: loop over different output (x, y) + * + */ + + int16_t i_out_y, i_out_x; + int16_t i_ker_y, i_ker_x; + q7_t *colBuffer = (q7_t *) bufferA; + q7_t *pBuffer = colBuffer; + const q7_t *pBias = bias; + q7_t *pOut = Im_out; + uint16_t rowCnt; + uint16_t row_shift; + + /* do some checking here, basically ch_im_in == ch_im_out */ + if (ch_im_in != ch_im_out) + { + return ARM_MATH_SIZE_MISMATCH; + } + + for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) + { + /* we first do im2col here */ + for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; + i_ker_y++) + { + for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; + i_ker_x++) + { + if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x) + { + /* arm_fill_q7(0, pBuffer, ch_im_in); */ + memset(pBuffer, 0, ch_im_in); + } else + { + /* arm_copy_q7((q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */ + memcpy(pBuffer, (q7_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, ch_im_in); + } + pBuffer += ch_im_in; + } + } + + /* we will do the computation here for each channel */ + rowCnt = ch_im_out >> 2; + row_shift = 0; + pBias = bias; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = (dim_kernel_x * dim_kernel_y) >> 1; + q7_t *pB = colBuffer + row_shift; + const q7_t *pA = wt + row_shift; + row_shift += 4; + +#ifdef USE_INTRINSIC + +#ifndef ARM_MATH_BIG_ENDIAN + + while (colCnt) + { + q31_t inA1, inA2, inB1, inB2, opA, opB; + + inB1 = arm_nn_read_q7x4(pB); + pB += ch_im_in; + opB = arm_nn_read_q7x4(pB); + pB += ch_im_in; + inB2 = __PKHTB(opB, inB1, 16); + inB1 = __PKHBT(inB1, opB, 16); + inA1 = arm_nn_read_q7x4(pA); + pA += ch_im_in; + opB = arm_nn_read_q7x4(pA); + pA += ch_im_in; + inA2 = __PKHTB(opB, inA1, 16); + inA1 = __PKHBT(inA1, opB, 16); + opA = __SXTB16(inA1); + opB = __SXTB16(inB1); + sum = __SMLAD(opA, opB, sum); + opA = __SXTB16(__ROR(inA1, 8)); + opB = __SXTB16(__ROR(inB1, 8)); + sum2 = __SMLAD(opA, opB, sum2); + opA = __SXTB16(inA2); + opB = __SXTB16(inB2); + sum3 = __SMLAD(opA, opB, sum3); + opA = __SXTB16(__ROR(inA2, 8)); + opB = __SXTB16(__ROR(inB2, 8)); + sum4 = __SMLAD(opA, opB, sum4); + colCnt--; + } +#else + + while (colCnt) + { + q31_t inA1, inA2, inB1, inB2, opA, opB; + + inB1 = arm_nn_read_q7x4(pB); + pB += ch_im_in; + opB = arm_nn_read_q7x4(pB); + pB += ch_im_in; + inB2 = __PKHBT(opB, inB1, 16); + inB1 = __PKHTB(inB1, opB, 16); + inA1 = arm_nn_read_q7x4(pA); + pA += ch_im_in; + opB = arm_nn_read_q7x4(pA); + pA += ch_im_in; + inA2 = __PKHBT(opB, inA1, 16); + inA1 = __PKHTB(inA1, opB, 16); + opA = __SXTB16(inA1); + opB = __SXTB16(inB1); + sum2 = __SMLAD(opA, opB, sum2); + opA = __SXTB16(__ROR(inA1, 8)); + opB = __SXTB16(__ROR(inB1, 8)); + sum = __SMLAD(opA, opB, sum); + opA = __SXTB16(inA2); + opB = __SXTB16(inB2); + sum4 = __SMLAD(opA, opB, sum4); + opA = __SXTB16(__ROR(inA2, 8)); + opB = __SXTB16(__ROR(inB2, 8)); + sum3 = __SMLAD(opA, opB, sum3); + colCnt--; + } + +#endif /* ARM_MATH_BIG_ENDIAN */ + +#else + +#ifndef ARM_MATH_BIG_ENDIAN + // r0 r1 r2 r3 r4 r5 + // inA1, inA2, inB1, inB2, opA, opB + asm volatile ("COL_LOOP:\n" + "ldr.w r2, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "ldr.w r5, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "pkhtb r3, r5, r2, ASR #16\n" + "pkhbt r2, r2, r5, LSL #16\n" + "ldr.w r0, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "ldr.w r5, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "pkhtb r1, r5, r0, ASR #16\n" + "pkhbt r0, r0, r5, LSL #16\n" + "sxtb16 r4, r0\n" + "sxtb16 r5, r2\n" + "smlad %[sum], r4, r5, %[sum]\n" + "mov.w r4, r0, ror #8\n" + "mov.w r5, r2, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum2], r4, r5, %[sum2]\n" + "sxtb16 r4, r1\n" + "sxtb16 r5, r3\n" + "smlad %[sum3], r4, r5, %[sum3]\n" + "mov.w r4, r1, ror #8\n" + "mov.w r5, r3, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum4], r4, r5, %[sum4]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP\n":[sum] "+r"(sum),[sum2] "+r"(sum2),[sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt), + [ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5"); +#else + // r0 r1 r2 r3 r4 r5 + // inA1, inA2, inB1, inB2, opA, opB + asm volatile ("COL_LOOP:\n" + "ldr.w r2, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "ldr.w r5, [%[pB], #0]\n" + "add.w %[pB], %[pB], %[ch_im_in]\n" + "pkhbt r3, r5, r2, LSL #16\n" + "pkhtb r2, r2, r5, ASR #16\n" + "ldr.w r0, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "ldr.w r5, [%[pA], #0]\n" + "add.w %[pA], %[pA], %[ch_im_in]\n" + "pkhbt r1, r5, r0, LSL #16\n" + "pkhtb r0, r0, r5, ASR #16\n" + "sxtb16 r4, r0\n" + "sxtb16 r5, r2\n" + "smlad %[sum2], r4, r5, %[sum2]\n" + "mov.w r4, r0, ror #8\n" + "mov.w r5, r2, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum], r4, r5, %[sum]\n" + "sxtb16 r4, r1\n" + "sxtb16 r5, r3\n" + "smlad %[sum4], r4, r5, %[sum4]\n" + "mov.w r4, r1, ror #8\n" + "mov.w r5, r3, ror #8\n" + "sxtb16 r4, r4\n" + "sxtb16 r5, r5\n" + "smlad %[sum3], r4, r5, %[sum3]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP\n":[sum] "+r"(sum),[sum2] "+r"(sum2),[sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt), + [ch_im_in] "r"(ch_im_in):"r0", "r1", "r2", "r3", "r4", "r5"); +#endif /*ARM_MATH_BIG_ENDIAN */ + +#endif /* USE_INTRINSIC */ + + colCnt = (dim_kernel_x * dim_kernel_y) & 0x1; + while (colCnt) + { + union arm_nnword inA, inB; + inA.word = arm_nn_read_q7x4(pA); + pA += ch_im_in; + inB.word = arm_nn_read_q7x4(pB); + pB += ch_im_in; + sum += inA.bytes[0] * inB.bytes[0]; + sum2 += inA.bytes[1] * inB.bytes[1]; + sum3 += inA.bytes[2] * inB.bytes[2]; + sum4 += inA.bytes[3] * inB.bytes[3]; + colCnt--; + } + + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum2 >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum3 >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum4 >> out_shift), 8); + + rowCnt--; + } + + rowCnt = ch_im_out & 0x3; + while (rowCnt) + { + q7_t *pB = colBuffer + row_shift; + const q7_t *pA = wt + row_shift; + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + uint16_t colCnt = (dim_kernel_x * dim_kernel_y); + + row_shift += 1; + + while (colCnt) + { + q7_t A1 = *pA; + q7_t B1 = *pB; + pA += ch_im_in; + pB += ch_im_in; + sum += A1 * B1; + + colCnt--; + } + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + rowCnt--; + } + + // clear counter and pointers + pBuffer = colBuffer; + } + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + int i_out_y, i_out_x, i_ch_out; + int i_ker_y, i_ker_x; + + /* do some checking here, basically ch_im_in == ch_im_out */ + if (ch_im_in != ch_im_out) + { + return ARM_MATH_SIZE_MISMATCH; + } + + for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++) + { + for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++) + { + for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++) + { + // for each output + int conv_out = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift); + for (i_ker_y = 0; i_ker_y < dim_kernel_y; i_ker_y++) + { + for (i_ker_x = 0; i_ker_x < dim_kernel_x; i_ker_x++) + { + int in_row = stride_y * i_out_y + i_ker_y - padding_y; + int in_col = stride_x * i_out_x + i_ker_x - padding_x; + if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x) + { + conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + i_ch_out] * + wt[(i_ker_y * dim_kernel_x + i_ker_x) * ch_im_out + i_ch_out]; + } + } + } + Im_out[(i_out_y * dim_im_out_x + i_out_x) * ch_im_out + i_ch_out] = + (q7_t) __SSAT((conv_out >> out_shift), 8); + } + } + } + +#endif /* ARM_MATH_DSP */ + + + /* Return to application */ + return ARM_MATH_SUCCESS; + +} + +/** + * @} end of NNConv group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_depthwise_conv_s8_core.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_depthwise_conv_s8_core.c new file mode 100644 index 0000000..e53b718 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_depthwise_conv_s8_core.c @@ -0,0 +1,222 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_depthwise_conv_s8_core.c + * Description: Depthwise convolution on im2col buffers. + * + * $Date: May 29, 2020 + * $Revision: V.1.0.3 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/* + * Depthwise conv on an im2col buffer where the input channel equals + * output channel. + * + * Refer header file for details. + * + */ + +q7_t *arm_nn_depthwise_conv_s8_core(const q7_t *row, + const q15_t *col, + const uint16_t num_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int32_t activation_min, + const int32_t activation_max, + const uint16_t kernel_size, + const int32_t *const output_bias, + q7_t *out) +{ +#if defined(ARM_MATH_MVEI) + int32_t ch_per_loop = num_ch / 4; + + const int32_t *bias = output_bias; + int8_t *out_tmp = out; + + int32_t idx = 0; + + while (ch_per_loop > 0) + { + int32x4_t ip_0; + int32x4_t ip_1; + int32_t ker_loop = kernel_size / 3; + int32x4_t out_0 = vldrwq_s32(bias); + int32x4_t out_1 = out_0; + bias += 4; + + const int32_t offset = idx * 4; + const int8_t *row_0 = row + offset; + const int16_t *col_0 = col + offset; + const int16_t *col_1 = col + kernel_size * num_ch + offset; + + int32x4_t ker_0 = vldrbq_s32(row_0); + + while (ker_loop > 0) + { + const int8_t *row_1 = row_0 + num_ch; + const int8_t *row_2 = row_0 + 2 * num_ch; + const int32x4_t ker_1 = vldrbq_s32(row_1); + const int32x4_t ker_2 = vldrbq_s32(row_2); + + ip_0 = vldrhq_s32(col_0); + ip_1 = vldrhq_s32(col_1); + col_0 += num_ch; + col_1 += num_ch; + + out_0 += vmulq_s32(ip_0, ker_0); + out_1 += vmulq_s32(ip_1, ker_0); + + ip_0 = vldrhq_s32(col_0); + ip_1 = vldrhq_s32(col_1); + col_0 += num_ch; + col_1 += num_ch; + + out_0 += vmulq_s32(ip_0, ker_1); + out_1 += vmulq_s32(ip_1, ker_1); + + ip_0 = vldrhq_s32(col_0); + ip_1 = vldrhq_s32(col_1); + col_0 += num_ch; + col_1 += num_ch; + + out_0 += vmulq_s32(ip_0, ker_2); + out_1 += vmulq_s32(ip_1, ker_2); + row_0 += 3 * num_ch; + + ker_0 = vldrbq_s32(row_0); + ker_loop--; + } + + idx++; + /* Handle tail kernel elements */ + ker_loop = kernel_size - ((kernel_size / 3) * 3); + while (ker_loop > 0) + { + ip_0 = vldrhq_s32(col_0); + ip_1 = vldrhq_s32(col_1); + + out_0 += vmulq_s32(ip_0, ker_0); + out_1 += vmulq_s32(ip_1, ker_0); + + col_0 += num_ch; + col_1 += num_ch; + + ip_0 = vldrhq_s32(col_0); + ip_1 = vldrhq_s32(col_1); + + row_0 += num_ch; + ker_0 = vldrbq_s32(row_0); + ker_loop--; + } + const int32x4_t mult = vldrwq_s32(out_mult); + const int32x4_t shift = vldrwq_s32(out_shift); + out_mult += 4; + out_shift += 4; + + out_0 = arm_requantize_mve_32x4(out_0, mult, shift); + out_1 = arm_requantize_mve_32x4(out_1, mult, shift); + + out_0 = vaddq_n_s32(out_0, out_offset); + out_0 = vmaxq_s32(out_0, vdupq_n_s32(activation_min)); + out_0 = vminq_s32(out_0, vdupq_n_s32(activation_max)); + vstrbq_s32(out_tmp, out_0); + + out_1 = vaddq_n_s32(out_1, out_offset); + out_1 = vmaxq_s32(out_1, vdupq_n_s32(activation_min)); + out_1 = vminq_s32(out_1, vdupq_n_s32(activation_max)); + vstrbq_s32(out_tmp + num_ch, out_1); + + out_tmp += 4; + ch_per_loop--; + } + + int32_t tail_ch = num_ch & 3; + if (tail_ch != 0) + { + int32_t ch_idx = (num_ch & ~3); + int32x4_t col_0_sum; + int32x4_t col_1_sum; + + const int32_t single_buffer_size = kernel_size * num_ch; + for (int i = 0; i < tail_ch; i++) + { + const int16_t *col_pos_0 = col + ch_idx; + const int16_t *col_pos_1 = col_pos_0 + single_buffer_size; + + const int8_t *row_pos = row + ch_idx; + int32_t sum_0 = bias[i]; + int32_t sum_1 = bias[i]; + + for (int j = 0; j < kernel_size; j++) + { + const int8_t row_val = row_pos[j * num_ch]; + sum_0 += row_val * col_pos_0[j * num_ch]; + sum_1 += row_val * col_pos_1[j * num_ch]; + } + col_0_sum[i] = sum_0; + col_1_sum[i] = sum_1; + + ch_idx++; + } + const mve_pred16_t p = vctp32q((uint32_t)tail_ch); + const int32x4_t mult = vldrwq_z_s32(out_mult, p); + const int32x4_t shift = vldrwq_z_s32(out_shift, p); + + col_0_sum = arm_requantize_mve_32x4(col_0_sum, mult, shift); + col_1_sum = arm_requantize_mve_32x4(col_1_sum, mult, shift); + + col_0_sum = vaddq_n_s32(col_0_sum, out_offset); + col_0_sum = vmaxq_s32(col_0_sum, vdupq_n_s32(activation_min)); + col_0_sum = vminq_s32(col_0_sum, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out_tmp, col_0_sum, p); + + col_1_sum = vaddq_n_s32(col_1_sum, out_offset); + col_1_sum = vmaxq_s32(col_1_sum, vdupq_n_s32(activation_min)); + col_1_sum = vminq_s32(col_1_sum, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out_tmp + num_ch, col_1_sum, p); + + out_tmp += tail_ch; + } + + return out_tmp + num_ch; +#else + (void)row; + (void)col; + (void)num_ch; + (void)out_shift; + (void)out_mult; + (void)out_offset; + (void)activation_min; + (void)activation_max; + (void)kernel_size; + (void)output_bias; + (void)out; + return NULL; +#endif +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c new file mode 100644 index 0000000..e71c2e5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c @@ -0,0 +1,182 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mult_kernel_q7_q15.c + * Description: Matrix-multiplication function for convolution + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + + /** + * @brief Matrix-multiplication function for convolution. + * + * @details Refer to header file for details. + * + */ + +q7_t *arm_nn_mat_mult_kernel_q7_q15(const q7_t * pA, + const q15_t * pInBuffer, + const uint16_t ch_im_out, + const uint16_t numCol_A, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q7_t * pOut) +{ +#if defined (ARM_MATH_DSP) + /* set up the second output pointers */ + q7_t *pOut2 = pOut + ch_im_out; + const q7_t *pBias = bias; + + uint16_t rowCnt = ch_im_out >> 1; + /* this loop over rows in A */ + while (rowCnt) + { + /* setup pointers for B */ + const q15_t *pB = pInBuffer; + const q15_t *pB2 = pB + numCol_A; + + /* align the second pointer for A */ + const q7_t *pA2 = pA + numCol_A; + + /* init the sum with bias */ + q31_t sum = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = numCol_A >> 2; + /* accumulate over the vector */ + while (colCnt) + { + q31_t inA11, inA12, inA21, inA22; + + q31_t inB1 = arm_nn_read_q15x2_ia(&pB); + q31_t inB2 = arm_nn_read_q15x2_ia(&pB2); + + pA = read_and_pad(pA, &inA11, &inA12); + pA2 = read_and_pad(pA2, &inA21, &inA22); + + sum = __SMLAD(inA11, inB1, sum); + sum2 = __SMLAD(inA11, inB2, sum2); + sum3 = __SMLAD(inA21, inB1, sum3); + sum4 = __SMLAD(inA21, inB2, sum4); + + inB1 = arm_nn_read_q15x2_ia(&pB); + inB2 = arm_nn_read_q15x2_ia(&pB2); + + sum = __SMLAD(inA12, inB1, sum); + sum2 = __SMLAD(inA12, inB2, sum2); + sum3 = __SMLAD(inA22, inB1, sum3); + sum4 = __SMLAD(inA22, inB2, sum4); + + colCnt--; + } /* while over colCnt */ + colCnt = numCol_A & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + q7_t inA2 = *pA2++; + q15_t inB2 = *pB2++; + + sum += inA1 * inB1; + sum2 += inA1 * inB2; + sum3 += inA2 * inB1; + sum4 += inA2 * inB2; + colCnt--; + } /* while over colCnt */ + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum3 >> out_shift), 8); + *pOut2++ = (q7_t) __SSAT((sum2 >> out_shift), 8); + *pOut2++ = (q7_t) __SSAT((sum4 >> out_shift), 8); + + /* skip the row computed with A2 */ + pA += numCol_A; + rowCnt--; + } /* for over ch_im_out */ + + /* compute left-over row if any */ + if (ch_im_out & 0x1) + { + /* setup pointers for B */ + const q15_t *pB = pInBuffer; + const q15_t *pB2 = pB + numCol_A; + + /* load the bias */ + q31_t sum = ((q31_t)(*pBias) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = numCol_A >> 2; + while (colCnt) + { + q31_t inA11, inA12; + + q31_t inB1 = arm_nn_read_q15x2_ia(&pB); + q31_t inB2 = arm_nn_read_q15x2_ia(&pB2); + + pA = read_and_pad(pA, &inA11, &inA12); + + sum = __SMLAD(inA11, inB1, sum); + sum2 = __SMLAD(inA11, inB2, sum2); + + inB1 = arm_nn_read_q15x2_ia(&pB); + inB2 = arm_nn_read_q15x2_ia(&pB2); + + sum = __SMLAD(inA12, inB1, sum); + sum2 = __SMLAD(inA12, inB2, sum2); + + colCnt--; + } + colCnt = numCol_A & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + q15_t inB2 = *pB2++; + + sum += inA1 * inB1; + sum2 += inA1 * inB2; + colCnt--; + } + + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + *pOut2++ = (q7_t) __SSAT((sum2 >> out_shift), 8); + } + + pOut += ch_im_out; + + /* return the new output pointer with offset */ + return pOut; +#else + /* To be completed */ + return NULL; +#endif /* ARM_MATH_DSP */ + +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c new file mode 100644 index 0000000..fc4ad76 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c @@ -0,0 +1,132 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mult_kernel_q7_q15_reordered.c + * Description: Matrix-multiplication function for convolution with reordered columns + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" + + /** + * @brief Matrix-multiplication function for convolution with re-ordered input. + * + * @details Refer to header file for details. + * + */ + +q7_t *arm_nn_mat_mult_kernel_q7_q15_reordered(const q7_t * pA, + const q15_t * pInBuffer, + const uint16_t ch_im_out, + const uint16_t numCol_A, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q7_t * pOut) +{ + +#if defined (ARM_MATH_DSP) + /* set up the second output pointers */ + q7_t *pOut2 = pOut + ch_im_out; + int i; + + /* this loop over rows in A */ + for (i = 0; i < ch_im_out; i += 2) + { + /* setup pointers for B */ + const q15_t *pB = pInBuffer; + const q15_t *pB2 = pB + numCol_A; + + /* align the second pointer for A */ + const q7_t *pA2 = pA + numCol_A; + + /* init the sum with bias */ + q31_t sum = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(bias[i + 1]) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(bias[i + 1]) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = numCol_A >> 2; + /* accumulate over the vector */ + while (colCnt) + { + q31_t inA11, inA12, inA21, inA22; + + q31_t inB1 = arm_nn_read_q15x2_ia(&pB); + q31_t inB2 = arm_nn_read_q15x2_ia(&pB2); + + pA = read_and_pad_reordered(pA, &inA11, &inA12); + pA2 = read_and_pad_reordered(pA2, &inA21, &inA22); + + sum = __SMLAD(inA11, inB1, sum); + sum2 = __SMLAD(inA11, inB2, sum2); + sum3 = __SMLAD(inA21, inB1, sum3); + sum4 = __SMLAD(inA21, inB2, sum4); + + inB1 = arm_nn_read_q15x2_ia(&pB); + inB2 = arm_nn_read_q15x2_ia(&pB2); + + sum = __SMLAD(inA12, inB1, sum); + sum2 = __SMLAD(inA12, inB2, sum2); + sum3 = __SMLAD(inA22, inB1, sum3); + sum4 = __SMLAD(inA22, inB2, sum4); + + colCnt--; + } /* while over colCnt */ + colCnt = numCol_A & 0x3; + while (colCnt) + { + q7_t inA1 = *pA++; + q15_t inB1 = *pB++; + q7_t inA2 = *pA2++; + q15_t inB2 = *pB2++; + + sum += inA1 * inB1; + sum2 += inA1 * inB2; + sum3 += inA2 * inB1; + sum4 += inA2 * inB2; + colCnt--; + } /* while over colCnt */ + *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8); + *pOut++ = (q7_t) __SSAT((sum3 >> out_shift), 8); + *pOut2++ = (q7_t) __SSAT((sum2 >> out_shift), 8); + *pOut2++ = (q7_t) __SSAT((sum4 >> out_shift), 8); + + /* skip the row computed with A2 */ + pA += numCol_A; + } /* for over ch_im_out */ + + pOut += ch_im_out; + + /* return the new output pointer with offset */ + return pOut; +#else + /* To be completed */ + return NULL; +#endif /* ARM_MATH_DSP */ +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16.c new file mode 100644 index 0000000..e2df8ca --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16.c @@ -0,0 +1,394 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mult_kernel_s8_s16.c + * Description: Matrix-multiplication function for convolution + * + * $Date: May 29, 2020 + * $Revision: V.1.0.2 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/* + * Matrix-multiplication function for convolution with per-channel requantization. + * + * Refer header file for details. + * + */ + +q7_t *arm_nn_mat_mult_kernel_s8_s16(const q7_t *input_a, + const q15_t *input_b, + const uint16_t output_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int16_t activation_min, + const int16_t activation_max, + const uint16_t num_col_a, + const int32_t *const output_bias, + q7_t *out_0) +{ +#if defined(ARM_MATH_MVEI) +#define ROW_PER_LOOP (4) +#define COL_PER_LOOP (8) + + const q7_t *ip_a0_s8 = input_a; + q7_t *out_1 = out_0 + output_ch; + + const int32_t *bias = output_bias; + + int32_t row_count = output_ch / ROW_PER_LOOP; + + while (row_count) + { + const q15_t *ip_b0_s16 = input_b; + const q15_t *ip_b1_s16 = input_b + num_col_a; + + const q7_t *ip_a1_s8 = ip_a0_s8 + num_col_a; + const q7_t *ip_a2_s8 = ip_a0_s8 + num_col_a * 2; + const q7_t *ip_a3_s8 = ip_a0_s8 + num_col_a * 3; + + q31_t ch_0_out_n = bias[0]; + q31_t ch_1_out_n = bias[1]; + q31_t ch_2_out_n = bias[2]; + q31_t ch_3_out_n = bias[3]; + + q31_t ch_0_out_n1 = ch_0_out_n; + q31_t ch_1_out_n1 = ch_1_out_n; + q31_t ch_2_out_n1 = ch_2_out_n; + q31_t ch_3_out_n1 = ch_3_out_n; + bias += 4; + + int32_t col_count = num_col_a / COL_PER_LOOP; + + while (col_count) + { + // Load inputs + const int16x8_t ip_b0 = vld1q_s16(ip_b0_s16); + ip_b0_s16 += COL_PER_LOOP; + const int16x8_t ip_b1 = vld1q_s16(ip_b1_s16); + ip_b1_s16 += COL_PER_LOOP; + + // Load filters + const int16x8_t ip_a0 = vldrbq_s16(ip_a0_s8); + ip_a0_s8 += COL_PER_LOOP; + const int16x8_t ip_a1 = vldrbq_s16(ip_a1_s8); + ip_a1_s8 += COL_PER_LOOP; + const int16x8_t ip_a2 = vldrbq_s16(ip_a2_s8); + ip_a2_s8 += COL_PER_LOOP; + const int16x8_t ip_a3 = vldrbq_s16(ip_a3_s8); + ip_a3_s8 += COL_PER_LOOP; + + // MAC + ch_0_out_n += vmladavq_s16(ip_b0, ip_a0); + ch_1_out_n += vmladavq_s16(ip_b0, ip_a1); + ch_2_out_n += vmladavq_s16(ip_b0, ip_a2); + ch_3_out_n += vmladavq_s16(ip_b0, ip_a3); + ch_0_out_n1 += vmladavq_s16(ip_b1, ip_a0); + ch_1_out_n1 += vmladavq_s16(ip_b1, ip_a1); + ch_2_out_n1 += vmladavq_s16(ip_b1, ip_a2); + ch_3_out_n1 += vmladavq_s16(ip_b1, ip_a3); + + col_count--; + } + + /* Handle tail */ + col_count = (num_col_a & (COL_PER_LOOP - 1)) - 1; + while (col_count >= 0) + { + const int32_t b0 = ip_b0_s16[col_count]; + const int32_t b1 = ip_b1_s16[col_count]; + + ch_0_out_n += b0 * ip_a0_s8[col_count]; + ch_1_out_n += b0 * ip_a1_s8[col_count]; + ch_2_out_n += b0 * ip_a2_s8[col_count]; + ch_3_out_n += b0 * ip_a3_s8[col_count]; + + ch_0_out_n1 += b1 * ip_a0_s8[col_count]; + ch_1_out_n1 += b1 * ip_a1_s8[col_count]; + ch_2_out_n1 += b1 * ip_a2_s8[col_count]; + ch_3_out_n1 += b1 * ip_a3_s8[col_count]; + col_count--; + } + ip_a0_s8 += (num_col_a & (COL_PER_LOOP - 1)); + + int32x4_t out_vec_0; + int32x4_t out_vec_1; + out_vec_0[0] = ch_0_out_n; + out_vec_0[1] = ch_1_out_n; + out_vec_0[2] = ch_2_out_n; + out_vec_0[3] = ch_3_out_n; + + out_vec_1[0] = ch_0_out_n1; + out_vec_1[1] = ch_1_out_n1; + out_vec_1[2] = ch_2_out_n1; + out_vec_1[3] = ch_3_out_n1; + + int32x4_t mult = vldrwq_s32(out_mult); + int32x4_t shift = vldrwq_s32(out_shift); + out_mult += ROW_PER_LOOP; + out_shift += ROW_PER_LOOP; + + out_vec_0 = arm_requantize_mve_32x4(out_vec_0, mult, shift); + out_vec_1 = arm_requantize_mve_32x4(out_vec_1, mult, shift); + + out_vec_0 = vaddq_n_s32(out_vec_0, out_offset); + out_vec_0 = vmaxq_s32(out_vec_0, vdupq_n_s32(activation_min)); + out_vec_0 = vminq_s32(out_vec_0, vdupq_n_s32(activation_max)); + vstrbq_s32(out_0, out_vec_0); + out_0 += ROW_PER_LOOP; + + out_vec_1 = vaddq_n_s32(out_vec_1, out_offset); + out_vec_1 = vmaxq_s32(out_vec_1, vdupq_n_s32(activation_min)); + out_vec_1 = vminq_s32(out_vec_1, vdupq_n_s32(activation_max)); + vstrbq_s32(out_1, out_vec_1); + out_1 += ROW_PER_LOOP; + row_count--; + ip_a0_s8 += (num_col_a * 3); + } + + row_count = output_ch & (ROW_PER_LOOP - 1); + + if (row_count) + { + ip_a0_s8 = input_a + num_col_a * (output_ch & ~3); + const mve_pred16_t p = vctp32q((uint32_t)row_count); + int32x4_t out_vec_0 = vdupq_n_s32(0); + int32x4_t out_vec_1 = vdupq_n_s32(0); + int32x4_t mult_tail; + int32x4_t shift_tail; + + for (int i_ch = 0; i_ch < row_count; i_ch++) + { + int32_t output_0 = bias[i_ch]; + int32_t output_1 = bias[i_ch]; + const q15_t *ip_b0_s16 = input_b; + const q15_t *ip_b1_s16 = input_b + num_col_a; + + for (int i_idx = 0; i_idx < num_col_a; i_idx++) + { + output_0 += ip_b0_s16[i_idx] * ip_a0_s8[i_idx]; + output_1 += ip_b1_s16[i_idx] * ip_a0_s8[i_idx]; + } + + ip_a0_s8 += num_col_a; + out_vec_0[i_ch] = output_0; + out_vec_1[i_ch] = output_1; + mult_tail[i_ch] = out_mult[i_ch]; + shift_tail[i_ch] = out_shift[i_ch]; + } + out_vec_0 = arm_requantize_mve_32x4(out_vec_0, mult_tail, shift_tail); + out_vec_1 = arm_requantize_mve_32x4(out_vec_1, mult_tail, shift_tail); + + out_vec_0 = vaddq_n_s32(out_vec_0, out_offset); + out_vec_0 = vmaxq_s32(out_vec_0, vdupq_n_s32(activation_min)); + out_vec_0 = vminq_s32(out_vec_0, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out_0, out_vec_0, p); + + out_vec_1 = vaddq_n_s32(out_vec_1, out_offset); + out_vec_1 = vmaxq_s32(out_vec_1, vdupq_n_s32(activation_min)); + out_vec_1 = vminq_s32(out_vec_1, vdupq_n_s32(activation_max)); + + vstrbq_p_s32(out_1, out_vec_1, p); + out_1 += row_count; + } + + return out_1; + +#elif defined(ARM_MATH_DSP) + /* set up the second output pointers */ + q7_t *out_1 = out_0 + output_ch; + const int32_t *bias = output_bias; + + uint16_t row_count = output_ch / 2; + const q7_t *ip_a0 = input_a; + /* this loop over rows in A */ + while (row_count) + { + /* setup pointers for B */ + const q15_t *ip_b0 = input_b; + const q15_t *ip_b1 = ip_b0 + num_col_a; + + /* align the second pointer for A */ + const q7_t *ip_a1 = ip_a0 + num_col_a; + + /* Init accumulator with bias for channel N and N + 1 */ + q31_t ch_0_out_0 = *bias; + q31_t ch_0_out_1 = *bias++; + q31_t ch_1_out_0 = *bias; + q31_t ch_1_out_1 = *bias++; + + uint16_t col_count = num_col_a / 4; + /* accumulate over the vector */ + while (col_count) + { + q31_t a01, a02, a11, a12; + q31_t b0 = arm_nn_read_q15x2_ia(&ip_b0); + q31_t b1 = arm_nn_read_q15x2_ia(&ip_b1); + + ip_a0 = read_and_pad(ip_a0, &a01, &a02); + ip_a1 = read_and_pad(ip_a1, &a11, &a12); + + ch_0_out_0 = __SMLAD(a01, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a01, b1, ch_0_out_1); + ch_1_out_0 = __SMLAD(a11, b0, ch_1_out_0); + ch_1_out_1 = __SMLAD(a11, b1, ch_1_out_1); + + b0 = arm_nn_read_q15x2_ia(&ip_b0); + b1 = arm_nn_read_q15x2_ia(&ip_b1); + + ch_0_out_0 = __SMLAD(a02, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a02, b1, ch_0_out_1); + ch_1_out_0 = __SMLAD(a12, b0, ch_1_out_0); + ch_1_out_1 = __SMLAD(a12, b1, ch_1_out_1); + + col_count--; + } /* while over col_count */ + col_count = num_col_a & 0x3; + while (col_count) + { + q7_t a0 = *ip_a0++; + q15_t b0 = *ip_b0++; + q7_t a1 = *ip_a1++; + q15_t b1 = *ip_b1++; + + ch_0_out_0 += a0 * b0; + ch_0_out_1 += a0 * b1; + ch_1_out_0 += a1 * b0; + ch_1_out_1 += a1 * b1; + col_count--; + } /* while over col_count */ + + ch_0_out_0 = arm_nn_requantize(ch_0_out_0, *out_mult, *out_shift); + ch_0_out_0 += out_offset; + ch_0_out_0 = MAX(ch_0_out_0, activation_min); + ch_0_out_0 = MIN(ch_0_out_0, activation_max); + *out_0++ = (q7_t)ch_0_out_0; + + ch_0_out_1 = arm_nn_requantize(ch_0_out_1, *out_mult, *out_shift); + ch_0_out_1 += out_offset; + ch_0_out_1 = MAX(ch_0_out_1, activation_min); + ch_0_out_1 = MIN(ch_0_out_1, activation_max); + *out_1++ = (q7_t)ch_0_out_1; + out_mult++; + out_shift++; + + ch_1_out_0 = arm_nn_requantize(ch_1_out_0, *out_mult, *out_shift); + ch_1_out_0 += out_offset; + ch_1_out_0 = MAX(ch_1_out_0, activation_min); + ch_1_out_0 = MIN(ch_1_out_0, activation_max); + *out_0++ = (q7_t)ch_1_out_0; + + ch_1_out_1 = arm_nn_requantize(ch_1_out_1, *out_mult, *out_shift); + ch_1_out_1 += out_offset; + ch_1_out_1 = MAX(ch_1_out_1, activation_min); + ch_1_out_1 = MIN(ch_1_out_1, activation_max); + *out_1++ = (q7_t)ch_1_out_1; + out_mult++; + out_shift++; + + /* skip row */ + ip_a0 += num_col_a; + row_count--; + } + + /* compute the last odd numbered row if any */ + if (output_ch & 0x1) + { + /* setup pointers for B */ + const q15_t *ip_b0 = input_b; + const q15_t *ip_b1 = ip_b0 + num_col_a; + + /* load the bias */ + q31_t ch_0_out_0 = *bias; + q31_t ch_0_out_1 = *bias++; + + uint16_t col_count = num_col_a >> 2; + while (col_count) + { + q31_t a01, a02; + q31_t b0 = arm_nn_read_q15x2_ia(&ip_b0); + q31_t b1 = arm_nn_read_q15x2_ia(&ip_b1); + + ip_a0 = read_and_pad(ip_a0, &a01, &a02); + + ch_0_out_0 = __SMLAD(a01, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a01, b1, ch_0_out_1); + + b0 = arm_nn_read_q15x2_ia(&ip_b0); + b1 = arm_nn_read_q15x2_ia(&ip_b1); + ch_0_out_0 = __SMLAD(a02, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a02, b1, ch_0_out_1); + + col_count--; + } + col_count = num_col_a & 0x3; + while (col_count) + { + q7_t a0 = *ip_a0++; + q15_t b0 = *ip_b0++; + q15_t b1 = *ip_b1++; + + ch_0_out_0 += a0 * b0; + ch_0_out_1 += a0 * b1; + col_count--; + } + ch_0_out_0 = arm_nn_requantize(ch_0_out_0, *out_mult, *out_shift); + ch_0_out_0 += out_offset; + ch_0_out_0 = MAX(ch_0_out_0, activation_min); + ch_0_out_0 = MIN(ch_0_out_0, activation_max); + *out_0++ = (q7_t)ch_0_out_0; + + ch_0_out_1 = arm_nn_requantize(ch_0_out_1, *out_mult, *out_shift); + ch_0_out_1 += out_offset; + ch_0_out_1 = MAX(ch_0_out_1, activation_min); + ch_0_out_1 = MIN(ch_0_out_1, activation_max); + *out_1++ = (q7_t)ch_0_out_1; + out_mult++; + out_shift++; + } + + out_0 += output_ch; + + /* return the new output pointer with offset */ + return out_0; +#else + (void)input_a; + (void)input_b; + (void)output_ch; + (void)out_shift; + (void)out_mult; + (void)out_offset; + (void)activation_min; + (void)activation_max; + (void)num_col_a; + (void)output_bias; + (void)out_0; + /* To be completed */ + return NULL; +#endif +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16_reordered.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16_reordered.c new file mode 100644 index 0000000..de5b300 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16_reordered.c @@ -0,0 +1,204 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mult_kernel_s8_s16_reordered.c + * Description: Matrix-multiplication function for convolution with reordered columns + * + * $Date: February 27, 2020 + * $Revision: V.1.0.2 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" + +/* + * Matrix-multiplication with re-ordered input and bias inputs for convolution with per-channel + * requantization. The re-ordering is a consequence of sign extension is done by the SXTB16 command. + * + * Refer header file for details. This function differs from arm_nn_mat_mult_kernel_s8_s16(), in that it uses + * read_and_pad_reordered() instead of arm_nn_mat_mult_kernel_s8_s16(). Investigating the cycles impact and + * unifying these two functions is a potential future improvement. + * + */ + +q7_t *arm_nn_mat_mult_kernel_s8_s16_reordered(const q7_t *input_a, + const q15_t *input_b, + const uint16_t output_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int16_t activation_min, + const int16_t activation_max, + const uint16_t num_col_a, + const int32_t *const output_bias, + q7_t *out_0) +{ +#if defined(ARM_MATH_DSP) + /* set up the second output pointers */ + q7_t *out_1 = out_0 + output_ch; + const int32_t *bias = output_bias; + + uint16_t row_count = output_ch / 2; + const q7_t *ip_a0 = input_a; + /* this loop over rows in A */ + while (row_count) + { + /* setup pointers for B */ + const q15_t *ip_b0 = input_b; + const q15_t *ip_b1 = ip_b0 + num_col_a; + + /* align the second pointer for A */ + const q7_t *ip_a1 = ip_a0 + num_col_a; + + /* Init accumulator with bias for channel N and N + 1 */ + q31_t ch_0_out_0 = *bias; + q31_t ch_0_out_1 = *bias++; + q31_t ch_1_out_0 = *bias; + q31_t ch_1_out_1 = *bias++; + + uint16_t col_count = num_col_a / 4; + /* accumulate over the vector */ + while (col_count) + { + q31_t a01, a02, a11, a12; + q31_t b0 = arm_nn_read_q15x2_ia(&ip_b0); + q31_t b1 = arm_nn_read_q15x2_ia(&ip_b1); + + ip_a0 = read_and_pad_reordered(ip_a0, &a01, &a02); + ip_a1 = read_and_pad_reordered(ip_a1, &a11, &a12); + + ch_0_out_0 = __SMLAD(a01, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a01, b1, ch_0_out_1); + ch_1_out_0 = __SMLAD(a11, b0, ch_1_out_0); + ch_1_out_1 = __SMLAD(a11, b1, ch_1_out_1); + + b0 = arm_nn_read_q15x2_ia(&ip_b0); + b1 = arm_nn_read_q15x2_ia(&ip_b1); + + ch_0_out_0 = __SMLAD(a02, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a02, b1, ch_0_out_1); + ch_1_out_0 = __SMLAD(a12, b0, ch_1_out_0); + ch_1_out_1 = __SMLAD(a12, b1, ch_1_out_1); + + col_count--; + } /* while over col_count */ + + ch_0_out_0 = arm_nn_requantize(ch_0_out_0, *out_mult, *out_shift); + ch_0_out_0 += out_offset; + ch_0_out_0 = MAX(ch_0_out_0, activation_min); + ch_0_out_0 = MIN(ch_0_out_0, activation_max); + *out_0++ = (q7_t)ch_0_out_0; + + ch_0_out_1 = arm_nn_requantize(ch_0_out_1, *out_mult, *out_shift); + ch_0_out_1 += out_offset; + ch_0_out_1 = MAX(ch_0_out_1, activation_min); + ch_0_out_1 = MIN(ch_0_out_1, activation_max); + *out_1++ = (q7_t)ch_0_out_1; + out_mult++; + out_shift++; + + ch_1_out_0 = arm_nn_requantize(ch_1_out_0, *out_mult, *out_shift); + ch_1_out_0 += out_offset; + ch_1_out_0 = MAX(ch_1_out_0, activation_min); + ch_1_out_0 = MIN(ch_1_out_0, activation_max); + *out_0++ = (q7_t)ch_1_out_0; + + ch_1_out_1 = arm_nn_requantize(ch_1_out_1, *out_mult, *out_shift); + ch_1_out_1 += out_offset; + ch_1_out_1 = MAX(ch_1_out_1, activation_min); + ch_1_out_1 = MIN(ch_1_out_1, activation_max); + *out_1++ = (q7_t)ch_1_out_1; + out_mult++; + out_shift++; + + /* skip row */ + ip_a0 += num_col_a; + row_count--; + } + + if (output_ch & 1) + { + /* setup pointers for B */ + const q15_t *ip_b0 = input_b; + const q15_t *ip_b1 = ip_b0 + num_col_a; + + /* Init accumulator with bias for channel N + 1 */ + q31_t ch_0_out_0 = *bias; + q31_t ch_0_out_1 = ch_0_out_0; + + int32_t col_count = num_col_a / 4; + while (col_count) + { + q31_t a01, a02; + q31_t b0 = arm_nn_read_q15x2_ia(&ip_b0); + q31_t b1 = arm_nn_read_q15x2_ia(&ip_b1); + + ip_a0 = read_and_pad_reordered(ip_a0, &a01, &a02); + + ch_0_out_0 = __SMLAD(a01, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a01, b1, ch_0_out_1); + + b0 = arm_nn_read_q15x2_ia(&ip_b0); + b1 = arm_nn_read_q15x2_ia(&ip_b1); + + ch_0_out_0 = __SMLAD(a02, b0, ch_0_out_0); + ch_0_out_1 = __SMLAD(a02, b1, ch_0_out_1); + + col_count--; + } /* while over col_count */ + + ch_0_out_0 = arm_nn_requantize(ch_0_out_0, *out_mult, *out_shift); + ch_0_out_0 += out_offset; + ch_0_out_0 = MAX(ch_0_out_0, activation_min); + ch_0_out_0 = MIN(ch_0_out_0, activation_max); + *out_0++ = (q7_t)ch_0_out_0; + + ch_0_out_1 = arm_nn_requantize(ch_0_out_1, *out_mult, *out_shift); + ch_0_out_1 += out_offset; + ch_0_out_1 = MAX(ch_0_out_1, activation_min); + ch_0_out_1 = MIN(ch_0_out_1, activation_max); + *out_1++ = (q7_t)ch_0_out_1; + } + + out_0 += output_ch; + + /* return the new output pointer with offset */ + return out_0; +#else + (void)input_a; + (void)input_b; + (void)output_ch; + (void)out_shift; + (void)out_mult; + (void)out_offset; + (void)activation_min; + (void)activation_max; + (void)num_col_a; + (void)output_bias; + (void)out_0; + /* To be completed */ + return NULL; +#endif +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_s8.c new file mode 100644 index 0000000..cabfe44 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_s8.c @@ -0,0 +1,184 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mult_s8.c + * Description: General Matrix-multiplication function + * + * $Date: July 27, 2020 + * $Revision: V.2.0.4 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/* + * s8 General matrix multiplication function with per-channel requantization for upto 4 column batches. + * + * Refer header file for details. + * + */ + +q7_t *arm_nn_mat_mult_s8(const q7_t *input_row, + const q7_t *input_col, + const uint16_t output_ch, + const uint16_t col_batches, + const int32_t *output_shift, + const int32_t *output_mult, + const int32_t out_offset, + const int32_t col_offset, + const int32_t row_offset, + const int16_t activation_min, + const int16_t activation_max, + const uint16_t row_len, + const int32_t *const bias, + q7_t *out) +{ +#if defined(ARM_MATH_MVEI) + (void)row_offset; + if (col_batches == 4) + { + for (int i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) + { + int32_t row_len_tmp = row_len; + const int8_t *ip_r0 = input_row + (i_out_ch * row_len); + const int8_t *ip_c0 = input_col; + const int8_t *ip_c1 = input_col + row_len; + const int8_t *ip_c2 = input_col + (2 * row_len); + const int8_t *ip_c3 = input_col + (3 * row_len); + + int32_t acc_0 = 0; + int32_t acc_1 = 0; + int32_t acc_2 = 0; + int32_t acc_3 = 0; + const int32_t row_loop_cnt = (row_len + 7) / 8; + + for (int i_row_loop = 0; i_row_loop < row_loop_cnt; i_row_loop++) + { + mve_pred16_t p = vctp16q((uint32_t)row_len_tmp); + const int16x8_t offset = vdupq_m_n_s16(vuninitializedq_s16(), col_offset, p); + row_len_tmp -= 8; + + int16x8_t r0 = vldrbq_z_s16(ip_r0, p); + ip_r0 += 8; + + int16x8_t c0 = vldrbq_z_s16(ip_c0, p); + ip_c0 += 8; + c0 = vaddq_m_s16(vuninitializedq_s16(), c0, offset, p); + + int16x8_t c1 = vldrbq_z_s16(ip_c1, p); + ip_c1 += 8; + c1 = vaddq_m_s16(vuninitializedq_s16(), c1, offset, p); + + int16x8_t c2 = vldrbq_z_s16(ip_c2, p); + ip_c2 += 8; + c2 = vaddq_m_s16(vuninitializedq_s16(), c2, offset, p); + + int16x8_t c3 = vldrbq_z_s16(ip_c3, p); + ip_c3 += 8; + c3 = vaddq_m_s16(vuninitializedq_s16(), c3, offset, p); + + acc_0 = vmladavaq_p_s16(acc_0, r0, c0, p); + acc_1 = vmladavaq_p_s16(acc_1, r0, c1, p); + acc_2 = vmladavaq_p_s16(acc_2, r0, c2, p); + acc_3 = vmladavaq_p_s16(acc_3, r0, c3, p); + } + + int32x4_t res = {acc_0, acc_1, acc_2, acc_3}; + if (bias) + { + res = vaddq_n_s32(res, bias[i_out_ch]); + } + res = arm_requantize_mve(res, output_mult[i_out_ch], output_shift[i_out_ch]); + res = vaddq_n_s32(res, out_offset); + + res = vmaxq_s32(res, vdupq_n_s32(activation_min)); + res = vminq_s32(res, vdupq_n_s32(activation_max)); + + const uint32x4_t scatter_offset = {0, output_ch, output_ch * 2, output_ch * 3}; + vstrbq_scatter_offset_s32(&out[i_out_ch], scatter_offset, res); + } + out += 4 * output_ch; + } + else + { + for (int i_col_batch = (col_batches & ~0x3); i_col_batch < (col_batches & 0x3); i_col_batch++) + { + for (int i_out_ch = 0; i_out_ch < output_ch; i_out_ch++) + { + int32_t row_len_tmp = row_len; + + const int8_t *ip_r0 = input_row + (i_out_ch * row_len); + const int8_t *ip_c0 = input_col + (i_col_batch * row_len); + int32_t acc_0 = 0; + const int32_t row_loop_cnt = (row_len + 7) / 8; + + for (int i_row_loop = 0; i_row_loop < row_loop_cnt; i_row_loop++) + { + const mve_pred16_t p = vctp16q((uint32_t)row_len_tmp); + const int16x8_t offset = vdupq_m_n_s16(vuninitializedq_s16(), col_offset, p); + row_len_tmp -= 8; + + int16x8_t r0 = vldrbq_z_s16(ip_r0, p); + ip_r0 += 8; + int16x8_t c0 = vldrbq_z_s16(ip_c0, p); + ip_c0 += 8; + + c0 = vaddq_m_s16(vuninitializedq_s16(), c0, offset, p); + acc_0 = vmladavaq_p_s16(acc_0, r0, c0, p); + } + + if (bias) + { + acc_0 += bias[i_out_ch]; + } + acc_0 = arm_nn_requantize(acc_0, output_mult[i_out_ch], output_shift[i_out_ch]); + acc_0 += out_offset; + acc_0 = MAX(acc_0, activation_min); + acc_0 = MIN(acc_0, activation_max); + out[i_out_ch] = (q7_t)acc_0; + } + out += output_ch; + } + } + return out; + +#else + (void)input_row; + (void)input_col; + (void)output_ch; + (void)col_batches; + (void)output_shift; + (void)output_mult; + (void)out_offset; + (void)col_offset; + (void)row_offset; + (void)activation_min; + (void)activation_max; + (void)row_len; + (void)bias; + (void)out; + return NULL; +#endif +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c new file mode 100644 index 0000000..b489632 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c @@ -0,0 +1,202 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_fully_connected_mat_q7_vec_q15.c + * Description: Mixed Q15-Q7 fully-connected layer function + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup FC + * @{ + */ + + /** + * @brief Mixed Q15-Q7 fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * + * Buffer size: + * + * vec_buffer size: 0 + * + * Q7_Q15 version of the fully connected layer + * + * Weights are in q7_t and Activations are in q15_t + * + */ + +arm_status +arm_fully_connected_mat_q7_vec_q15(const q15_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q15_t * pOut, + q15_t * vec_buffer) +{ + (void)vec_buffer; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + const q7_t *pB = pM; + const q7_t *pB2; + q15_t *pO = pOut; + const q7_t *pBias = bias; + const q15_t *pA = pV; + + uint16_t rowCnt = num_of_rows >> 1; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + uint16_t colCnt = dim_vec >> 2; + + pA = pV; + pB2 = pB + dim_vec; + + while (colCnt) + { + q31_t inV, inM11, inM12, inM21, inM22; + pB = read_and_pad(pB, &inM11, &inM12); + pB2 = read_and_pad(pB2, &inM21, &inM22); + + inV = arm_nn_read_q15x2_ia(&pA); + + sum = __SMLAD(inV, inM11, sum); + sum2 = __SMLAD(inV, inM21, sum2); + + inV = arm_nn_read_q15x2_ia(&pA); + + sum = __SMLAD(inV, inM12, sum); + sum2 = __SMLAD(inV, inM22, sum2); + + colCnt--; + } + colCnt = dim_vec & 0x3; + while (colCnt) + { + q15_t inV = *pA++; + q7_t inM = *pB++; + q7_t inM2 = *pB2++; + + sum += inV * inM; + sum2 += inV * inM2; + colCnt--; + } /* while over colCnt */ + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum2 >> out_shift), 16)); + + /*adjust the pointers and counters */ + pB += dim_vec; + rowCnt--; + } + + /* left-over part of the rows */ + rowCnt = num_of_rows & 0x1; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + uint16_t colCnt = dim_vec >> 2; + + pA = pV; + + while (colCnt) + { + q31_t inV1, inV2, inM11, inM12; + + pB = read_and_pad(pB, &inM11, &inM12); + + inV1 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV1, inM11, sum); + + inV2 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV2, inM12, sum); + + colCnt--; + } + + /* left-over of the vector */ + colCnt = dim_vec & 0x3; + while (colCnt) + { + q15_t inV = *pA++; + q7_t inM = *pB++; + sum += inV * inM; + colCnt--; + } + + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + + rowCnt--; + } + +#else + int i, j; + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + for (i = 0; i < num_of_rows; i++) + { + int ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + for (j = 0; j < dim_vec; j++) + { + ip_out += pV[j] * pM[i * dim_vec + j]; + } + pOut[i] = (q15_t) __SSAT((ip_out >> out_shift), 16); + } + +#endif /* ARM_MATH_DSP */ + + /* Return to ARM_MATH_SUCCESS */ + return (ARM_MATH_SUCCESS); + +} + +/** + * @} end of FC group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c new file mode 100644 index 0000000..32d3fb6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c @@ -0,0 +1,407 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_fully_connected_mat_q7_vec_q15_opt.c + * Description: Mixed Q15-Q7 opt fully-connected layer function + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup FC + * @{ + */ + + /** + * @brief Mixed Q15-Q7 opt fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * + * Buffer size: + * + * vec_buffer size: 0 + * + * Q7_Q15 version of the fully connected layer + * + * Weights are in q7_t and Activations are in q15_t + * + * Limitation: x4 version requires weight reordering to work + * + * Here we use only one pointer to read 4 rows in the weight + * matrix. So if the original q7_t matrix looks like this: + * + * | a11 | a12 | a13 | a14 | a15 | a16 | a17 | + * + * | a21 | a22 | a23 | a24 | a25 | a26 | a27 | + * + * | a31 | a32 | a33 | a34 | a35 | a36 | a37 | + * + * | a41 | a42 | a43 | a44 | a45 | a46 | a47 | + * + * | a51 | a52 | a53 | a54 | a55 | a56 | a57 | + * + * | a61 | a62 | a63 | a64 | a65 | a66 | a67 | + * + * We operates on multiple-of-4 rows, so the first four rows becomes + * + * | a11 | a21 | a12 | a22 | a31 | a41 | a32 | a42 | + * + * | a13 | a23 | a14 | a24 | a33 | a43 | a34 | a44 | + * + * | a15 | a25 | a16 | a26 | a35 | a45 | a36 | a46 | + * + * The column left over will be in-order. + * which is: + * | a17 | a27 | a37 | a47 | + * + * For the left-over rows, we do 1x1 computation, so the data remains + * as its original order. + * + * So the stored weight matrix looks like this: + * + * | a11 | a21 | a12 | a22 | a31 | a41 | + * + * | a32 | a42 | a13 | a23 | a14 | a24 | + * + * | a33 | a43 | a34 | a44 | a15 | a25 | + * + * | a16 | a26 | a35 | a45 | a36 | a46 | + * + * | a17 | a27 | a37 | a47 | a51 | a52 | + * + * | a53 | a54 | a55 | a56 | a57 | a61 | + * + * | a62 | a63 | a64 | a65 | a66 | a67 | + * + */ + +arm_status +arm_fully_connected_mat_q7_vec_q15_opt(const q15_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, const q7_t * bias, q15_t * pOut, q15_t * vec_buffer) +{ + + (void)vec_buffer; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + const q7_t *pB = pM; + q15_t *pO = pOut; + const q7_t *pBias = bias; + const q15_t *pA = pV; + + uint16_t rowCnt = num_of_rows >> 2; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 1; + + pA = pV; + +#ifdef USE_INTRINSIC + +#ifndef ARM_MATH_BIG_ENDIAN + + while (colCnt) + { + q31_t inM11, inM12, inM13, inM14; + q31_t inV; + + inV = arm_nn_read_q15x2_ia(&pA); + inM11 = arm_nn_read_q7x4_ia(&pB); + inM12 = __SXTB16(__ROR(inM11, 8)); + inM11 = __SXTB16(inM11); + sum = __SMLAD(inM11, inV, sum); + sum2 = __SMLAD(inM12, inV, sum2); + inM13 = arm_nn_read_q7x4_ia(&pB); + inM14 = __SXTB16(__ROR(inM13, 8)); + inM13 = __SXTB16(inM13); + sum3 = __SMLAD(inM13, inV, sum3); + sum4 = __SMLAD(inM14, inV, sum4); + colCnt--; + } + +#else + + while (colCnt) + { + q31_t inM11, inM12, inM13, inM14; + q31_t inV; + + inV = *__SIMD32(pA)++; + inM11 = arm_nn_read_q7x4_ia(&pB); + inM12 = __SXTB16(__ROR(inM11, 8)); + inM11 = __SXTB16(inM11); + sum = __SMLAD(inM12, inV, sum); + sum2 = __SMLAD(inM11, inV, sum2); + inM13 = arm_nn_read_q7x4_ia(&pB); + inM14 = __SXTB16(__ROR(inM13, 8)); + inM13 = __SXTB16(inM13); + sum3 = __SMLAD(inM14, inV, sum3); + sum4 = __SMLAD(inM13, inV, sum4); + colCnt--; + } + +#endif /* ARM_MATH_BIG_ENDIAN */ + +#else + + /* + * register needed: + * loop counter: colCnt + * accumulators: sum, sum2, sum3, sum4 + * pointers: pB, pA + * weight data: inM11, inM12, inM13, inM14 + * activation data: inV + */ + +#ifndef ARM_MATH_BIG_ENDIAN + asm volatile ("COL_LOOP_%=:\n" + "ldr.w r4, [%[pA]], #4\n" + "ldr.w r1, [%[pB]], #8\n" + "mov.w r0, r1, ror #8\n" + "sxtb16 r0, r0\n" + "sxtb16 r1, r1\n" + "smlad %[sum], r4, r1, %[sum]\n" + "smlad %[sum2], r4, r0, %[sum2]\n" + "ldr.w r3, [%[pB], #-4]\n" + "mov.w r2, r3, ror #8\n" + "sxtb16 r2, r2\n" + "sxtb16 r3, r3\n" + "smlad %[sum3], r4, r3, %[sum3]\n" + "smlad %[sum4], r4, r2, %[sum4]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP_%=\n":[sum] "+r"(sum), + [sum2] "+r"(sum2),[sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4"); +#else + asm volatile ("COL_LOOP_%=:\n" + "ldr.w r4, [%[pA]], #4\n" + "ldr.w r1, [%[pB]], #8\n" + "mov.w r0, r1, ror #8\n" + "sxtb16 r0, r0\n" + "sxtb16 r1, r1\n" + "smlad %[sum], r4, r0, %[sum]\n" + "smlad %[sum2], r4, r1, %[sum2]\n" + "ldr.w r3, [%[pB], #-4]\n" + "mov.w r2, r3, ror #8\n" + "sxtb16 r2, r2\n" + "sxtb16 r3, r3\n" + "smlad %[sum3], r4, r2, %[sum3]\n" + "smlad %[sum4], r4, r3, %[sum4]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP_%=\n":[sum] "+r"(sum), + [sum2] "+r"(sum2),[sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4"); +#endif /* ARM_MATH_BIG_ENDIAN */ + +#endif /* USE_INTRINSIC */ + + colCnt = dim_vec & 0x1; + while (colCnt) + { + q15_t inV = *pA++; + q7_t inM = *pB++; + q7_t inM2 = *pB++; + q7_t inM3 = *pB++; + q7_t inM4 = *pB++; + + sum += inV * inM; + sum2 += inV * inM2; + sum3 += inV * inM3; + sum4 += inV * inM4; + colCnt--; + } /* while over colCnt */ + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum2 >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum3 >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum4 >> out_shift), 16)); + + /* adjust the pointers and counters */ + rowCnt--; + } + + /* left-over part of the rows */ + rowCnt = num_of_rows & 0x3; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 2; + + pA = pV; + + while (colCnt) + { + q31_t inV1, inV2, inM11, inM12; + + pB = read_and_pad(pB, &inM11, &inM12); + + inV1 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV1, inM11, sum); + + inV2 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV2, inM12, sum); + + colCnt--; + } + + /* left-over of the vector */ + colCnt = dim_vec & 0x3; + while (colCnt) + { + q15_t inV = *pA++; + q7_t inM = *pB++; + sum += inV * inM; + colCnt--; + } + + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + + rowCnt--; + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + uint16_t rowCnt = num_of_rows >> 2; + const q7_t *pB = pM; + const q15_t *pA; + q15_t *pO = pOut; + const q7_t *pBias = bias; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + uint16_t colCnt = dim_vec >> 1; + + pA = pV; + + while (colCnt) + { + q15_t inA1 = *pA++; + q15_t inA2 = *pA++; + + q7_t inB1 = *pB++; + q7_t inB3 = *pB++; + q7_t inB2 = *pB++; + q7_t inB4 = *pB++; + + sum += inA1 * inB1 + inA2 * inB2; + sum2 += inA1 * inB3 + inA2 * inB4; + + inB1 = *pB++; + inB3 = *pB++; + inB2 = *pB++; + inB4 = *pB++; + + sum3 += inA1 * inB1 + inA2 * inB2; + sum4 += inA1 * inB3 + inA2 * inB4; + + colCnt--; + } + + colCnt = dim_vec & 0x1; + while (colCnt) + { + q15_t inA = *pA++; + q7_t inB = *pB++; + sum += inA * inB; + inB = *pB++; + sum2 += inA * inB; + inB = *pB++; + sum3 += inA * inB; + inB = *pB++; + sum4 += inA * inB; + + colCnt--; + } + *pO++ = (q15_t) __SSAT((sum >> out_shift), 16); + *pO++ = (q15_t) __SSAT((sum2 >> out_shift), 16); + *pO++ = (q15_t) __SSAT((sum3 >> out_shift), 16); + *pO++ = (q15_t) __SSAT((sum4 >> out_shift), 16); + + rowCnt--; + } + + rowCnt = num_of_rows & 0x3; + + while (rowCnt) + { + int ip_out = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + int j; + + pA = pV; + for (j = 0; j < dim_vec; j++) + { + q15_t inA = *pA++; + q7_t inB = *pB++; + ip_out += inA * inB; + } + *pO++ = (q15_t) __SSAT((ip_out >> out_shift), 16); + + rowCnt--; + } + +#endif /* ARM_MATH_DSP */ + + /* Return to ARM_MATH_SUCCESS */ + return (ARM_MATH_SUCCESS); + +} + +/** + * @} end of FC group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c new file mode 100644 index 0000000..04dc076 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c @@ -0,0 +1,196 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_fully_connected_q15.c + * Description: Q15 basic fully-connected layer function + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup FC + * @{ + */ + + /** + * @brief Q15 opt fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + * + * @details + * + * Buffer size: + * + * vec_buffer size: 0 + * + */ + +arm_status +arm_fully_connected_q15(const q15_t * pV, + const q15_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q15_t * bias, + q15_t * pOut, + q15_t * vec_buffer) +{ + (void)vec_buffer; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + const q15_t *pB = pM; + const q15_t *pB2 = pB + dim_vec; + q15_t *pO = pOut; + const q15_t *pA; + const q15_t *pBias = bias; + uint16_t rowCnt = num_of_rows >> 1; + + /* this loop loops over different output */ + while (rowCnt) { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 2; + + pA = pV; + pB2 = pB + dim_vec; + + while (colCnt) + { + q31_t inV1, inM1, inM2; + inV1 = arm_nn_read_q15x2_ia(&pA); + inM1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inV1, inM1, sum); + inM2 = arm_nn_read_q15x2_ia(&pB2); + sum2 = __SMLAD(inV1, inM2, sum2); + + inV1 = arm_nn_read_q15x2_ia(&pA); + inM1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inV1, inM1, sum); + inM2 = arm_nn_read_q15x2_ia(&pB2); + sum2 = __SMLAD(inV1, inM2, sum2); + + colCnt--; + } + colCnt = dim_vec & 0x3; + while (colCnt) + { + q15_t inV = *pA++; + q15_t inM = *pB++; + q15_t inM2 = *pB2++; + + sum += inV * inM; + sum2 += inV * inM2; + colCnt--; + } /* while over colCnt */ + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum2>> out_shift), 16)); + + /* adjust the pointers and counters */ + pB = pB + dim_vec; + rowCnt --; + } + + rowCnt = num_of_rows & 0x1; + + while (rowCnt) { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 2; + + pA = pV; + + while (colCnt) { + q31_t inV1, inM1; + inV1 = arm_nn_read_q15x2_ia(&pA); + inM1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inV1, inM1, sum); + + inV1 = arm_nn_read_q15x2_ia(&pA); + inM1 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inV1, inM1, sum); + + colCnt--; + } + + /* left-over of the vector */ + colCnt = dim_vec & 0x3; + while(colCnt) { + q15_t inV = *pA++; + q15_t inM = *pB++; + + sum += inV * inM; + + colCnt--; + } + + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + + rowCnt --; + } + +#else + int i, j; + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + for (i = 0; i < num_of_rows; i++) + { + int ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + for (j = 0; j < dim_vec; j++) + { + ip_out += pV[j] * pM[i * dim_vec + j]; + } + pOut[i] = (q15_t) __SSAT((ip_out >> out_shift), 16); + } + +#endif /* ARM_MATH_DSP */ + + /* Return to application */ + return (ARM_MATH_SUCCESS); + +} + +/** + * @} end of FC group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c new file mode 100644 index 0000000..9fd44da --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c @@ -0,0 +1,335 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_fully_connected_q15_opt.c + * Description: Q15 opt fully-connected layer function + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup FC + * @{ + */ + + /** + * @brief Q15 opt fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + * + * @details + * + * Buffer size: + * + * vec_buffer size: 0 + * + * Here we use only one pointer to read 4 rows in the weight + * matrix. So if the original matrix looks like this: + * + * | a11 | a12 | a13 | + * + * | a21 | a22 | a23 | + * + * | a31 | a32 | a33 | + * + * | a41 | a42 | a43 | + * + * | a51 | a52 | a53 | + * + * | a61 | a62 | a63 | + * + * We operates on multiple-of-4 rows, so the first four rows becomes + * + * | a11 | a12 | a21 | a22 | a31 | a32 | a41 | a42 | + * + * | a13 | a23 | a33 | a43 | + * + * Remaining rows are kept the same original order. + * + * So the stored weight matrix looks like this: + * + * + * | a11 | a12 | a21 | a22 | a31 | a32 | a41 | a42 | + * + * | a13 | a23 | a33 | a43 | a51 | a52 | a53 | a61 | + * + * | a62 | a63 | + */ + +arm_status +arm_fully_connected_q15_opt(const q15_t * pV, + const q15_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q15_t * bias, + q15_t * pOut, + q15_t * vec_buffer) +{ + (void)vec_buffer; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + const q15_t *pB = pM; + q15_t *pO = pOut; + const q15_t *pBias = bias; + const q15_t *pA = pV; + + uint16_t rowCnt = num_of_rows >> 2; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 1; + + pA = pV; + +#ifdef USE_INTRINSIC + + while (colCnt) + { + q31_t inM11, inM12, inM13, inM14; + q31_t inV; + + inV = arm_nn_read_q15x2_ia(&pA); + inM11 = arm_nn_read_q15x2_ia(&pB); + sum = __SMLAD(inV, inM11, sum); + inM12 = arm_nn_read_q15x2_ia(&pB); + sum2 = __SMLAD(inV, inM12, sum2); + inM13 = arm_nn_read_q15x2_ia(&pB); + sum3 = __SMLAD(inV, inM13, sum3); + inM14 = arm_nn_read_q15x2_ia(&pB); + sum4 = __SMLAD(inV, inM14, sum4); + colCnt--; + } + +#else + + /* + * register needed: + * loop counter: colCnt + * accumulators: sum, sum2, sum3, sum4 + * pointers: pB, pA + * weight data: inM11, inM12, inM13, inM14 + * activation data: inV + */ + + asm volatile ("COL_LOOP_%=:\n" + "ldr.w r4, [%[pA]], #4\n" + "ldr.w r0, [%[pB]], #16\n" + "smlad %[sum], r4, r0, %[sum]\n" + "ldr.w r1, [%[pB] , #-12]\n" + "smlad %[sum2], r4, r1, %[sum2]\n" + "ldr.w r2, [%[pB] , #-8]\n" + "smlad %[sum3], r4, r2, %[sum3]\n" + "ldr.w r3, [%[pB] , #-4]\n" + "smlad %[sum4], r4, r3, %[sum4]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP_%=\n":[sum] "+r"(sum), + [sum2] "+r"(sum2),[sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4"); + +#endif /* USE_INTRINSIC */ + + colCnt = dim_vec & 0x1; + while (colCnt) + { + + q15_t inV = *pA++; + q15_t inM = *pB++; + q15_t inM2 = *pB++; + q15_t inM3 = *pB++; + q15_t inM4 = *pB++; + + sum += inV * inM; + sum2 += inV * inM2; + sum3 += inV * inM3; + sum4 += inV * inM4; + colCnt--; + } /* while over colCnt */ + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum2 >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum3 >> out_shift), 16)); + *pO++ = (q15_t) (__SSAT((sum4 >> out_shift), 16)); + + /* adjust the pointers and counters */ + rowCnt--; + } + + /* left-over part of the rows */ + rowCnt = num_of_rows & 0x3; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 2; + + pA = pV; + + while (colCnt) + { + q31_t inV1, inV2, inM1, inM2; + + inM1 = arm_nn_read_q15x2_ia(&pB); + inV1 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV1, inM1, sum); + + inM2 = arm_nn_read_q15x2_ia(&pB); + inV2 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV2, inM2, sum); + + colCnt--; + } + + /* left-over of the vector */ + colCnt = dim_vec & 0x3; + while (colCnt) + { + q15_t inV = *pA++; + q15_t inM = *pB++; + sum += inV * inM; + colCnt--; + } + + *pO++ = (q15_t) (__SSAT((sum >> out_shift), 16)); + + rowCnt--; + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + uint16_t rowCnt = num_of_rows >> 2; + const q15_t *pB = pM; + const q15_t *pA; + q15_t *pO = pOut; + const q15_t *pBias = bias; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 1; + + pA = pV; + while (colCnt) + { + q15_t inA1 = *pA++; + q15_t inA2 = *pA++; + + q15_t inB1 = *pB++; + q15_t inB2 = *pB++; + sum += inA1 * inB1 + inA2 * inB2; + + inB1 = *pB++; + inB2 = *pB++; + sum2 += inA1 * inB1 + inA2 * inB2; + + inB1 = *pB++; + inB2 = *pB++; + sum3 += inA1 * inB1 + inA2 * inB2; + + inB1 = *pB++; + inB2 = *pB++; + sum4 += inA1 * inB1 + inA2 * inB2; + + colCnt--; + } + colCnt = dim_vec & 0x1; + while (colCnt) + { + q15_t inA = *pA++; + q15_t inB = *pB++; + sum += inA * inB; + inB = *pB++; + sum2 += inA * inB; + inB = *pB++; + sum3 += inA * inB; + inB = *pB++; + sum4 += inA * inB; + colCnt--; + } + *pO++ = (q15_t) __SSAT((sum >> out_shift), 16); + *pO++ = (q15_t) __SSAT((sum2 >> out_shift), 16); + *pO++ = (q15_t) __SSAT((sum3 >> out_shift), 16); + *pO++ = (q15_t) __SSAT((sum4 >> out_shift), 16); + + rowCnt--; + } + rowCnt = num_of_rows & 0x3; + + while (rowCnt) + { + int ip_out = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + int j; + + pA = pV; + for (j = 0; j < dim_vec; j++) + { + q15_t inA = *pA++; + q15_t inB = *pB++; + ip_out += inA * inB; + } + *pO++ = (q15_t) __SSAT((ip_out >> out_shift), 16); + + rowCnt--; + } + +#endif /* ARM_MATH_DSP */ + + /* Return to ARM_MATH_SUCCESS */ + return (ARM_MATH_SUCCESS); + +} + +/** + * @} end of FC group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c new file mode 100644 index 0000000..d3feb08 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c @@ -0,0 +1,201 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_fully_connected_q7.c + * Description: Q7 basic fully-connected layer function + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup FC + * @{ + */ + + /** + * @brief Q7 basic fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * + * Buffer size: + * + * vec_buffer size: dim_vec + * + * This basic function is designed to work with regular weight + * matrix without interleaving. + * + */ + +arm_status +arm_fully_connected_q7(const q7_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, const q7_t * bias, q7_t * pOut, q15_t * vec_buffer) +{ + +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + const q7_t *pB = pM; + const q7_t *pB2; + q7_t *pO = pOut; + const q7_t *pBias = bias; + const q15_t *pA; + uint16_t rowCnt = num_of_rows >> 1; + + /* expand the vector into the buffer */ + arm_q7_to_q15_reordered_no_shift(pV, vec_buffer, dim_vec); + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + uint16_t colCnt = dim_vec >> 2; + + pA = vec_buffer; + pB2 = pB + dim_vec; + + while (colCnt) + { + q31_t inV, inM11, inM12, inM21, inM22; + pB = read_and_pad_reordered(pB, &inM11, &inM12); + pB2 = read_and_pad_reordered(pB2, &inM21, &inM22); + + inV = arm_nn_read_q15x2_ia(&pA); + + sum = __SMLAD(inV, inM11, sum); + sum2 = __SMLAD(inV, inM21, sum2); + + inV = arm_nn_read_q15x2_ia(&pA); + + sum = __SMLAD(inV, inM12, sum); + sum2 = __SMLAD(inV, inM22, sum2); + + colCnt--; + } + colCnt = dim_vec & 0x3; + while (colCnt) + { + q7_t inV = *pA++; + q15_t inM = *pB++; + q15_t inM2 = *pB2++; + + sum += inV * inM; + sum2 += inV * inM2; + colCnt--; + } /* while over colCnt */ + *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8)); + *pO++ = (q7_t) (__SSAT((sum2 >> out_shift), 8)); + + /* adjust the pointers and counters */ + pB += dim_vec; + rowCnt--; + } + + /* left-over part of the rows */ + rowCnt = num_of_rows & 0x1; + + while (rowCnt) + { + uint16_t colCnt = dim_vec >> 2; + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + pA = vec_buffer; + + while (colCnt) + { + q31_t inV1, inV2, inM11, inM12; + + pB = read_and_pad_reordered(pB, &inM11, &inM12); + + inV1 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV1, inM11, sum); + + inV2 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV2, inM12, sum); + + colCnt--; + } + + /* left-over of the vector */ + colCnt = dim_vec & 0x3; + while (colCnt) + { + q7_t inV = *pA++; + q15_t inM = *pB++; + sum += inV * inM; + colCnt--; + } + + *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8)); + + rowCnt--; + } + +#else + int i, j; + + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + for (i = 0; i < num_of_rows; i++) + { + int ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift); + for (j = 0; j < dim_vec; j++) + { + ip_out += pV[j] * pM[i * dim_vec + j]; + } + pOut[i] = (q7_t) __SSAT((ip_out >> out_shift), 8); + } + +#endif /* ARM_MATH_DSP */ + + /* Return to ARM_MATH_SUCCESS */ + return (ARM_MATH_SUCCESS); + +} + +/** + * @} end of FC group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c new file mode 100644 index 0000000..6d868c6 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c @@ -0,0 +1,487 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_fully_connected_q7_opt.c + * Description: Q7 basic fully-connected layer function + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup FC + * @{ + */ + + /** + * @brief Q7 opt fully-connected layer function + * @param[in] pV pointer to input vector + * @param[in] pM pointer to matrix weights + * @param[in] dim_vec length of the vector + * @param[in] num_of_rows number of rows in weight matrix + * @param[in] bias_shift amount of left-shift for bias + * @param[in] out_shift amount of right-shift for output + * @param[in] bias pointer to bias + * @param[in,out] pOut pointer to output vector + * @param[in,out] vec_buffer pointer to buffer space for input + * @return The function returns ARM_MATH_SUCCESS + * + * @details + * + * Buffer size: + * + * vec_buffer size: dim_vec + * + * This opt function is designed to work with interleaved weight + * matrix. The vector input is assumed in q7_t format, we call + * arm_q7_to_q15_no_shift_shuffle function to expand into + * q15_t format with certain weight re-ordering, refer to the function + * comments for more details. + * Here we use only one pointer to read 4 rows in the weight + * matrix. So if the original q7_t matrix looks like this: + * + * | a11 | a12 | a13 | a14 | a15 | a16 | a17 | + * + * | a21 | a22 | a23 | a24 | a25 | a26 | a27 | + * + * | a31 | a32 | a33 | a34 | a35 | a36 | a37 | + * + * | a41 | a42 | a43 | a44 | a45 | a46 | a47 | + * + * | a51 | a52 | a53 | a54 | a55 | a56 | a57 | + * + * | a61 | a62 | a63 | a64 | a65 | a66 | a67 | + * + * + * We operates on multiple-of-4 rows, so the first four rows becomes + * + * | a11 | a21 | a13 | a23 | a31 | a41 | a33 | a43 | + * + * | a12 | a22 | a14 | a24 | a32 | a42 | a34 | a44 | + * + * | a15 | a25 | a35 | a45 | a16 | a26 | a36 | a46 | + * + * So within the kernel, we first read the re-ordered vector in as: + * + * | b1 | b3 | and | b2 | b4 | + * + * the four q31_t weights will look like + * + * | a11 | a13 |, | a21 | a23 |, | a31 | a33 |, | a41 | a43 | + * + * | a12 | a14 |, | a22 | a24 |, | a32 | a34 |, | a42 | a44 | + * + * The column left over will be in-order. + * which is: + * + * | a17 | a27 | a37 | a47 | + * + * For the left-over rows, we do 1x1 computation, so the data remains + * as its original order. + * + * So the stored weight matrix looks like this: + * + * | a11 | a21 | a13 | a23 | a31 | a41 | + * + * | a33 | a43 | a12 | a22 | a14 | a24 | + * + * | a32 | a42 | a34 | a44 | a15 | a25 | + * + * | a35 | a45 | a16 | a26 | a36 | a46 | + * + * | a17 | a27 | a37 | a47 | a51 | a52 | + * + * | a53 | a54 | a55 | a56 | a57 | a61 | + * + * | a62 | a63 | a64 | a65 | a66 | a67 | + * + * + */ + +arm_status +arm_fully_connected_q7_opt(const q7_t * pV, + const q7_t * pM, + const uint16_t dim_vec, + const uint16_t num_of_rows, + const uint16_t bias_shift, + const uint16_t out_shift, + const q7_t * bias, + q7_t * pOut, + q15_t * vec_buffer) +{ + +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + const q7_t *pB = pM; + q7_t *pO = pOut; + const q7_t *pBias = bias; + const q15_t *pA; + uint16_t rowCnt = num_of_rows >> 2; + + arm_q7_to_q15_reordered_no_shift(pV, vec_buffer, dim_vec); + + while (rowCnt) + { + + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 2; + + pA = vec_buffer; + +#ifdef USE_INTRINSIC + +#ifndef ARM_MATH_BIG_ENDIAN + while (colCnt) + { + q31_t inM11, inM12, inM13, inM14; + q31_t inV; + + inV = arm_nn_read_q15x2_ia(&pA); + inM11 = arm_nn_read_q7x4_ia(&pB); + inM12 = __SXTB16(__ROR(inM11, 8)); + inM11 = __SXTB16(inM11); + sum = __SMLAD(inM11, inV, sum); + sum2 = __SMLAD(inM12, inV, sum2); + inM13 = arm_nn_read_q7x4_ia(&pB); + inM14 = __SXTB16(__ROR(inM13, 8)); + inM13 = __SXTB16(inM13); + sum3 = __SMLAD(inM13, inV, sum3); + sum4 = __SMLAD(inM14, inV, sum4); + + inV = arm_nn_read_q15x2_ia(&pA); + inM11 = arm_nn_read_q7x4_ia(&pB); + inM12 = __SXTB16(__ROR(inM11, 8)); + inM11 = __SXTB16(inM11); + sum = __SMLAD(inM11, inV, sum); + sum2 = __SMLAD(inM12, inV, sum2); + inM13 = arm_nn_read_q7x4_ia(&pB); + inM14 = __SXTB16(__ROR(inM13, 8)); + inM13 = __SXTB16(inM13); + sum3 = __SMLAD(inM13, inV, sum3); + sum4 = __SMLAD(inM14, inV, sum4); + colCnt--; + } +#else + while (colCnt) + { + q31_t inM11, inM12, inM13, inM14; + q31_t inV; + + inV = arm_nn_read_q15x2_ia(&pA); + inM11 = arm_nn_read_q7x4_ia(&pB); + inM12 = __SXTB16(__ROR(inM11, 8)); + inM11 = __SXTB16(inM11); + sum = __SMLAD(inM12, inV, sum); + sum2 = __SMLAD(inM11, inV, sum2); + inM13 = arm_nn_read_q7x4_ia(&pB); + inM14 = __SXTB16(__ROR(inM13, 8)); + inM13 = __SXTB16(inM13); + sum3 = __SMLAD(inM14, inV, sum3); + sum4 = __SMLAD(inM13, inV, sum4); + + inV = arm_nn_read_q15x2_ia(&pA); + inM11 = arm_nn_read_q7x4_ia(&pB); + inM12 = __SXTB16(__ROR(inM11, 8)); + inM11 = __SXTB16(inM11); + sum = __SMLAD(inM12, inV, sum); + sum2 = __SMLAD(inM11, inV, sum2); + inM13 = arm_nn_read_q7x4_ia(&pB); + inM14 = __SXTB16(__ROR(inM13, 8)); + inM13 = __SXTB16(inM13); + sum3 = __SMLAD(inM14, inV, sum3); + sum4 = __SMLAD(inM13, inV, sum4); + colCnt--; + } +#endif /* ARM_MATH_BIG_ENDIAN */ + +#else + + /* + * register needed: + * loop counter: colCnt + * accumulators: sum, sum2, sum3, sum4 + * pointers: pB, pA + * weight data: inM11, inM12, inM13, inM14 + * activation data: inV + */ + +#ifndef ARM_MATH_BIG_ENDIAN + asm volatile ("COL_LOOP_%=:\n" + "ldr.w r4, [%[pA]], #8\n" + "ldr.w r1, [%[pB]], #16\n" + "mov.w r0, r1, ror #8\n" + "sxtb16 r0, r0\n" + "sxtb16 r1, r1\n" + "smlad %[sum], r4, r1, %[sum]\n" + "smlad %[sum2], r4, r0, %[sum2]\n" + "ldr.w r3, [%[pB], #-12]\n" + "mov.w r2, r3, ror #8\n" + "sxtb16 r2, r2\n" + "sxtb16 r3, r3\n" + "smlad %[sum3], r4, r3, %[sum3]\n" + "smlad %[sum4], r4, r2, %[sum4]\n" + "ldr.w r4, [%[pA], #-4]\n" + "ldr.w r1, [%[pB], #-8]\n" + "mov.w r0, r1, ror #8\n" + "sxtb16 r0, r0\n" + "sxtb16 r1, r1\n" + "smlad %[sum], r4, r1, %[sum]\n" + "smlad %[sum2], r4, r0, %[sum2]\n" + "ldr.w r3, [%[pB], #-4]\n" + "mov.w r2, r3, ror #8\n" + "sxtb16 r2, r2\n" + "sxtb16 r3, r3\n" + "smlad %[sum3], r4, r3, %[sum3]\n" + "smlad %[sum4], r4, r2, %[sum4]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP_%=\n":[sum] "+r"(sum), + [sum2] "+r"(sum2),[sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4"); +#else + asm volatile ("COL_LOOP_%=:\n" + "ldr.w r4, [%[pA]], #8\n" + "ldr.w r1, [%[pB]], #16\n" + "mov.w r0, r1, ror #8\n" + "sxtb16 r0, r0\n" + "sxtb16 r1, r1\n" + "smlad %[sum], r4, r0, %[sum]\n" + "smlad %[sum2], r4, r1, %[sum2]\n" + "ldr.w r3, [%[pB], #-12]\n" + "mov.w r2, r3, ror #8\n" + "sxtb16 r2, r2\n" + "sxtb16 r3, r3\n" + "smlad %[sum3], r4, r2, %[sum3]\n" + "smlad %[sum4], r4, r3, %[sum4]\n" + "ldr.w r4, [%[pA], #-4]\n" + "ldr.w r1, [%[pB], #-8]\n" + "mov.w r0, r1, ror #8\n" + "sxtb16 r0, r0\n" + "sxtb16 r1, r1\n" + "smlad %[sum], r4, r0, %[sum]\n" + "smlad %[sum2], r4, r1, %[sum2]\n" + "ldr.w r3, [%[pB], #-4]\n" + "mov.w r2, r3, ror #8\n" + "sxtb16 r2, r2\n" + "sxtb16 r3, r3\n" + "smlad %[sum3], r4, r2, %[sum3]\n" + "smlad %[sum4], r4, r3, %[sum4]\n" + "subs %[colCnt], #1\n" + "bne COL_LOOP_%=\n":[sum] "+r"(sum), + [sum2] "+r"(sum2),[sum3] "+r"(sum3), + [sum4] "+r"(sum4),[pB] "+r"(pB),[pA] "+r"(pA):[colCnt] "r"(colCnt):"r0", "r1", "r2", "r3", "r4"); +#endif /* ARM_MATH_BIG_ENDIAN */ + +#endif /* USE_INTRINSIC */ + + colCnt = dim_vec & 0x3; + while (colCnt) + { + q15_t inV = *pA++; + q7_t inM = *pB++; + q7_t inM2 = *pB++; + q7_t inM3 = *pB++; + q7_t inM4 = *pB++; + + sum += inV * inM; + sum2 += inV * inM2; + sum3 += inV * inM3; + sum4 += inV * inM4; + colCnt--; + } /* while over colCnt */ + *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8)); + *pO++ = (q7_t) (__SSAT((sum2 >> out_shift), 8)); + *pO++ = (q7_t) (__SSAT((sum3 >> out_shift), 8)); + *pO++ = (q7_t) (__SSAT((sum4 >> out_shift), 8)); + + /* adjust the pointers and counters */ + rowCnt--; + } + + /* left-over part of the rows */ + rowCnt = num_of_rows & 0x3; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + uint16_t colCnt = dim_vec >> 2; + + pA = vec_buffer; + + while (colCnt) + { + q31_t inV1, inV2, inM11, inM12; + + pB = read_and_pad_reordered(pB, &inM11, &inM12); + + inV1 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV1, inM11, sum); + + inV2 = arm_nn_read_q15x2_ia(&pA); + sum = __SMLAD(inV2, inM12, sum); + + colCnt--; + } + + /* left-over of the vector */ + colCnt = dim_vec & 0x3; + while (colCnt) + { + q15_t inV = *pA++; + q7_t inM = *pB++; + sum += inV * inM; + colCnt--; + } + + *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8)); + + rowCnt--; + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + uint16_t rowCnt = num_of_rows >> 2; + const q7_t *pB = pM; + const q7_t *pA; + q7_t *pO = pOut; + const q7_t *pBias = bias; + + while (rowCnt) + { + q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum3 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + q31_t sum4 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + uint16_t colCnt = dim_vec >> 2; + + pA = pV; + + while (colCnt) + { + q7_t inA1 = *pA++; + q7_t inA3 = *pA++; + q7_t inA2 = *pA++; + q7_t inA4 = *pA++; + + q7_t inB1 = *pB++; + q7_t inB3 = *pB++; + q7_t inB2 = *pB++; + q7_t inB4 = *pB++; + + sum += inA1 * inB1 + inA2 * inB2; + sum2 += inA1 * inB3 + inA2 * inB4; + + inB1 = *pB++; + inB3 = *pB++; + inB2 = *pB++; + inB4 = *pB++; + + sum3 += inA1 * inB1 + inA2 * inB2; + sum4 += inA1 * inB3 + inA2 * inB4; + + inB1 = *pB++; + inB3 = *pB++; + inB2 = *pB++; + inB4 = *pB++; + + sum += inA3 * inB1 + inA4 * inB2; + sum2 += inA3 * inB3 + inA4 * inB4; + + inB1 = *pB++; + inB3 = *pB++; + inB2 = *pB++; + inB4 = *pB++; + + sum3 += inA3 * inB1 + inA4 * inB2; + sum4 += inA3 * inB3 + inA4 * inB4; + + colCnt--; + } + colCnt = dim_vec & 0x3; + while (colCnt) + { + q7_t inA = *pA++; + q7_t inB = *pB++; + sum += inA * inB; + inB = *pB++; + sum2 += inA * inB; + inB = *pB++; + sum3 += inA * inB; + inB = *pB++; + sum4 += inA * inB; + + colCnt--; + } + *pO++ = (q7_t) __SSAT((sum >> out_shift), 8); + *pO++ = (q7_t) __SSAT((sum2 >> out_shift), 8); + *pO++ = (q7_t) __SSAT((sum3 >> out_shift), 8); + *pO++ = (q7_t) __SSAT((sum4 >> out_shift), 8); + + rowCnt--; + } + + rowCnt = num_of_rows & 0x3; + + while (rowCnt) + { + int ip_out = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift); + + int j; + + pA = pV; + for (j = 0; j < dim_vec; j++) + { + q7_t inA = *pA++; + q7_t inB = *pB++; + ip_out += inA * inB; + } + *pO++ = (q7_t) __SSAT((ip_out >> out_shift), 8); + + rowCnt--; + } + +#endif /* ARM_MATH_DSP */ + + /* Return to ARM_MATH_SUCCESS */ + return (ARM_MATH_SUCCESS); + +} + +/** + * @} end of FC group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_s8.c new file mode 100644 index 0000000..ff1c652 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_s8.c @@ -0,0 +1,100 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_fully_connected_s8 + * Description: Fully connected function compatible with TF Lite. + * + * $Date: May 2, 2020 + * $Revision: V.2.0.0 + * + * Target Processor: Cortex-M and Cortex-A cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup FC + * @{ + */ + +/* + * S8 basic fully-connected and matrix multiplication layer function for TensorFlow Lite + * + * Refer header file for details. + * + */ + +arm_status +arm_fully_connected_s8(const cmsis_nn_context *ctx, + const cmsis_nn_fc_params *fc_params, + const cmsis_nn_per_tensor_quant_params *quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input, + const cmsis_nn_dims *filter_dims, + const q7_t *kernel, + const cmsis_nn_dims *bias_dims, + const int32_t *bias, + const cmsis_nn_dims *output_dims, + q7_t *output) +{ + (void)bias_dims; + (void)ctx; + int32_t batch_cnt = input_dims->n; + + while (batch_cnt) + { + arm_nn_vec_mat_mult_t_s8(input, + kernel, + bias, + output, + fc_params->input_offset, + fc_params->filter_offset, + fc_params->output_offset, + quant_params->multiplier, + quant_params->shift, + filter_dims->n, /* col_dim or accum_depth */ + output_dims->c, /* row_dim or output_depth */ + fc_params->activation.min, + fc_params->activation.max); + input += filter_dims->n; + output += output_dims->c; + batch_cnt--; + } + return (ARM_MATH_SUCCESS); +} + +int32_t arm_fully_connected_s8_get_buffer_size(const cmsis_nn_dims *filter_dims) +{ + (void)filter_dims; + return 0; +} + +/** + * @} end of FC group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_accumulate_q7_to_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_accumulate_q7_to_q15.c new file mode 100644 index 0000000..5eeb0e1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_accumulate_q7_to_q15.c @@ -0,0 +1,83 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_accumulate_q7_to_q15.c + * Description: Accumulate q7 vector into q15 one. + * + * $Date: May 29, 2020 + * $Revision: V.1.0.1 + * + * pSrc Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +void arm_nn_accumulate_q7_to_q15(q15_t *pDst, const q7_t *pSrc, uint32_t length) +{ + q15_t *pCnt = pDst; + const q7_t *pV = pSrc; + q31_t v1, v2, vo1, vo2; + int32_t cnt = length >> 2; + q31_t in; + + while (cnt > 0l) + { + q31_t value = arm_nn_read_q7x4_ia(&pV); + v1 = __SXTB16(__ROR((uint32_t)value, 8)); + v2 = __SXTB16(value); +#ifndef ARM_MATH_BIG_ENDIAN + vo2 = (q31_t)__PKHTB(v1, v2, 16); + vo1 = (q31_t)__PKHBT(v2, v1, 16); +#else + vo1 = (q31_t)__PKHTB(v1, v2, 16); + vo2 = (q31_t)__PKHBT(v2, v1, 16); +#endif + + in = arm_nn_read_q15x2(pCnt); + write_q15x2_ia(&pCnt, __QADD16(vo1, in)); + + in = arm_nn_read_q15x2(pCnt); + write_q15x2_ia(&pCnt, __QADD16(vo2, in)); + + cnt--; + } + cnt = length & 0x3; + while (cnt > 0l) + { + *pCnt++ += *pV++; + cnt--; + } +} + +/** + * @} end of NNBasicMath group + */ +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_add_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_add_q7.c new file mode 100644 index 0000000..91da6fd --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_add_q7.c @@ -0,0 +1,84 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_add_q7.c + * Description: Non saturating addition of elements of a q7 vector. + * + * $Date: July 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +void arm_nn_add_q7(const q7_t *input, q31_t *output, uint32_t block_size) +{ + uint32_t block_count; + q31_t result = 0; +#if defined(ARM_MATH_DSP) + /* Loop unrolling: Compute 4 outputs at a time */ + block_count = block_size >> 2U; + + while (block_count > 0U) + { + const int32_t mult_q15x2 = (1UL << 16) | 1UL; + q31_t in_q7x4 = arm_nn_read_q7x4_ia(&input); + q31_t temp_q15x2 = __SXTAB16(__SXTB16(in_q7x4), + __ROR((uint32_t)in_q7x4, 8)); + + result = __SMLAD(temp_q15x2, mult_q15x2, result); + + /* Decrement loop counter */ + block_count--; + } + + /* Loop unrolling: Compute remaining outputs */ + block_count = block_size & 0x3; +#else + block_count = block_size; +#endif + while (block_count > 0U) + { + /* Add and store result in destination buffer. */ + result += *input++; + + /* Decrement loop counter */ + block_count--; + } + + *output = result; +} + +/** + * @} end of NNBasicMath group + */ +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_padded_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_padded_s8.c new file mode 100644 index 0000000..ba9be42 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_padded_s8.c @@ -0,0 +1,172 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_depthwise_conv_nt_t_padded_s8.c + * Description: Depthwise convolution with padded matrices. + * + * $Date: March 17, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M processors with MVE extension + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +/* + * Depthwise convolution of transposed rhs matrix with 4 lhs matrices. One or more of the rhs matrices are padded. + * Dimensions are the same for lhs and rhs. + * + * Refer header file for details. + * + */ + +q7_t *arm_nn_depthwise_conv_nt_t_padded_s8(const q7_t *lhs, + const q7_t *rhs, + const int32_t input_offset, + const uint16_t num_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int32_t activation_min, + const int32_t activation_max, + const uint16_t row_x_col, + const int32_t *const output_bias, + q7_t *out) +{ +#if defined(ARM_MATH_MVEI) + int32_t loop_count = (num_ch + 3) / 4; + const int32_t *bias = output_bias; + uint32_t num_ch_to_process = num_ch; + + for (int i_loop_cnt = 0, offset = 0; i_loop_cnt < loop_count; + num_ch_to_process -= 4, out += 4, offset += 4, i_loop_cnt++) + { + int32x4_t out_0 = vldrwq_s32(bias); + int32x4_t out_1 = out_0; + int32x4_t out_2 = out_0; + int32x4_t out_3 = out_0; + bias += 4; + + const int8_t *rhs_0 = rhs + offset; + const int8_t *lhs_0 = lhs + offset; + const int8_t *lhs_1 = lhs + row_x_col * num_ch + offset; + const int8_t *lhs_2 = lhs + (row_x_col * num_ch * 2) + offset; + const int8_t *lhs_3 = lhs + (row_x_col * num_ch * 3) + offset; + + for (int i_row_x_col = 0; i_row_x_col < row_x_col; i_row_x_col++) + { + const int32x4_t ker_0 = vldrbq_s32(rhs_0); + + int32x4_t ip_0 = vldrbq_s32(lhs_0); + ip_0 = vaddq_n_s32(ip_0, input_offset); + out_0 += vmulq_s32(ip_0, ker_0); + + int32x4_t ip_1 = vldrbq_s32(lhs_1); + ip_1 = vaddq_n_s32(ip_1, input_offset); + out_1 += vmulq_s32(ip_1, ker_0); + + int32x4_t ip_2 = vldrbq_s32(lhs_2); + ip_2 = vaddq_n_s32(ip_2, input_offset); + out_2 += vmulq_s32(ip_2, ker_0); + + int32x4_t ip_3 = vldrbq_s32(lhs_3); + ip_3 = vaddq_n_s32(ip_3, input_offset); + + out_3 += vmulq_s32(ip_3, ker_0); + + lhs_0 += num_ch; + lhs_1 += num_ch; + lhs_2 += num_ch; + lhs_3 += num_ch; + + rhs_0 += num_ch; + } + + const int32x4_t mult = vldrwq_s32(out_mult); + const int32x4_t shift = vldrwq_s32(out_shift); + out_mult += 4; + out_shift += 4; + + out_0 = arm_requantize_mve_32x4(out_0, mult, shift); + out_0 = vaddq_n_s32(out_0, out_offset); + out_0 = vmaxq_s32(out_0, vdupq_n_s32(activation_min)); + out_0 = vminq_s32(out_0, vdupq_n_s32(activation_max)); + mve_pred16_t p = vctp32q(num_ch_to_process); + vstrbq_p_s32(out, out_0, p); + + out_1 = arm_requantize_mve_32x4(out_1, mult, shift); + out_1 = vaddq_n_s32(out_1, out_offset); + out_1 = vmaxq_s32(out_1, vdupq_n_s32(activation_min)); + out_1 = vminq_s32(out_1, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out + num_ch, out_1, p); + + out_2 = arm_requantize_mve_32x4(out_2, mult, shift); + out_2 = vaddq_n_s32(out_2, out_offset); + out_2 = vmaxq_s32(out_2, vdupq_n_s32(activation_min)); + out_2 = vminq_s32(out_2, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out + 2 * num_ch, out_2, p); + + out_3 = arm_requantize_mve_32x4(out_3, mult, shift); + out_3 = vaddq_n_s32(out_3, out_offset); + out_3 = vmaxq_s32(out_3, vdupq_n_s32(activation_min)); + out_3 = vminq_s32(out_3, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out + 3 * num_ch, out_3, p); + } + + const int tail_ch = num_ch & 0x3; + if (tail_ch != 0) + { + out -= (4 - tail_ch); + } + return out + (3 * num_ch); + +#else + (void)lhs; + (void)rhs; + (void)input_offset; + (void)num_ch; + (void)out_shift; + (void)out_mult; + (void)out_offset; + (void)activation_min; + (void)activation_max; + (void)row_x_col; + (void)output_bias; + (void)out; + return NULL; +#endif +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_s8.c new file mode 100644 index 0000000..c4e74d7 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_s8.c @@ -0,0 +1,174 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_depthwise_conv_nt_t_s8.c + * Description: Depthwise convolution on matrices with no padding. + * + * $Date: March 17, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M processors with MVE extension. + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +/* + * Depthwise convolution of rhs matrix with 4 lhs matrices with no padding. Dimensions are the same for lhs and rhs. + * + * Refer header file for details. + * + */ + +q7_t *arm_nn_depthwise_conv_nt_t_s8(const q7_t *lhs, + const q7_t *rhs, + const int32_t input_offset, + const uint16_t num_ch, + const int32_t *out_shift, + const int32_t *out_mult, + const int32_t out_offset, + const int32_t activation_min, + const int32_t activation_max, + const uint16_t row_x_col, + const int32_t *const output_bias, + q7_t *out) +{ +#if defined(ARM_MATH_MVEI) + const int32_t *bias = output_bias; + int32_t loop_count = (num_ch + 3) / 4; + uint32_t num_ch_to_process = num_ch; + + for (int i_loop_cnt = 0, offset = 0; i_loop_cnt < loop_count; + num_ch_to_process -= 4, offset += 4, out += 4, i_loop_cnt++) + { + int32x4_t out_0 = vldrwq_s32(bias); + int32x4_t out_1 = out_0; + int32x4_t out_2 = out_0; + int32x4_t out_3 = out_0; + bias += 4; + + const int8_t *rhs_0 = rhs + offset; + const int8_t *lhs_0 = lhs + offset; + const int8_t *lhs_1 = lhs + row_x_col * num_ch + offset; + const int8_t *lhs_2 = lhs + (row_x_col * num_ch * 2) + offset; + const int8_t *lhs_3 = lhs + (row_x_col * num_ch * 3) + offset; + int32x4_t ker_sum = vdupq_n_s32(0); + + for (int i_row_x_col = 0; i_row_x_col < row_x_col; i_row_x_col++) + { + const int32x4_t ker_0 = vldrbq_s32(rhs_0); + ker_sum = vaddq_s32(ker_sum, ker_0); + + int32x4_t ip_0 = vldrbq_s32(lhs_0); + out_0 += vmulq_s32(ip_0, ker_0); + + int32x4_t ip_1 = vldrbq_s32(lhs_1); + out_1 += vmulq_s32(ip_1, ker_0); + + int32x4_t ip_2 = vldrbq_s32(lhs_2); + out_2 += vmulq_s32(ip_2, ker_0); + + int32x4_t ip_3 = vldrbq_s32(lhs_3); + out_3 += vmulq_s32(ip_3, ker_0); + + lhs_0 += num_ch; + lhs_1 += num_ch; + lhs_2 += num_ch; + lhs_3 += num_ch; + + rhs_0 += num_ch; + } + + ker_sum = vmulq_n_s32(ker_sum, input_offset); + out_0 = ker_sum + out_0; + out_1 = ker_sum + out_1; + out_2 = ker_sum + out_2; + out_3 = ker_sum + out_3; + + const int32x4_t mult = vldrwq_s32(out_mult); + const int32x4_t shift = vldrwq_s32(out_shift); + out_mult += 4; + out_shift += 4; + mve_pred16_t p = vctp32q(num_ch_to_process); + + out_0 = arm_requantize_mve_32x4(out_0, mult, shift); + out_0 = vaddq_n_s32(out_0, out_offset); + out_0 = vmaxq_s32(out_0, vdupq_n_s32(activation_min)); + out_0 = vminq_s32(out_0, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out, out_0, p); + + out_1 = arm_requantize_mve_32x4(out_1, mult, shift); + out_1 = vaddq_n_s32(out_1, out_offset); + out_1 = vmaxq_s32(out_1, vdupq_n_s32(activation_min)); + out_1 = vminq_s32(out_1, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out + num_ch, out_1, p); + + out_2 = arm_requantize_mve_32x4(out_2, mult, shift); + out_2 = vaddq_n_s32(out_2, out_offset); + out_2 = vmaxq_s32(out_2, vdupq_n_s32(activation_min)); + out_2 = vminq_s32(out_2, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out + 2 * num_ch, out_2, p); + + out_3 = arm_requantize_mve_32x4(out_3, mult, shift); + out_3 = vaddq_n_s32(out_3, out_offset); + out_3 = vmaxq_s32(out_3, vdupq_n_s32(activation_min)); + out_3 = vminq_s32(out_3, vdupq_n_s32(activation_max)); + vstrbq_p_s32(out + 3 * num_ch, out_3, p); + } + + const int tail_ch = num_ch & 0x3; + if (tail_ch != 0) + { + out -= (4 - tail_ch); + } + + return out + (3 * num_ch); +#else + (void)lhs; + (void)rhs; + (void)input_offset; + (void)num_ch; + (void)out_shift; + (void)out_mult; + (void)out_offset; + (void)activation_min; + (void)activation_max; + (void)row_x_col; + (void)output_bias; + (void)out; + return NULL; +#endif +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_1x_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_1x_s8.c new file mode 100644 index 0000000..988e9ae --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_1x_s8.c @@ -0,0 +1,94 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mul_core_1x_s8.c + * Description: General Matrix-multiplication function + * + * $Date: January 20, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +/* + * s8 matrix multiplication to process 1 row + * + * Refer header file for details. + * + */ + +arm_status arm_nn_mat_mul_core_1x_s8(int32_t row_elements, + const int8_t *row_base, + const int8_t *col_base, + int32_t *const sum_col, + int32_t *const output) +{ + int32_t acc_n0 = 0; + int32_t sum_tmp = 0; + +#if defined(ARM_MATH_MVEI) && !defined(ARM_MATH_AUTOVECTORIZE) + + __asm volatile ( + " vldrb.8 q0, [%[col]], 16 \n" + " wlstp.8 lr, %[cnt], 1f \n" + "2: \n" + " vaddva.s8 %[sum], q0 \n" + " vldrb.8 q1, [%[row0]], 16 \n" + " vmladava.s8 %[out0], q0, q1 \n" + " vldrb.8 q0, [%[col]], 16 \n" + " letp lr, 2b \n" + "1: \n" + :[col] "+r"(col_base) + ,[sum] "+Te"(sum_tmp) + ,[row0] "+r"(row_base) + ,[out0] "+Te"(acc_n0) + :[cnt] "r"(row_elements) + :"q0","q1", "memory", "r14"); +#else + for (int i = 0; i < row_elements; i++) + { + sum_tmp += col_base[i]; + acc_n0 += row_base[i] * col_base[i]; + } +#endif + + *sum_col = sum_tmp; + *output = acc_n0; + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_4x_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_4x_s8.c new file mode 100644 index 0000000..d1af1fa --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_4x_s8.c @@ -0,0 +1,121 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mul_core_4x_s8.c + * Description: General matrix multiplication function for MVE extension + * + * $Date: January 20, 2020 + * $Revision: V.2.0.0 + * + * Target Processor: Cortex-M cores + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +/* + * s8 matrix multiplication to process 4 rows and one column + * + * Refer header file for details. + * + */ +arm_status arm_nn_mat_mul_core_4x_s8(const int32_t row_elements, + const int32_t offset, + const int8_t *row_base, + const int8_t *col_base, + int32_t *const sum_col, + int32_t *const output) +{ + int32_t acc_n0 = 0; + int32_t acc_n1 = 0; + int32_t acc_n2 = 0; + int32_t acc_n3 = 0; + + const int8_t *ip_row_0 = row_base; + const int8_t *ip_row_1 = row_base + offset; + const int8_t *ip_row_2 = row_base + (2 * offset); + const int8_t *ip_row_3 = row_base + (3 * offset); + int32_t sum_tmp = 0; + +#if defined(ARM_MATH_MVEI) && !defined(ARM_MATH_AUTOVECTORIZE) + __asm volatile( + " vldrb.8 q0, [%[col]], 16 \n" + " wlstp.8 lr, %[cnt], 1f \n" + "2: \n" + " vaddva.s8 %[sum], q0 \n" + " vldrb.8 q1, [%[row0]], 16 \n" + " vmladava.s8 %[out0], q0, q1 \n" + " vldrb.8 q2, [%[row1]], 16 \n" + " vmladava.s8 %[out1], q0, q2 \n" + " vldrb.8 q3, [%[row2]], 16 \n" + " vmladava.s8 %[out2], q0, q3 \n" + " vldrb.8 q4, [%[row3]], 16 \n" + " vmladava.s8 %[out3], q0, q4 \n" + " vldrb.8 q0, [%[col]], 16 \n" + " letp lr, 2b \n" + "1: \n" + :[col] "+r"(col_base) + ,[sum] "+Te"(sum_tmp) + ,[row0] "+r"(ip_row_0) + ,[row1] "+r"(ip_row_1) + ,[row2] "+r"(ip_row_2) + ,[row3] "+r"(ip_row_3) + ,[out0] "+Te"(acc_n0) + ,[out1] "+Te"(acc_n1) + ,[out2] "+Te"(acc_n2) + ,[out3] "+Te"(acc_n3) + : [cnt] "r"(row_elements) + : "q0", "q1", "q2", "q3", "q4", "memory", "r14"); +#else + for (int i = 0; i < row_elements; i++) + { + int32_t col = col_base[i]; + sum_tmp += col; + acc_n0 += ip_row_0[i] * col; + acc_n1 += ip_row_1[i] * col; + acc_n2 += ip_row_2[i] * col; + acc_n3 += ip_row_3[i] * col; + } +#endif + output[0] = acc_n0; + output[1] = acc_n1; + output[2] = acc_n2; + output[3] = acc_n3; + + *sum_col = sum_tmp; + + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c new file mode 100644 index 0000000..0c040b1 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c @@ -0,0 +1,587 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mat_mult_s8_nt_t_s8 + * Description: Matrix multiplication support function with the right-hand-side (rhs) matrix transposed + * + * $Date: July 27 2020 + * $Revision: V.1.0.2 + * + * Target Processor: Cortex-M + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +/* + * s8 matrix multiplication with the right-hand-side matrix transposed + * + * Refer header file for details. + * + */ +arm_status arm_nn_mat_mult_nt_t_s8(const q7_t *lhs, + const q7_t *rhs, + const q31_t *bias, + q7_t *dst, + const int32_t *dst_multipliers, + const int32_t *dst_shifts, + const int32_t lhs_rows, + const int32_t rhs_rows, + const int32_t rhs_cols, + const int32_t lhs_offset, + const int32_t dst_offset, + const int32_t activation_min, + const int32_t activation_max) +{ +#if defined(ARM_MATH_DSP) + const int32_t off0 = rhs_cols - 4; + + for (int32_t rhs_rows_idx = 0; rhs_rows_idx <= (rhs_rows - 2); rhs_rows_idx += 2) + { + const q7_t *lhs_ptr = &lhs[0]; + q7_t *dst_ptr = &dst[0]; + + q31_t lhs_offset_contribution0 = 0; + q31_t lhs_offset_contribution1 = 0; + + for (int32_t x = 0; x < rhs_cols; ++x) + { + lhs_offset_contribution0 += rhs[x]; + lhs_offset_contribution1 += rhs[x + rhs_cols]; + } + + lhs_offset_contribution0 *= lhs_offset; + lhs_offset_contribution1 *= lhs_offset; + if (bias) + { + lhs_offset_contribution0 += bias[rhs_rows_idx]; + lhs_offset_contribution1 += bias[rhs_rows_idx + 1]; + } + + int32_t lhs_rows_idx = lhs_rows >> 1; + + while (lhs_rows_idx) + { + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = lhs_offset_contribution0; + q31_t res01 = lhs_offset_contribution1; + q31_t res10 = lhs_offset_contribution0; + q31_t res11 = lhs_offset_contribution1; + + int32_t rhs_cols_idx = 0; + + q31_t val0, val1, val2, val3, val4, val5; + + for (; rhs_cols_idx <= (rhs_cols - 16); rhs_cols_idx += 16) + { + val1 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val2 = __SXTB16(val1); + val0 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val4 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val1 = __SXTB16_RORn(val1, 8); + val0 = __SXTB16_RORn(val0, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTB16(val4); + res00 = __SMLAD(val0, val1, res00); + val4 = __SXTB16_RORn(val4, 8); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val0, val4, res01); + + // 4 x MAC res10, res11 + val0 = arm_nn_read_q7x4((const q7_t *)&lhs_ptr[off0]); + val3 = __SXTB16(val0); + val0 = __SXTB16_RORn(val0, 8); + res10 = __SMLAD(val3, val2, res10); + res11 = __SMLAD(val3, val5, res11); + res10 = __SMLAD(val0, val1, res10); + val1 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + res11 = __SMLAD(val0, val4, res11); + + val4 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val2 = __SXTB16(val1); + val0 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val1 = __SXTB16_RORn(val1, 8); + val0 = __SXTB16_RORn(val0, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTB16(val4); + res00 = __SMLAD(val0, val1, res00); + val4 = __SXTB16_RORn(val4, 8); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val0, val4, res01); + + // 4 x MAC res10, res11 + val0 = arm_nn_read_q7x4((const q7_t *)&lhs_ptr[off0]); + val3 = __SXTB16(val0); + val0 = __SXTB16_RORn(val0, 8); + res10 = __SMLAD(val3, val2, res10); + res11 = __SMLAD(val3, val5, res11); + res10 = __SMLAD(val0, val1, res10); + val1 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + res11 = __SMLAD(val0, val4, res11); + + val4 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val2 = __SXTB16(val1); + val0 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val1 = __SXTB16_RORn(val1, 8); + val0 = __SXTB16_RORn(val0, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTB16(val4); + res00 = __SMLAD(val0, val1, res00); + val4 = __SXTB16_RORn(val4, 8); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val0, val4, res01); + + // 4 x MAC res10, res11 + val0 = arm_nn_read_q7x4((const q7_t *)&lhs_ptr[off0]); + val3 = __SXTB16(val0); + val0 = __SXTB16_RORn(val0, 8); + res10 = __SMLAD(val3, val2, res10); + res11 = __SMLAD(val3, val5, res11); + res10 = __SMLAD(val0, val1, res10); + val1 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + res11 = __SMLAD(val0, val4, res11); + + val4 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val2 = __SXTB16(val1); + val0 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val1 = __SXTB16_RORn(val1, 8); + val0 = __SXTB16_RORn(val0, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTB16(val4); + res00 = __SMLAD(val0, val1, res00); + val4 = __SXTB16_RORn(val4, 8); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val0, val4, res01); + + // 4 x MAC res10, res11 + val0 = arm_nn_read_q7x4((const q7_t *)&lhs_ptr[off0]); + val3 = __SXTB16(val0); + val0 = __SXTB16_RORn(val0, 8); + res10 = __SMLAD(val3, val2, res10); + res11 = __SMLAD(val3, val5, res11); + res10 = __SMLAD(val0, val1, res10); + res11 = __SMLAD(val0, val4, res11); + } + + for (; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q7_t rhs_value0 = rhs_ptr[0]; + q7_t rhs_value1 = rhs_ptr[rhs_cols]; + q7_t lhs_value = lhs_ptr[0]; + + res00 += lhs_value * rhs_value0; + res01 += lhs_value * rhs_value1; + + lhs_value = lhs_ptr[rhs_cols]; + res10 += lhs_value * rhs_value0; + res11 += lhs_value * rhs_value1; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multipliers[rhs_rows_idx], dst_shifts[rhs_rows_idx]); + res01 = arm_nn_requantize(res01, dst_multipliers[rhs_rows_idx + 1], dst_shifts[rhs_rows_idx + 1]); + res10 = arm_nn_requantize(res10, dst_multipliers[rhs_rows_idx], dst_shifts[rhs_rows_idx]); + res11 = arm_nn_requantize(res11, dst_multipliers[rhs_rows_idx + 1], dst_shifts[rhs_rows_idx + 1]); + + // Add offset + res00 += dst_offset; + res01 += dst_offset; + res10 += dst_offset; + res11 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + res01 = MAX(res01, activation_min); + res01 = MIN(res01, activation_max); + res10 = MAX(res10, activation_min); + res10 = MIN(res10, activation_max); + res11 = MAX(res11, activation_min); + res11 = MIN(res11, activation_max); + + dst_ptr[0] = (q7_t)res00; + dst_ptr[1] = (q7_t)res01; + dst_ptr += rhs_rows; + dst_ptr[0] = (q7_t)res10; + dst_ptr[1] = (q7_t)res11; + dst_ptr += rhs_rows; + + lhs_ptr += rhs_cols; + + lhs_rows_idx--; + } + + // Left-over rows + if (lhs_rows % 2) + { + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = lhs_offset_contribution0; + q31_t res01 = lhs_offset_contribution1; + + int32_t rhs_cols_idx = 0; + + q31_t val0, val1, val2, val3, val4, val5; + for (; rhs_cols_idx <= (rhs_cols - 16); rhs_cols_idx += 16) + { + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val5 = __SXTB16(val2); + val4 = __SXTB16(val1); + val0 = __SXTB16_RORn(val0, 8); + val2 = __SXTB16_RORn(val2, 8); + val1 = __SXTB16_RORn(val1, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val5, val3, res00); + res00 = __SMLAD(val2, val0, res00); + res01 = __SMLAD(val5, val4, res01); + res01 = __SMLAD(val2, val1, res01); + + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val5 = __SXTB16(val2); + val4 = __SXTB16(val1); + val0 = __SXTB16_RORn(val0, 8); + val2 = __SXTB16_RORn(val2, 8); + val1 = __SXTB16_RORn(val1, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val5, val3, res00); + res00 = __SMLAD(val2, val0, res00); + res01 = __SMLAD(val5, val4, res01); + res01 = __SMLAD(val2, val1, res01); + + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val5 = __SXTB16(val2); + val4 = __SXTB16(val1); + val0 = __SXTB16_RORn(val0, 8); + val2 = __SXTB16_RORn(val2, 8); + val1 = __SXTB16_RORn(val1, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val5, val3, res00); + res00 = __SMLAD(val2, val0, res00); + res01 = __SMLAD(val5, val4, res01); + res01 = __SMLAD(val2, val1, res01); + + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = arm_nn_read_q7x4((const q7_t *)&rhs_ptr[off0]); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val3 = __SXTB16(val0); + val5 = __SXTB16(val2); + val4 = __SXTB16(val1); + val0 = __SXTB16_RORn(val0, 8); + val2 = __SXTB16_RORn(val2, 8); + val1 = __SXTB16_RORn(val1, 8); + + // 4 x MAC res00, res01 + res00 = __SMLAD(val5, val3, res00); + res00 = __SMLAD(val2, val0, res00); + res01 = __SMLAD(val5, val4, res01); + res01 = __SMLAD(val2, val1, res01); + } + + // Left-over accumulations + for (; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q7_t rhs_value0 = rhs_ptr[0]; + q7_t rhs_value1 = rhs_ptr[rhs_cols]; + q7_t lhs_value = lhs_ptr[0]; + + res00 += lhs_value * rhs_value0; + res01 += lhs_value * rhs_value1; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multipliers[rhs_rows_idx], dst_shifts[rhs_rows_idx]); + res01 = arm_nn_requantize(res01, dst_multipliers[rhs_rows_idx + 1], dst_shifts[rhs_rows_idx + 1]); + + // Add offset + res00 += dst_offset; + res01 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + res01 = MAX(res01, activation_min); + res01 = MIN(res01, activation_max); + + dst_ptr[0] = (q7_t)res00; + dst_ptr[1] = (q7_t)res01; + } + + rhs += 2 * rhs_cols; + dst += 2; + } + + if (rhs_rows % 2) + { + const q7_t *lhs_ptr = &lhs[0]; + q7_t *dst_ptr = &dst[0]; + + for (int32_t lhs_rows_idx = 0; lhs_rows_idx < lhs_rows; ++lhs_rows_idx) + { + const q7_t *rhs_ptr = &rhs[0]; + q31_t res00 = 0; + if (bias) + { + res00 = bias[rhs_rows - 1]; + } + + for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q31_t rhs_value = rhs_ptr[0]; + q31_t lhs_value = lhs_ptr[0] + lhs_offset; + + res00 += lhs_value * rhs_value; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multipliers[rhs_rows - 1], dst_shifts[rhs_rows - 1]); + + // Add offset + res00 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + + dst_ptr[0] = (q7_t)res00; + dst_ptr += rhs_rows; + } + } +#else + for (int32_t rhs_rows_idx = 0; rhs_rows_idx <= (rhs_rows - 2); rhs_rows_idx += 2) + { + const q7_t *lhs_ptr = &lhs[0]; + q7_t *dst_ptr = &dst[0]; + + q31_t lhs_offset_contribution0 = 0; + q31_t lhs_offset_contribution1 = 0; + + for (int32_t x = 0; x < rhs_cols; ++x) + { + lhs_offset_contribution0 += rhs[x]; + lhs_offset_contribution1 += rhs[x + rhs_cols]; + } + + lhs_offset_contribution0 *= lhs_offset; + lhs_offset_contribution1 *= lhs_offset; + if (bias) + { + lhs_offset_contribution0 += bias[rhs_rows_idx]; + lhs_offset_contribution1 += bias[rhs_rows_idx + 1]; + } + + int32_t lhs_rows_idx = lhs_rows >> 1; + + while (lhs_rows_idx) + { + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = lhs_offset_contribution0; + q31_t res01 = lhs_offset_contribution1; + q31_t res10 = lhs_offset_contribution0; + q31_t res11 = lhs_offset_contribution1; + + for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q7_t rhs_value0 = rhs_ptr[0]; + q7_t rhs_value1 = rhs_ptr[rhs_cols]; + q7_t lhs_value = lhs_ptr[0]; + + res00 += lhs_value * rhs_value0; + res01 += lhs_value * rhs_value1; + + lhs_value = lhs_ptr[rhs_cols]; + res10 += lhs_value * rhs_value0; + res11 += lhs_value * rhs_value1; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multipliers[rhs_rows_idx], dst_shifts[rhs_rows_idx]); + res01 = arm_nn_requantize(res01, dst_multipliers[rhs_rows_idx + 1], dst_shifts[rhs_rows_idx + 1]); + res10 = arm_nn_requantize(res10, dst_multipliers[rhs_rows_idx], dst_shifts[rhs_rows_idx]); + res11 = arm_nn_requantize(res11, dst_multipliers[rhs_rows_idx + 1], dst_shifts[rhs_rows_idx + 1]); + + // Add offset + res00 += dst_offset; + res01 += dst_offset; + res10 += dst_offset; + res11 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + res01 = MAX(res01, activation_min); + res01 = MIN(res01, activation_max); + res10 = MAX(res10, activation_min); + res10 = MIN(res10, activation_max); + res11 = MAX(res11, activation_min); + res11 = MIN(res11, activation_max); + + dst_ptr[0] = (q7_t)res00; + dst_ptr[1] = (q7_t)res01; + dst_ptr += rhs_rows; + dst_ptr[0] = (q7_t)res10; + dst_ptr[1] = (q7_t)res11; + dst_ptr += rhs_rows; + + lhs_ptr += rhs_cols; + + lhs_rows_idx--; + } + + // Left-over rows + if (lhs_rows % 2) + { + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = lhs_offset_contribution0; + q31_t res01 = lhs_offset_contribution1; + + for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q7_t rhs_value0 = rhs_ptr[0]; + q7_t rhs_value1 = rhs_ptr[rhs_cols]; + q7_t lhs_value = lhs_ptr[0]; + + res00 += lhs_value * rhs_value0; + res01 += lhs_value * rhs_value1; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multipliers[rhs_rows_idx], dst_shifts[rhs_rows_idx]); + res01 = arm_nn_requantize(res01, dst_multipliers[rhs_rows_idx + 1], dst_shifts[rhs_rows_idx + 1]); + + // Add offset + res00 += dst_offset; + res01 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + res01 = MAX(res01, activation_min); + res01 = MIN(res01, activation_max); + + dst_ptr[0] = (q7_t)res00; + dst_ptr[1] = (q7_t)res01; + } + + rhs += 2 * rhs_cols; + dst += 2; + } + + if (rhs_rows % 2) + { + const q7_t *lhs_ptr = &lhs[0]; + q7_t *dst_ptr = &dst[0]; + + for (int32_t lhs_rows_idx = 0; lhs_rows_idx < lhs_rows; ++lhs_rows_idx) + { + const q7_t *rhs_ptr = &rhs[0]; + q31_t res00 = 0; + if (bias) + { + res00 = bias[rhs_rows - 1]; + } + + for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q31_t rhs_value = rhs_ptr[0]; + q31_t lhs_value = lhs_ptr[0] + lhs_offset; + + res00 += lhs_value * rhs_value; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multipliers[rhs_rows - 1], dst_shifts[rhs_rows - 1]); + + // Add offset + res00 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + + dst_ptr[0] = (q7_t)res00; + dst_ptr += rhs_rows; + } + } +#endif + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c new file mode 100644 index 0000000..ef9b870 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c @@ -0,0 +1,148 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mult_q15.c + * Description: Q15 vector multiplication with variable output shifts + * + * $Date: 29. April 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + + +/** + * @brief Q7 vector multiplication with variable output shifts + * @param[in] *pSrcA pointer to the first input vector + * @param[in] *pSrcB pointer to the second input vector + * @param[out] *pDst pointer to the output vector + * @param[in] out_shift amount of right-shift for output + * @param[in] blockSize number of samples in each vector + * + * Scaling and Overflow Behavior: + * \par + * The function uses saturating arithmetic. + * Results outside of the allowable Q15 range [0x8000 0x7FFF] will be saturated. + */ + +void arm_nn_mult_q15( + q15_t * pSrcA, + q15_t * pSrcB, + q15_t * pDst, + const uint16_t out_shift, + uint32_t blockSize) +{ + uint32_t blkCnt; /* loop counters */ + +#if defined (ARM_MATH_DSP) + +/* Run the below code for Cortex-M4 and Cortex-M3 */ + q31_t inA1, inA2, inB1, inB2; /* temporary input variables */ + q15_t out1, out2, out3, out4; /* temporary output variables */ + q31_t mul1, mul2, mul3, mul4; /* temporary variables */ + + /* loop Unrolling */ + blkCnt = blockSize >> 2U; + + /* First part of the processing with loop unrolling. Compute 4 outputs at a time. + ** a second loop below computes the remaining 1 to 3 samples. */ + while (blkCnt > 0U) + { + /* read two samples at a time from sourceA */ + inA1 = arm_nn_read_q15x2_ia((const q15_t **)&pSrcA); + /* read two samples at a time from sourceB */ + inB1 = arm_nn_read_q15x2_ia((const q15_t **)&pSrcB); + /* read two samples at a time from sourceA */ + inA2 = arm_nn_read_q15x2_ia((const q15_t **)&pSrcA); + /* read two samples at a time from sourceB */ + inB2 = arm_nn_read_q15x2_ia((const q15_t **)&pSrcB); + + /* multiply mul = sourceA * sourceB */ + mul1 = (q31_t) ((q15_t) (inA1 >> 16) * (q15_t) (inB1 >> 16)); + mul2 = (q31_t) ((q15_t) inA1 * (q15_t) inB1); + mul3 = (q31_t) ((q15_t) (inA2 >> 16) * (q15_t) (inB2 >> 16)); + mul4 = (q31_t) ((q15_t) inA2 * (q15_t) inB2); + + /* saturate result to 16 bit */ + out1 = (q15_t) __SSAT((q31_t) (mul1 + NN_ROUND(out_shift)) >> out_shift, 16); + out2 = (q15_t) __SSAT((q31_t) (mul2 + NN_ROUND(out_shift)) >> out_shift, 16); + out3 = (q15_t) __SSAT((q31_t) (mul3 + NN_ROUND(out_shift)) >> out_shift, 16); + out4 = (q15_t) __SSAT((q31_t) (mul4 + NN_ROUND(out_shift)) >> out_shift, 16); + + /* store the result */ +#ifndef ARM_MATH_BIG_ENDIAN + + *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16); + *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16); + +#else + + *__SIMD32(pDst)++ = __PKHBT(out2, out1, 16); + *__SIMD32(pDst)++ = __PKHBT(out4, out3, 16); + +#endif /* #ifndef ARM_MATH_BIG_ENDIAN */ + + /* Decrement the blockSize loop counter */ + blkCnt--; + } + + /* If the blockSize is not a multiple of 4, compute any remaining output samples here. + ** No loop unrolling is used. */ + blkCnt = blockSize % 0x4U; + +#else + + /* Run the below code for Cortex-M0 */ + + /* Initialize blkCnt with number of samples */ + blkCnt = blockSize; + +#endif /* #if defined (ARM_MATH_DSP) */ + + + while (blkCnt > 0U) + { + /* C = A * B */ + /* Multiply the inputs and store the result in the destination buffer */ + *pDst++ = (q15_t) __SSAT(((q31_t) ((q31_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 16); + + /* Decrement the blockSize loop counter */ + blkCnt--; + } +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c new file mode 100644 index 0000000..8c627fc --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c @@ -0,0 +1,121 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_mult_q7.c + * Description: Q7 vector multiplication with variable output shifts + * + * $Date: 29. April 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +/** + * @brief Q7 vector multiplication with variable output shifts + * @param[in] *pSrcA pointer to the first input vector + * @param[in] *pSrcB pointer to the second input vector + * @param[out] *pDst pointer to the output vector + * @param[in] out_shift amount of right-shift for output + * @param[in] blockSize number of samples in each vector + * + * Scaling and Overflow Behavior: + * \par + * The function uses saturating arithmetic. + * Results outside of the allowable Q7 range [0x80 0x7F] will be saturated. + */ + +void arm_nn_mult_q7( + q7_t * pSrcA, + q7_t * pSrcB, + q7_t * pDst, + const uint16_t out_shift, + uint32_t blockSize) +{ + uint32_t blkCnt; /* loop counters */ + +#if defined (ARM_MATH_DSP) + +/* Run the below code for Cortex-M4 and Cortex-M3 */ + q7_t out1, out2, out3, out4; /* Temporary variables to store the product */ + + /* loop Unrolling */ + blkCnt = blockSize >> 2U; + + /* First part of the processing with loop unrolling. Compute 4 outputs at a time. + ** a second loop below computes the remaining 1 to 3 samples. */ + while (blkCnt > 0U) + { + /* C = A * B */ + /* Multiply the inputs and store the results in temporary variables */ + out1 = (q7_t) __SSAT(((q15_t) ((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); + out2 = (q7_t) __SSAT(((q15_t) ((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); + out3 = (q7_t) __SSAT(((q15_t) ((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); + out4 = (q7_t) __SSAT(((q15_t) ((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); + + /* Store the results of 4 inputs in the destination buffer in single cycle by packing */ + *__SIMD32(pDst)++ = __PACKq7(out1, out2, out3, out4); + + /* Decrement the blockSize loop counter */ + blkCnt--; + } + + /* If the blockSize is not a multiple of 4, compute any remaining output samples here. + ** No loop unrolling is used. */ + blkCnt = blockSize % 0x4U; + +#else + + /* Run the below code for Cortex-M0 */ + + /* Initialize blkCnt with number of samples */ + blkCnt = blockSize; + +#endif /* #if defined (ARM_MATH_DSP) */ + + + while (blkCnt > 0U) + { + /* C = A * B */ + /* Multiply the inputs and store the result in the destination buffer */ + *pDst++ = (q7_t) __SSAT(((q15_t) ((q15_t) (*pSrcA++) * (*pSrcB++) + NN_ROUND(out_shift)) >> out_shift), 8); + + /* Decrement the blockSize loop counter */ + blkCnt--; + } +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_vec_mat_mult_t_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_vec_mat_mult_t_s8.c new file mode 100644 index 0000000..05d8834 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_vec_mat_mult_t_s8.c @@ -0,0 +1,465 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nn_vec_mat_mult_t_s8 + * Description: s8 vector by matrix (transposed) multiplication + * + * $Date: April 2, 2020 + * $Revision: V.1.5.0 + * + * Target Processor: Cortex-M + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup NNBasicMath + * @{ + */ + +/* + * s8 vector(lhs) by matrix (transposed) multiplication + * + * Refer header file for details. + * + */ +arm_status arm_nn_vec_mat_mult_t_s8(const q7_t *lhs, + const q7_t *rhs, + const q31_t *bias, + q7_t *dst, + const int32_t lhs_offset, + const int32_t rhs_offset, + const int32_t dst_offset, + const int32_t dst_multiplier, + const int32_t dst_shift, + const int32_t rhs_cols, + const int32_t rhs_rows, + const int32_t activation_min, + const int32_t activation_max) +{ +#if defined(ARM_MATH_MVEI) + + const int16x8_t rhs_offset_vec = vdupq_n_s16((int16_t)rhs_offset); + const int16x8_t lhs_offset_vec = vdupq_n_s16((int16_t)lhs_offset); + + int32_t row_loop_cnt = rhs_rows / 4; + + for (int i_row_loop_cnt = 0; i_row_loop_cnt < row_loop_cnt; i_row_loop_cnt++) + { + int32_t acc1 = bias[0]; + int32_t acc2 = bias[1]; + int32_t acc3 = bias[2]; + int32_t acc4 = bias[3]; + bias += 4; + + int32x4_t acc; + const int32_t col_loop_cnt = (rhs_cols + 7) / 8; + + const int8_t *vec = lhs; + const int8_t *rhs_0 = rhs; + const int8_t *rhs_1 = rhs + rhs_cols; + const int8_t *rhs_2 = rhs + 2 * rhs_cols; + const int8_t *rhs_3 = rhs + 3 * rhs_cols; + + uint32_t col_cnt = (uint32_t)rhs_cols; + + for (int i = 0; i < col_loop_cnt; i++) + { + mve_pred16_t p = vctp16q(col_cnt); + col_cnt -= 8; + const int16x8_t tmp_b = vaddq_m_s16(vuninitializedq_s16(), + vldrbq_z_s16(vec, p), lhs_offset_vec, p); + + const int16x8_t tmp_a0 = vaddq_m_s16(vuninitializedq_s16(), + vldrbq_z_s16(rhs_0, p), rhs_offset_vec, p); + acc1 = vmladavaq_p_s16(acc1, tmp_a0, tmp_b, p); + + const int16x8_t tmp_a1 = vaddq_m_s16(vuninitializedq_s16(), + vldrbq_z_s16(rhs_1, p), rhs_offset_vec, p); + acc2 = vmladavaq_p_s16(acc2, tmp_a1, tmp_b, p); + + const int16x8_t tmp_a2 = vaddq_m_s16(vuninitializedq_s16(), + vldrbq_z_s16(rhs_2, p), rhs_offset_vec, p); + acc3 = vmladavaq_p_s16(acc3, tmp_a2, tmp_b, p); + + const int16x8_t tmp_a3 = vaddq_m_s16(vuninitializedq_s16(), + vldrbq_z_s16(rhs_3, p), rhs_offset_vec, p); + acc4 = vmladavaq_p_s16(acc4, tmp_a3, tmp_b, p); + + vec += 8; + rhs_0 += 8; + rhs_1 += 8; + rhs_2 += 8; + rhs_3 += 8; + } + rhs += 4 * rhs_cols; + + acc[0] = acc1; + acc[1] = acc2; + acc[2] = acc3; + acc[3] = acc4; + + acc = arm_requantize_mve(acc, dst_multiplier, dst_shift); + acc = vaddq_s32(acc, vdupq_n_s32(dst_offset)); + acc = vmaxq_s32(acc, vdupq_n_s32(activation_min)); + acc = vminq_s32(acc, vdupq_n_s32(activation_max)); + + vstrbq_s32(dst, acc); + dst += 4; + } + + row_loop_cnt = rhs_rows & 3; + + for (int i_row_loop_cnt = 0; i_row_loop_cnt < row_loop_cnt; + i_row_loop_cnt++) + { + int32_t acc = *bias++; + const int32_t col_loop_cnt = (rhs_cols + 7) / 8; + const int8_t *vec = lhs; + const int8_t *kernel_cur = rhs; + + uint32_t col_cnt = (uint32_t)rhs_cols; + + for (int i = 0; i < col_loop_cnt; i++) + { + mve_pred16_t p = vctp16q(col_cnt); + col_cnt -= 8; + const int16x8_t tmp_b = vaddq_m_s16(vuninitializedq_s16(), + vldrbq_z_s16(vec, p), lhs_offset_vec, p); + + const int16x8_t tmp_a = vaddq_m_s16(vuninitializedq_s16(), + vldrbq_z_s16(kernel_cur, p), rhs_offset_vec, p); + acc = vmladavaq_p_s16(acc, tmp_a, tmp_b, p); + vec += 8; + kernel_cur += 8; + } + rhs += rhs_cols; + + acc = arm_nn_requantize(acc, dst_multiplier, dst_shift); + acc += dst_offset; + + acc = MAX(acc, activation_min); + acc = MIN(acc, activation_max); + *dst++ = (int8_t)(acc); + } + +#elif defined(ARM_MATH_DSP) + const int32_t off0 = rhs_cols - 4; + const int16_t lhs_offset_s16 = lhs_offset; + const int16_t rhs_offset_s16 = rhs_offset; + + const uint32_t lhs_offset_s16x2 = __PKHBT(lhs_offset_s16, lhs_offset_s16, 16); + const uint32_t rhs_offset_s16x2 = __PKHBT(rhs_offset_s16, rhs_offset_s16, 16); + + for (int32_t rhs_rows_idx = 0; rhs_rows_idx <= (rhs_rows - 2); rhs_rows_idx += 2) + { + const q7_t *lhs_ptr = &lhs[0]; + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = *bias++; + q31_t res01 = *bias++; + + int32_t rhs_cols_idx = 0; + + q31_t val0, val1, val2, val3, val4, val5; + for (; rhs_cols_idx <= (rhs_cols - 16); rhs_cols_idx += 16) + { + // Read 4 x int8 values from the RHS matrix + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val2 = __SXTAB16(rhs_offset_s16x2, val0); + // Read 4 x int8 values from the LHS vector + val1 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val1); + // Read 4 x int8 values from the RHS matrix + val4 = arm_nn_read_q7x4((const q7_t *)rhs_ptr + off0); + val1 = __SXTAB16(lhs_offset_s16x2, __ROR(val1, 8)); + + // Perform the accumulations + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTAB16(rhs_offset_s16x2, val4); + res00 = __SMLAD(val1, val0, res00); + val4 = __SXTAB16(rhs_offset_s16x2, __ROR(val4, 8)); + // Read 4 x int8 values from the RHS matrix + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val1, val4, res01); + + val2 = __SXTAB16(rhs_offset_s16x2, val0); + // Read 4 x int8 values from the LHS vector + val1 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val1); + // Read 4 x int8 values from the RHS matrix + val4 = arm_nn_read_q7x4((const q7_t *)rhs_ptr + off0); + val1 = __SXTAB16(lhs_offset_s16x2, __ROR(val1, 8)); + + // Perform the accumulations + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTAB16(rhs_offset_s16x2, val4); + res00 = __SMLAD(val1, val0, res00); + val4 = __SXTAB16(rhs_offset_s16x2, __ROR(val4, 8)); + // Read 4 x int8 values from the RHS matrix + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val1, val4, res01); + + val2 = __SXTAB16(rhs_offset_s16x2, val0); + // Read 4 x int8 values from the LHS vector + val1 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val1); + // Read 4 x int8 values from the RHS matrix + val4 = arm_nn_read_q7x4((const q7_t *)rhs_ptr + off0); + val1 = __SXTAB16(lhs_offset_s16x2, __ROR(val1, 8)); + + // Perform the accumulations + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTAB16(rhs_offset_s16x2, val4); + res00 = __SMLAD(val1, val0, res00); + val4 = __SXTAB16(rhs_offset_s16x2, __ROR(val4, 8)); + // Read 4 x int8 values from the RHS matrix + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val1, val4, res01); + + val2 = __SXTAB16(rhs_offset_s16x2, val0); + // Read 4 x int8 values from the LHS vector + val1 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val1); + // Read 4 x int8 values from the RHS matrix + val4 = arm_nn_read_q7x4((const q7_t *)rhs_ptr + off0); + val1 = __SXTAB16(lhs_offset_s16x2, __ROR(val1, 8)); + + // Perform the accumulations + res00 = __SMLAD(val3, val2, res00); + val5 = __SXTAB16(rhs_offset_s16x2, val4); + res00 = __SMLAD(val1, val0, res00); + val4 = __SXTAB16(rhs_offset_s16x2, __ROR(val4, 8)); + res01 = __SMLAD(val3, val5, res01); + res01 = __SMLAD(val1, val4, res01); + } + + for (; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q31_t rhs_value0 = rhs_ptr[0] + rhs_offset; + q31_t rhs_value1 = rhs_ptr[rhs_cols] + rhs_offset; + q31_t lhs_value = lhs_ptr[0] + lhs_offset; + + res00 += lhs_value * rhs_value0; + res01 += lhs_value * rhs_value1; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multiplier, dst_shift); + res01 = arm_nn_requantize(res01, dst_multiplier, dst_shift); + + // Add offset + res00 += dst_offset; + res01 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + res01 = MAX(res01, activation_min); + res01 = MIN(res01, activation_max); + + *dst++ = (q7_t)res00; + *dst++ = (q7_t)res01; + + rhs += 2 * rhs_cols; + } + + if (rhs_rows % 2) + { + const q7_t *lhs_ptr = &lhs[0]; + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = *bias++; + + int32_t rhs_cols_idx = 0; + + q31_t val0, val1, val2, val3; + for (; rhs_cols_idx <= (rhs_cols - 16); rhs_cols_idx += 16) + { + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = __SXTAB16(rhs_offset_s16x2, val0); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val2); + val2 = __SXTAB16(lhs_offset_s16x2, __ROR(val2, 8)); + + // Partial accumulations + res00 = __SMLAD(val3, val1, res00); + res00 = __SMLAD(val2, val0, res00); + + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = __SXTAB16(rhs_offset_s16x2, val0); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val2); + val2 = __SXTAB16(lhs_offset_s16x2, __ROR(val2, 8)); + + // Partial accumulations + res00 = __SMLAD(val3, val1, res00); + res00 = __SMLAD(val2, val0, res00); + + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = __SXTAB16(rhs_offset_s16x2, val0); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val2); + val2 = __SXTAB16(lhs_offset_s16x2, __ROR(val2, 8)); + + // Partial accumulations + res00 = __SMLAD(val3, val1, res00); + res00 = __SMLAD(val2, val0, res00); + + val0 = arm_nn_read_q7x4_ia((const q7_t **)&rhs_ptr); + val1 = __SXTAB16(rhs_offset_s16x2, val0); + val2 = arm_nn_read_q7x4_ia((const q7_t **)&lhs_ptr); + val0 = __SXTAB16(rhs_offset_s16x2, __ROR(val0, 8)); + val3 = __SXTAB16(lhs_offset_s16x2, val2); + val2 = __SXTAB16(lhs_offset_s16x2, __ROR(val2, 8)); + + // Partial accumulations + res00 = __SMLAD(val3, val1, res00); + res00 = __SMLAD(val2, val0, res00); + } + + for (; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q31_t rhs_value0 = rhs_ptr[0] + rhs_offset; + q31_t lhs_value = lhs_ptr[0] + lhs_offset; + + res00 += lhs_value * rhs_value0; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multiplier, dst_shift); + + // Add offset + res00 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + + *dst = (q7_t)res00; + } + +#else + + for (int32_t rhs_rows_idx = 0; rhs_rows_idx <= (rhs_rows - 2); rhs_rows_idx += 2) + { + const q7_t *lhs_ptr = &lhs[0]; + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = *bias++; + q31_t res01 = *bias++; + + for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q31_t rhs_value0 = rhs_ptr[0] + rhs_offset; + q31_t rhs_value1 = rhs_ptr[rhs_cols] + rhs_offset; + q31_t lhs_value = lhs_ptr[0] + lhs_offset; + + res00 += lhs_value * rhs_value0; + res01 += lhs_value * rhs_value1; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multiplier, dst_shift); + res01 = arm_nn_requantize(res01, dst_multiplier, dst_shift); + + // Add offset + res00 += dst_offset; + res01 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + res01 = MAX(res01, activation_min); + res01 = MIN(res01, activation_max); + + *dst++ = (q7_t)res00; + *dst++ = (q7_t)res01; + + rhs += 2 * rhs_cols; + } + + if (rhs_rows % 2) + { + const q7_t *lhs_ptr = &lhs[0]; + const q7_t *rhs_ptr = &rhs[0]; + + q31_t res00 = *bias++; + + for (int32_t rhs_cols_idx = 0; rhs_cols_idx < rhs_cols; ++rhs_cols_idx) + { + q31_t rhs_value0 = rhs_ptr[0] + rhs_offset; + q31_t lhs_value = lhs_ptr[0] + lhs_offset; + + res00 += lhs_value * rhs_value0; + + ++rhs_ptr; + ++lhs_ptr; + } + + // Quantize down + res00 = arm_nn_requantize(res00, dst_multiplier, dst_shift); + + // Add offset + res00 += dst_offset; + + // Clamp the result + res00 = MAX(res00, activation_min); + res00 = MIN(res00, activation_max); + + *dst = (q7_t)res00; + } +#endif + + return ARM_MATH_SUCCESS; +} + +/** + * @} end of NNBasicMath group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nntables.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nntables.c new file mode 100644 index 0000000..a82b8b2 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nntables.c @@ -0,0 +1,300 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_nntables.c + * Description: Converts the elements of the Q7 vector to Q15 vector without left-shift + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @brief tables for various activation functions + * + * This file include the declaration of common tables. + * Most of them are used for activation functions + * + * Assumption: + * Unified table: input is 3.x format, i.e, range of [-8, 8) + * sigmoid(8) = 0.9996646498695336 + * tanh(8) = 0.9999997749296758 + * The accuracy here should be good enough + * + * 2-stage HL table: + * + * The entire input range is divided into two parts: + * + * Low range table: 0x000x xxxx or 0x111x xxxx + * table entry will be the binary number excluding the first + * two digits, i.e., 0x0x xxxx or 0x1x xxxx + * + * + * + * High range table 0x0010 0000 -- 0x0111 1111 + * 0x1000 0000 -- 0x1101 1111 + * + * For positive numbers, table entry will be + * 0x0010 0000 -- 0x0111 1111 minus 0x0010 0000 + * i.e., 0x0000 0000 - 0x0101 11111 + * + * same thing for the negative numbers, table entry will be + * 0x1000 0000 -- 0x1101 1111 minux 0x0010 0000 + * i.e., 0x0110 0000 - 0x1011 1111 + */ + +const q7_t sigmoidTable_q7[256] = { + 0x40, 0x42, 0x44, 0x46, 0x48, 0x4a, 0x4c, 0x4e, + 0x50, 0x52, 0x53, 0x55, 0x57, 0x59, 0x5a, 0x5c, + 0x5e, 0x5f, 0x61, 0x62, 0x63, 0x65, 0x66, 0x67, + 0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70, + 0x71, 0x72, 0x72, 0x73, 0x74, 0x74, 0x75, 0x76, + 0x76, 0x77, 0x77, 0x78, 0x78, 0x79, 0x79, 0x7a, + 0x7a, 0x7a, 0x7b, 0x7b, 0x7b, 0x7c, 0x7c, 0x7c, + 0x7c, 0x7c, 0x7d, 0x7d, 0x7d, 0x7d, 0x7d, 0x7e, + 0x7e, 0x7e, 0x7e, 0x7e, 0x7e, 0x7e, 0x7e, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x01, 0x01, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, + 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x04, + 0x04, 0x04, 0x04, 0x04, 0x05, 0x05, 0x05, 0x06, + 0x06, 0x06, 0x07, 0x07, 0x08, 0x08, 0x09, 0x09, + 0x0a, 0x0a, 0x0b, 0x0c, 0x0c, 0x0d, 0x0e, 0x0e, + 0x0f, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, + 0x17, 0x19, 0x1a, 0x1b, 0x1d, 0x1e, 0x1f, 0x21, + 0x22, 0x24, 0x26, 0x27, 0x29, 0x2b, 0x2d, 0x2e, + 0x30, 0x32, 0x34, 0x36, 0x38, 0x3a, 0x3c, 0x3e, +}; + +const q15_t sigmoidTable_q15[256] = { + 0x4000, 0x4200, 0x43ff, 0x45fc, 0x47f5, 0x49eb, 0x4bdc, 0x4dc8, + 0x4fad, 0x518a, 0x5360, 0x552c, 0x56ef, 0x58a8, 0x5a57, 0x5bfb, + 0x5d93, 0x5f20, 0x60a1, 0x6216, 0x637f, 0x64db, 0x662b, 0x676f, + 0x68a6, 0x69d2, 0x6af1, 0x6c05, 0x6d0d, 0x6e09, 0x6efb, 0x6fe2, + 0x70be, 0x7190, 0x7258, 0x7316, 0x73cc, 0x7478, 0x751b, 0x75b7, + 0x764a, 0x76d6, 0x775b, 0x77d8, 0x784f, 0x78c0, 0x792a, 0x798f, + 0x79ee, 0x7a48, 0x7a9d, 0x7aed, 0x7b39, 0x7b80, 0x7bc4, 0x7c03, + 0x7c3f, 0x7c78, 0x7cad, 0x7ce0, 0x7d0f, 0x7d3c, 0x7d66, 0x7d8d, + 0x7db3, 0x7dd6, 0x7df7, 0x7e16, 0x7e33, 0x7e4f, 0x7e69, 0x7e81, + 0x7e98, 0x7eae, 0x7ec2, 0x7ed5, 0x7ee7, 0x7ef8, 0x7f08, 0x7f17, + 0x7f25, 0x7f32, 0x7f3e, 0x7f4a, 0x7f55, 0x7f5f, 0x7f69, 0x7f72, + 0x7f7b, 0x7f83, 0x7f8a, 0x7f91, 0x7f98, 0x7f9e, 0x7fa4, 0x7faa, + 0x7faf, 0x7fb4, 0x7fb8, 0x7fbd, 0x7fc1, 0x7fc5, 0x7fc8, 0x7fcc, + 0x7fcf, 0x7fd2, 0x7fd5, 0x7fd7, 0x7fda, 0x7fdc, 0x7fde, 0x7fe0, + 0x7fe2, 0x7fe4, 0x7fe6, 0x7fe7, 0x7fe9, 0x7fea, 0x7feb, 0x7fed, + 0x7fee, 0x7fef, 0x7ff0, 0x7ff1, 0x7ff2, 0x7ff3, 0x7ff4, 0x7ff4, + 0x000b, 0x000c, 0x000c, 0x000d, 0x000e, 0x000f, 0x0010, 0x0011, + 0x0012, 0x0013, 0x0015, 0x0016, 0x0017, 0x0019, 0x001a, 0x001c, + 0x001e, 0x0020, 0x0022, 0x0024, 0x0026, 0x0029, 0x002b, 0x002e, + 0x0031, 0x0034, 0x0038, 0x003b, 0x003f, 0x0043, 0x0048, 0x004c, + 0x0051, 0x0056, 0x005c, 0x0062, 0x0068, 0x006f, 0x0076, 0x007d, + 0x0085, 0x008e, 0x0097, 0x00a1, 0x00ab, 0x00b6, 0x00c2, 0x00ce, + 0x00db, 0x00e9, 0x00f8, 0x0108, 0x0119, 0x012b, 0x013e, 0x0152, + 0x0168, 0x017f, 0x0197, 0x01b1, 0x01cd, 0x01ea, 0x0209, 0x022a, + 0x024d, 0x0273, 0x029a, 0x02c4, 0x02f1, 0x0320, 0x0353, 0x0388, + 0x03c1, 0x03fd, 0x043c, 0x0480, 0x04c7, 0x0513, 0x0563, 0x05b8, + 0x0612, 0x0671, 0x06d6, 0x0740, 0x07b1, 0x0828, 0x08a5, 0x092a, + 0x09b6, 0x0a49, 0x0ae5, 0x0b88, 0x0c34, 0x0cea, 0x0da8, 0x0e70, + 0x0f42, 0x101e, 0x1105, 0x11f7, 0x12f3, 0x13fb, 0x150f, 0x162e, + 0x175a, 0x1891, 0x19d5, 0x1b25, 0x1c81, 0x1dea, 0x1f5f, 0x20e0, + 0x226d, 0x2405, 0x25a9, 0x2758, 0x2911, 0x2ad4, 0x2ca0, 0x2e76, + 0x3053, 0x3238, 0x3424, 0x3615, 0x380b, 0x3a04, 0x3c01, 0x3e00, +}; + +const q15_t sigmoidLTable_q15[128] = { + 0x4000, 0x4100, 0x4200, 0x42ff, 0x43ff, 0x44fd, 0x45fc, 0x46f9, + 0x47f5, 0x48f1, 0x49eb, 0x4ae5, 0x4bdc, 0x4cd3, 0x4dc8, 0x4ebb, + 0x4fad, 0x509c, 0x518a, 0x5276, 0x5360, 0x5447, 0x552c, 0x560f, + 0x56ef, 0x57cd, 0x58a8, 0x5981, 0x5a57, 0x5b2a, 0x5bfb, 0x5cc9, + 0x5d93, 0x5e5b, 0x5f20, 0x5fe2, 0x60a1, 0x615d, 0x6216, 0x62cc, + 0x637f, 0x642e, 0x64db, 0x6584, 0x662b, 0x66ce, 0x676f, 0x680c, + 0x68a6, 0x693d, 0x69d2, 0x6a63, 0x6af1, 0x6b7c, 0x6c05, 0x6c8a, + 0x6d0d, 0x6d8d, 0x6e09, 0x6e84, 0x6efb, 0x6f70, 0x6fe2, 0x7051, + 0x0f42, 0x0faf, 0x101e, 0x1090, 0x1105, 0x117c, 0x11f7, 0x1273, + 0x12f3, 0x1376, 0x13fb, 0x1484, 0x150f, 0x159d, 0x162e, 0x16c3, + 0x175a, 0x17f4, 0x1891, 0x1932, 0x19d5, 0x1a7c, 0x1b25, 0x1bd2, + 0x1c81, 0x1d34, 0x1dea, 0x1ea3, 0x1f5f, 0x201e, 0x20e0, 0x21a5, + 0x226d, 0x2337, 0x2405, 0x24d6, 0x25a9, 0x267f, 0x2758, 0x2833, + 0x2911, 0x29f1, 0x2ad4, 0x2bb9, 0x2ca0, 0x2d8a, 0x2e76, 0x2f64, + 0x3053, 0x3145, 0x3238, 0x332d, 0x3424, 0x351b, 0x3615, 0x370f, + 0x380b, 0x3907, 0x3a04, 0x3b03, 0x3c01, 0x3d01, 0x3e00, 0x3f00, +}; + +const q15_t sigmoidHTable_q15[192] = { + 0x70be, 0x7190, 0x7258, 0x7316, 0x73cc, 0x7478, 0x751b, 0x75b7, + 0x764a, 0x76d6, 0x775b, 0x77d8, 0x784f, 0x78c0, 0x792a, 0x798f, + 0x79ee, 0x7a48, 0x7a9d, 0x7aed, 0x7b39, 0x7b80, 0x7bc4, 0x7c03, + 0x7c3f, 0x7c78, 0x7cad, 0x7ce0, 0x7d0f, 0x7d3c, 0x7d66, 0x7d8d, + 0x7db3, 0x7dd6, 0x7df7, 0x7e16, 0x7e33, 0x7e4f, 0x7e69, 0x7e81, + 0x7e98, 0x7eae, 0x7ec2, 0x7ed5, 0x7ee7, 0x7ef8, 0x7f08, 0x7f17, + 0x7f25, 0x7f32, 0x7f3e, 0x7f4a, 0x7f55, 0x7f5f, 0x7f69, 0x7f72, + 0x7f7b, 0x7f83, 0x7f8a, 0x7f91, 0x7f98, 0x7f9e, 0x7fa4, 0x7faa, + 0x7faf, 0x7fb4, 0x7fb8, 0x7fbd, 0x7fc1, 0x7fc5, 0x7fc8, 0x7fcc, + 0x7fcf, 0x7fd2, 0x7fd5, 0x7fd7, 0x7fda, 0x7fdc, 0x7fde, 0x7fe0, + 0x7fe2, 0x7fe4, 0x7fe6, 0x7fe7, 0x7fe9, 0x7fea, 0x7feb, 0x7fed, + 0x7fee, 0x7fef, 0x7ff0, 0x7ff1, 0x7ff2, 0x7ff3, 0x7ff4, 0x7ff4, + 0x000b, 0x000c, 0x000c, 0x000d, 0x000e, 0x000f, 0x0010, 0x0011, + 0x0012, 0x0013, 0x0015, 0x0016, 0x0017, 0x0019, 0x001a, 0x001c, + 0x001e, 0x0020, 0x0022, 0x0024, 0x0026, 0x0029, 0x002b, 0x002e, + 0x0031, 0x0034, 0x0038, 0x003b, 0x003f, 0x0043, 0x0048, 0x004c, + 0x0051, 0x0056, 0x005c, 0x0062, 0x0068, 0x006f, 0x0076, 0x007d, + 0x0085, 0x008e, 0x0097, 0x00a1, 0x00ab, 0x00b6, 0x00c2, 0x00ce, + 0x00db, 0x00e9, 0x00f8, 0x0108, 0x0119, 0x012b, 0x013e, 0x0152, + 0x0168, 0x017f, 0x0197, 0x01b1, 0x01cd, 0x01ea, 0x0209, 0x022a, + 0x024d, 0x0273, 0x029a, 0x02c4, 0x02f1, 0x0320, 0x0353, 0x0388, + 0x03c1, 0x03fd, 0x043c, 0x0480, 0x04c7, 0x0513, 0x0563, 0x05b8, + 0x0612, 0x0671, 0x06d6, 0x0740, 0x07b1, 0x0828, 0x08a5, 0x092a, + 0x09b6, 0x0a49, 0x0ae5, 0x0b88, 0x0c34, 0x0cea, 0x0da8, 0x0e70, +}; + +const q7_t tanhTable_q7[256] = { + 0x00, 0x08, 0x10, 0x18, 0x1f, 0x27, 0x2e, 0x35, + 0x3b, 0x41, 0x47, 0x4c, 0x51, 0x56, 0x5a, 0x5e, + 0x61, 0x65, 0x68, 0x6a, 0x6d, 0x6f, 0x71, 0x72, + 0x74, 0x75, 0x76, 0x78, 0x78, 0x79, 0x7a, 0x7b, + 0x7b, 0x7c, 0x7c, 0x7d, 0x7d, 0x7e, 0x7e, 0x7e, + 0x7e, 0x7e, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, + 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x81, + 0x81, 0x81, 0x81, 0x81, 0x81, 0x81, 0x81, 0x82, + 0x82, 0x82, 0x82, 0x82, 0x83, 0x83, 0x84, 0x84, + 0x85, 0x85, 0x86, 0x87, 0x88, 0x88, 0x8a, 0x8b, + 0x8c, 0x8e, 0x8f, 0x91, 0x93, 0x96, 0x98, 0x9b, + 0x9f, 0xa2, 0xa6, 0xaa, 0xaf, 0xb4, 0xb9, 0xbf, + 0xc5, 0xcb, 0xd2, 0xd9, 0xe1, 0xe8, 0xf0, 0xf8, +}; + +const q15_t tanhTable_q15[256] = { + 0x0000, 0x07fd, 0x0feb, 0x17b9, 0x1f59, 0x26bf, 0x2ddf, 0x34ae, + 0x3b27, 0x4142, 0x46fd, 0x4c56, 0x514d, 0x55e2, 0x5a1a, 0x5df6, + 0x617c, 0x64b0, 0x6797, 0x6a37, 0x6c95, 0x6eb5, 0x709e, 0x7254, + 0x73dc, 0x753a, 0x7672, 0x7788, 0x787f, 0x795b, 0x7a1e, 0x7acb, + 0x7b65, 0x7bee, 0x7c66, 0x7cd1, 0x7d30, 0x7d84, 0x7dce, 0x7e0f, + 0x7e49, 0x7e7d, 0x7eaa, 0x7ed2, 0x7ef5, 0x7f14, 0x7f30, 0x7f48, + 0x7f5e, 0x7f71, 0x7f82, 0x7f91, 0x7f9e, 0x7fa9, 0x7fb3, 0x7fbc, + 0x7fc4, 0x7fcb, 0x7fd1, 0x7fd7, 0x7fdc, 0x7fe0, 0x7fe4, 0x7fe7, + 0x7fea, 0x7fed, 0x7fef, 0x7ff1, 0x7ff3, 0x7ff4, 0x7ff6, 0x7ff7, + 0x7ff8, 0x7ff9, 0x7ffa, 0x7ffa, 0x7ffb, 0x7ffc, 0x7ffc, 0x7ffd, + 0x7ffd, 0x7ffd, 0x7ffe, 0x7ffe, 0x7ffe, 0x7ffe, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001, + 0x8001, 0x8001, 0x8001, 0x8002, 0x8002, 0x8002, 0x8002, 0x8003, + 0x8003, 0x8003, 0x8004, 0x8004, 0x8005, 0x8006, 0x8006, 0x8007, + 0x8008, 0x8009, 0x800a, 0x800c, 0x800d, 0x800f, 0x8011, 0x8013, + 0x8016, 0x8019, 0x801c, 0x8020, 0x8024, 0x8029, 0x802f, 0x8035, + 0x803c, 0x8044, 0x804d, 0x8057, 0x8062, 0x806f, 0x807e, 0x808f, + 0x80a2, 0x80b8, 0x80d0, 0x80ec, 0x810b, 0x812e, 0x8156, 0x8183, + 0x81b7, 0x81f1, 0x8232, 0x827c, 0x82d0, 0x832f, 0x839a, 0x8412, + 0x849b, 0x8535, 0x85e2, 0x86a5, 0x8781, 0x8878, 0x898e, 0x8ac6, + 0x8c24, 0x8dac, 0x8f62, 0x914b, 0x936b, 0x95c9, 0x9869, 0x9b50, + 0x9e84, 0xa20a, 0xa5e6, 0xaa1e, 0xaeb3, 0xb3aa, 0xb903, 0xbebe, + 0xc4d9, 0xcb52, 0xd221, 0xd941, 0xe0a7, 0xe847, 0xf015, 0xf803, +}; + +const q15_t tanhLTable_q15[128] = { + 0x0000, 0x0400, 0x07fd, 0x0bf7, 0x0feb, 0x13d7, 0x17b9, 0x1b90, + 0x1f59, 0x2314, 0x26bf, 0x2a58, 0x2ddf, 0x3151, 0x34ae, 0x37f6, + 0x3b27, 0x3e40, 0x4142, 0x442c, 0x46fd, 0x49b6, 0x4c56, 0x4edd, + 0x514d, 0x53a3, 0x55e2, 0x580a, 0x5a1a, 0x5c13, 0x5df6, 0x5fc4, + 0x617c, 0x6320, 0x64b0, 0x662d, 0x6797, 0x68f0, 0x6a37, 0x6b6e, + 0x6c95, 0x6dac, 0x6eb5, 0x6fb0, 0x709e, 0x717f, 0x7254, 0x731e, + 0x73dc, 0x7490, 0x753a, 0x75da, 0x7672, 0x7701, 0x7788, 0x7807, + 0x787f, 0x78f0, 0x795b, 0x79bf, 0x7a1e, 0x7a77, 0x7acb, 0x7b1b, + 0x849b, 0x84e5, 0x8535, 0x8589, 0x85e2, 0x8641, 0x86a5, 0x8710, + 0x8781, 0x87f9, 0x8878, 0x88ff, 0x898e, 0x8a26, 0x8ac6, 0x8b70, + 0x8c24, 0x8ce2, 0x8dac, 0x8e81, 0x8f62, 0x9050, 0x914b, 0x9254, + 0x936b, 0x9492, 0x95c9, 0x9710, 0x9869, 0x99d3, 0x9b50, 0x9ce0, + 0x9e84, 0xa03c, 0xa20a, 0xa3ed, 0xa5e6, 0xa7f6, 0xaa1e, 0xac5d, + 0xaeb3, 0xb123, 0xb3aa, 0xb64a, 0xb903, 0xbbd4, 0xbebe, 0xc1c0, + 0xc4d9, 0xc80a, 0xcb52, 0xceaf, 0xd221, 0xd5a8, 0xd941, 0xdcec, + 0xe0a7, 0xe470, 0xe847, 0xec29, 0xf015, 0xf409, 0xf803, 0xfc00, +}; + +const q15_t tanhHTable_q15[192] = { + 0x7b65, 0x7bee, 0x7c66, 0x7cd1, 0x7d30, 0x7d84, 0x7dce, 0x7e0f, + 0x7e49, 0x7e7d, 0x7eaa, 0x7ed2, 0x7ef5, 0x7f14, 0x7f30, 0x7f48, + 0x7f5e, 0x7f71, 0x7f82, 0x7f91, 0x7f9e, 0x7fa9, 0x7fb3, 0x7fbc, + 0x7fc4, 0x7fcb, 0x7fd1, 0x7fd7, 0x7fdc, 0x7fe0, 0x7fe4, 0x7fe7, + 0x7fea, 0x7fed, 0x7fef, 0x7ff1, 0x7ff3, 0x7ff4, 0x7ff6, 0x7ff7, + 0x7ff8, 0x7ff9, 0x7ffa, 0x7ffa, 0x7ffb, 0x7ffc, 0x7ffc, 0x7ffd, + 0x7ffd, 0x7ffd, 0x7ffe, 0x7ffe, 0x7ffe, 0x7ffe, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, 0x7fff, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001, 0x8001, + 0x8001, 0x8001, 0x8001, 0x8002, 0x8002, 0x8002, 0x8002, 0x8003, + 0x8003, 0x8003, 0x8004, 0x8004, 0x8005, 0x8006, 0x8006, 0x8007, + 0x8008, 0x8009, 0x800a, 0x800c, 0x800d, 0x800f, 0x8011, 0x8013, + 0x8016, 0x8019, 0x801c, 0x8020, 0x8024, 0x8029, 0x802f, 0x8035, + 0x803c, 0x8044, 0x804d, 0x8057, 0x8062, 0x806f, 0x807e, 0x808f, + 0x80a2, 0x80b8, 0x80d0, 0x80ec, 0x810b, 0x812e, 0x8156, 0x8183, + 0x81b7, 0x81f1, 0x8232, 0x827c, 0x82d0, 0x832f, 0x839a, 0x8412, +}; + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c new file mode 100644 index 0000000..212880a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c @@ -0,0 +1,125 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_q7_to_q15_no_shift.c + * Description: Converts the elements of the Q7 vector to Q15 vector without left-shift + * + * $Date: May 29, 2020 + * $Revision: V.1.0.2 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup nndata_convert + * @{ + */ + +/** + * @brief Converts the elements of the Q7 vector to Q15 vector without left-shift + * @param[in] *pSrc points to the Q7 input vector + * @param[out] *pDst points to the Q15 output vector + * @param[in] blockSize length of the input vector + * + * \par Description: + * + * The equation used for the conversion process is: + * + * 
+ * 	pDst[n] = (q15_t) pSrc[n];   0 <= n < blockSize.
+ *
+ * + */ + +void arm_q7_to_q15_no_shift(const q7_t * pSrc, q15_t * pDst, uint32_t blockSize) +{ + const q7_t *pIn = pSrc; + uint32_t blkCnt; + +#if defined(ARM_MATH_DSP) + q31_t in; + q31_t in1, in2; + q31_t out1, out2; + + /*loop Unrolling */ + blkCnt = blockSize >> 2u; + + /* First part of the processing with loop unrolling. Compute 4 outputs at a time. */ + while (blkCnt > 0u) + { + in = arm_nn_read_q7x4_ia(&pIn); + + /* rotatate in by 8 and extend two q7_t values to q15_t values */ + in1 = __SXTB16(__ROR((uint32_t)in, 8)); + + /* extend remaining two q7_t values to q15_t values */ + in2 = __SXTB16(in); + +#ifndef ARM_MATH_BIG_ENDIAN + out2 = (int32_t)__PKHTB(in1, in2, 16); + out1 = (int32_t)__PKHBT(in2, in1, 16); +#else + out1 = (int32_t)__PKHTB(in1, in2, 16); + out2 = (int32_t)__PKHBT(in2, in1, 16); +#endif + write_q15x2_ia(&pDst, out1); + write_q15x2_ia(&pDst, out2); + + /* Decrement the loop counter */ + blkCnt--; + } + + /* If the blockSize is not a multiple of 4, compute any remaining output samples here. + ** No loop unrolling is used. */ + blkCnt = blockSize % 0x4u; + +#else + + /* Run the below code for Cortex-M0 */ + + /* Loop over blockSize number of values */ + blkCnt = blockSize; + +#endif /* #ifndef ARM_MATH_CM0_FAMILY */ + + while (blkCnt > 0u) + { + /* convert from q7 to q15 and then store the results in the destination buffer */ + *pDst++ = (q15_t)*pIn++; + + /* Decrement the loop counter */ + blkCnt--; + } + +} + +/** + * @} end of nndata_convert group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c new file mode 100644 index 0000000..6d35f50 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c @@ -0,0 +1,147 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_q7_to_q15_reordered_no_shift.c + * Description: Converts the elements of the Q7 vector to reordered Q15 vector without left-shift + * + * $Date: May 29, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup nndata_convert + * @{ + */ + +/** + * @brief Converts the elements of the Q7 vector to reordered Q15 vector without left-shift + * @param[in] *pSrc points to the Q7 input vector + * @param[out] *pDst points to the Q15 output vector + * @param[in] blockSize length of the input vector + * + * @details + * + * This function does the q7 to q15 expansion with re-ordering + * + * 
+ *                          |   A1   |   A2   |   A3   |   A4   |
+ *
+ *                           0      7 8     15 16    23 24    31
+ *
+ * + * is converted into: + * + * 
+ *  |       A1       |       A3       |   and  |       A2       |       A4       |
+ *
+ *   0             15 16            31          0             15 16            31
+ *
+ * + * + * This looks strange but is natural considering how sign-extension is done at + * assembly level. + * + * The expansion of other other oprand will follow the same rule so that the end + * results are the same. + * + * The tail (i.e., last (N % 4) elements) will still be in original order. + * + */ + +void arm_q7_to_q15_reordered_no_shift(const q7_t * pSrc, q15_t * pDst, uint32_t blockSize) +{ + const q7_t *pIn = pSrc; /* Src pointer */ + uint32_t blkCnt; /* loop counter */ + +#ifndef ARM_MATH_CM0_FAMILY + q31_t in; + q31_t in1, in2; + + /* Run the below code for Cortex-M4 and Cortex-M3 */ + + /*loop Unrolling */ + blkCnt = blockSize >> 2u; + + /* First part of the processing with loop unrolling. Compute 4 outputs at a time. + ** a second loop below computes the remaining 1 to 3 samples. */ + while (blkCnt > 0u) + { + /* C = (q15_t) A << 8 */ + /* convert from q7 to q15 and then store the results in the destination buffer */ + in = arm_nn_read_q7x4_ia(&pIn); + + /* rotatate in by 8 and extend two q7_t values to q15_t values */ + in1 = __SXTB16(__ROR((uint32_t)in, 8)); + + /* extend remainig two q7_t values to q15_t values */ + in2 = __SXTB16(in); + +#ifndef ARM_MATH_BIG_ENDIAN + *__SIMD32(pDst)++ = in2; + *__SIMD32(pDst)++ = in1; +#else + *__SIMD32(pDst)++ = in1; + *__SIMD32(pDst)++ = in2; +#endif + + /* Decrement the loop counter */ + blkCnt--; + } + + /* If the blockSize is not a multiple of 4, compute any remaining output samples here. + ** No loop unrolling is used. */ + blkCnt = blockSize % 0x4u; + +#else + + /* Run the below code for Cortex-M0 */ + + /* Loop over blockSize number of values */ + blkCnt = blockSize; + +#endif /* #ifndef ARM_MATH_CM0_FAMILY */ + + while (blkCnt > 0u) + { + /* C = (q15_t) A << 8 */ + /* convert from q7 to q15 and then store the results in the destination buffer */ + *pDst++ = (q15_t) * pIn++; + + /* Decrement the loop counter */ + blkCnt--; + } + +} + +/** + * @} end of q7_to_x group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_with_offset.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_with_offset.c new file mode 100644 index 0000000..567417d --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_with_offset.c @@ -0,0 +1,103 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_q7_to_q15_reordered_with_offset.c + * Description: Converts the elements of the Q7 vector to a reordered Q15 vector with an added offset. The re-ordering + * is a signature of sign extension intrinsic(DSP extension). + * + * $Date: May 29, 2020 + * $Revision: V.2.0.3 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup nndata_convert + * @{ + */ + +/** + * @brief Converts the elements of the Q7 vector to a reordered Q15 vector with an added offset. + * + * @note Refer header file for details. + * + */ + +void arm_q7_to_q15_reordered_with_offset(const q7_t *src, q15_t *dst, uint32_t block_size, q15_t offset) +{ + +#if defined(ARM_MATH_DSP) + uint32_t block_cnt; + /* Run the below code for cores that support SIMD instructions */ + q31_t in_q7x4; + q31_t out_q15x2_1; + q31_t out_q15x2_2; + + /*loop unrolling */ + block_cnt = block_size >> 2u; + + /* First part of the processing with loop unrolling. Compute 4 outputs at a time. */ + const q31_t offset_q15x2 = (q31_t)__PKHBT(offset, offset, 16); + while (block_cnt > 0u) + { + /* convert from q7 to q15 and then store the results in the destination buffer */ + in_q7x4 = arm_nn_read_q7x4_ia(&src); + + /* Extract and sign extend each of the four q7 values to q15 */ + out_q15x2_1 = __SXTAB16(offset_q15x2, __ROR((uint32_t)in_q7x4, 8)); + out_q15x2_2 = __SXTAB16(offset_q15x2, in_q7x4); + + write_q15x2_ia(&dst, out_q15x2_2); + write_q15x2_ia(&dst, out_q15x2_1); + + block_cnt--; + } + /* Handle left over samples */ + block_cnt = block_size % 0x4u; + + while (block_cnt > 0u) + { + *dst++ = (q15_t)*src++ + offset; + + /* Decrement the loop counter */ + block_cnt--; + } +#else + (void)src; + (void)dst; + (void)block_size; + (void)offset; + /* Not available */ +#endif +} + +/** + * @} end of nndata_convert group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_with_offset.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_with_offset.c new file mode 100644 index 0000000..9e0867c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_with_offset.c @@ -0,0 +1,120 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in_q7x4 compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in_q7x4 writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_q7_to_q15_with_offset.c + * Description: Converts the elements of the Q7 vector to Q15 vector with an added offset + * + * $Date: March 3, 2020 + * $Revision: V.2.0.2 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupSupport + */ + +/** + * @addtogroup nndata_convert + * @{ + */ + +void arm_q7_to_q15_with_offset(const q7_t *src, + q15_t *dst, + uint32_t block_size, + q15_t offset) +{ + int block_cnt; + +#if defined(ARM_MATH_MVEI) + + int16x8_t source; + const int16x8_t source_offset = vdupq_n_s16(offset); + block_cnt = block_size / 8; + + while (block_cnt > 0) + { + source = vldrbq_s16(src); + source = vaddq_s16(source, source_offset); + vstrhq_s16(dst, source); + dst += 8; + src += 8; + block_cnt--; + } + + block_cnt = block_size & 0x7; + +#elif defined(ARM_MATH_DSP) + /* Run the below code for cores that support SIMD instructions */ + q31_t in_q7x4; + q31_t in_q15x2_1; + q31_t in_q15x2_2; + q31_t out_q15x2_1; + q31_t out_q15x2_2; + + /*loop unrolling */ + block_cnt = block_size >> 2; + + /* First part of the processing with loop unrolling. Compute 4 outputs at a time. */ + const q31_t offset_q15x2 = __PKHBT(offset, offset, 16); + while (block_cnt > 0) + { + /* convert from q7 to q15 and then store the results in the destination buffer */ + in_q7x4 = arm_nn_read_q7x4_ia(&src); + + /* Extract and sign extend each of the four q7 values to q15 */ + in_q15x2_1 = __SXTAB16(offset_q15x2, __ROR(in_q7x4, 8)); + in_q15x2_2 = __SXTAB16(offset_q15x2, in_q7x4); + + out_q15x2_2 = __PKHTB(in_q15x2_1, in_q15x2_2, 16); + out_q15x2_1 = __PKHBT(in_q15x2_2, in_q15x2_1, 16); + + write_q15x2_ia(&dst, out_q15x2_1); + write_q15x2_ia(&dst, out_q15x2_2); + + block_cnt--; + } + /* Handle left over samples */ + block_cnt = block_size % 0x4; + +#else + /* Run the below code for Cortex-M0 */ + /* Loop over block_size number of values */ + block_cnt = block_size; +#endif + + while (block_cnt > 0) + { + *dst++ = (q15_t)*src++ + offset; + + /* Decrement the loop counter */ + block_cnt--; + } +} + +/** + * @} end of nndata_convert group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_avgpool_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_avgpool_s8.c new file mode 100644 index 0000000..761b95b --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_avgpool_s8.c @@ -0,0 +1,366 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_avgpool_s8.c + * Description: Pooling function implementations + * + * $Date: September 3,2020 + * $Revision: V.2.0.2 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +#if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI) + +static void scale_q31_to_q7_and_clamp(const q31_t *buffer, + q7_t *target, + int32_t length, + const int32_t count, + const int act_min, + const int act_max) +{ + const int half_count = count / 2; + for (int i = 0; i < length; i++) + { + int32_t sum = buffer[i] > 0 ? (buffer[i] + half_count) : (buffer[i] - half_count); + sum = sum / count; + sum = MAX(sum, act_min); + sum = MIN(sum, act_max); + + target[i] = (q7_t)sum; + } +} +#endif + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Pooling + * @{ + */ + +/* + * s8 average pooling function + * + * Refer to header file for details. + * + */ + +#if defined(ARM_MATH_MVEI) + +arm_status arm_avgpool_s8(const cmsis_nn_context *ctx, + const cmsis_nn_pool_params *pool_params, + const cmsis_nn_dims *input_dims, + const q7_t *src, + const cmsis_nn_dims *filter_dims, + const cmsis_nn_dims *output_dims, + q7_t *dst) +{ + (void)ctx; + const int32_t input_y = input_dims->h; + const int32_t input_x = input_dims->w; + const int32_t output_y = output_dims->h; + const int32_t output_x = output_dims->w; + const int32_t stride_y = pool_params->stride.h; + const int32_t stride_x = pool_params->stride.w; + const int32_t kernel_y = filter_dims->h; + const int32_t kernel_x = filter_dims->w; + const int32_t pad_y = pool_params->padding.h; + const int32_t pad_x = pool_params->padding.w; + const int32_t act_min = pool_params->activation.min; + const int32_t act_max = pool_params->activation.max; + const int32_t ch_src = input_dims->c; + + int32_t i_x, i_y; + int32_t k_x, k_y; + + for (i_y = 0; i_y < output_y; i_y++) + { + for (i_x = 0; i_x < output_x; i_x++) + { + + int32_t k_y_start, k_y_end; + int32_t k_x_start, k_x_end; + int32_t chCnt; + const int8_t *pTmp, *pTmpInner; + int8_t *pDst; + + k_y_start = MAX(0, i_y * stride_y - pad_y); + k_y_end = MIN(i_y * stride_y - pad_y + kernel_y, input_y); + + k_x_start = MAX(0, i_x * stride_x - pad_x); + k_x_end = MIN(i_x * stride_x - pad_x + kernel_x, input_x); + + pTmp = src; + pDst = &dst[ch_src * (i_x + i_y * output_x)]; + + chCnt = ch_src >> 4; + while (chCnt > 0) + { + int32x4_t sumV1, sumV2, sumV3, sumV4; + + int8x16_t tempV; + int16x8_t tempVLO, tempVHI; + int32x4_t tempVLOLO, tempVLOHI, tempVHILO, tempVHIHI; + int32_t count = 0; + + sumV1 = vdupq_n_s32(0); + sumV2 = vdupq_n_s32(0); + sumV3 = vdupq_n_s32(0); + sumV4 = vdupq_n_s32(0); + + for (k_y = k_y_start; k_y < k_y_end; k_y++) + { + for (k_x = k_x_start; k_x < k_x_end; k_x++) + { + pTmpInner = pTmp + (ch_src * (k_x + k_y * input_x)); + tempV = vldrbq_s8(pTmpInner); + + tempVLO = vmovlbq_s8(tempV); + tempVHI = vmovltq_s8(tempV); + + tempVLOLO = vmovlbq_s16(tempVLO); + tempVLOHI = vmovltq_s16(tempVLO); + + tempVHILO = vmovlbq_s16(tempVHI); + tempVHIHI = vmovltq_s16(tempVHI); + + sumV1 = vaddq_s32(sumV1, tempVLOLO); + sumV2 = vaddq_s32(sumV2, tempVLOHI); + sumV3 = vaddq_s32(sumV3, tempVHILO); + sumV4 = vaddq_s32(sumV4, tempVHIHI); + + count++; + } + } + + sumV1[0] = sumV1[0] > 0 ? (sumV1[0] + count / 2) / count : (sumV1[0] - count / 2) / count; + sumV1[1] = sumV1[1] > 0 ? (sumV1[1] + count / 2) / count : (sumV1[1] - count / 2) / count; + sumV1[2] = sumV1[2] > 0 ? (sumV1[2] + count / 2) / count : (sumV1[2] - count / 2) / count; + sumV1[3] = sumV1[3] > 0 ? (sumV1[3] + count / 2) / count : (sumV1[3] - count / 2) / count; + + sumV2[0] = sumV2[0] > 0 ? (sumV2[0] + count / 2) / count : (sumV2[0] - count / 2) / count; + sumV2[1] = sumV2[1] > 0 ? (sumV2[1] + count / 2) / count : (sumV2[1] - count / 2) / count; + sumV2[2] = sumV2[2] > 0 ? (sumV2[2] + count / 2) / count : (sumV2[2] - count / 2) / count; + sumV2[3] = sumV2[3] > 0 ? (sumV2[3] + count / 2) / count : (sumV2[3] - count / 2) / count; + + sumV3[0] = sumV3[0] > 0 ? (sumV3[0] + count / 2) / count : (sumV3[0] - count / 2) / count; + sumV3[1] = sumV3[1] > 0 ? (sumV3[1] + count / 2) / count : (sumV3[1] - count / 2) / count; + sumV3[2] = sumV3[2] > 0 ? (sumV3[2] + count / 2) / count : (sumV3[2] - count / 2) / count; + sumV3[3] = sumV3[3] > 0 ? (sumV3[3] + count / 2) / count : (sumV3[3] - count / 2) / count; + + sumV4[0] = sumV4[0] > 0 ? (sumV4[0] + count / 2) / count : (sumV4[0] - count / 2) / count; + sumV4[1] = sumV4[1] > 0 ? (sumV4[1] + count / 2) / count : (sumV4[1] - count / 2) / count; + sumV4[2] = sumV4[2] > 0 ? (sumV4[2] + count / 2) / count : (sumV4[2] - count / 2) / count; + sumV4[3] = sumV4[3] > 0 ? (sumV4[3] + count / 2) / count : (sumV4[3] - count / 2) / count; + + sumV1 = vmaxq_s32(sumV1, vdupq_n_s32(act_min)); + sumV1 = vminq_s32(sumV1, vdupq_n_s32(act_max)); + + sumV2 = vmaxq_s32(sumV2, vdupq_n_s32(act_min)); + sumV2 = vminq_s32(sumV2, vdupq_n_s32(act_max)); + + sumV3 = vmaxq_s32(sumV3, vdupq_n_s32(act_min)); + sumV3 = vminq_s32(sumV3, vdupq_n_s32(act_max)); + + sumV4 = vmaxq_s32(sumV4, vdupq_n_s32(act_min)); + sumV4 = vminq_s32(sumV4, vdupq_n_s32(act_max)); + + tempVLO = vmovnbq_s32(tempVLO, sumV1); + tempVLO = vmovntq_s32(tempVLO, sumV2); + + tempVHI = vmovnbq_s32(tempVHI, sumV3); + tempVHI = vmovntq_s32(tempVHI, sumV4); + + tempV = vmovnbq_s16(tempV, tempVLO); + tempV = vmovntq_s16(tempV, tempVHI); + + vstrbq_s8(pDst, tempV); + pDst += 16; + + chCnt--; + pTmp += 16; + } + + chCnt = ch_src & 0xF; + while (chCnt > 0) + { + int32_t sum = 0; + int32_t count = 0; + + for (k_y = k_y_start; k_y < k_y_end; k_y++) + { + for (k_x = k_x_start; k_x < k_x_end; k_x++) + { + sum += pTmp[ch_src * (k_x + k_y * input_x)]; + count++; + } + } + sum = sum > 0 ? (sum + count / 2) / count : (sum - count / 2) / count; + sum = MAX(sum, act_min); + sum = MIN(sum, act_max); + + *pDst++ = sum; + + chCnt--; + pTmp++; + } + } + } + return ARM_MATH_SUCCESS; +} + +#else +arm_status arm_avgpool_s8(const cmsis_nn_context *ctx, + const cmsis_nn_pool_params *pool_params, + const cmsis_nn_dims *input_dims, + const q7_t *src, + const cmsis_nn_dims *filter_dims, + const cmsis_nn_dims *output_dims, + q7_t *dst) +{ + const int32_t input_y = input_dims->h; + const int32_t input_x = input_dims->w; + const int32_t output_y = output_dims->h; + const int32_t output_x = output_dims->w; + const int32_t stride_y = pool_params->stride.h; + const int32_t stride_x = pool_params->stride.w; + const int32_t kernel_y = filter_dims->h; + const int32_t kernel_x = filter_dims->w; + const int32_t pad_y = pool_params->padding.h; + const int32_t pad_x = pool_params->padding.w; + const int32_t act_min = pool_params->activation.min; + const int32_t act_max = pool_params->activation.max; + const int32_t ch_src = input_dims->c; + q31_t *buffer = (q31_t *)ctx->buf; + +#if defined(ARM_MATH_DSP) + + /* Run the following code for CPU's with DSP extension + */ + for (int i_y = 0, idx_y = -pad_y; i_y < output_y; idx_y += stride_y, i_y++) + { + for (int i_x = 0, idx_x = -pad_x; i_x < output_x; idx_x += stride_x, i_x++) + { + /* Condition for kernel start dimension: + (base_idx_ + kernel__start) >= 0 */ + const int32_t kernel_y_start = MAX(0, -idx_y); + const int32_t kernel_x_start = MAX(0, -idx_x); + + /* Condition for kernel end dimension: + (base_idx_ + kernel__end) < dim_src_ */ + const int32_t kernel_y_end = MIN(kernel_y, input_y - idx_y); + const int32_t kernel_x_end = MIN(kernel_x, input_x - idx_x); + + int count = 0; + + for (int k_y = kernel_y_start; k_y < kernel_y_end; k_y++) + { + for (int k_x = kernel_x_start; k_x < kernel_x_end; k_x++) + { + const q7_t *start = src + ch_src * (k_x + idx_x + (k_y + idx_y) * input_x); + + if (count == 0) + { + for (int i = 0; i < ch_src; i++) + { + buffer[i] = start[i]; + } + } + else + { + for (int i = 0; i < ch_src; i++) + { + buffer[i] = __QADD(start[i], buffer[i]); + } + } + count++; + } + } + scale_q31_to_q7_and_clamp(buffer, dst, ch_src, count, act_min, act_max); + dst += ch_src; + } + } +#else + + /* Reference C code adapted from CMSIS-NN arm_avepool_q7_HWC. + */ + (void)buffer; + int16_t i_ch_in, i_x, i_y; + int16_t k_x, k_y; + + for (i_y = 0; i_y < output_y; i_y++) + { + for (i_x = 0; i_x < output_x; i_x++) + { + for (i_ch_in = 0; i_ch_in < ch_src; i_ch_in++) + { + int sum = 0; + int count = 0; + for (k_y = i_y * stride_y - pad_y; k_y < i_y * stride_y - pad_y + kernel_y; k_y++) + { + for (k_x = i_x * stride_x - pad_x; k_x < i_x * stride_x - pad_x + kernel_x; k_x++) + { + if (k_y >= 0 && k_x >= 0 && k_y < input_y && k_x < input_x) + { + sum += src[i_ch_in + ch_src * (k_x + k_y * input_x)]; + count++; + } + } + } + sum = sum > 0 ? (sum + count / 2) / count : (sum - count / 2) / count; + sum = MAX(sum, act_min); + sum = MIN(sum, act_max); + + dst[i_ch_in + ch_src * (i_x + i_y * output_x)] = sum; + } + } + } + +#endif + return ARM_MATH_SUCCESS; +} + +#endif /* ARM_MATH_MVEI */ + +int32_t arm_avgpool_s8_get_buffer_size(const int output_x, + const int ch_src) +{ + (void)output_x; + +#if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI) + return (ch_src * sizeof(int32_t)); +#else + (void)ch_src; + return 0; +#endif +} +/** + * @} end of Pooling group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_max_pool_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_max_pool_s8.c new file mode 100644 index 0000000..31f2071 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_max_pool_s8.c @@ -0,0 +1,238 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_max_pool_s8.c + * Description: Pooling function implementations + * + * $Date: June 11, 2020 + * $Revision: V.2.0.0 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +static void compare_and_replace_if_larger_q7(q7_t *base, + const q7_t *target, + int32_t length) +{ +#if defined(ARM_MATH_MVEI) + int32_t loop_count = (length + 15) / 16; + for (int i = 0; i < loop_count; i++) + { + mve_pred16_t p = vctp16q((uint32_t)length); + const int8x16_t op_1 = vldrbq_z_s8(base, p); + const int8x16_t op_2 = vldrbq_z_s8(target, p); + const int8x16_t max = vmaxq_m_s8(vuninitializedq_s8(), op_1, op_2, p); + vstrbq_p_s8(base, max, p); + base += 16; + target += 16; + length -= 16; + } +#else + q7_t *dst = base; + const q7_t *src = target; + union arm_nnword ref_max; + union arm_nnword comp_max; + int32_t cnt = length >> 2; + + while (cnt > 0l) + { + ref_max.word = arm_nn_read_q7x4(dst); + comp_max.word = arm_nn_read_q7x4_ia(&src); + + if (comp_max.bytes[0] > ref_max.bytes[0]) + { + ref_max.bytes[0] = comp_max.bytes[0]; + } + if (comp_max.bytes[1] > ref_max.bytes[1]) + { + ref_max.bytes[1] = comp_max.bytes[1]; + } + if (comp_max.bytes[2] > ref_max.bytes[2]) + { + ref_max.bytes[2] = comp_max.bytes[2]; + } + if (comp_max.bytes[3] > ref_max.bytes[3]) + { + ref_max.bytes[3] = comp_max.bytes[3]; + } + + write_q7x4_ia(&dst, ref_max.word); + + cnt--; + } + + cnt = length & 0x3; + while (cnt > 0l) + { + if (*src > *dst) + { + *dst = *src; + } + dst++; + src++; + cnt--; + } +#endif +} + +static void +clamp_output(q7_t *source, int32_t length, const int32_t act_min, const int32_t act_max) +{ +#if defined(ARM_MATH_MVEI) + int32_t + loop_count = (length + 15) / 16; + for (int i = 0; i < loop_count; i++) + { + mve_pred16_t p = vctp16q((uint32_t)length); + length -= 16; + const int8x16_t src = vldrbq_z_s8(source, p); + const int8x16_t predicated_min = vdupq_m_n_s8(vuninitializedq_s8(), (int8_t)act_min, p); + const int8x16_t predicated_max = vdupq_m_n_s8(vuninitializedq_s8(), (int8_t)act_max, p); + int8x16_t + res = vmaxq_m_s8(vuninitializedq_s8(), src, predicated_min, p); + res = vminq_m_s8(vuninitializedq_s8(), src, predicated_max, p); + vstrbq_p_s8(source, res, p); + source += 16; + } +#else + union arm_nnword in; + int32_t cnt = length >> 2; + + while (cnt > 0l) + { + in.word = arm_nn_read_q7x4(source); + + in.bytes[0] = MAX(in.bytes[0], act_min); + in.bytes[0] = MIN(in.bytes[0], act_max); + in.bytes[1] = MAX(in.bytes[1], act_min); + in.bytes[1] = MIN(in.bytes[1], act_max); + in.bytes[2] = MAX(in.bytes[2], act_min); + in.bytes[2] = MIN(in.bytes[2], act_max); + in.bytes[3] = MAX(in.bytes[3], act_min); + in.bytes[3] = MIN(in.bytes[3], act_max); + + write_q7x4_ia(&source, in.word); + cnt--; + } + + cnt = length & 0x3; + while (cnt > 0l) + { + int32_t comp = *source; + comp = MAX(comp, act_min); + comp = MIN(comp, act_max); + *source++ = (int8_t)comp; + cnt--; + } +#endif +} + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Pooling + * @{ + */ + +/* + * Optimized s8 max pooling function + * + * Refer to header file for details. + * + */ + +arm_status +arm_max_pool_s8(const cmsis_nn_context *ctx, + const cmsis_nn_pool_params *pool_params, + const cmsis_nn_dims *input_dims, + const q7_t *src, + const cmsis_nn_dims *filter_dims, + const cmsis_nn_dims *output_dims, + q7_t *dst) +{ + const int32_t input_y = input_dims->h; + const int32_t input_x = input_dims->w; + const int32_t output_y = output_dims->h; + const int32_t output_x = output_dims->w; + const int32_t stride_y = pool_params->stride.h; + const int32_t stride_x = pool_params->stride.w; + const int32_t kernel_y = filter_dims->h; + const int32_t kernel_x = filter_dims->w; + const int32_t pad_y = pool_params->padding.h; + const int32_t pad_x = pool_params->padding.w; + const int32_t act_min = pool_params->activation.min; + const int32_t act_max = pool_params->activation.max; + const int32_t channel_in = input_dims->c; + (void)ctx; + q7_t *dst_base = dst; + + for (int i_y = 0, base_idx_y = -pad_y; i_y < output_y; base_idx_y += stride_y, i_y++) + { + for (int i_x = 0, base_idx_x = -pad_x; i_x < output_x; base_idx_x += stride_x, i_x++) + { + /* Condition for kernel start dimension: (base_idx_ + kernel__start) >= 0 */ + const int32_t ker_y_start = MAX(0, -base_idx_y); + const int32_t ker_x_start = MAX(0, -base_idx_x); + + /* Condition for kernel end dimension: (base_idx_ + kernel__end) < dim_src_ */ + const int32_t kernel_y_end = MIN(kernel_y, input_y - base_idx_y); + const int32_t kernel_x_end = MIN(kernel_x, input_x - base_idx_x); + + int count = 0; + + for (int k_y = ker_y_start; k_y < kernel_y_end; k_y++) + { + for (int k_x = ker_x_start; k_x < kernel_x_end; k_x++) + { + const q7_t *start = src + channel_in * (k_x + base_idx_x + (k_y + base_idx_y) * input_x); + + if (count == 0) + { + memcpy(dst, start, channel_in); + count++; + } + else + { + compare_and_replace_if_larger_q7(dst, start, channel_in); + } + } + } + /* 'count' is expected to be non-zero here. */ + dst += channel_in; + } + } + + clamp_output(dst_base, output_x * output_y * channel_in, act_min, act_max); + + return ARM_MATH_SUCCESS; +} + +/** + * @} end of Pooling group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c new file mode 100644 index 0000000..7649b0a --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c @@ -0,0 +1,457 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_pool_q7_HWC.c + * Description: Pooling function implementations + * + * $Date: 17. January 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +#if defined (ARM_MATH_DSP) + +/** + * @brief A few utility functions used by pooling functions + * + * + */ + +static void buffer_scale_back_q15_to_q7(q15_t * buffer, q7_t * target, uint16_t length, uint16_t scale) +{ + int i; + + for (i = 0; i < length; i++) + { + target[i] = (q7_t) (buffer[i] / scale); + } +} + +static void compare_and_replace_if_larger_q7(q7_t * base, // base data + const q7_t * target, // compare target + const uint16_t length // data length + ) +{ + q7_t *pIn = base; + const q7_t *pCom = target; + union arm_nnword in; + union arm_nnword com; + uint16_t cnt = length >> 2; + + while (cnt > 0u) + { + in.word = arm_nn_read_q7x4((const q7_t*)pIn); + com.word = arm_nn_read_q7x4_ia((const q7_t**)&pCom); + + // if version + if (com.bytes[0] > in.bytes[0]) + in.bytes[0] = com.bytes[0]; + if (com.bytes[1] > in.bytes[1]) + in.bytes[1] = com.bytes[1]; + if (com.bytes[2] > in.bytes[2]) + in.bytes[2] = com.bytes[2]; + if (com.bytes[3] > in.bytes[3]) + in.bytes[3] = com.bytes[3]; + + *__SIMD32(pIn)++ = in.word; + + cnt--; + } + + cnt = length & 0x3; + while (cnt > 0u) + { + if (*pCom > *pIn) + { + *pIn = *pCom; + } + pIn++; + pCom++; + cnt--; + } +} + +static void accumulate_q7_to_q15(q15_t * base, q7_t * target, const uint16_t length) +{ + q15_t *pCnt = base; + q7_t *pV = target; + q31_t v1, v2, vo1, vo2; + uint16_t cnt = length >> 2; + q31_t in; + + while (cnt > 0u) + { + q31_t value = arm_nn_read_q7x4_ia((const q7_t**)&pV); + v1 = __SXTB16(__ROR(value, 8)); + v2 = __SXTB16(value); +#ifndef ARM_MATH_BIG_ENDIAN + + vo2 = __PKHTB(v1, v2, 16); + vo1 = __PKHBT(v2, v1, 16); + +#else + + vo1 = __PKHTB(v1, v2, 16); + vo2 = __PKHBT(v2, v1, 16); + +#endif + + in = arm_nn_read_q15x2(pCnt); + *__SIMD32(pCnt)++ = __QADD16(vo1, in); + + in = arm_nn_read_q15x2(pCnt); + *__SIMD32(pCnt)++ = __QADD16(vo2, in); + + cnt--; + } + cnt = length & 0x3; + while (cnt > 0u) + { + *pCnt++ += *pV++; + cnt--; + } +} + +#endif // ARM_MATH_DSP + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Pooling + * @{ + */ + + /** + * @brief Q7 max pooling function + * @param[in, out] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA Not used + * @param[in,out] Im_out pointer to output tensor + * + * @details + * + * The pooling function is implemented as split x-pooling then + * y-pooling. + * + * This pooling function is input-destructive. Input data is undefined + * after calling this function. + * + */ + +void +arm_maxpool_q7_HWC(q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, const uint16_t dim_im_out, q7_t * bufferA, q7_t * Im_out) +{ + (void)bufferA; +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + int16_t i_x, i_y; + + /* first does the pooling along x axis */ + for (i_y = 0; i_y < dim_im_in; i_y++) + { + + for (i_x = 0; i_x < dim_im_out; i_x++) + { + /* for each output pixel */ + q7_t *target = Im_in + (i_y * dim_im_in + i_x) * ch_im_in; + q7_t *win_start; + q7_t *win_stop; + if (i_x * stride - padding < 0) + { + win_start = target; + } else + { + win_start = Im_in + (i_y * dim_im_in + i_x * stride - padding) * ch_im_in; + } + + if (i_x * stride - padding + dim_kernel >= dim_im_in) + { + win_stop = Im_in + (i_y * dim_im_in + dim_im_in) * ch_im_in; + } else + { + win_stop = Im_in + (i_y * dim_im_in + i_x * stride - padding + dim_kernel) * ch_im_in; + } + + /* first step is to copy over initial data */ + /* arm_copy_q7(win_start, target, ch_im_in); */ + memmove(target, win_start, ch_im_in); + + /* start the max operation from the second part */ + win_start += ch_im_in; + for (; win_start < win_stop; win_start += ch_im_in) + { + compare_and_replace_if_larger_q7(target, win_start, ch_im_in); + } + } + } + + /* then does the pooling along y axis */ + for (i_y = 0; i_y < dim_im_out; i_y++) + { + + /* for each output row */ + q7_t *target = Im_out + i_y * dim_im_out * ch_im_in; + q7_t *row_start; + q7_t *row_end; + /* setting the starting row */ + if (i_y * stride - padding < 0) + { + row_start = Im_in; + } else + { + row_start = Im_in + (i_y * stride - padding) * dim_im_in * ch_im_in; + } + /* setting the stopping row */ + if (i_y * stride - padding + dim_kernel >= dim_im_in) + { + row_end = Im_in + dim_im_in * dim_im_in * ch_im_in; + } else + { + row_end = Im_in + (i_y * stride - padding + dim_kernel) * dim_im_in * ch_im_in; + } + + /* copy over the first row */ + /* arm_copy_q7(row_start, target, dim_im_out * ch_im_in); */ + memmove(target, row_start, dim_im_out * ch_im_in); + + /* move over to next row */ + row_start += ch_im_in * dim_im_in; + + for (; row_start < row_end; row_start += dim_im_in * ch_im_in) + { + compare_and_replace_if_larger_q7(target, row_start, dim_im_out * ch_im_in); + } + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + int16_t i_ch_in, i_x, i_y; + int16_t k_x, k_y; + + for (i_ch_in = 0; i_ch_in < ch_im_in; i_ch_in++) + { + for (i_y = 0; i_y < dim_im_out; i_y++) + { + for (i_x = 0; i_x < dim_im_out; i_x++) + { + int max = -129; + for (k_y = i_y * stride - padding; k_y < i_y * stride - padding + dim_kernel; k_y++) + { + for (k_x = i_x * stride - padding; k_x < i_x * stride - padding + dim_kernel; k_x++) + { + if (k_y >= 0 && k_x >= 0 && k_y < dim_im_in && k_x < dim_im_in) + { + if (Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)] > max) + { + max = Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)]; + } + } + } + } + Im_out[i_ch_in + ch_im_in * (i_x + i_y * dim_im_out)] = max; + } + } + } + +#endif /* ARM_MATH_DSP */ + +} + + /** + * @brief Q7 average pooling function + * @param[in,out] Im_in pointer to input tensor + * @param[in] dim_im_in input tensor dimention + * @param[in] ch_im_in number of input tensor channels + * @param[in] dim_kernel filter kernel size + * @param[in] padding padding sizes + * @param[in] stride convolution stride + * @param[in] dim_im_out output tensor dimension + * @param[in,out] bufferA pointer to buffer space for input + * @param[in,out] Im_out pointer to output tensor + * + * @details + * + * Buffer size: + * + * bufferA size: 2*dim_im_out*ch_im_in + * + * The pooling function is implemented as split x-pooling then + * y-pooling. + * + * This pooling function is input-destructive. Input data is undefined + * after calling this function. + * + */ + +void +arm_avepool_q7_HWC(q7_t * Im_in, + const uint16_t dim_im_in, + const uint16_t ch_im_in, + const uint16_t dim_kernel, + const uint16_t padding, + const uint16_t stride, const uint16_t dim_im_out, q7_t * bufferA, q7_t * Im_out) +{ + +#if defined (ARM_MATH_DSP) + /* Run the following code for Cortex-M4 and Cortex-M7 */ + + q15_t *buffer = (q15_t *) bufferA; + int16_t i_x, i_y; + int16_t count = 0; + + /* first does the pooling along x axis */ + for (i_y = 0; i_y < dim_im_in; i_y++) + { + + for (i_x = 0; i_x < dim_im_out; i_x++) + { + /* for each output pixel */ + q7_t *target = Im_in + (i_y * dim_im_in + i_x) * ch_im_in; + q7_t *win_start; + q7_t *win_stop; + if (i_x * stride - padding < 0) + { + win_start = target; + } else + { + win_start = Im_in + (i_y * dim_im_in + i_x * stride - padding) * ch_im_in; + } + + if (i_x * stride - padding + dim_kernel >= dim_im_in) + { + win_stop = Im_in + (i_y * dim_im_in + dim_im_in) * ch_im_in; + } else + { + win_stop = Im_in + (i_y * dim_im_in + i_x * stride - padding + dim_kernel) * ch_im_in; + } + + /* first step is to copy over initial data */ + arm_q7_to_q15_no_shift(win_start, buffer, ch_im_in); + count = 1; + + /* start the max operation from the second part */ + win_start += ch_im_in; + for (; win_start < win_stop; win_start += ch_im_in) + { + accumulate_q7_to_q15(buffer, win_start, ch_im_in); + count++; + } + buffer_scale_back_q15_to_q7(buffer, target, ch_im_in, count); + } + } + + /* then does the pooling along y axis */ + for (i_y = 0; i_y < dim_im_out; i_y++) + { + /* for each output row */ + q7_t *target = Im_out + i_y * dim_im_out * ch_im_in; + q7_t *row_start; + q7_t *row_end; + /* setting the starting row */ + if (i_y * stride - padding < 0) + { + row_start = Im_in; + } else + { + row_start = Im_in + (i_y * stride - padding) * dim_im_in * ch_im_in; + } + /* setting the stopping row */ + if (i_y * stride - padding + dim_kernel >= dim_im_in) + { + row_end = Im_in + dim_im_in * dim_im_in * ch_im_in; + } else + { + row_end = Im_in + (i_y * stride - padding + dim_kernel) * dim_im_in * ch_im_in; + } + + /* copy over the first row */ + arm_q7_to_q15_no_shift(row_start, buffer, dim_im_out * ch_im_in); + count = 1; + + /* move over to next row */ + row_start += ch_im_in * dim_im_in; + + for (; row_start < row_end; row_start += dim_im_in * ch_im_in) + { + accumulate_q7_to_q15(buffer, row_start, dim_im_out * ch_im_in); + count++; + } + buffer_scale_back_q15_to_q7(buffer, target, dim_im_out * ch_im_in, count); + } + +#else + /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */ + + (void)bufferA; + int16_t i_ch_in, i_x, i_y; + int16_t k_x, k_y; + + for (i_ch_in = 0; i_ch_in < ch_im_in; i_ch_in++) + { + for (i_y = 0; i_y < dim_im_out; i_y++) + { + for (i_x = 0; i_x < dim_im_out; i_x++) + { + int sum = 0; + int count = 0; + for (k_y = i_y * stride - padding; k_y < i_y * stride - padding + dim_kernel; k_y++) + { + for (k_x = i_x * stride - padding; k_x < i_x * stride - padding + dim_kernel; k_x++) + { + if (k_y >= 0 && k_x >= 0 && k_y < dim_im_in && k_x < dim_im_in) + { + sum += Im_in[i_ch_in + ch_im_in * (k_x + k_y * dim_im_in)]; + count++; + } + } + } + Im_out[i_ch_in + ch_im_in * (i_x + i_y * dim_im_out)] = sum / count; + } + } + } + +#endif /* ARM_MATH_DSP */ + +} + +/** + * @} end of Pooling group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ReshapeFunctions/arm_reshape_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ReshapeFunctions/arm_reshape_s8.c new file mode 100644 index 0000000..199ab56 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ReshapeFunctions/arm_reshape_s8.c @@ -0,0 +1,60 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_reshape_s8.c + * Description: Reshape a s8 vector + * + * $Date: September 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Reshape + * @{ + */ + +/** + * Basic s8 reshape function. + * + * Refer header file for details. + * + */ + +void arm_reshape_s8(const int8_t *input, + int8_t *output, + const uint32_t total_size) +{ + memcpy(output, input, total_size); +} + +/** + * @} end of Reshape group + */ +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SVDFunctions/arm_svdf_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SVDFunctions/arm_svdf_s8.c new file mode 100644 index 0000000..bbdaee5 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SVDFunctions/arm_svdf_s8.c @@ -0,0 +1,228 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_svdf_s8.c + * Description: S8 basic SVDF layer function + * + * $Date: 17. August 2020 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M processors + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nn_types.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup SVDF + * @{ + */ + +/* + * S8 SVDF layer function for TensorFlow Lite + * + * Refer to header file for details. + * + */ + +arm_status +arm_svdf_s8(const cmsis_nn_context *input_ctx, + const cmsis_nn_context *output_ctx, + const cmsis_nn_svdf_params *svdf_params, + const cmsis_nn_per_tensor_quant_params *input_quant_params, + const cmsis_nn_per_tensor_quant_params *output_quant_params, + const cmsis_nn_dims *input_dims, + const q7_t *input_data, + const cmsis_nn_dims *state_dims, + q15_t *state_data, + const cmsis_nn_dims *weights_feature_dims, + const q7_t *weights_feature_data, + const cmsis_nn_dims *weights_time_dims, + const q15_t *weights_time_data, + const cmsis_nn_dims *bias_dims, + const q31_t *bias_data, + const cmsis_nn_dims *output_dims, + q7_t *output_data) +{ + (void)bias_dims; + (void)state_dims; + (void)output_dims; + + const q31_t multiplier_in = input_quant_params->multiplier; + const q31_t shift_in = input_quant_params->shift; + const q31_t multiplier_out = output_quant_params->multiplier; + const q31_t shift_2 = output_quant_params->shift; + const int32_t zp_in = svdf_params->input_offset; + const int32_t zp_out = svdf_params->output_offset; + const int32_t in_activation_min = svdf_params->input_activation.min; + const int32_t in_activation_max = svdf_params->input_activation.max; + const int32_t out_activation_min = svdf_params->output_activation.min; + const int32_t out_activation_max = svdf_params->output_activation.max; + const int16_t rank = svdf_params->rank; + + int32_t zp_32 = (-zp_in & 0xffff) | + ((-zp_in & 0xffff) << 16); + + const int32_t input_batches = input_dims->n; + const int32_t input_height = input_dims->h; + const int32_t feature_batches = weights_feature_dims->n; + const int32_t time_batches = weights_time_dims->h; + const int32_t unit_count = feature_batches / rank; + + q31_t *buffer_a = (q31_t *)input_ctx->buf; + q31_t *buffer_b = (q31_t *)output_ctx->buf; + + memmove((q15_t *)state_data, (q15_t *)state_data + 1, + (size_t)(input_batches * feature_batches * time_batches * (int32_t)sizeof(int16_t))); + + q15_t *res_ptr = state_data + (time_batches - 1); + for (int i_batch = 0; i_batch < input_batches; i_batch++) + { + const q7_t *buffer_1 = weights_feature_data; + for (int r = 0; r < feature_batches; r++) + { + q31_t dot_prod = 0; + + const q7_t *buffer_2 = input_data + i_batch * input_height; + +#if defined(ARM_MATH_DSP) + int c = 0; + int32_t block_count = input_height >> 2; + for (int i = 0; i < block_count; i++) + { + c += 4; + + q31_t r1 = arm_nn_read_q7x4_ia(&buffer_1); + q31_t r1_a = __SXTB16(r1); + q31_t r1_b = __SXTB16(__ROR((uint32_t)r1, 8)); + + q31_t r2 = arm_nn_read_q7x4_ia(&buffer_2); + q31_t r2_a = __SXTAB16(zp_32, r2); + q31_t r2_b = __SXTAB16(zp_32, __ROR((uint32_t)r2, 8)); + + dot_prod = __SMLAD(r1_a, r2_a, dot_prod); + dot_prod = __SMLAD(r1_b, r2_b, dot_prod); + } + + for (; c < input_height; c++) + { + dot_prod += *buffer_1 * (*buffer_2 - zp_in); + buffer_1++; + buffer_2++; + } +#else + for (int c = 0; c < input_height; c++) + { + dot_prod += *buffer_1 * (*buffer_2 - zp_in); + buffer_1++; + buffer_2++; + } +#endif + + dot_prod = arm_nn_requantize(dot_prod, + multiplier_in, + shift_in); + dot_prod = CLAMP(dot_prod, in_activation_max, in_activation_min); + *res_ptr = dot_prod; + res_ptr += time_batches; + } + } + + for (int i_batch = 0; i_batch < input_batches; i_batch++) + { + q31_t *ptr_a = buffer_a + i_batch * feature_batches; + + const q15_t *v1 = weights_time_data; + const q15_t *v2 = state_data + i_batch * time_batches * feature_batches; + for (int i_feature_batch = 0; i_feature_batch < feature_batches; i_feature_batch++) + { + *ptr_a = 0; + + int32_t sum = 0; +#if defined(ARM_MATH_DSP) + int j = 0; + int32_t block_count = time_batches >> 1; + for (int i = 0; i < block_count; i++) + { + j += 2; + q31_t r1 = arm_nn_read_q15x2_ia(&v1); + q31_t r2 = arm_nn_read_q15x2_ia(&v2); + + sum = __SMLAD(r1, r2, sum); + } + + // Process the remaining data + for (; j < time_batches; j++) + { + sum += *v1 * *v2; + v1++; + v2++; + } +#else + for (int j = 0; j < time_batches; j++) + { + sum += *v1 * *v2; + v1++; + v2++; + } +#endif + + *ptr_a = sum; + ptr_a++; + } + } + + for (int i_batch = 0; i_batch < input_batches; i_batch++) + { + q31_t *output_data_temp = buffer_b + i_batch * unit_count; + q31_t *ptr_a = buffer_a + i_batch * feature_batches; + + for (int i = 0; i < unit_count; i++) + { + output_data_temp[i] = bias_data[i]; + for (int j = 0; j < rank; j++) + { + output_data_temp[i] += *ptr_a; + ptr_a++; + } + } + } + + for (int i = 0; i < input_batches * unit_count; i++) + { + output_data[i] = (q7_t)CLAMP(arm_nn_requantize(buffer_b[i], multiplier_out, shift_2) + zp_out, + out_activation_max, out_activation_min); + } + + return (ARM_MATH_SUCCESS); +} + +/** + * @} end of SVDF group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c new file mode 100644 index 0000000..65fb649 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c @@ -0,0 +1,122 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_softmax_q15.c + * Description: Q15 softmax function + * + * $Date: 20. February 2018 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Softmax + * @{ + */ + + /** + * @brief Q15 softmax function + * @param[in] vec_in pointer to input vector + * @param[in] dim_vec input vector dimention + * @param[out] p_out pointer to output vector + * + * @details + * + * Here, instead of typical e based softmax, we use + * 2-based softmax, i.e.,: + * + * y_i = 2^(x_i) / sum(2^x_j) + * + * The relative output will be different here. + * But mathematically, the gradient will be the same + * with a log(2) scaling factor. + * + */ + +void arm_softmax_q15(const q15_t * vec_in, const uint16_t dim_vec, q15_t * p_out) +{ + q31_t sum; + int16_t i; + uint8_t shift; + q31_t base; + base = -1 * 0x100000; + for (i = 0; i < dim_vec; i++) + { + if (vec_in[i] > base) + { + base = vec_in[i]; + } + } + + /* we ignore really small values + * anyway, they will be 0 after shrinking + * to q15_t + */ + base = base - 16; + + sum = 0; + + for (i = 0; i < dim_vec; i++) + { + if (vec_in[i] > base) + { + shift = (uint8_t)__USAT(vec_in[i] - base, 5); + sum += 0x1 << shift; + } + } + + /* This is effectively (0x1 << 32) / sum */ + int64_t div_base = 0x100000000LL; + int output_base = (int32_t)(div_base / sum); + + /* Final confidence will be output_base >> ( 17 - (vec_in[i] - base) ) + * so 32768 (0x1<<15) -> 100% confidence when sum = 0x1 << 16, output_base = 0x1 << 16 + * and vec_in[i]-base = 16 + */ + for (i = 0; i < dim_vec; i++) + { + if (vec_in[i] > base) + { + /* Here minimum value of 17+base-vec[i] will be 1 */ + shift = (uint8_t)__USAT(17+base-vec_in[i], 5); + p_out[i] = (q15_t) __SSAT((output_base >> shift), 16); + } else + { + p_out[i] = 0; + } + } + +} + +/** + * @} end of Softmax group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c new file mode 100644 index 0000000..8e30cf8 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c @@ -0,0 +1,111 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_softmax_q7.c + * Description: Q7 softmax function + * + * $Date: June 8, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Softmax + * @{ + */ + + /** + * @brief Q7 softmax function + * @param[in] vec_in pointer to input vector + * @param[in] dim_vec input vector dimention + * @param[out] p_out pointer to output vector + * + * @details + * + * Here, instead of typical natural logarithm e based softmax, we use + * 2-based softmax here, i.e.,: + * + * y_i = 2^(x_i) / sum(2^x_j) + * + * The relative output will be different here. + * But mathematically, the gradient will be the same + * with a log(2) scaling factor. + * + */ + +void arm_softmax_q7(const q7_t * vec_in, const uint16_t dim_vec, q7_t * p_out ) +{ + q31_t sum; + int16_t i; + uint8_t shift; + q15_t base; + base = -128; + + /* We first search for the maximum */ + for (i = 0; i < dim_vec; i++) + { + if (vec_in[i] > base) + { + base = vec_in[i]; + } + } + + /* + * So the base is set to max-8, meaning + * that we ignore really small values. + * anyway, they will be 0 after shrinking to q7_t. + */ + base = base - (1 << 3); + + sum = 0; + + for (i = 0; i < dim_vec; i++) + { + shift = (uint8_t)__USAT(vec_in[i] - base, 3); + sum += 0x1 << shift; + } + + /* This is effectively (0x1 << 20) / sum */ + int output_base = (1 << 20) / sum; + + for (i = 0; i < dim_vec; i++) + { + + /* Here minimum value of 13+base-vec_in[i] will be 5 */ + shift = (uint8_t)__USAT(13 + base - vec_in[i], 5); + p_out[i] = (q7_t)__SSAT((output_base >> shift), 8); + } +} + +/** + * @} end of Softmax group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c new file mode 100644 index 0000000..07a4b75 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c @@ -0,0 +1,260 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_softmax_s8.c + * Description: S8 softmax function + * + * $Date: April 6, 2020 + * $Revision: V.2.0.0 + * + * Target Processor: Cortex-M cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnsupportfunctions.h" + +#define ACCUM_BITS 12 + +#ifdef ARM_MATH_MVEI +static int32x4_t arm_exp_on_negative_values_mve_32x4(int32x4_t val) +{ +#define SHIFT_START (24) + int32_t shift = SHIFT_START; + int32x4_t mask; + + const int32x4_t val_mod_minus_quarter = vandq_s32(val, vdupq_n_s32((1 << SHIFT_START) - 1)) - vdupq_n_s32(1 << SHIFT_START); + const int32x4_t remainder = vsubq_s32(val_mod_minus_quarter, val); + const int32x4_t x = vaddq_n_s32(val_mod_minus_quarter << 5, 1 << 28); + const int32x4_t x2 = MUL_SAT_MVE(x, x); + const int32x4_t op_1 = DIV_POW2_MVE(MUL_SAT_MVE(x2, x2), 2) + MUL_SAT_MVE(x2, x); + const int32x4_t op_2 = x + DIV_POW2_MVE(MUL_SAT_MVE(op_1, vdupq_n_s32(715827883)) + x2, 1); + int32x4_t result = vdupq_n_s32(1895147668) + MUL_SAT_MVE(vdupq_n_s32(1895147668), op_2); + +#define SELECT_IF_NON_ZERO(x) \ + { \ + mve_pred16_t p = vcmpneq_n_s32(remainder & vdupq_n_s32(1 << shift++), 0); \ + mask = vmvnq_m_s32(vdupq_n_s32(0), vdupq_n_s32(0), p); \ + result = SELECT_USING_MASK(mask, MUL_SAT_MVE(result, vdupq_n_s32(x)), result); \ + } + + SELECT_IF_NON_ZERO(1672461947) + SELECT_IF_NON_ZERO(1302514674) + SELECT_IF_NON_ZERO(790015084) + SELECT_IF_NON_ZERO(290630308) + SELECT_IF_NON_ZERO(39332535) + SELECT_IF_NON_ZERO(720401) + SELECT_IF_NON_ZERO(242) + +#undef SELECT_IF_NON_ZERO + + mve_pred16_t p = vcmpeqq_n_s32(val, 0); + mask = vmvnq_m_s32(vdupq_n_s32(0), vdupq_n_s32(0), p); + + result = SELECT_USING_MASK(mask, vdupq_n_s32(Q31_MAX), result); + return result; +} +#endif + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Softmax + * @{ + */ + +void arm_softmax_s8(const int8_t *input, + const int32_t num_rows, + const int32_t row_size, + const int32_t mult, + const int32_t shift, + const int32_t diff_min, + int8_t *output) +{ +#ifdef ARM_MATH_MVEI + +#define ACT_MIN ((int8_t)Q7_MIN) +#define ACT_MAX ((int8_t)Q7_MAX) + + const int32_t mask = (1 << shift); + + for (int i_num_rows = 0; i_num_rows < num_rows; ++i_num_rows) + { + int8_t max = ACT_MIN; + + int32_t vec_count = (row_size + 15) / 16; + uint32_t r_count = (uint32_t)row_size; + for (int i = 0; i < vec_count; i++) + { + mve_pred16_t p = vctp8q(r_count); + const int8x16_t ip = vldrbq_z_s8(&input[i * 16], p); + max = vmaxvq_p_s8(max, ip, p); + r_count -= 16; + } + + vec_count = row_size / 4; + int32_t idx = 0; + int32_t sum = 0; + + while (vec_count) + { + int32x4_t ip = vldrbq_s32(&input[idx * 4]); + ip = vsubq_n_s32(ip, max); + mve_pred16_t p = vcmpgeq_n_s32(ip, diff_min); + if (p != 0) + { + ip = vmulq_n_s32(ip, mask); + + int32x4_t res = MUL_SAT_MVE(ip, vdupq_n_s32(mult)); + + res = arm_exp_on_negative_values_mve_32x4(res); + res = DIV_POW2_MVE(res, ACCUM_BITS); + res = vpselq_s32(res, vdupq_n_s32(0), p); + sum += vaddvq_s32(res); + } + + vec_count--; + idx++; + } + + const int32_t tail_idx = row_size & ~3; + for (int i = 0; i < (row_size & 3); i++) + { + const int32_t diff = input[tail_idx + i] - max; + if (diff >= diff_min) + { + sum += DIV_POW2(EXP_ON_NEG(MUL_SAT(diff * mask, mult)), ACCUM_BITS); + } + } + + const int32_t headroom = __CLZ((uint32_t)sum); + const int32_t bits_over_unit = ACCUM_BITS - headroom + 23; + const int32_t shifted_scale = ONE_OVER1((sum << headroom) - (1 << 31)); + + vec_count = row_size / 4; + idx = 0; + + while (vec_count) + { + int32x4_t ip = vldrbq_s32(&input[idx]); + ip = vsubq_n_s32(ip, max); + + mve_pred16_t p = vcmpgeq_n_s32(ip, diff_min); + + int32x4_t tmp_res; + + if (p != 0) + { + ip = vmulq_n_s32(ip, mask); + + tmp_res = MUL_SAT_MVE(ip, vdupq_n_s32(mult)); + tmp_res = arm_exp_on_negative_values_mve_32x4(tmp_res); + tmp_res = MUL_SAT_MVE(vdupq_n_s32(shifted_scale), tmp_res); + tmp_res = DIV_POW2_MVE(tmp_res, bits_over_unit); + tmp_res += vdupq_n_s32(ACT_MIN); + + tmp_res = vmaxq_s32(tmp_res, vdupq_n_s32(ACT_MIN)); + tmp_res = vminq_s32(tmp_res, vdupq_n_s32(ACT_MAX)); + tmp_res = vpselq_s32(tmp_res, vdupq_n_s32(ACT_MIN), p); + } + else + { + tmp_res = vdupq_n_s32(ACT_MIN); + } + vstrbq_s32(&output[idx], tmp_res); + vec_count--; + idx += 4; + } + + for (int i = 0; i < (row_size & 3); i++) + { + int32_t diff = input[tail_idx + i] - max; + if (diff >= diff_min) + { + const int32_t res = DIV_POW2(MUL_SAT(shifted_scale, EXP_ON_NEG(MUL_SAT(diff * mask, mult))), bits_over_unit) - 128; + output[tail_idx + i] = (int8_t)CLAMP(res, (int32_t)ACT_MAX, (int32_t)ACT_MIN); + } + else + { + output[tail_idx + i] = ACT_MIN; + } + } + + input += row_size; + output += row_size; + } +#else + const int32_t mask = (1 << shift); + + int32_t col = 0; + int32_t row_idx; + + for (row_idx = 0; row_idx < num_rows; ++row_idx) + { + // Find the maximum value in order to ensure numerical stability + int8_t max = *input; + + for (col = 1; col < row_size; ++col) + { + max = MAX(max, input[col]); + } + + int32_t diff = 0; + int32_t sum = 0; + + for (col = 0; col < row_size; ++col) + { + diff = input[col] - max; + if (diff >= diff_min) + { + sum += DIV_POW2(EXP_ON_NEG(MUL_SAT(diff * mask, mult)), ACCUM_BITS); + } + } + + const int32_t headroom = __CLZ(sum); + const int32_t bits_over_unit = ACCUM_BITS - headroom + 23; + const int32_t shifted_scale = ONE_OVER1((sum << headroom) - (1 << 31)); + + for (col = 0; col < row_size; ++col) + { + diff = input[col] - max; + if (diff >= diff_min) + { + const int32_t res = DIV_POW2(MUL_SAT(shifted_scale, EXP_ON_NEG(MUL_SAT(diff * mask, mult))), bits_over_unit) - 128; + output[col] = (int8_t)CLAMP(res, (int32_t)127, (int32_t)-128); + } + else + { + output[col] = -128; + } + } + input += row_size; + output += row_size; + } + +#endif +} +/** + * @} end of Softmax group + */ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c new file mode 100644 index 0000000..5dffb6f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c @@ -0,0 +1,103 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2020 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_softmax_u8.c + * Description: U8 softmax function + * + * $Date: May 29, 2020 + * $Revision: V.1.0.1 + * + * Target Processor: Cortex-M CPUs + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +#define ACCUM_BITS 12 + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Softmax + * @{ + */ +void arm_softmax_u8(const uint8_t *input, + const int32_t num_rows, + const int32_t row_size, + const int32_t mult, + const int32_t shift, + const int32_t diff_min, + uint8_t *output) +{ + const int32_t mask = (1 << shift); + + int32_t col = 0; + int32_t row_idx; + + for(row_idx = 0; row_idx < num_rows; ++row_idx) + { + // Find the maximum value in order to ensure numerical stability + uint8_t max = *input; + + for (col = 1; col < row_size; ++col) + { + max = MAX(max, input[col]); + } + + int32_t diff = 0; + int32_t sum = 0; + + for (col = 0; col < row_size; ++col) + { + diff = input[col] - max; + if(diff >= diff_min) + { + sum += DIV_POW2(EXP_ON_NEG(MUL_SAT(diff * mask, mult)), ACCUM_BITS); + } + } + + const int32_t headroom = __CLZ((uint32_t)sum); + const int32_t bits_over_unit = ACCUM_BITS - headroom + 23; + const int32_t shifted_scale = ONE_OVER1((sum << headroom) - (1 << 31)); + + for (col = 0; col < row_size; ++col) + { + diff = input[col] - max; + if (diff >= diff_min) + { + const int32_t res = DIV_POW2(MUL_SAT(shifted_scale, EXP_ON_NEG(MUL_SAT(diff * mask, mult))), bits_over_unit); + output[col] = (uint8_t) CLAMP(res, (int32_t)255, (int32_t)0); + } + else + { + output[col] = 0; + } + } + input += row_size; + output += row_size; + } +} +/** + * @} end of Softmax group + */ +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c new file mode 100644 index 0000000..622f167 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c @@ -0,0 +1,78 @@ +#if !defined(ESP32) +/* + * Copyright (C) 2010-2019 Arm Limited or its affiliates. All rights reserved. + * + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the License); you may + * not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an AS IS BASIS, WITHOUT + * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* ---------------------------------------------------------------------- + * Project: CMSIS NN Library + * Title: arm_softmax_with_batch_q7.c + * Description: Q7 softmax function + * + * $Date: 05. August 2019 + * $Revision: V.1.0.0 + * + * Target Processor: Cortex-M and Cortex-A cores + * + * -------------------------------------------------------------------- */ + +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include/arm_math.h" +#include "tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include/arm_nnfunctions.h" + +/** + * @ingroup groupNN + */ + +/** + * @addtogroup Softmax + * @{ + */ + + /** + * @brief Q7 softmax function with batch parameter + * @param[in] vec_in pointer to input vector + * @param[in] nb_batches number of batches + * @param[in] dim_vec input vector dimention + * @param[out] p_out pointer to output vector + * + * @details + * + * Here, instead of typical natural logarithm e based softmax, we use + * 2-based softmax here, i.e.,: + * + * y_i = 2^(x_i) / sum(2^x_j) + * + * The relative output will be different here. + * But mathematically, the gradient will be the same + * with a log(2) scaling factor. + * + */ + +void arm_softmax_with_batch_q7(const q7_t * vec_in, const uint16_t nb_batches,const uint16_t dim_vec, q7_t * p_out ) +{ + for(int i=0; i + +#define MAXFACTORS 32 +/* e.g. an fft of length 128 has 4 factors + as far as kissfft is concerned + 4*4*4*2 + */ + +struct kiss_fft_state{ + int nfft; + int inverse; + int factors[2*MAXFACTORS]; + kiss_fft_cpx twiddles[1]; +}; + +/* + Explanation of macros dealing with complex math: + + C_MUL(m,a,b) : m = a*b + C_FIXDIV( c , div ) : if a fixed point impl., c /= div. noop otherwise + C_SUB( res, a,b) : res = a - b + C_SUBFROM( res , a) : res -= a + C_ADDTO( res , a) : res += a + * */ +#ifdef FIXED_POINT +#if (FIXED_POINT==32) +# define FRACBITS 31 +# define SAMPPROD int64_t +#define SAMP_MAX 2147483647 +#else +# define FRACBITS 15 +# define SAMPPROD int32_t +#define SAMP_MAX 32767 +#endif + +#define SAMP_MIN -SAMP_MAX + +#if defined(CHECK_OVERFLOW) +# define CHECK_OVERFLOW_OP(a,op,b) \ + if ( (SAMPPROD)(a) op (SAMPPROD)(b) > SAMP_MAX || (SAMPPROD)(a) op (SAMPPROD)(b) < SAMP_MIN ) { \ + fprintf(stderr,"WARNING:overflow @ " __FILE__ "(%d): (%d " #op" %d) = %ld\n",__LINE__,(a),(b),(SAMPPROD)(a) op (SAMPPROD)(b) ); } +#endif + + +# define smul(a,b) ( (SAMPPROD)(a)*(b) ) +# define sround( x ) (kiss_fft_scalar)( ( (x) + (1<<(FRACBITS-1)) ) >> FRACBITS ) + +# define S_MUL(a,b) sround( smul(a,b) ) + +# define C_MUL(m,a,b) \ + do{ (m).r = sround( smul((a).r,(b).r) - smul((a).i,(b).i) ); \ + (m).i = sround( smul((a).r,(b).i) + smul((a).i,(b).r) ); }while(0) + +# define DIVSCALAR(x,k) \ + (x) = sround( smul( x, SAMP_MAX/k ) ) + +# define C_FIXDIV(c,div) \ + do { DIVSCALAR( (c).r , div); \ + DIVSCALAR( (c).i , div); }while (0) + +# define C_MULBYSCALAR( c, s ) \ + do{ (c).r = sround( smul( (c).r , s ) ) ;\ + (c).i = sround( smul( (c).i , s ) ) ; }while(0) + +#else /* not FIXED_POINT*/ + +# define S_MUL(a,b) ( (a)*(b) ) +#define C_MUL(m,a,b) \ + do{ (m).r = (a).r*(b).r - (a).i*(b).i;\ + (m).i = (a).r*(b).i + (a).i*(b).r; }while(0) +# define C_FIXDIV(c,div) /* NOOP */ +# define C_MULBYSCALAR( c, s ) \ + do{ (c).r *= (s);\ + (c).i *= (s); }while(0) +#endif + +#ifndef CHECK_OVERFLOW_OP +# define CHECK_OVERFLOW_OP(a,op,b) /* noop */ +#endif + +#define C_ADD( res, a,b)\ + do { \ + CHECK_OVERFLOW_OP((a).r,+,(b).r)\ + CHECK_OVERFLOW_OP((a).i,+,(b).i)\ + (res).r=(a).r+(b).r; (res).i=(a).i+(b).i; \ + }while(0) +#define C_SUB( res, a,b)\ + do { \ + CHECK_OVERFLOW_OP((a).r,-,(b).r)\ + CHECK_OVERFLOW_OP((a).i,-,(b).i)\ + (res).r=(a).r-(b).r; (res).i=(a).i-(b).i; \ + }while(0) +#define C_ADDTO( res , a)\ + do { \ + CHECK_OVERFLOW_OP((res).r,+,(a).r)\ + CHECK_OVERFLOW_OP((res).i,+,(a).i)\ + (res).r += (a).r; (res).i += (a).i;\ + }while(0) + +#define C_SUBFROM( res , a)\ + do {\ + CHECK_OVERFLOW_OP((res).r,-,(a).r)\ + CHECK_OVERFLOW_OP((res).i,-,(a).i)\ + (res).r -= (a).r; (res).i -= (a).i; \ + }while(0) + + +#ifdef FIXED_POINT +# define KISS_FFT_COS(phase) floor(.5+SAMP_MAX * cos (phase)) +# define KISS_FFT_SIN(phase) floor(.5+SAMP_MAX * sin (phase)) +# define HALF_OF(x) ((x)>>1) +#elif defined(USE_SIMD) +# define KISS_FFT_COS(phase) _mm_set1_ps( cos(phase) ) +# define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) ) +# define HALF_OF(x) ((x)*_mm_set1_ps(.5)) +#else +# define KISS_FFT_COS(phase) (kiss_fft_scalar) cos(phase) +# define KISS_FFT_SIN(phase) (kiss_fft_scalar) sin(phase) +# define HALF_OF(x) ((x)*.5) +#endif + +#define kf_cexp(x,phase) \ + do{ \ + (x)->r = KISS_FFT_COS(phase);\ + (x)->i = KISS_FFT_SIN(phase);\ + }while(0) + + +/* a debugging function */ +#define pcpx(c)\ + fprintf(stderr,"%g + %gi\n",(double)((c)->r),(double)((c)->i) ) + + +#ifdef KISS_FFT_USE_ALLOCA +// define this to allow use of alloca instead of malloc for temporary buffers +// Temporary buffers are used in two case: +// 1. FFT sizes that have "bad" factors. i.e. not 2,3 and 5 +// 2. "in-place" FFTs. Notice the quotes, since kissfft does not really do an in-place transform. +#include +#define KISS_FFT_TMP_ALLOC(nbytes) alloca(nbytes) +#define KISS_FFT_TMP_FREE(ptr) +#else +#define KISS_FFT_TMP_ALLOC(nbytes) KISS_FFT_MALLOC(nbytes) +#define KISS_FFT_TMP_FREE(ptr) KISS_FFT_FREE(ptr) +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.c new file mode 100644 index 0000000..78bcbab --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.c @@ -0,0 +1,411 @@ +#if !defined(ESP32) +/* +Copyright (c) 2003-2010, Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + + +#include "tensorflow_arm/third_party/kissfft/_kiss_fft_guts.h" +/* The guts header contains all the multiplication and addition macros that are defined for + fixed or floating point complex numbers. It also delares the kf_ internal functions. + */ + +static void kf_bfly2( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m + ) +{ + kiss_fft_cpx * Fout2; + kiss_fft_cpx * tw1 = st->twiddles; + kiss_fft_cpx t; + Fout2 = Fout + m; + do{ + C_FIXDIV(*Fout,2); C_FIXDIV(*Fout2,2); + + C_MUL (t, *Fout2 , *tw1); + tw1 += fstride; + C_SUB( *Fout2 , *Fout , t ); + C_ADDTO( *Fout , t ); + ++Fout2; + ++Fout; + }while (--m); +} + +static void kf_bfly4( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + const size_t m + ) +{ + kiss_fft_cpx *tw1,*tw2,*tw3; + kiss_fft_cpx scratch[6]; + size_t k=m; + const size_t m2=2*m; + const size_t m3=3*m; + + + tw3 = tw2 = tw1 = st->twiddles; + + do { + C_FIXDIV(*Fout,4); C_FIXDIV(Fout[m],4); C_FIXDIV(Fout[m2],4); C_FIXDIV(Fout[m3],4); + + C_MUL(scratch[0],Fout[m] , *tw1 ); + C_MUL(scratch[1],Fout[m2] , *tw2 ); + C_MUL(scratch[2],Fout[m3] , *tw3 ); + + C_SUB( scratch[5] , *Fout, scratch[1] ); + C_ADDTO(*Fout, scratch[1]); + C_ADD( scratch[3] , scratch[0] , scratch[2] ); + C_SUB( scratch[4] , scratch[0] , scratch[2] ); + C_SUB( Fout[m2], *Fout, scratch[3] ); + tw1 += fstride; + tw2 += fstride*2; + tw3 += fstride*3; + C_ADDTO( *Fout , scratch[3] ); + + if(st->inverse) { + Fout[m].r = scratch[5].r - scratch[4].i; + Fout[m].i = scratch[5].i + scratch[4].r; + Fout[m3].r = scratch[5].r + scratch[4].i; + Fout[m3].i = scratch[5].i - scratch[4].r; + }else{ + Fout[m].r = scratch[5].r + scratch[4].i; + Fout[m].i = scratch[5].i - scratch[4].r; + Fout[m3].r = scratch[5].r - scratch[4].i; + Fout[m3].i = scratch[5].i + scratch[4].r; + } + ++Fout; + }while(--k); +} + +static void kf_bfly3( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + size_t m + ) +{ + size_t k=m; + const size_t m2 = 2*m; + kiss_fft_cpx *tw1,*tw2; + kiss_fft_cpx scratch[5]; + kiss_fft_cpx epi3; + epi3 = st->twiddles[fstride*m]; + + tw1=tw2=st->twiddles; + + do{ + C_FIXDIV(*Fout,3); C_FIXDIV(Fout[m],3); C_FIXDIV(Fout[m2],3); + + C_MUL(scratch[1],Fout[m] , *tw1); + C_MUL(scratch[2],Fout[m2] , *tw2); + + C_ADD(scratch[3],scratch[1],scratch[2]); + C_SUB(scratch[0],scratch[1],scratch[2]); + tw1 += fstride; + tw2 += fstride*2; + + Fout[m].r = Fout->r - HALF_OF(scratch[3].r); + Fout[m].i = Fout->i - HALF_OF(scratch[3].i); + + C_MULBYSCALAR( scratch[0] , epi3.i ); + + C_ADDTO(*Fout,scratch[3]); + + Fout[m2].r = Fout[m].r + scratch[0].i; + Fout[m2].i = Fout[m].i - scratch[0].r; + + Fout[m].r -= scratch[0].i; + Fout[m].i += scratch[0].r; + + ++Fout; + }while(--k); +} + +static void kf_bfly5( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m + ) +{ + kiss_fft_cpx *Fout0,*Fout1,*Fout2,*Fout3,*Fout4; + int u; + kiss_fft_cpx scratch[13]; + kiss_fft_cpx * twiddles = st->twiddles; + kiss_fft_cpx *tw; + kiss_fft_cpx ya,yb; + ya = twiddles[fstride*m]; + yb = twiddles[fstride*2*m]; + + Fout0=Fout; + Fout1=Fout0+m; + Fout2=Fout0+2*m; + Fout3=Fout0+3*m; + Fout4=Fout0+4*m; + + tw=st->twiddles; + for ( u=0; ur += scratch[7].r + scratch[8].r; + Fout0->i += scratch[7].i + scratch[8].i; + + scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r); + scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r); + + scratch[6].r = S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i); + scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i); + + C_SUB(*Fout1,scratch[5],scratch[6]); + C_ADD(*Fout4,scratch[5],scratch[6]); + + scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r); + scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r); + scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i); + scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i); + + C_ADD(*Fout2,scratch[11],scratch[12]); + C_SUB(*Fout3,scratch[11],scratch[12]); + + ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4; + } +} + +/* perform the butterfly for one stage of a mixed radix FFT */ +static void kf_bfly_generic( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m, + int p + ) +{ + int u,k,q1,q; + kiss_fft_cpx * twiddles = st->twiddles; + kiss_fft_cpx t; + int Norig = st->nfft; + + kiss_fft_cpx * scratch = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC(sizeof(kiss_fft_cpx)*p); + + for ( u=0; u=Norig) twidx-=Norig; + C_MUL(t,scratch[q] , twiddles[twidx] ); + C_ADDTO( Fout[ k ] ,t); + } + k += m; + } + } + KISS_FFT_TMP_FREE(scratch); +} + +static +void kf_work( + kiss_fft_cpx * Fout, + const kiss_fft_cpx * f, + const size_t fstride, + int in_stride, + int * factors, + const kiss_fft_cfg st + ) +{ + kiss_fft_cpx * Fout_beg=Fout; + const int p=*factors++; /* the radix */ + const int m=*factors++; /* stage's fft length/p */ + const kiss_fft_cpx * Fout_end = Fout + p*m; + +#ifdef _OPENMP + // use openmp extensions at the + // top-level (not recursive) + if (fstride==1 && p<=5) + { + int k; + + // execute the p different work units in different threads +# pragma omp parallel for + for (k=0;k floor_sqrt) + p = n; /* no more factors, skip to end */ + } + n /= p; + *facbuf++ = p; + *facbuf++ = n; + } while (n > 1); +} + +/* + * + * User-callable function to allocate all necessary storage space for the fft. + * + * The return value is a contiguous block of memory, allocated with malloc. As such, + * It can be freed with free(), rather than a kiss_fft-specific function. + * */ +kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem ) +{ + kiss_fft_cfg st=NULL; + size_t memneeded = sizeof(struct kiss_fft_state) + + sizeof(kiss_fft_cpx)*(nfft-1); /* twiddle factors*/ + + if ( lenmem==NULL ) { + st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded ); + }else{ + if (mem != NULL && *lenmem >= memneeded) + st = (kiss_fft_cfg)mem; + *lenmem = memneeded; + } + if (st) { + int i; + st->nfft=nfft; + st->inverse = inverse_fft; + + for (i=0;iinverse) + phase *= -1; + kf_cexp(st->twiddles+i, phase ); + } + + kf_factor(nfft,st->factors); + } + return st; +} + + +void kiss_fft_stride(kiss_fft_cfg st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int in_stride) +{ + if (fin == fout) { + //NOTE: this is not really an in-place FFT algorithm. + //It just performs an out-of-place FFT into a temp buffer + kiss_fft_cpx * tmpbuf = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC( sizeof(kiss_fft_cpx)*st->nfft); + kf_work(tmpbuf,fin,1,in_stride, st->factors,st); + memcpy(fout,tmpbuf,sizeof(kiss_fft_cpx)*st->nfft); + KISS_FFT_TMP_FREE(tmpbuf); + }else{ + kf_work( fout, fin, 1,in_stride, st->factors,st ); + } +} + +void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout) +{ + kiss_fft_stride(cfg,fin,fout,1); +} + + +void kiss_fft_cleanup(void) +{ + // nothing needed any more +} + +int kiss_fft_next_fast_size(int n) +{ + while(1) { + int m=n; + while ( (m%2) == 0 ) m/=2; + while ( (m%3) == 0 ) m/=3; + while ( (m%5) == 0 ) m/=5; + if (m<=1) + break; /* n is completely factorable by twos, threes, and fives */ + n++; + } + return n; +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.h new file mode 100644 index 0000000..676bcac --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/kiss_fft.h @@ -0,0 +1,133 @@ +#if !defined(ESP32) +#ifndef KISS_FFT_H +#define KISS_FFT_H + +#include +#include +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +/* + ATTENTION! + If you would like a : + -- a utility that will handle the caching of fft objects + -- real-only (no imaginary time component ) FFT + -- a multi-dimensional FFT + -- a command-line utility to perform ffts + -- a command-line utility to perform fast-convolution filtering + + Then see kfc.h kiss_fftr.h kiss_fftnd.h fftutil.c kiss_fastfir.c + in the tools/ directory. +*/ + +#ifdef USE_SIMD +# include +# define kiss_fft_scalar __m128 +#define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes,16) +#define KISS_FFT_FREE _mm_free +#else +#define KISS_FFT_MALLOC(X) (void*)(0) /* Patched. */ +#define KISS_FFT_FREE(X) /* Patched. */ +#endif + + +// Patched automatically by download_dependencies.sh so default is 16 bit. +#ifndef FIXED_POINT +#define FIXED_POINT (16) +#endif +// End patch. + +#ifdef FIXED_POINT +#include /* Patched. */ +# if (FIXED_POINT == 32) +# define kiss_fft_scalar int32_t +# else +# define kiss_fft_scalar int16_t +# endif +#else +# ifndef kiss_fft_scalar +/* default is float */ +# define kiss_fft_scalar float +# endif +#endif + +typedef struct { + kiss_fft_scalar r; + kiss_fft_scalar i; +}kiss_fft_cpx; + +typedef struct kiss_fft_state* kiss_fft_cfg; + +/* + * kiss_fft_alloc + * + * Initialize a FFT (or IFFT) algorithm's cfg/state buffer. + * + * typical usage: kiss_fft_cfg mycfg=kiss_fft_alloc(1024,0,NULL,NULL); + * + * The return value from fft_alloc is a cfg buffer used internally + * by the fft routine or NULL. + * + * If lenmem is NULL, then kiss_fft_alloc will allocate a cfg buffer using malloc. + * The returned value should be free()d when done to avoid memory leaks. + * + * The state can be placed in a user supplied buffer 'mem': + * If lenmem is not NULL and mem is not NULL and *lenmem is large enough, + * then the function places the cfg in mem and the size used in *lenmem + * and returns mem. + * + * If lenmem is not NULL and ( mem is NULL or *lenmem is not large enough), + * then the function returns NULL and places the minimum cfg + * buffer size in *lenmem. + * */ + +kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem); + +/* + * kiss_fft(cfg,in_out_buf) + * + * Perform an FFT on a complex input buffer. + * for a forward FFT, + * fin should be f[0] , f[1] , ... ,f[nfft-1] + * fout will be F[0] , F[1] , ... ,F[nfft-1] + * Note that each element is complex and can be accessed like + f[k].r and f[k].i + * */ +void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout); + +/* + A more generic version of the above function. It reads its input from every Nth sample. + * */ +void kiss_fft_stride(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int fin_stride); + +/* If kiss_fft_alloc allocated a buffer, it is one contiguous + buffer and can be simply free()d when no longer needed*/ +#define kiss_fft_free free + +/* + Cleans up some memory that gets managed internally. Not necessary to call, but it might clean up + your compiler output to call this before you exit. +*/ +void kiss_fft_cleanup(void); + + +/* + * Returns the smallest integer k, such that k>=n and k has only "fast" factors (2,3,5) + */ +int kiss_fft_next_fast_size(int n); + +/* for real ffts, we need an even size */ +#define kiss_fftr_next_fast_size_real(n) \ + (kiss_fft_next_fast_size( ((n)+1)>>1)<<1) + +#ifdef __cplusplus +} +#endif + +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/tools/kiss_fftr.c b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/tools/kiss_fftr.c new file mode 100644 index 0000000..bce7b9c --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/tools/kiss_fftr.c @@ -0,0 +1,162 @@ +#if !defined(ESP32) +/* +Copyright (c) 2003-2004, Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#include "tensorflow_arm/third_party/kissfft/tools/kiss_fftr.h" +#include "tensorflow_arm/third_party/kissfft/_kiss_fft_guts.h" + +struct kiss_fftr_state{ + kiss_fft_cfg substate; + kiss_fft_cpx * tmpbuf; + kiss_fft_cpx * super_twiddles; +#ifdef USE_SIMD + void * pad; +#endif +}; + +kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem) +{ + int i; + kiss_fftr_cfg st = NULL; + size_t subsize, memneeded; + + if (nfft & 1) { + /* fprintf(stderr,"Real FFT optimization must be even.\n"); */ + return NULL; + } + nfft >>= 1; + + kiss_fft_alloc (nfft, inverse_fft, NULL, &subsize); + memneeded = sizeof(struct kiss_fftr_state) + subsize + sizeof(kiss_fft_cpx) * ( nfft * 3 / 2); + + if (lenmem == NULL) { + st = (kiss_fftr_cfg) KISS_FFT_MALLOC (memneeded); + } else { + if (*lenmem >= memneeded) + st = (kiss_fftr_cfg) mem; + *lenmem = memneeded; + } + if (!st) + return NULL; + + st->substate = (kiss_fft_cfg) (st + 1); /*just beyond kiss_fftr_state struct */ + st->tmpbuf = (kiss_fft_cpx *) (((char *) st->substate) + subsize); + st->super_twiddles = st->tmpbuf + nfft; + kiss_fft_alloc(nfft, inverse_fft, st->substate, &subsize); + + for (i = 0; i < nfft/2; ++i) { + double phase = + -3.14159265358979323846264338327 * ((double) (i+1) / nfft + .5); + if (inverse_fft) + phase *= -1; + kf_cexp (st->super_twiddles+i,phase); + } + return st; +} + +void kiss_fftr(kiss_fftr_cfg st,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata) +{ + /* input buffer timedata is stored row-wise */ + int k,ncfft; + kiss_fft_cpx fpnk,fpk,f1k,f2k,tw,tdc; + + if ( st->substate->inverse) { + /* fprintf(stderr,"kiss fft usage error: improper alloc\n"); */ + return; /* exit(1); */ + } + + ncfft = st->substate->nfft; + + /*perform the parallel fft of two real signals packed in real,imag*/ + kiss_fft( st->substate , (const kiss_fft_cpx*)timedata, st->tmpbuf ); + /* The real part of the DC element of the frequency spectrum in st->tmpbuf + * contains the sum of the even-numbered elements of the input time sequence + * The imag part is the sum of the odd-numbered elements + * + * The sum of tdc.r and tdc.i is the sum of the input time sequence. + * yielding DC of input time sequence + * The difference of tdc.r - tdc.i is the sum of the input (dot product) [1,-1,1,-1... + * yielding Nyquist bin of input time sequence + */ + + tdc.r = st->tmpbuf[0].r; + tdc.i = st->tmpbuf[0].i; + C_FIXDIV(tdc,2); + CHECK_OVERFLOW_OP(tdc.r ,+, tdc.i); + CHECK_OVERFLOW_OP(tdc.r ,-, tdc.i); + freqdata[0].r = tdc.r + tdc.i; + freqdata[ncfft].r = tdc.r - tdc.i; +#ifdef USE_SIMD + freqdata[ncfft].i = freqdata[0].i = _mm_set1_ps(0); +#else + freqdata[ncfft].i = freqdata[0].i = 0; +#endif + + for ( k=1;k <= ncfft/2 ; ++k ) { + fpk = st->tmpbuf[k]; + fpnk.r = st->tmpbuf[ncfft-k].r; + fpnk.i = - st->tmpbuf[ncfft-k].i; + C_FIXDIV(fpk,2); + C_FIXDIV(fpnk,2); + + C_ADD( f1k, fpk , fpnk ); + C_SUB( f2k, fpk , fpnk ); + C_MUL( tw , f2k , st->super_twiddles[k-1]); + + freqdata[k].r = HALF_OF(f1k.r + tw.r); + freqdata[k].i = HALF_OF(f1k.i + tw.i); + freqdata[ncfft-k].r = HALF_OF(f1k.r - tw.r); + freqdata[ncfft-k].i = HALF_OF(tw.i - f1k.i); + } +} + +void kiss_fftri(kiss_fftr_cfg st,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata) +{ + /* input buffer timedata is stored row-wise */ + int k, ncfft; + + if (st->substate->inverse == 0) { + /* fprintf (stderr, "kiss fft usage error: improper alloc\n"); */ + return; /* exit (1); */ + } + + ncfft = st->substate->nfft; + + st->tmpbuf[0].r = freqdata[0].r + freqdata[ncfft].r; + st->tmpbuf[0].i = freqdata[0].r - freqdata[ncfft].r; + C_FIXDIV(st->tmpbuf[0],2); + + for (k = 1; k <= ncfft / 2; ++k) { + kiss_fft_cpx fk, fnkc, fek, fok, tmp; + fk = freqdata[k]; + fnkc.r = freqdata[ncfft - k].r; + fnkc.i = -freqdata[ncfft - k].i; + C_FIXDIV( fk , 2 ); + C_FIXDIV( fnkc , 2 ); + + C_ADD (fek, fk, fnkc); + C_SUB (tmp, fk, fnkc); + C_MUL (fok, tmp, st->super_twiddles[k-1]); + C_ADD (st->tmpbuf[k], fek, fok); + C_SUB (st->tmpbuf[ncfft - k], fek, fok); +#ifdef USE_SIMD + st->tmpbuf[ncfft - k].i *= _mm_set1_ps(-1.0); +#else + st->tmpbuf[ncfft - k].i *= -1; +#endif + } + kiss_fft (st->substate, st->tmpbuf, (kiss_fft_cpx *) timedata); +} + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/kissfft/tools/kiss_fftr.h b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/tools/kiss_fftr.h similarity index 88% rename from src/kissfft/tools/kiss_fftr.h rename to src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/tools/kiss_fftr.h index 72e5a57..159bbf8 100644 --- a/src/kissfft/tools/kiss_fftr.h +++ b/src/tensorflow_arm/tensorflow/lite/micro/tools/make/downloads/kissfft/tools/kiss_fftr.h @@ -1,7 +1,8 @@ +#if !defined(ESP32) #ifndef KISS_FTR_H #define KISS_FTR_H -#include "kiss_fft.h" +#include "tensorflow_arm/third_party/kissfft/kiss_fft.h" #ifdef __cplusplus extern "C" { #endif @@ -44,3 +45,5 @@ void kiss_fftri(kiss_fftr_cfg cfg,const kiss_fft_cpx *freqdata,kiss_fft_scalar * } #endif #endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/portable_type_to_tflitetype.h b/src/tensorflow_arm/tensorflow/lite/portable_type_to_tflitetype.h new file mode 100644 index 0000000..602fcd9 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/portable_type_to_tflitetype.h @@ -0,0 +1,74 @@ +#if !defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_PORTABLE_TYPE_TO_TFLITETYPE_H_ +#define TENSORFLOW_LITE_PORTABLE_TYPE_TO_TFLITETYPE_H_ + +// Most of the definitions have been moved to this subheader so that Micro +// can include it without relying on and , which isn't +// available on all platforms. + +// Arduino build defines abs as a macro here. That is invalid C++, and breaks +// libc++'s header, undefine it. +#ifdef abs +#undef abs +#endif + +#include "tensorflow_arm/tensorflow/lite/c/common.h" + +namespace tflite { + +// Map statically from a C++ type to a TfLiteType. Used in interpreter for +// safe casts. +// Example: +// typeToTfLiteType() -> kTfLiteBool +template +constexpr TfLiteType typeToTfLiteType() { + return kTfLiteNoType; +} +// Map from TfLiteType to the corresponding C++ type. +// Example: +// TfLiteTypeToType::Type -> bool +template +struct TfLiteTypeToType {}; // Specializations below + +// Template specialization for both typeToTfLiteType and TfLiteTypeToType. +#define MATCH_TYPE_AND_TFLITE_TYPE(CPP_TYPE, TFLITE_TYPE_ENUM) \ + template <> \ + constexpr TfLiteType typeToTfLiteType() { \ + return TFLITE_TYPE_ENUM; \ + } \ + template <> \ + struct TfLiteTypeToType { \ + using Type = CPP_TYPE; \ + } + +// No string mapping is included here, since the TF Lite packed representation +// doesn't correspond to a C++ type well. +MATCH_TYPE_AND_TFLITE_TYPE(int32_t, kTfLiteInt32); +MATCH_TYPE_AND_TFLITE_TYPE(int16_t, kTfLiteInt16); +MATCH_TYPE_AND_TFLITE_TYPE(int64_t, kTfLiteInt64); +MATCH_TYPE_AND_TFLITE_TYPE(float, kTfLiteFloat32); +MATCH_TYPE_AND_TFLITE_TYPE(unsigned char, kTfLiteUInt8); +MATCH_TYPE_AND_TFLITE_TYPE(int8_t, kTfLiteInt8); +MATCH_TYPE_AND_TFLITE_TYPE(bool, kTfLiteBool); +MATCH_TYPE_AND_TFLITE_TYPE(TfLiteFloat16, kTfLiteFloat16); +MATCH_TYPE_AND_TFLITE_TYPE(double, kTfLiteFloat64); +MATCH_TYPE_AND_TFLITE_TYPE(uint64_t, kTfLiteUInt64); + +} // namespace tflite +#endif // TENSORFLOW_LITE_PORTABLE_TYPE_TO_TFLITETYPE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/schema/schema_generated.h b/src/tensorflow_arm/tensorflow/lite/schema/schema_generated.h new file mode 100644 index 0000000..da1d575 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/schema/schema_generated.h @@ -0,0 +1,17150 @@ +#if !defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +// automatically generated by the FlatBuffers compiler, do not modify + + +#ifndef FLATBUFFERS_GENERATED_SCHEMA_TFLITE_H_ +#define FLATBUFFERS_GENERATED_SCHEMA_TFLITE_H_ + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" + +namespace tflite { + +struct CustomQuantization; +struct CustomQuantizationT; + +struct QuantizationParameters; +struct QuantizationParametersT; + +struct Int32Vector; +struct Int32VectorT; + +struct Uint16Vector; +struct Uint16VectorT; + +struct Uint8Vector; +struct Uint8VectorT; + +struct DimensionMetadata; +struct DimensionMetadataT; + +struct SparsityParameters; +struct SparsityParametersT; + +struct Tensor; +struct TensorT; + +struct Conv2DOptions; +struct Conv2DOptionsT; + +struct Pool2DOptions; +struct Pool2DOptionsT; + +struct DepthwiseConv2DOptions; +struct DepthwiseConv2DOptionsT; + +struct ConcatEmbeddingsOptions; +struct ConcatEmbeddingsOptionsT; + +struct LSHProjectionOptions; +struct LSHProjectionOptionsT; + +struct SVDFOptions; +struct SVDFOptionsT; + +struct RNNOptions; +struct RNNOptionsT; + +struct SequenceRNNOptions; +struct SequenceRNNOptionsT; + +struct BidirectionalSequenceRNNOptions; +struct BidirectionalSequenceRNNOptionsT; + +struct FullyConnectedOptions; +struct FullyConnectedOptionsT; + +struct SoftmaxOptions; +struct SoftmaxOptionsT; + +struct ConcatenationOptions; +struct ConcatenationOptionsT; + +struct AddOptions; +struct AddOptionsT; + +struct MulOptions; +struct MulOptionsT; + +struct L2NormOptions; +struct L2NormOptionsT; + +struct LocalResponseNormalizationOptions; +struct LocalResponseNormalizationOptionsT; + +struct LSTMOptions; +struct LSTMOptionsT; + +struct UnidirectionalSequenceLSTMOptions; +struct UnidirectionalSequenceLSTMOptionsT; + +struct BidirectionalSequenceLSTMOptions; +struct BidirectionalSequenceLSTMOptionsT; + +struct ResizeBilinearOptions; +struct ResizeBilinearOptionsT; + +struct ResizeNearestNeighborOptions; +struct ResizeNearestNeighborOptionsT; + +struct CallOptions; +struct CallOptionsT; + +struct PadOptions; +struct PadOptionsT; + +struct PadV2Options; +struct PadV2OptionsT; + +struct ReshapeOptions; +struct ReshapeOptionsT; + +struct SpaceToBatchNDOptions; +struct SpaceToBatchNDOptionsT; + +struct BatchToSpaceNDOptions; +struct BatchToSpaceNDOptionsT; + +struct SkipGramOptions; +struct SkipGramOptionsT; + +struct SpaceToDepthOptions; +struct SpaceToDepthOptionsT; + +struct DepthToSpaceOptions; +struct DepthToSpaceOptionsT; + +struct SubOptions; +struct SubOptionsT; + +struct DivOptions; +struct DivOptionsT; + +struct TopKV2Options; +struct TopKV2OptionsT; + +struct EmbeddingLookupSparseOptions; +struct EmbeddingLookupSparseOptionsT; + +struct GatherOptions; +struct GatherOptionsT; + +struct TransposeOptions; +struct TransposeOptionsT; + +struct ExpOptions; +struct ExpOptionsT; + +struct CosOptions; +struct CosOptionsT; + +struct ReducerOptions; +struct ReducerOptionsT; + +struct SqueezeOptions; +struct SqueezeOptionsT; + +struct SplitOptions; +struct SplitOptionsT; + +struct SplitVOptions; +struct SplitVOptionsT; + +struct StridedSliceOptions; +struct StridedSliceOptionsT; + +struct LogSoftmaxOptions; +struct LogSoftmaxOptionsT; + +struct CastOptions; +struct CastOptionsT; + +struct DequantizeOptions; +struct DequantizeOptionsT; + +struct MaximumMinimumOptions; +struct MaximumMinimumOptionsT; + +struct TileOptions; +struct TileOptionsT; + +struct ArgMaxOptions; +struct ArgMaxOptionsT; + +struct ArgMinOptions; +struct ArgMinOptionsT; + +struct GreaterOptions; +struct GreaterOptionsT; + +struct GreaterEqualOptions; +struct GreaterEqualOptionsT; + +struct LessOptions; +struct LessOptionsT; + +struct LessEqualOptions; +struct LessEqualOptionsT; + +struct NegOptions; +struct NegOptionsT; + +struct SelectOptions; +struct SelectOptionsT; + +struct SliceOptions; +struct SliceOptionsT; + +struct TransposeConvOptions; +struct TransposeConvOptionsT; + +struct ExpandDimsOptions; +struct ExpandDimsOptionsT; + +struct SparseToDenseOptions; +struct SparseToDenseOptionsT; + +struct EqualOptions; +struct EqualOptionsT; + +struct NotEqualOptions; +struct NotEqualOptionsT; + +struct ShapeOptions; +struct ShapeOptionsT; + +struct RankOptions; +struct RankOptionsT; + +struct PowOptions; +struct PowOptionsT; + +struct FakeQuantOptions; +struct FakeQuantOptionsT; + +struct PackOptions; +struct PackOptionsT; + +struct LogicalOrOptions; +struct LogicalOrOptionsT; + +struct OneHotOptions; +struct OneHotOptionsT; + +struct AbsOptions; +struct AbsOptionsT; + +struct HardSwishOptions; +struct HardSwishOptionsT; + +struct LogicalAndOptions; +struct LogicalAndOptionsT; + +struct LogicalNotOptions; +struct LogicalNotOptionsT; + +struct UnpackOptions; +struct UnpackOptionsT; + +struct FloorDivOptions; +struct FloorDivOptionsT; + +struct SquareOptions; +struct SquareOptionsT; + +struct ZerosLikeOptions; +struct ZerosLikeOptionsT; + +struct FillOptions; +struct FillOptionsT; + +struct FloorModOptions; +struct FloorModOptionsT; + +struct RangeOptions; +struct RangeOptionsT; + +struct LeakyReluOptions; +struct LeakyReluOptionsT; + +struct SquaredDifferenceOptions; +struct SquaredDifferenceOptionsT; + +struct MirrorPadOptions; +struct MirrorPadOptionsT; + +struct UniqueOptions; +struct UniqueOptionsT; + +struct ReverseV2Options; +struct ReverseV2OptionsT; + +struct AddNOptions; +struct AddNOptionsT; + +struct GatherNdOptions; +struct GatherNdOptionsT; + +struct WhereOptions; +struct WhereOptionsT; + +struct ReverseSequenceOptions; +struct ReverseSequenceOptionsT; + +struct MatrixDiagOptions; +struct MatrixDiagOptionsT; + +struct QuantizeOptions; +struct QuantizeOptionsT; + +struct MatrixSetDiagOptions; +struct MatrixSetDiagOptionsT; + +struct IfOptions; +struct IfOptionsT; + +struct CallOnceOptions; +struct CallOnceOptionsT; + +struct WhileOptions; +struct WhileOptionsT; + +struct NonMaxSuppressionV4Options; +struct NonMaxSuppressionV4OptionsT; + +struct NonMaxSuppressionV5Options; +struct NonMaxSuppressionV5OptionsT; + +struct ScatterNdOptions; +struct ScatterNdOptionsT; + +struct SelectV2Options; +struct SelectV2OptionsT; + +struct DensifyOptions; +struct DensifyOptionsT; + +struct SegmentSumOptions; +struct SegmentSumOptionsT; + +struct BatchMatMulOptions; +struct BatchMatMulOptionsT; + +struct CumsumOptions; +struct CumsumOptionsT; + +struct BroadcastToOptions; +struct BroadcastToOptionsT; + +struct Rfft2dOptions; +struct Rfft2dOptionsT; + +struct OperatorCode; +struct OperatorCodeT; + +struct Operator; +struct OperatorT; + +struct SubGraph; +struct SubGraphT; + +struct Buffer; +struct BufferT; + +struct Metadata; +struct MetadataT; + +struct TensorMap; +struct TensorMapT; + +struct SignatureDef; +struct SignatureDefT; + +struct Model; +struct ModelT; + +enum TensorType { + TensorType_FLOAT32 = 0, + TensorType_FLOAT16 = 1, + TensorType_INT32 = 2, + TensorType_UINT8 = 3, + TensorType_INT64 = 4, + TensorType_STRING = 5, + TensorType_BOOL = 6, + TensorType_INT16 = 7, + TensorType_COMPLEX64 = 8, + TensorType_INT8 = 9, + TensorType_FLOAT64 = 10, + TensorType_COMPLEX128 = 11, + TensorType_UINT64 = 12, + TensorType_MIN = TensorType_FLOAT32, + TensorType_MAX = TensorType_UINT64 +}; + +inline const TensorType (&EnumValuesTensorType())[13] { + static const TensorType values[] = { + TensorType_FLOAT32, + TensorType_FLOAT16, + TensorType_INT32, + TensorType_UINT8, + TensorType_INT64, + TensorType_STRING, + TensorType_BOOL, + TensorType_INT16, + TensorType_COMPLEX64, + TensorType_INT8, + TensorType_FLOAT64, + TensorType_COMPLEX128, + TensorType_UINT64 + }; + return values; +} + +inline const char * const *EnumNamesTensorType() { + static const char * const names[14] = { + "FLOAT32", + "FLOAT16", + "INT32", + "UINT8", + "INT64", + "STRING", + "BOOL", + "INT16", + "COMPLEX64", + "INT8", + "FLOAT64", + "COMPLEX128", + "UINT64", + nullptr + }; + return names; +} + +inline const char *EnumNameTensorType(TensorType e) { + if (flatbuffers::IsOutRange(e, TensorType_FLOAT32, TensorType_UINT64)) return ""; + const size_t index = static_cast(e); + return EnumNamesTensorType()[index]; +} + +enum QuantizationDetails { + QuantizationDetails_NONE = 0, + QuantizationDetails_CustomQuantization = 1, + QuantizationDetails_MIN = QuantizationDetails_NONE, + QuantizationDetails_MAX = QuantizationDetails_CustomQuantization +}; + +inline const QuantizationDetails (&EnumValuesQuantizationDetails())[2] { + static const QuantizationDetails values[] = { + QuantizationDetails_NONE, + QuantizationDetails_CustomQuantization + }; + return values; +} + +inline const char * const *EnumNamesQuantizationDetails() { + static const char * const names[3] = { + "NONE", + "CustomQuantization", + nullptr + }; + return names; +} + +inline const char *EnumNameQuantizationDetails(QuantizationDetails e) { + if (flatbuffers::IsOutRange(e, QuantizationDetails_NONE, QuantizationDetails_CustomQuantization)) return ""; + const size_t index = static_cast(e); + return EnumNamesQuantizationDetails()[index]; +} + +template struct QuantizationDetailsTraits { + static const QuantizationDetails enum_value = QuantizationDetails_NONE; +}; + +template<> struct QuantizationDetailsTraits { + static const QuantizationDetails enum_value = QuantizationDetails_CustomQuantization; +}; + +struct QuantizationDetailsUnion { + QuantizationDetails type; + void *value; + + QuantizationDetailsUnion() : type(QuantizationDetails_NONE), value(nullptr) {} + QuantizationDetailsUnion(QuantizationDetailsUnion&& u) FLATBUFFERS_NOEXCEPT : + type(QuantizationDetails_NONE), value(nullptr) + { std::swap(type, u.type); std::swap(value, u.value); } + QuantizationDetailsUnion(const QuantizationDetailsUnion &) FLATBUFFERS_NOEXCEPT; + QuantizationDetailsUnion &operator=(const QuantizationDetailsUnion &u) FLATBUFFERS_NOEXCEPT + { QuantizationDetailsUnion t(u); std::swap(type, t.type); std::swap(value, t.value); return *this; } + QuantizationDetailsUnion &operator=(QuantizationDetailsUnion &&u) FLATBUFFERS_NOEXCEPT + { std::swap(type, u.type); std::swap(value, u.value); return *this; } + ~QuantizationDetailsUnion() { Reset(); } + + void Reset(); + +#ifndef FLATBUFFERS_CPP98_STL + template + void Set(T&& val) { + using RT = typename std::remove_reference::type; + Reset(); + type = QuantizationDetailsTraits::enum_value; + if (type != QuantizationDetails_NONE) { + value = new RT(std::forward(val)); + } + } +#endif // FLATBUFFERS_CPP98_STL + + static void *UnPack(const void *obj, QuantizationDetails type, const flatbuffers::resolver_function_t *resolver); + flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const flatbuffers::rehasher_function_t *_rehasher = nullptr) const; + + tflite::CustomQuantizationT *AsCustomQuantization() { + return type == QuantizationDetails_CustomQuantization ? + reinterpret_cast(value) : nullptr; + } + const tflite::CustomQuantizationT *AsCustomQuantization() const { + return type == QuantizationDetails_CustomQuantization ? + reinterpret_cast(value) : nullptr; + } +}; + +bool VerifyQuantizationDetails(flatbuffers::Verifier &verifier, const void *obj, QuantizationDetails type); +bool VerifyQuantizationDetailsVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector> *values, const flatbuffers::Vector *types); + +enum DimensionType { + DimensionType_DENSE = 0, + DimensionType_SPARSE_CSR = 1, + DimensionType_MIN = DimensionType_DENSE, + DimensionType_MAX = DimensionType_SPARSE_CSR +}; + +inline const DimensionType (&EnumValuesDimensionType())[2] { + static const DimensionType values[] = { + DimensionType_DENSE, + DimensionType_SPARSE_CSR + }; + return values; +} + +inline const char * const *EnumNamesDimensionType() { + static const char * const names[3] = { + "DENSE", + "SPARSE_CSR", + nullptr + }; + return names; +} + +inline const char *EnumNameDimensionType(DimensionType e) { + if (flatbuffers::IsOutRange(e, DimensionType_DENSE, DimensionType_SPARSE_CSR)) return ""; + const size_t index = static_cast(e); + return EnumNamesDimensionType()[index]; +} + +enum SparseIndexVector { + SparseIndexVector_NONE = 0, + SparseIndexVector_Int32Vector = 1, + SparseIndexVector_Uint16Vector = 2, + SparseIndexVector_Uint8Vector = 3, + SparseIndexVector_MIN = SparseIndexVector_NONE, + SparseIndexVector_MAX = SparseIndexVector_Uint8Vector +}; + +inline const SparseIndexVector (&EnumValuesSparseIndexVector())[4] { + static const SparseIndexVector values[] = { + SparseIndexVector_NONE, + SparseIndexVector_Int32Vector, + SparseIndexVector_Uint16Vector, + SparseIndexVector_Uint8Vector + }; + return values; +} + +inline const char * const *EnumNamesSparseIndexVector() { + static const char * const names[5] = { + "NONE", + "Int32Vector", + "Uint16Vector", + "Uint8Vector", + nullptr + }; + return names; +} + +inline const char *EnumNameSparseIndexVector(SparseIndexVector e) { + if (flatbuffers::IsOutRange(e, SparseIndexVector_NONE, SparseIndexVector_Uint8Vector)) return ""; + const size_t index = static_cast(e); + return EnumNamesSparseIndexVector()[index]; +} + +template struct SparseIndexVectorTraits { + static const SparseIndexVector enum_value = SparseIndexVector_NONE; +}; + +template<> struct SparseIndexVectorTraits { + static const SparseIndexVector enum_value = SparseIndexVector_Int32Vector; +}; + +template<> struct SparseIndexVectorTraits { + static const SparseIndexVector enum_value = SparseIndexVector_Uint16Vector; +}; + +template<> struct SparseIndexVectorTraits { + static const SparseIndexVector enum_value = SparseIndexVector_Uint8Vector; +}; + +struct SparseIndexVectorUnion { + SparseIndexVector type; + void *value; + + SparseIndexVectorUnion() : type(SparseIndexVector_NONE), value(nullptr) {} + SparseIndexVectorUnion(SparseIndexVectorUnion&& u) FLATBUFFERS_NOEXCEPT : + type(SparseIndexVector_NONE), value(nullptr) + { std::swap(type, u.type); std::swap(value, u.value); } + SparseIndexVectorUnion(const SparseIndexVectorUnion &) FLATBUFFERS_NOEXCEPT; + SparseIndexVectorUnion &operator=(const SparseIndexVectorUnion &u) FLATBUFFERS_NOEXCEPT + { SparseIndexVectorUnion t(u); std::swap(type, t.type); std::swap(value, t.value); return *this; } + SparseIndexVectorUnion &operator=(SparseIndexVectorUnion &&u) FLATBUFFERS_NOEXCEPT + { std::swap(type, u.type); std::swap(value, u.value); return *this; } + ~SparseIndexVectorUnion() { Reset(); } + + void Reset(); + +#ifndef FLATBUFFERS_CPP98_STL + template + void Set(T&& val) { + using RT = typename std::remove_reference::type; + Reset(); + type = SparseIndexVectorTraits::enum_value; + if (type != SparseIndexVector_NONE) { + value = new RT(std::forward(val)); + } + } +#endif // FLATBUFFERS_CPP98_STL + + static void *UnPack(const void *obj, SparseIndexVector type, const flatbuffers::resolver_function_t *resolver); + flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const flatbuffers::rehasher_function_t *_rehasher = nullptr) const; + + tflite::Int32VectorT *AsInt32Vector() { + return type == SparseIndexVector_Int32Vector ? + reinterpret_cast(value) : nullptr; + } + const tflite::Int32VectorT *AsInt32Vector() const { + return type == SparseIndexVector_Int32Vector ? + reinterpret_cast(value) : nullptr; + } + tflite::Uint16VectorT *AsUint16Vector() { + return type == SparseIndexVector_Uint16Vector ? + reinterpret_cast(value) : nullptr; + } + const tflite::Uint16VectorT *AsUint16Vector() const { + return type == SparseIndexVector_Uint16Vector ? + reinterpret_cast(value) : nullptr; + } + tflite::Uint8VectorT *AsUint8Vector() { + return type == SparseIndexVector_Uint8Vector ? + reinterpret_cast(value) : nullptr; + } + const tflite::Uint8VectorT *AsUint8Vector() const { + return type == SparseIndexVector_Uint8Vector ? + reinterpret_cast(value) : nullptr; + } +}; + +bool VerifySparseIndexVector(flatbuffers::Verifier &verifier, const void *obj, SparseIndexVector type); +bool VerifySparseIndexVectorVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector> *values, const flatbuffers::Vector *types); + +enum BuiltinOperator { + BuiltinOperator_ADD = 0, + BuiltinOperator_AVERAGE_POOL_2D = 1, + BuiltinOperator_CONCATENATION = 2, + BuiltinOperator_CONV_2D = 3, + BuiltinOperator_DEPTHWISE_CONV_2D = 4, + BuiltinOperator_DEPTH_TO_SPACE = 5, + BuiltinOperator_DEQUANTIZE = 6, + BuiltinOperator_EMBEDDING_LOOKUP = 7, + BuiltinOperator_FLOOR = 8, + BuiltinOperator_FULLY_CONNECTED = 9, + BuiltinOperator_HASHTABLE_LOOKUP = 10, + BuiltinOperator_L2_NORMALIZATION = 11, + BuiltinOperator_L2_POOL_2D = 12, + BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION = 13, + BuiltinOperator_LOGISTIC = 14, + BuiltinOperator_LSH_PROJECTION = 15, + BuiltinOperator_LSTM = 16, + BuiltinOperator_MAX_POOL_2D = 17, + BuiltinOperator_MUL = 18, + BuiltinOperator_RELU = 19, + BuiltinOperator_RELU_N1_TO_1 = 20, + BuiltinOperator_RELU6 = 21, + BuiltinOperator_RESHAPE = 22, + BuiltinOperator_RESIZE_BILINEAR = 23, + BuiltinOperator_RNN = 24, + BuiltinOperator_SOFTMAX = 25, + BuiltinOperator_SPACE_TO_DEPTH = 26, + BuiltinOperator_SVDF = 27, + BuiltinOperator_TANH = 28, + BuiltinOperator_CONCAT_EMBEDDINGS = 29, + BuiltinOperator_SKIP_GRAM = 30, + BuiltinOperator_CALL = 31, + BuiltinOperator_CUSTOM = 32, + BuiltinOperator_EMBEDDING_LOOKUP_SPARSE = 33, + BuiltinOperator_PAD = 34, + BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN = 35, + BuiltinOperator_GATHER = 36, + BuiltinOperator_BATCH_TO_SPACE_ND = 37, + BuiltinOperator_SPACE_TO_BATCH_ND = 38, + BuiltinOperator_TRANSPOSE = 39, + BuiltinOperator_MEAN = 40, + BuiltinOperator_SUB = 41, + BuiltinOperator_DIV = 42, + BuiltinOperator_SQUEEZE = 43, + BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM = 44, + BuiltinOperator_STRIDED_SLICE = 45, + BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN = 46, + BuiltinOperator_EXP = 47, + BuiltinOperator_TOPK_V2 = 48, + BuiltinOperator_SPLIT = 49, + BuiltinOperator_LOG_SOFTMAX = 50, + BuiltinOperator_DELEGATE = 51, + BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM = 52, + BuiltinOperator_CAST = 53, + BuiltinOperator_PRELU = 54, + BuiltinOperator_MAXIMUM = 55, + BuiltinOperator_ARG_MAX = 56, + BuiltinOperator_MINIMUM = 57, + BuiltinOperator_LESS = 58, + BuiltinOperator_NEG = 59, + BuiltinOperator_PADV2 = 60, + BuiltinOperator_GREATER = 61, + BuiltinOperator_GREATER_EQUAL = 62, + BuiltinOperator_LESS_EQUAL = 63, + BuiltinOperator_SELECT = 64, + BuiltinOperator_SLICE = 65, + BuiltinOperator_SIN = 66, + BuiltinOperator_TRANSPOSE_CONV = 67, + BuiltinOperator_SPARSE_TO_DENSE = 68, + BuiltinOperator_TILE = 69, + BuiltinOperator_EXPAND_DIMS = 70, + BuiltinOperator_EQUAL = 71, + BuiltinOperator_NOT_EQUAL = 72, + BuiltinOperator_LOG = 73, + BuiltinOperator_SUM = 74, + BuiltinOperator_SQRT = 75, + BuiltinOperator_RSQRT = 76, + BuiltinOperator_SHAPE = 77, + BuiltinOperator_POW = 78, + BuiltinOperator_ARG_MIN = 79, + BuiltinOperator_FAKE_QUANT = 80, + BuiltinOperator_REDUCE_PROD = 81, + BuiltinOperator_REDUCE_MAX = 82, + BuiltinOperator_PACK = 83, + BuiltinOperator_LOGICAL_OR = 84, + BuiltinOperator_ONE_HOT = 85, + BuiltinOperator_LOGICAL_AND = 86, + BuiltinOperator_LOGICAL_NOT = 87, + BuiltinOperator_UNPACK = 88, + BuiltinOperator_REDUCE_MIN = 89, + BuiltinOperator_FLOOR_DIV = 90, + BuiltinOperator_REDUCE_ANY = 91, + BuiltinOperator_SQUARE = 92, + BuiltinOperator_ZEROS_LIKE = 93, + BuiltinOperator_FILL = 94, + BuiltinOperator_FLOOR_MOD = 95, + BuiltinOperator_RANGE = 96, + BuiltinOperator_RESIZE_NEAREST_NEIGHBOR = 97, + BuiltinOperator_LEAKY_RELU = 98, + BuiltinOperator_SQUARED_DIFFERENCE = 99, + BuiltinOperator_MIRROR_PAD = 100, + BuiltinOperator_ABS = 101, + BuiltinOperator_SPLIT_V = 102, + BuiltinOperator_UNIQUE = 103, + BuiltinOperator_CEIL = 104, + BuiltinOperator_REVERSE_V2 = 105, + BuiltinOperator_ADD_N = 106, + BuiltinOperator_GATHER_ND = 107, + BuiltinOperator_COS = 108, + BuiltinOperator_WHERE = 109, + BuiltinOperator_RANK = 110, + BuiltinOperator_ELU = 111, + BuiltinOperator_REVERSE_SEQUENCE = 112, + BuiltinOperator_MATRIX_DIAG = 113, + BuiltinOperator_QUANTIZE = 114, + BuiltinOperator_MATRIX_SET_DIAG = 115, + BuiltinOperator_ROUND = 116, + BuiltinOperator_HARD_SWISH = 117, + BuiltinOperator_IF = 118, + BuiltinOperator_WHILE = 119, + BuiltinOperator_NON_MAX_SUPPRESSION_V4 = 120, + BuiltinOperator_NON_MAX_SUPPRESSION_V5 = 121, + BuiltinOperator_SCATTER_ND = 122, + BuiltinOperator_SELECT_V2 = 123, + BuiltinOperator_DENSIFY = 124, + BuiltinOperator_SEGMENT_SUM = 125, + BuiltinOperator_BATCH_MATMUL = 126, + BuiltinOperator_PLACEHOLDER_FOR_GREATER_OP_CODES = 127, + BuiltinOperator_CUMSUM = 128, + BuiltinOperator_CALL_ONCE = 129, + BuiltinOperator_BROADCAST_TO = 130, + BuiltinOperator_RFFT2D = 131, + BuiltinOperator_MIN = BuiltinOperator_ADD, + BuiltinOperator_MAX = BuiltinOperator_RFFT2D +}; + +inline const BuiltinOperator (&EnumValuesBuiltinOperator())[132] { + static const BuiltinOperator values[] = { + BuiltinOperator_ADD, + BuiltinOperator_AVERAGE_POOL_2D, + BuiltinOperator_CONCATENATION, + BuiltinOperator_CONV_2D, + BuiltinOperator_DEPTHWISE_CONV_2D, + BuiltinOperator_DEPTH_TO_SPACE, + BuiltinOperator_DEQUANTIZE, + BuiltinOperator_EMBEDDING_LOOKUP, + BuiltinOperator_FLOOR, + BuiltinOperator_FULLY_CONNECTED, + BuiltinOperator_HASHTABLE_LOOKUP, + BuiltinOperator_L2_NORMALIZATION, + BuiltinOperator_L2_POOL_2D, + BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION, + BuiltinOperator_LOGISTIC, + BuiltinOperator_LSH_PROJECTION, + BuiltinOperator_LSTM, + BuiltinOperator_MAX_POOL_2D, + BuiltinOperator_MUL, + BuiltinOperator_RELU, + BuiltinOperator_RELU_N1_TO_1, + BuiltinOperator_RELU6, + BuiltinOperator_RESHAPE, + BuiltinOperator_RESIZE_BILINEAR, + BuiltinOperator_RNN, + BuiltinOperator_SOFTMAX, + BuiltinOperator_SPACE_TO_DEPTH, + BuiltinOperator_SVDF, + BuiltinOperator_TANH, + BuiltinOperator_CONCAT_EMBEDDINGS, + BuiltinOperator_SKIP_GRAM, + BuiltinOperator_CALL, + BuiltinOperator_CUSTOM, + BuiltinOperator_EMBEDDING_LOOKUP_SPARSE, + BuiltinOperator_PAD, + BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN, + BuiltinOperator_GATHER, + BuiltinOperator_BATCH_TO_SPACE_ND, + BuiltinOperator_SPACE_TO_BATCH_ND, + BuiltinOperator_TRANSPOSE, + BuiltinOperator_MEAN, + BuiltinOperator_SUB, + BuiltinOperator_DIV, + BuiltinOperator_SQUEEZE, + BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM, + BuiltinOperator_STRIDED_SLICE, + BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN, + BuiltinOperator_EXP, + BuiltinOperator_TOPK_V2, + BuiltinOperator_SPLIT, + BuiltinOperator_LOG_SOFTMAX, + BuiltinOperator_DELEGATE, + BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM, + BuiltinOperator_CAST, + BuiltinOperator_PRELU, + BuiltinOperator_MAXIMUM, + BuiltinOperator_ARG_MAX, + BuiltinOperator_MINIMUM, + BuiltinOperator_LESS, + BuiltinOperator_NEG, + BuiltinOperator_PADV2, + BuiltinOperator_GREATER, + BuiltinOperator_GREATER_EQUAL, + BuiltinOperator_LESS_EQUAL, + BuiltinOperator_SELECT, + BuiltinOperator_SLICE, + BuiltinOperator_SIN, + BuiltinOperator_TRANSPOSE_CONV, + BuiltinOperator_SPARSE_TO_DENSE, + BuiltinOperator_TILE, + BuiltinOperator_EXPAND_DIMS, + BuiltinOperator_EQUAL, + BuiltinOperator_NOT_EQUAL, + BuiltinOperator_LOG, + BuiltinOperator_SUM, + BuiltinOperator_SQRT, + BuiltinOperator_RSQRT, + BuiltinOperator_SHAPE, + BuiltinOperator_POW, + BuiltinOperator_ARG_MIN, + BuiltinOperator_FAKE_QUANT, + BuiltinOperator_REDUCE_PROD, + BuiltinOperator_REDUCE_MAX, + BuiltinOperator_PACK, + BuiltinOperator_LOGICAL_OR, + BuiltinOperator_ONE_HOT, + BuiltinOperator_LOGICAL_AND, + BuiltinOperator_LOGICAL_NOT, + BuiltinOperator_UNPACK, + BuiltinOperator_REDUCE_MIN, + BuiltinOperator_FLOOR_DIV, + BuiltinOperator_REDUCE_ANY, + BuiltinOperator_SQUARE, + BuiltinOperator_ZEROS_LIKE, + BuiltinOperator_FILL, + BuiltinOperator_FLOOR_MOD, + BuiltinOperator_RANGE, + BuiltinOperator_RESIZE_NEAREST_NEIGHBOR, + BuiltinOperator_LEAKY_RELU, + BuiltinOperator_SQUARED_DIFFERENCE, + BuiltinOperator_MIRROR_PAD, + BuiltinOperator_ABS, + BuiltinOperator_SPLIT_V, + BuiltinOperator_UNIQUE, + BuiltinOperator_CEIL, + BuiltinOperator_REVERSE_V2, + BuiltinOperator_ADD_N, + BuiltinOperator_GATHER_ND, + BuiltinOperator_COS, + BuiltinOperator_WHERE, + BuiltinOperator_RANK, + BuiltinOperator_ELU, + BuiltinOperator_REVERSE_SEQUENCE, + BuiltinOperator_MATRIX_DIAG, + BuiltinOperator_QUANTIZE, + BuiltinOperator_MATRIX_SET_DIAG, + BuiltinOperator_ROUND, + BuiltinOperator_HARD_SWISH, + BuiltinOperator_IF, + BuiltinOperator_WHILE, + BuiltinOperator_NON_MAX_SUPPRESSION_V4, + BuiltinOperator_NON_MAX_SUPPRESSION_V5, + BuiltinOperator_SCATTER_ND, + BuiltinOperator_SELECT_V2, + BuiltinOperator_DENSIFY, + BuiltinOperator_SEGMENT_SUM, + BuiltinOperator_BATCH_MATMUL, + BuiltinOperator_PLACEHOLDER_FOR_GREATER_OP_CODES, + BuiltinOperator_CUMSUM, + BuiltinOperator_CALL_ONCE, + BuiltinOperator_BROADCAST_TO, + BuiltinOperator_RFFT2D}; + return values; +} + +inline const char * const *EnumNamesBuiltinOperator() { + static const char *const names[133] = {"ADD", + "AVERAGE_POOL_2D", + "CONCATENATION", + "CONV_2D", + "DEPTHWISE_CONV_2D", + "DEPTH_TO_SPACE", + "DEQUANTIZE", + "EMBEDDING_LOOKUP", + "FLOOR", + "FULLY_CONNECTED", + "HASHTABLE_LOOKUP", + "L2_NORMALIZATION", + "L2_POOL_2D", + "LOCAL_RESPONSE_NORMALIZATION", + "LOGISTIC", + "LSH_PROJECTION", + "LSTM", + "MAX_POOL_2D", + "MUL", + "RELU", + "RELU_N1_TO_1", + "RELU6", + "RESHAPE", + "RESIZE_BILINEAR", + "RNN", + "SOFTMAX", + "SPACE_TO_DEPTH", + "SVDF", + "TANH", + "CONCAT_EMBEDDINGS", + "SKIP_GRAM", + "CALL", + "CUSTOM", + "EMBEDDING_LOOKUP_SPARSE", + "PAD", + "UNIDIRECTIONAL_SEQUENCE_RNN", + "GATHER", + "BATCH_TO_SPACE_ND", + "SPACE_TO_BATCH_ND", + "TRANSPOSE", + "MEAN", + "SUB", + "DIV", + "SQUEEZE", + "UNIDIRECTIONAL_SEQUENCE_LSTM", + "STRIDED_SLICE", + "BIDIRECTIONAL_SEQUENCE_RNN", + "EXP", + "TOPK_V2", + "SPLIT", + "LOG_SOFTMAX", + "DELEGATE", + "BIDIRECTIONAL_SEQUENCE_LSTM", + "CAST", + "PRELU", + "MAXIMUM", + "ARG_MAX", + "MINIMUM", + "LESS", + "NEG", + "PADV2", + "GREATER", + "GREATER_EQUAL", + "LESS_EQUAL", + "SELECT", + "SLICE", + "SIN", + "TRANSPOSE_CONV", + "SPARSE_TO_DENSE", + "TILE", + "EXPAND_DIMS", + "EQUAL", + "NOT_EQUAL", + "LOG", + "SUM", + "SQRT", + "RSQRT", + "SHAPE", + "POW", + "ARG_MIN", + "FAKE_QUANT", + "REDUCE_PROD", + "REDUCE_MAX", + "PACK", + "LOGICAL_OR", + "ONE_HOT", + "LOGICAL_AND", + "LOGICAL_NOT", + "UNPACK", + "REDUCE_MIN", + "FLOOR_DIV", + "REDUCE_ANY", + "SQUARE", + "ZEROS_LIKE", + "FILL", + "FLOOR_MOD", + "RANGE", + "RESIZE_NEAREST_NEIGHBOR", + "LEAKY_RELU", + "SQUARED_DIFFERENCE", + "MIRROR_PAD", + "ABS", + "SPLIT_V", + "UNIQUE", + "CEIL", + "REVERSE_V2", + "ADD_N", + "GATHER_ND", + "COS", + "WHERE", + "RANK", + "ELU", + "REVERSE_SEQUENCE", + "MATRIX_DIAG", + "QUANTIZE", + "MATRIX_SET_DIAG", + "ROUND", + "HARD_SWISH", + "IF", + "WHILE", + "NON_MAX_SUPPRESSION_V4", + "NON_MAX_SUPPRESSION_V5", + "SCATTER_ND", + "SELECT_V2", + "DENSIFY", + "SEGMENT_SUM", + "BATCH_MATMUL", + "PLACEHOLDER_FOR_GREATER_OP_CODES", + "CUMSUM", + "CALL_ONCE", + "BROADCAST_TO", + "RFFT2D", + nullptr}; + return names; +} + +inline const char *EnumNameBuiltinOperator(BuiltinOperator e) { + if (flatbuffers::IsOutRange(e, BuiltinOperator_ADD, BuiltinOperator_RFFT2D)) + return ""; + const size_t index = static_cast(e); + return EnumNamesBuiltinOperator()[index]; +} + +enum BuiltinOptions { + BuiltinOptions_NONE = 0, + BuiltinOptions_Conv2DOptions = 1, + BuiltinOptions_DepthwiseConv2DOptions = 2, + BuiltinOptions_ConcatEmbeddingsOptions = 3, + BuiltinOptions_LSHProjectionOptions = 4, + BuiltinOptions_Pool2DOptions = 5, + BuiltinOptions_SVDFOptions = 6, + BuiltinOptions_RNNOptions = 7, + BuiltinOptions_FullyConnectedOptions = 8, + BuiltinOptions_SoftmaxOptions = 9, + BuiltinOptions_ConcatenationOptions = 10, + BuiltinOptions_AddOptions = 11, + BuiltinOptions_L2NormOptions = 12, + BuiltinOptions_LocalResponseNormalizationOptions = 13, + BuiltinOptions_LSTMOptions = 14, + BuiltinOptions_ResizeBilinearOptions = 15, + BuiltinOptions_CallOptions = 16, + BuiltinOptions_ReshapeOptions = 17, + BuiltinOptions_SkipGramOptions = 18, + BuiltinOptions_SpaceToDepthOptions = 19, + BuiltinOptions_EmbeddingLookupSparseOptions = 20, + BuiltinOptions_MulOptions = 21, + BuiltinOptions_PadOptions = 22, + BuiltinOptions_GatherOptions = 23, + BuiltinOptions_BatchToSpaceNDOptions = 24, + BuiltinOptions_SpaceToBatchNDOptions = 25, + BuiltinOptions_TransposeOptions = 26, + BuiltinOptions_ReducerOptions = 27, + BuiltinOptions_SubOptions = 28, + BuiltinOptions_DivOptions = 29, + BuiltinOptions_SqueezeOptions = 30, + BuiltinOptions_SequenceRNNOptions = 31, + BuiltinOptions_StridedSliceOptions = 32, + BuiltinOptions_ExpOptions = 33, + BuiltinOptions_TopKV2Options = 34, + BuiltinOptions_SplitOptions = 35, + BuiltinOptions_LogSoftmaxOptions = 36, + BuiltinOptions_CastOptions = 37, + BuiltinOptions_DequantizeOptions = 38, + BuiltinOptions_MaximumMinimumOptions = 39, + BuiltinOptions_ArgMaxOptions = 40, + BuiltinOptions_LessOptions = 41, + BuiltinOptions_NegOptions = 42, + BuiltinOptions_PadV2Options = 43, + BuiltinOptions_GreaterOptions = 44, + BuiltinOptions_GreaterEqualOptions = 45, + BuiltinOptions_LessEqualOptions = 46, + BuiltinOptions_SelectOptions = 47, + BuiltinOptions_SliceOptions = 48, + BuiltinOptions_TransposeConvOptions = 49, + BuiltinOptions_SparseToDenseOptions = 50, + BuiltinOptions_TileOptions = 51, + BuiltinOptions_ExpandDimsOptions = 52, + BuiltinOptions_EqualOptions = 53, + BuiltinOptions_NotEqualOptions = 54, + BuiltinOptions_ShapeOptions = 55, + BuiltinOptions_PowOptions = 56, + BuiltinOptions_ArgMinOptions = 57, + BuiltinOptions_FakeQuantOptions = 58, + BuiltinOptions_PackOptions = 59, + BuiltinOptions_LogicalOrOptions = 60, + BuiltinOptions_OneHotOptions = 61, + BuiltinOptions_LogicalAndOptions = 62, + BuiltinOptions_LogicalNotOptions = 63, + BuiltinOptions_UnpackOptions = 64, + BuiltinOptions_FloorDivOptions = 65, + BuiltinOptions_SquareOptions = 66, + BuiltinOptions_ZerosLikeOptions = 67, + BuiltinOptions_FillOptions = 68, + BuiltinOptions_BidirectionalSequenceLSTMOptions = 69, + BuiltinOptions_BidirectionalSequenceRNNOptions = 70, + BuiltinOptions_UnidirectionalSequenceLSTMOptions = 71, + BuiltinOptions_FloorModOptions = 72, + BuiltinOptions_RangeOptions = 73, + BuiltinOptions_ResizeNearestNeighborOptions = 74, + BuiltinOptions_LeakyReluOptions = 75, + BuiltinOptions_SquaredDifferenceOptions = 76, + BuiltinOptions_MirrorPadOptions = 77, + BuiltinOptions_AbsOptions = 78, + BuiltinOptions_SplitVOptions = 79, + BuiltinOptions_UniqueOptions = 80, + BuiltinOptions_ReverseV2Options = 81, + BuiltinOptions_AddNOptions = 82, + BuiltinOptions_GatherNdOptions = 83, + BuiltinOptions_CosOptions = 84, + BuiltinOptions_WhereOptions = 85, + BuiltinOptions_RankOptions = 86, + BuiltinOptions_ReverseSequenceOptions = 87, + BuiltinOptions_MatrixDiagOptions = 88, + BuiltinOptions_QuantizeOptions = 89, + BuiltinOptions_MatrixSetDiagOptions = 90, + BuiltinOptions_HardSwishOptions = 91, + BuiltinOptions_IfOptions = 92, + BuiltinOptions_WhileOptions = 93, + BuiltinOptions_DepthToSpaceOptions = 94, + BuiltinOptions_NonMaxSuppressionV4Options = 95, + BuiltinOptions_NonMaxSuppressionV5Options = 96, + BuiltinOptions_ScatterNdOptions = 97, + BuiltinOptions_SelectV2Options = 98, + BuiltinOptions_DensifyOptions = 99, + BuiltinOptions_SegmentSumOptions = 100, + BuiltinOptions_BatchMatMulOptions = 101, + BuiltinOptions_CumsumOptions = 102, + BuiltinOptions_CallOnceOptions = 103, + BuiltinOptions_BroadcastToOptions = 104, + BuiltinOptions_Rfft2dOptions = 105, + BuiltinOptions_MIN = BuiltinOptions_NONE, + BuiltinOptions_MAX = BuiltinOptions_Rfft2dOptions +}; + +inline const BuiltinOptions (&EnumValuesBuiltinOptions())[106] { + static const BuiltinOptions values[] = { + BuiltinOptions_NONE, + BuiltinOptions_Conv2DOptions, + BuiltinOptions_DepthwiseConv2DOptions, + BuiltinOptions_ConcatEmbeddingsOptions, + BuiltinOptions_LSHProjectionOptions, + BuiltinOptions_Pool2DOptions, + BuiltinOptions_SVDFOptions, + BuiltinOptions_RNNOptions, + BuiltinOptions_FullyConnectedOptions, + BuiltinOptions_SoftmaxOptions, + BuiltinOptions_ConcatenationOptions, + BuiltinOptions_AddOptions, + BuiltinOptions_L2NormOptions, + BuiltinOptions_LocalResponseNormalizationOptions, + BuiltinOptions_LSTMOptions, + BuiltinOptions_ResizeBilinearOptions, + BuiltinOptions_CallOptions, + BuiltinOptions_ReshapeOptions, + BuiltinOptions_SkipGramOptions, + BuiltinOptions_SpaceToDepthOptions, + BuiltinOptions_EmbeddingLookupSparseOptions, + BuiltinOptions_MulOptions, + BuiltinOptions_PadOptions, + BuiltinOptions_GatherOptions, + BuiltinOptions_BatchToSpaceNDOptions, + BuiltinOptions_SpaceToBatchNDOptions, + BuiltinOptions_TransposeOptions, + BuiltinOptions_ReducerOptions, + BuiltinOptions_SubOptions, + BuiltinOptions_DivOptions, + BuiltinOptions_SqueezeOptions, + BuiltinOptions_SequenceRNNOptions, + BuiltinOptions_StridedSliceOptions, + BuiltinOptions_ExpOptions, + BuiltinOptions_TopKV2Options, + BuiltinOptions_SplitOptions, + BuiltinOptions_LogSoftmaxOptions, + BuiltinOptions_CastOptions, + BuiltinOptions_DequantizeOptions, + BuiltinOptions_MaximumMinimumOptions, + BuiltinOptions_ArgMaxOptions, + BuiltinOptions_LessOptions, + BuiltinOptions_NegOptions, + BuiltinOptions_PadV2Options, + BuiltinOptions_GreaterOptions, + BuiltinOptions_GreaterEqualOptions, + BuiltinOptions_LessEqualOptions, + BuiltinOptions_SelectOptions, + BuiltinOptions_SliceOptions, + BuiltinOptions_TransposeConvOptions, + BuiltinOptions_SparseToDenseOptions, + BuiltinOptions_TileOptions, + BuiltinOptions_ExpandDimsOptions, + BuiltinOptions_EqualOptions, + BuiltinOptions_NotEqualOptions, + BuiltinOptions_ShapeOptions, + BuiltinOptions_PowOptions, + BuiltinOptions_ArgMinOptions, + BuiltinOptions_FakeQuantOptions, + BuiltinOptions_PackOptions, + BuiltinOptions_LogicalOrOptions, + BuiltinOptions_OneHotOptions, + BuiltinOptions_LogicalAndOptions, + BuiltinOptions_LogicalNotOptions, + BuiltinOptions_UnpackOptions, + BuiltinOptions_FloorDivOptions, + BuiltinOptions_SquareOptions, + BuiltinOptions_ZerosLikeOptions, + BuiltinOptions_FillOptions, + BuiltinOptions_BidirectionalSequenceLSTMOptions, + BuiltinOptions_BidirectionalSequenceRNNOptions, + BuiltinOptions_UnidirectionalSequenceLSTMOptions, + BuiltinOptions_FloorModOptions, + BuiltinOptions_RangeOptions, + BuiltinOptions_ResizeNearestNeighborOptions, + BuiltinOptions_LeakyReluOptions, + BuiltinOptions_SquaredDifferenceOptions, + BuiltinOptions_MirrorPadOptions, + BuiltinOptions_AbsOptions, + BuiltinOptions_SplitVOptions, + BuiltinOptions_UniqueOptions, + BuiltinOptions_ReverseV2Options, + BuiltinOptions_AddNOptions, + BuiltinOptions_GatherNdOptions, + BuiltinOptions_CosOptions, + BuiltinOptions_WhereOptions, + BuiltinOptions_RankOptions, + BuiltinOptions_ReverseSequenceOptions, + BuiltinOptions_MatrixDiagOptions, + BuiltinOptions_QuantizeOptions, + BuiltinOptions_MatrixSetDiagOptions, + BuiltinOptions_HardSwishOptions, + BuiltinOptions_IfOptions, + BuiltinOptions_WhileOptions, + BuiltinOptions_DepthToSpaceOptions, + BuiltinOptions_NonMaxSuppressionV4Options, + BuiltinOptions_NonMaxSuppressionV5Options, + BuiltinOptions_ScatterNdOptions, + BuiltinOptions_SelectV2Options, + BuiltinOptions_DensifyOptions, + BuiltinOptions_SegmentSumOptions, + BuiltinOptions_BatchMatMulOptions, + BuiltinOptions_CumsumOptions, + BuiltinOptions_CallOnceOptions, + BuiltinOptions_BroadcastToOptions, + BuiltinOptions_Rfft2dOptions}; + return values; +} + +inline const char * const *EnumNamesBuiltinOptions() { + static const char *const names[107] = {"NONE", + "Conv2DOptions", + "DepthwiseConv2DOptions", + "ConcatEmbeddingsOptions", + "LSHProjectionOptions", + "Pool2DOptions", + "SVDFOptions", + "RNNOptions", + "FullyConnectedOptions", + "SoftmaxOptions", + "ConcatenationOptions", + "AddOptions", + "L2NormOptions", + "LocalResponseNormalizationOptions", + "LSTMOptions", + "ResizeBilinearOptions", + "CallOptions", + "ReshapeOptions", + "SkipGramOptions", + "SpaceToDepthOptions", + "EmbeddingLookupSparseOptions", + "MulOptions", + "PadOptions", + "GatherOptions", + "BatchToSpaceNDOptions", + "SpaceToBatchNDOptions", + "TransposeOptions", + "ReducerOptions", + "SubOptions", + "DivOptions", + "SqueezeOptions", + "SequenceRNNOptions", + "StridedSliceOptions", + "ExpOptions", + "TopKV2Options", + "SplitOptions", + "LogSoftmaxOptions", + "CastOptions", + "DequantizeOptions", + "MaximumMinimumOptions", + "ArgMaxOptions", + "LessOptions", + "NegOptions", + "PadV2Options", + "GreaterOptions", + "GreaterEqualOptions", + "LessEqualOptions", + "SelectOptions", + "SliceOptions", + "TransposeConvOptions", + "SparseToDenseOptions", + "TileOptions", + "ExpandDimsOptions", + "EqualOptions", + "NotEqualOptions", + "ShapeOptions", + "PowOptions", + "ArgMinOptions", + "FakeQuantOptions", + "PackOptions", + "LogicalOrOptions", + "OneHotOptions", + "LogicalAndOptions", + "LogicalNotOptions", + "UnpackOptions", + "FloorDivOptions", + "SquareOptions", + "ZerosLikeOptions", + "FillOptions", + "BidirectionalSequenceLSTMOptions", + "BidirectionalSequenceRNNOptions", + "UnidirectionalSequenceLSTMOptions", + "FloorModOptions", + "RangeOptions", + "ResizeNearestNeighborOptions", + "LeakyReluOptions", + "SquaredDifferenceOptions", + "MirrorPadOptions", + "AbsOptions", + "SplitVOptions", + "UniqueOptions", + "ReverseV2Options", + "AddNOptions", + "GatherNdOptions", + "CosOptions", + "WhereOptions", + "RankOptions", + "ReverseSequenceOptions", + "MatrixDiagOptions", + "QuantizeOptions", + "MatrixSetDiagOptions", + "HardSwishOptions", + "IfOptions", + "WhileOptions", + "DepthToSpaceOptions", + "NonMaxSuppressionV4Options", + "NonMaxSuppressionV5Options", + "ScatterNdOptions", + "SelectV2Options", + "DensifyOptions", + "SegmentSumOptions", + "BatchMatMulOptions", + "CumsumOptions", + "CallOnceOptions", + "BroadcastToOptions", + "Rfft2dOptions", + nullptr}; + return names; +} + +inline const char *EnumNameBuiltinOptions(BuiltinOptions e) { + if (flatbuffers::IsOutRange(e, BuiltinOptions_NONE, + BuiltinOptions_Rfft2dOptions)) + return ""; + const size_t index = static_cast(e); + return EnumNamesBuiltinOptions()[index]; +} + +template struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_NONE; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_Conv2DOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_DepthwiseConv2DOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ConcatEmbeddingsOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LSHProjectionOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_Pool2DOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SVDFOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_RNNOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_FullyConnectedOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SoftmaxOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ConcatenationOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_AddOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_L2NormOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LocalResponseNormalizationOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LSTMOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ResizeBilinearOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_CallOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ReshapeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SkipGramOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SpaceToDepthOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_EmbeddingLookupSparseOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_MulOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_PadOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_GatherOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_BatchToSpaceNDOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SpaceToBatchNDOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_TransposeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ReducerOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SubOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_DivOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SqueezeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SequenceRNNOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_StridedSliceOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ExpOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_TopKV2Options; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SplitOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LogSoftmaxOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_CastOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_DequantizeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_MaximumMinimumOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ArgMaxOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LessOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_NegOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_PadV2Options; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_GreaterOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_GreaterEqualOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LessEqualOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SelectOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SliceOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_TransposeConvOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SparseToDenseOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_TileOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ExpandDimsOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_EqualOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_NotEqualOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ShapeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_PowOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ArgMinOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_FakeQuantOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_PackOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LogicalOrOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_OneHotOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LogicalAndOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LogicalNotOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_UnpackOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_FloorDivOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SquareOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ZerosLikeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_FillOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_BidirectionalSequenceLSTMOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_BidirectionalSequenceRNNOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_UnidirectionalSequenceLSTMOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_FloorModOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_RangeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ResizeNearestNeighborOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_LeakyReluOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SquaredDifferenceOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_MirrorPadOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_AbsOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SplitVOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_UniqueOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ReverseV2Options; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_AddNOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_GatherNdOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_CosOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_WhereOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_RankOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ReverseSequenceOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_MatrixDiagOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_QuantizeOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_MatrixSetDiagOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_HardSwishOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_IfOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_WhileOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_DepthToSpaceOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_NonMaxSuppressionV4Options; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_NonMaxSuppressionV5Options; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ScatterNdOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SelectV2Options; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_DensifyOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_SegmentSumOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_BatchMatMulOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_CumsumOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_CallOnceOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_BroadcastToOptions; +}; + +template <> +struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_Rfft2dOptions; +}; + +struct BuiltinOptionsUnion { + BuiltinOptions type; + void *value; + + BuiltinOptionsUnion() : type(BuiltinOptions_NONE), value(nullptr) {} + BuiltinOptionsUnion(BuiltinOptionsUnion&& u) FLATBUFFERS_NOEXCEPT : + type(BuiltinOptions_NONE), value(nullptr) + { std::swap(type, u.type); std::swap(value, u.value); } + BuiltinOptionsUnion(const BuiltinOptionsUnion &) FLATBUFFERS_NOEXCEPT; + BuiltinOptionsUnion &operator=(const BuiltinOptionsUnion &u) FLATBUFFERS_NOEXCEPT + { BuiltinOptionsUnion t(u); std::swap(type, t.type); std::swap(value, t.value); return *this; } + BuiltinOptionsUnion &operator=(BuiltinOptionsUnion &&u) FLATBUFFERS_NOEXCEPT + { std::swap(type, u.type); std::swap(value, u.value); return *this; } + ~BuiltinOptionsUnion() { Reset(); } + + void Reset(); + +#ifndef FLATBUFFERS_CPP98_STL + template + void Set(T&& val) { + using RT = typename std::remove_reference::type; + Reset(); + type = BuiltinOptionsTraits::enum_value; + if (type != BuiltinOptions_NONE) { + value = new RT(std::forward(val)); + } + } +#endif // FLATBUFFERS_CPP98_STL + + static void *UnPack(const void *obj, BuiltinOptions type, const flatbuffers::resolver_function_t *resolver); + flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const flatbuffers::rehasher_function_t *_rehasher = nullptr) const; + + tflite::Conv2DOptionsT *AsConv2DOptions() { + return type == BuiltinOptions_Conv2DOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::Conv2DOptionsT *AsConv2DOptions() const { + return type == BuiltinOptions_Conv2DOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::DepthwiseConv2DOptionsT *AsDepthwiseConv2DOptions() { + return type == BuiltinOptions_DepthwiseConv2DOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::DepthwiseConv2DOptionsT *AsDepthwiseConv2DOptions() const { + return type == BuiltinOptions_DepthwiseConv2DOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ConcatEmbeddingsOptionsT *AsConcatEmbeddingsOptions() { + return type == BuiltinOptions_ConcatEmbeddingsOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ConcatEmbeddingsOptionsT *AsConcatEmbeddingsOptions() const { + return type == BuiltinOptions_ConcatEmbeddingsOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LSHProjectionOptionsT *AsLSHProjectionOptions() { + return type == BuiltinOptions_LSHProjectionOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LSHProjectionOptionsT *AsLSHProjectionOptions() const { + return type == BuiltinOptions_LSHProjectionOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::Pool2DOptionsT *AsPool2DOptions() { + return type == BuiltinOptions_Pool2DOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::Pool2DOptionsT *AsPool2DOptions() const { + return type == BuiltinOptions_Pool2DOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SVDFOptionsT *AsSVDFOptions() { + return type == BuiltinOptions_SVDFOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SVDFOptionsT *AsSVDFOptions() const { + return type == BuiltinOptions_SVDFOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::RNNOptionsT *AsRNNOptions() { + return type == BuiltinOptions_RNNOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::RNNOptionsT *AsRNNOptions() const { + return type == BuiltinOptions_RNNOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::FullyConnectedOptionsT *AsFullyConnectedOptions() { + return type == BuiltinOptions_FullyConnectedOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::FullyConnectedOptionsT *AsFullyConnectedOptions() const { + return type == BuiltinOptions_FullyConnectedOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SoftmaxOptionsT *AsSoftmaxOptions() { + return type == BuiltinOptions_SoftmaxOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SoftmaxOptionsT *AsSoftmaxOptions() const { + return type == BuiltinOptions_SoftmaxOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ConcatenationOptionsT *AsConcatenationOptions() { + return type == BuiltinOptions_ConcatenationOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ConcatenationOptionsT *AsConcatenationOptions() const { + return type == BuiltinOptions_ConcatenationOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::AddOptionsT *AsAddOptions() { + return type == BuiltinOptions_AddOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::AddOptionsT *AsAddOptions() const { + return type == BuiltinOptions_AddOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::L2NormOptionsT *AsL2NormOptions() { + return type == BuiltinOptions_L2NormOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::L2NormOptionsT *AsL2NormOptions() const { + return type == BuiltinOptions_L2NormOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LocalResponseNormalizationOptionsT *AsLocalResponseNormalizationOptions() { + return type == BuiltinOptions_LocalResponseNormalizationOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LocalResponseNormalizationOptionsT *AsLocalResponseNormalizationOptions() const { + return type == BuiltinOptions_LocalResponseNormalizationOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LSTMOptionsT *AsLSTMOptions() { + return type == BuiltinOptions_LSTMOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LSTMOptionsT *AsLSTMOptions() const { + return type == BuiltinOptions_LSTMOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ResizeBilinearOptionsT *AsResizeBilinearOptions() { + return type == BuiltinOptions_ResizeBilinearOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ResizeBilinearOptionsT *AsResizeBilinearOptions() const { + return type == BuiltinOptions_ResizeBilinearOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::CallOptionsT *AsCallOptions() { + return type == BuiltinOptions_CallOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::CallOptionsT *AsCallOptions() const { + return type == BuiltinOptions_CallOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ReshapeOptionsT *AsReshapeOptions() { + return type == BuiltinOptions_ReshapeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ReshapeOptionsT *AsReshapeOptions() const { + return type == BuiltinOptions_ReshapeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SkipGramOptionsT *AsSkipGramOptions() { + return type == BuiltinOptions_SkipGramOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SkipGramOptionsT *AsSkipGramOptions() const { + return type == BuiltinOptions_SkipGramOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SpaceToDepthOptionsT *AsSpaceToDepthOptions() { + return type == BuiltinOptions_SpaceToDepthOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SpaceToDepthOptionsT *AsSpaceToDepthOptions() const { + return type == BuiltinOptions_SpaceToDepthOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::EmbeddingLookupSparseOptionsT *AsEmbeddingLookupSparseOptions() { + return type == BuiltinOptions_EmbeddingLookupSparseOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::EmbeddingLookupSparseOptionsT *AsEmbeddingLookupSparseOptions() const { + return type == BuiltinOptions_EmbeddingLookupSparseOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::MulOptionsT *AsMulOptions() { + return type == BuiltinOptions_MulOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::MulOptionsT *AsMulOptions() const { + return type == BuiltinOptions_MulOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::PadOptionsT *AsPadOptions() { + return type == BuiltinOptions_PadOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::PadOptionsT *AsPadOptions() const { + return type == BuiltinOptions_PadOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::GatherOptionsT *AsGatherOptions() { + return type == BuiltinOptions_GatherOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::GatherOptionsT *AsGatherOptions() const { + return type == BuiltinOptions_GatherOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::BatchToSpaceNDOptionsT *AsBatchToSpaceNDOptions() { + return type == BuiltinOptions_BatchToSpaceNDOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::BatchToSpaceNDOptionsT *AsBatchToSpaceNDOptions() const { + return type == BuiltinOptions_BatchToSpaceNDOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SpaceToBatchNDOptionsT *AsSpaceToBatchNDOptions() { + return type == BuiltinOptions_SpaceToBatchNDOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SpaceToBatchNDOptionsT *AsSpaceToBatchNDOptions() const { + return type == BuiltinOptions_SpaceToBatchNDOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::TransposeOptionsT *AsTransposeOptions() { + return type == BuiltinOptions_TransposeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::TransposeOptionsT *AsTransposeOptions() const { + return type == BuiltinOptions_TransposeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ReducerOptionsT *AsReducerOptions() { + return type == BuiltinOptions_ReducerOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ReducerOptionsT *AsReducerOptions() const { + return type == BuiltinOptions_ReducerOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SubOptionsT *AsSubOptions() { + return type == BuiltinOptions_SubOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SubOptionsT *AsSubOptions() const { + return type == BuiltinOptions_SubOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::DivOptionsT *AsDivOptions() { + return type == BuiltinOptions_DivOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::DivOptionsT *AsDivOptions() const { + return type == BuiltinOptions_DivOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SqueezeOptionsT *AsSqueezeOptions() { + return type == BuiltinOptions_SqueezeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SqueezeOptionsT *AsSqueezeOptions() const { + return type == BuiltinOptions_SqueezeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SequenceRNNOptionsT *AsSequenceRNNOptions() { + return type == BuiltinOptions_SequenceRNNOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SequenceRNNOptionsT *AsSequenceRNNOptions() const { + return type == BuiltinOptions_SequenceRNNOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::StridedSliceOptionsT *AsStridedSliceOptions() { + return type == BuiltinOptions_StridedSliceOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::StridedSliceOptionsT *AsStridedSliceOptions() const { + return type == BuiltinOptions_StridedSliceOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ExpOptionsT *AsExpOptions() { + return type == BuiltinOptions_ExpOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ExpOptionsT *AsExpOptions() const { + return type == BuiltinOptions_ExpOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::TopKV2OptionsT *AsTopKV2Options() { + return type == BuiltinOptions_TopKV2Options ? + reinterpret_cast(value) : nullptr; + } + const tflite::TopKV2OptionsT *AsTopKV2Options() const { + return type == BuiltinOptions_TopKV2Options ? + reinterpret_cast(value) : nullptr; + } + tflite::SplitOptionsT *AsSplitOptions() { + return type == BuiltinOptions_SplitOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SplitOptionsT *AsSplitOptions() const { + return type == BuiltinOptions_SplitOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LogSoftmaxOptionsT *AsLogSoftmaxOptions() { + return type == BuiltinOptions_LogSoftmaxOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LogSoftmaxOptionsT *AsLogSoftmaxOptions() const { + return type == BuiltinOptions_LogSoftmaxOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::CastOptionsT *AsCastOptions() { + return type == BuiltinOptions_CastOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::CastOptionsT *AsCastOptions() const { + return type == BuiltinOptions_CastOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::DequantizeOptionsT *AsDequantizeOptions() { + return type == BuiltinOptions_DequantizeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::DequantizeOptionsT *AsDequantizeOptions() const { + return type == BuiltinOptions_DequantizeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::MaximumMinimumOptionsT *AsMaximumMinimumOptions() { + return type == BuiltinOptions_MaximumMinimumOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::MaximumMinimumOptionsT *AsMaximumMinimumOptions() const { + return type == BuiltinOptions_MaximumMinimumOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ArgMaxOptionsT *AsArgMaxOptions() { + return type == BuiltinOptions_ArgMaxOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ArgMaxOptionsT *AsArgMaxOptions() const { + return type == BuiltinOptions_ArgMaxOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LessOptionsT *AsLessOptions() { + return type == BuiltinOptions_LessOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LessOptionsT *AsLessOptions() const { + return type == BuiltinOptions_LessOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::NegOptionsT *AsNegOptions() { + return type == BuiltinOptions_NegOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::NegOptionsT *AsNegOptions() const { + return type == BuiltinOptions_NegOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::PadV2OptionsT *AsPadV2Options() { + return type == BuiltinOptions_PadV2Options ? + reinterpret_cast(value) : nullptr; + } + const tflite::PadV2OptionsT *AsPadV2Options() const { + return type == BuiltinOptions_PadV2Options ? + reinterpret_cast(value) : nullptr; + } + tflite::GreaterOptionsT *AsGreaterOptions() { + return type == BuiltinOptions_GreaterOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::GreaterOptionsT *AsGreaterOptions() const { + return type == BuiltinOptions_GreaterOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::GreaterEqualOptionsT *AsGreaterEqualOptions() { + return type == BuiltinOptions_GreaterEqualOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::GreaterEqualOptionsT *AsGreaterEqualOptions() const { + return type == BuiltinOptions_GreaterEqualOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LessEqualOptionsT *AsLessEqualOptions() { + return type == BuiltinOptions_LessEqualOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LessEqualOptionsT *AsLessEqualOptions() const { + return type == BuiltinOptions_LessEqualOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SelectOptionsT *AsSelectOptions() { + return type == BuiltinOptions_SelectOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SelectOptionsT *AsSelectOptions() const { + return type == BuiltinOptions_SelectOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SliceOptionsT *AsSliceOptions() { + return type == BuiltinOptions_SliceOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SliceOptionsT *AsSliceOptions() const { + return type == BuiltinOptions_SliceOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::TransposeConvOptionsT *AsTransposeConvOptions() { + return type == BuiltinOptions_TransposeConvOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::TransposeConvOptionsT *AsTransposeConvOptions() const { + return type == BuiltinOptions_TransposeConvOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SparseToDenseOptionsT *AsSparseToDenseOptions() { + return type == BuiltinOptions_SparseToDenseOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SparseToDenseOptionsT *AsSparseToDenseOptions() const { + return type == BuiltinOptions_SparseToDenseOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::TileOptionsT *AsTileOptions() { + return type == BuiltinOptions_TileOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::TileOptionsT *AsTileOptions() const { + return type == BuiltinOptions_TileOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ExpandDimsOptionsT *AsExpandDimsOptions() { + return type == BuiltinOptions_ExpandDimsOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ExpandDimsOptionsT *AsExpandDimsOptions() const { + return type == BuiltinOptions_ExpandDimsOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::EqualOptionsT *AsEqualOptions() { + return type == BuiltinOptions_EqualOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::EqualOptionsT *AsEqualOptions() const { + return type == BuiltinOptions_EqualOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::NotEqualOptionsT *AsNotEqualOptions() { + return type == BuiltinOptions_NotEqualOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::NotEqualOptionsT *AsNotEqualOptions() const { + return type == BuiltinOptions_NotEqualOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ShapeOptionsT *AsShapeOptions() { + return type == BuiltinOptions_ShapeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ShapeOptionsT *AsShapeOptions() const { + return type == BuiltinOptions_ShapeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::PowOptionsT *AsPowOptions() { + return type == BuiltinOptions_PowOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::PowOptionsT *AsPowOptions() const { + return type == BuiltinOptions_PowOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ArgMinOptionsT *AsArgMinOptions() { + return type == BuiltinOptions_ArgMinOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ArgMinOptionsT *AsArgMinOptions() const { + return type == BuiltinOptions_ArgMinOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::FakeQuantOptionsT *AsFakeQuantOptions() { + return type == BuiltinOptions_FakeQuantOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::FakeQuantOptionsT *AsFakeQuantOptions() const { + return type == BuiltinOptions_FakeQuantOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::PackOptionsT *AsPackOptions() { + return type == BuiltinOptions_PackOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::PackOptionsT *AsPackOptions() const { + return type == BuiltinOptions_PackOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LogicalOrOptionsT *AsLogicalOrOptions() { + return type == BuiltinOptions_LogicalOrOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LogicalOrOptionsT *AsLogicalOrOptions() const { + return type == BuiltinOptions_LogicalOrOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::OneHotOptionsT *AsOneHotOptions() { + return type == BuiltinOptions_OneHotOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::OneHotOptionsT *AsOneHotOptions() const { + return type == BuiltinOptions_OneHotOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LogicalAndOptionsT *AsLogicalAndOptions() { + return type == BuiltinOptions_LogicalAndOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LogicalAndOptionsT *AsLogicalAndOptions() const { + return type == BuiltinOptions_LogicalAndOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LogicalNotOptionsT *AsLogicalNotOptions() { + return type == BuiltinOptions_LogicalNotOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LogicalNotOptionsT *AsLogicalNotOptions() const { + return type == BuiltinOptions_LogicalNotOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::UnpackOptionsT *AsUnpackOptions() { + return type == BuiltinOptions_UnpackOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::UnpackOptionsT *AsUnpackOptions() const { + return type == BuiltinOptions_UnpackOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::FloorDivOptionsT *AsFloorDivOptions() { + return type == BuiltinOptions_FloorDivOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::FloorDivOptionsT *AsFloorDivOptions() const { + return type == BuiltinOptions_FloorDivOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SquareOptionsT *AsSquareOptions() { + return type == BuiltinOptions_SquareOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SquareOptionsT *AsSquareOptions() const { + return type == BuiltinOptions_SquareOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ZerosLikeOptionsT *AsZerosLikeOptions() { + return type == BuiltinOptions_ZerosLikeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ZerosLikeOptionsT *AsZerosLikeOptions() const { + return type == BuiltinOptions_ZerosLikeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::FillOptionsT *AsFillOptions() { + return type == BuiltinOptions_FillOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::FillOptionsT *AsFillOptions() const { + return type == BuiltinOptions_FillOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::BidirectionalSequenceLSTMOptionsT *AsBidirectionalSequenceLSTMOptions() { + return type == BuiltinOptions_BidirectionalSequenceLSTMOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::BidirectionalSequenceLSTMOptionsT *AsBidirectionalSequenceLSTMOptions() const { + return type == BuiltinOptions_BidirectionalSequenceLSTMOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::BidirectionalSequenceRNNOptionsT *AsBidirectionalSequenceRNNOptions() { + return type == BuiltinOptions_BidirectionalSequenceRNNOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::BidirectionalSequenceRNNOptionsT *AsBidirectionalSequenceRNNOptions() const { + return type == BuiltinOptions_BidirectionalSequenceRNNOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::UnidirectionalSequenceLSTMOptionsT *AsUnidirectionalSequenceLSTMOptions() { + return type == BuiltinOptions_UnidirectionalSequenceLSTMOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::UnidirectionalSequenceLSTMOptionsT *AsUnidirectionalSequenceLSTMOptions() const { + return type == BuiltinOptions_UnidirectionalSequenceLSTMOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::FloorModOptionsT *AsFloorModOptions() { + return type == BuiltinOptions_FloorModOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::FloorModOptionsT *AsFloorModOptions() const { + return type == BuiltinOptions_FloorModOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::RangeOptionsT *AsRangeOptions() { + return type == BuiltinOptions_RangeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::RangeOptionsT *AsRangeOptions() const { + return type == BuiltinOptions_RangeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ResizeNearestNeighborOptionsT *AsResizeNearestNeighborOptions() { + return type == BuiltinOptions_ResizeNearestNeighborOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ResizeNearestNeighborOptionsT *AsResizeNearestNeighborOptions() const { + return type == BuiltinOptions_ResizeNearestNeighborOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::LeakyReluOptionsT *AsLeakyReluOptions() { + return type == BuiltinOptions_LeakyReluOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::LeakyReluOptionsT *AsLeakyReluOptions() const { + return type == BuiltinOptions_LeakyReluOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SquaredDifferenceOptionsT *AsSquaredDifferenceOptions() { + return type == BuiltinOptions_SquaredDifferenceOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SquaredDifferenceOptionsT *AsSquaredDifferenceOptions() const { + return type == BuiltinOptions_SquaredDifferenceOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::MirrorPadOptionsT *AsMirrorPadOptions() { + return type == BuiltinOptions_MirrorPadOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::MirrorPadOptionsT *AsMirrorPadOptions() const { + return type == BuiltinOptions_MirrorPadOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::AbsOptionsT *AsAbsOptions() { + return type == BuiltinOptions_AbsOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::AbsOptionsT *AsAbsOptions() const { + return type == BuiltinOptions_AbsOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SplitVOptionsT *AsSplitVOptions() { + return type == BuiltinOptions_SplitVOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SplitVOptionsT *AsSplitVOptions() const { + return type == BuiltinOptions_SplitVOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::UniqueOptionsT *AsUniqueOptions() { + return type == BuiltinOptions_UniqueOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::UniqueOptionsT *AsUniqueOptions() const { + return type == BuiltinOptions_UniqueOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ReverseV2OptionsT *AsReverseV2Options() { + return type == BuiltinOptions_ReverseV2Options ? + reinterpret_cast(value) : nullptr; + } + const tflite::ReverseV2OptionsT *AsReverseV2Options() const { + return type == BuiltinOptions_ReverseV2Options ? + reinterpret_cast(value) : nullptr; + } + tflite::AddNOptionsT *AsAddNOptions() { + return type == BuiltinOptions_AddNOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::AddNOptionsT *AsAddNOptions() const { + return type == BuiltinOptions_AddNOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::GatherNdOptionsT *AsGatherNdOptions() { + return type == BuiltinOptions_GatherNdOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::GatherNdOptionsT *AsGatherNdOptions() const { + return type == BuiltinOptions_GatherNdOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::CosOptionsT *AsCosOptions() { + return type == BuiltinOptions_CosOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::CosOptionsT *AsCosOptions() const { + return type == BuiltinOptions_CosOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::WhereOptionsT *AsWhereOptions() { + return type == BuiltinOptions_WhereOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::WhereOptionsT *AsWhereOptions() const { + return type == BuiltinOptions_WhereOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::RankOptionsT *AsRankOptions() { + return type == BuiltinOptions_RankOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::RankOptionsT *AsRankOptions() const { + return type == BuiltinOptions_RankOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::ReverseSequenceOptionsT *AsReverseSequenceOptions() { + return type == BuiltinOptions_ReverseSequenceOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ReverseSequenceOptionsT *AsReverseSequenceOptions() const { + return type == BuiltinOptions_ReverseSequenceOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::MatrixDiagOptionsT *AsMatrixDiagOptions() { + return type == BuiltinOptions_MatrixDiagOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::MatrixDiagOptionsT *AsMatrixDiagOptions() const { + return type == BuiltinOptions_MatrixDiagOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::QuantizeOptionsT *AsQuantizeOptions() { + return type == BuiltinOptions_QuantizeOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::QuantizeOptionsT *AsQuantizeOptions() const { + return type == BuiltinOptions_QuantizeOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::MatrixSetDiagOptionsT *AsMatrixSetDiagOptions() { + return type == BuiltinOptions_MatrixSetDiagOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::MatrixSetDiagOptionsT *AsMatrixSetDiagOptions() const { + return type == BuiltinOptions_MatrixSetDiagOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::HardSwishOptionsT *AsHardSwishOptions() { + return type == BuiltinOptions_HardSwishOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::HardSwishOptionsT *AsHardSwishOptions() const { + return type == BuiltinOptions_HardSwishOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::IfOptionsT *AsIfOptions() { + return type == BuiltinOptions_IfOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::IfOptionsT *AsIfOptions() const { + return type == BuiltinOptions_IfOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::WhileOptionsT *AsWhileOptions() { + return type == BuiltinOptions_WhileOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::WhileOptionsT *AsWhileOptions() const { + return type == BuiltinOptions_WhileOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::DepthToSpaceOptionsT *AsDepthToSpaceOptions() { + return type == BuiltinOptions_DepthToSpaceOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::DepthToSpaceOptionsT *AsDepthToSpaceOptions() const { + return type == BuiltinOptions_DepthToSpaceOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::NonMaxSuppressionV4OptionsT *AsNonMaxSuppressionV4Options() { + return type == BuiltinOptions_NonMaxSuppressionV4Options ? + reinterpret_cast(value) : nullptr; + } + const tflite::NonMaxSuppressionV4OptionsT *AsNonMaxSuppressionV4Options() const { + return type == BuiltinOptions_NonMaxSuppressionV4Options ? + reinterpret_cast(value) : nullptr; + } + tflite::NonMaxSuppressionV5OptionsT *AsNonMaxSuppressionV5Options() { + return type == BuiltinOptions_NonMaxSuppressionV5Options ? + reinterpret_cast(value) : nullptr; + } + const tflite::NonMaxSuppressionV5OptionsT *AsNonMaxSuppressionV5Options() const { + return type == BuiltinOptions_NonMaxSuppressionV5Options ? + reinterpret_cast(value) : nullptr; + } + tflite::ScatterNdOptionsT *AsScatterNdOptions() { + return type == BuiltinOptions_ScatterNdOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::ScatterNdOptionsT *AsScatterNdOptions() const { + return type == BuiltinOptions_ScatterNdOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SelectV2OptionsT *AsSelectV2Options() { + return type == BuiltinOptions_SelectV2Options ? + reinterpret_cast(value) : nullptr; + } + const tflite::SelectV2OptionsT *AsSelectV2Options() const { + return type == BuiltinOptions_SelectV2Options ? + reinterpret_cast(value) : nullptr; + } + tflite::DensifyOptionsT *AsDensifyOptions() { + return type == BuiltinOptions_DensifyOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::DensifyOptionsT *AsDensifyOptions() const { + return type == BuiltinOptions_DensifyOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::SegmentSumOptionsT *AsSegmentSumOptions() { + return type == BuiltinOptions_SegmentSumOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::SegmentSumOptionsT *AsSegmentSumOptions() const { + return type == BuiltinOptions_SegmentSumOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::BatchMatMulOptionsT *AsBatchMatMulOptions() { + return type == BuiltinOptions_BatchMatMulOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::BatchMatMulOptionsT *AsBatchMatMulOptions() const { + return type == BuiltinOptions_BatchMatMulOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::CumsumOptionsT *AsCumsumOptions() { + return type == BuiltinOptions_CumsumOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::CumsumOptionsT *AsCumsumOptions() const { + return type == BuiltinOptions_CumsumOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::CallOnceOptionsT *AsCallOnceOptions() { + return type == BuiltinOptions_CallOnceOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::CallOnceOptionsT *AsCallOnceOptions() const { + return type == BuiltinOptions_CallOnceOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::BroadcastToOptionsT *AsBroadcastToOptions() { + return type == BuiltinOptions_BroadcastToOptions ? + reinterpret_cast(value) : nullptr; + } + const tflite::BroadcastToOptionsT *AsBroadcastToOptions() const { + return type == BuiltinOptions_BroadcastToOptions ? + reinterpret_cast(value) : nullptr; + } + tflite::Rfft2dOptionsT *AsRfft2dOptions() { + return type == BuiltinOptions_Rfft2dOptions + ? reinterpret_cast(value) + : nullptr; + } + const tflite::Rfft2dOptionsT *AsRfft2dOptions() const { + return type == BuiltinOptions_Rfft2dOptions + ? reinterpret_cast(value) + : nullptr; + } +}; + +bool VerifyBuiltinOptions(flatbuffers::Verifier &verifier, const void *obj, BuiltinOptions type); +bool VerifyBuiltinOptionsVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector> *values, const flatbuffers::Vector *types); + +enum Padding { + Padding_SAME = 0, + Padding_VALID = 1, + Padding_MIN = Padding_SAME, + Padding_MAX = Padding_VALID +}; + +inline const Padding (&EnumValuesPadding())[2] { + static const Padding values[] = { + Padding_SAME, + Padding_VALID + }; + return values; +} + +inline const char * const *EnumNamesPadding() { + static const char * const names[3] = { + "SAME", + "VALID", + nullptr + }; + return names; +} + +inline const char *EnumNamePadding(Padding e) { + if (flatbuffers::IsOutRange(e, Padding_SAME, Padding_VALID)) return ""; + const size_t index = static_cast(e); + return EnumNamesPadding()[index]; +} + +enum ActivationFunctionType { + ActivationFunctionType_NONE = 0, + ActivationFunctionType_RELU = 1, + ActivationFunctionType_RELU_N1_TO_1 = 2, + ActivationFunctionType_RELU6 = 3, + ActivationFunctionType_TANH = 4, + ActivationFunctionType_SIGN_BIT = 5, + ActivationFunctionType_MIN = ActivationFunctionType_NONE, + ActivationFunctionType_MAX = ActivationFunctionType_SIGN_BIT +}; + +inline const ActivationFunctionType (&EnumValuesActivationFunctionType())[6] { + static const ActivationFunctionType values[] = { + ActivationFunctionType_NONE, + ActivationFunctionType_RELU, + ActivationFunctionType_RELU_N1_TO_1, + ActivationFunctionType_RELU6, + ActivationFunctionType_TANH, + ActivationFunctionType_SIGN_BIT + }; + return values; +} + +inline const char * const *EnumNamesActivationFunctionType() { + static const char * const names[7] = { + "NONE", + "RELU", + "RELU_N1_TO_1", + "RELU6", + "TANH", + "SIGN_BIT", + nullptr + }; + return names; +} + +inline const char *EnumNameActivationFunctionType(ActivationFunctionType e) { + if (flatbuffers::IsOutRange(e, ActivationFunctionType_NONE, ActivationFunctionType_SIGN_BIT)) return ""; + const size_t index = static_cast(e); + return EnumNamesActivationFunctionType()[index]; +} + +enum LSHProjectionType { + LSHProjectionType_UNKNOWN = 0, + LSHProjectionType_SPARSE = 1, + LSHProjectionType_DENSE = 2, + LSHProjectionType_MIN = LSHProjectionType_UNKNOWN, + LSHProjectionType_MAX = LSHProjectionType_DENSE +}; + +inline const LSHProjectionType (&EnumValuesLSHProjectionType())[3] { + static const LSHProjectionType values[] = { + LSHProjectionType_UNKNOWN, + LSHProjectionType_SPARSE, + LSHProjectionType_DENSE + }; + return values; +} + +inline const char * const *EnumNamesLSHProjectionType() { + static const char * const names[4] = { + "UNKNOWN", + "SPARSE", + "DENSE", + nullptr + }; + return names; +} + +inline const char *EnumNameLSHProjectionType(LSHProjectionType e) { + if (flatbuffers::IsOutRange(e, LSHProjectionType_UNKNOWN, LSHProjectionType_DENSE)) return ""; + const size_t index = static_cast(e); + return EnumNamesLSHProjectionType()[index]; +} + +enum FullyConnectedOptionsWeightsFormat { + FullyConnectedOptionsWeightsFormat_DEFAULT = 0, + FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8 = 1, + FullyConnectedOptionsWeightsFormat_MIN = FullyConnectedOptionsWeightsFormat_DEFAULT, + FullyConnectedOptionsWeightsFormat_MAX = FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8 +}; + +inline const FullyConnectedOptionsWeightsFormat (&EnumValuesFullyConnectedOptionsWeightsFormat())[2] { + static const FullyConnectedOptionsWeightsFormat values[] = { + FullyConnectedOptionsWeightsFormat_DEFAULT, + FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8 + }; + return values; +} + +inline const char * const *EnumNamesFullyConnectedOptionsWeightsFormat() { + static const char * const names[3] = { + "DEFAULT", + "SHUFFLED4x16INT8", + nullptr + }; + return names; +} + +inline const char *EnumNameFullyConnectedOptionsWeightsFormat(FullyConnectedOptionsWeightsFormat e) { + if (flatbuffers::IsOutRange(e, FullyConnectedOptionsWeightsFormat_DEFAULT, FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8)) return ""; + const size_t index = static_cast(e); + return EnumNamesFullyConnectedOptionsWeightsFormat()[index]; +} + +enum LSTMKernelType { + LSTMKernelType_FULL = 0, + LSTMKernelType_BASIC = 1, + LSTMKernelType_MIN = LSTMKernelType_FULL, + LSTMKernelType_MAX = LSTMKernelType_BASIC +}; + +inline const LSTMKernelType (&EnumValuesLSTMKernelType())[2] { + static const LSTMKernelType values[] = { + LSTMKernelType_FULL, + LSTMKernelType_BASIC + }; + return values; +} + +inline const char * const *EnumNamesLSTMKernelType() { + static const char * const names[3] = { + "FULL", + "BASIC", + nullptr + }; + return names; +} + +inline const char *EnumNameLSTMKernelType(LSTMKernelType e) { + if (flatbuffers::IsOutRange(e, LSTMKernelType_FULL, LSTMKernelType_BASIC)) return ""; + const size_t index = static_cast(e); + return EnumNamesLSTMKernelType()[index]; +} + +enum CombinerType { + CombinerType_SUM = 0, + CombinerType_MEAN = 1, + CombinerType_SQRTN = 2, + CombinerType_MIN = CombinerType_SUM, + CombinerType_MAX = CombinerType_SQRTN +}; + +inline const CombinerType (&EnumValuesCombinerType())[3] { + static const CombinerType values[] = { + CombinerType_SUM, + CombinerType_MEAN, + CombinerType_SQRTN + }; + return values; +} + +inline const char * const *EnumNamesCombinerType() { + static const char * const names[4] = { + "SUM", + "MEAN", + "SQRTN", + nullptr + }; + return names; +} + +inline const char *EnumNameCombinerType(CombinerType e) { + if (flatbuffers::IsOutRange(e, CombinerType_SUM, CombinerType_SQRTN)) return ""; + const size_t index = static_cast(e); + return EnumNamesCombinerType()[index]; +} + +enum MirrorPadMode { + MirrorPadMode_REFLECT = 0, + MirrorPadMode_SYMMETRIC = 1, + MirrorPadMode_MIN = MirrorPadMode_REFLECT, + MirrorPadMode_MAX = MirrorPadMode_SYMMETRIC +}; + +inline const MirrorPadMode (&EnumValuesMirrorPadMode())[2] { + static const MirrorPadMode values[] = { + MirrorPadMode_REFLECT, + MirrorPadMode_SYMMETRIC + }; + return values; +} + +inline const char * const *EnumNamesMirrorPadMode() { + static const char * const names[3] = { + "REFLECT", + "SYMMETRIC", + nullptr + }; + return names; +} + +inline const char *EnumNameMirrorPadMode(MirrorPadMode e) { + if (flatbuffers::IsOutRange(e, MirrorPadMode_REFLECT, MirrorPadMode_SYMMETRIC)) return ""; + const size_t index = static_cast(e); + return EnumNamesMirrorPadMode()[index]; +} + +enum CustomOptionsFormat { + CustomOptionsFormat_FLEXBUFFERS = 0, + CustomOptionsFormat_MIN = CustomOptionsFormat_FLEXBUFFERS, + CustomOptionsFormat_MAX = CustomOptionsFormat_FLEXBUFFERS +}; + +inline const CustomOptionsFormat (&EnumValuesCustomOptionsFormat())[1] { + static const CustomOptionsFormat values[] = { + CustomOptionsFormat_FLEXBUFFERS + }; + return values; +} + +inline const char * const *EnumNamesCustomOptionsFormat() { + static const char * const names[2] = { + "FLEXBUFFERS", + nullptr + }; + return names; +} + +inline const char *EnumNameCustomOptionsFormat(CustomOptionsFormat e) { + if (flatbuffers::IsOutRange(e, CustomOptionsFormat_FLEXBUFFERS, CustomOptionsFormat_FLEXBUFFERS)) return ""; + const size_t index = static_cast(e); + return EnumNamesCustomOptionsFormat()[index]; +} + +struct CustomQuantizationT : public flatbuffers::NativeTable { + typedef CustomQuantization TableType; + std::vector custom; + CustomQuantizationT() { + } +}; + +struct CustomQuantization FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef CustomQuantizationT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_CUSTOM = 4 + }; + const flatbuffers::Vector *custom() const { + return GetPointer *>(VT_CUSTOM); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_CUSTOM) && + verifier.VerifyVector(custom()) && + verifier.EndTable(); + } + CustomQuantizationT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(CustomQuantizationT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const CustomQuantizationT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct CustomQuantizationBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_custom(flatbuffers::Offset> custom) { + fbb_.AddOffset(CustomQuantization::VT_CUSTOM, custom); + } + explicit CustomQuantizationBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + CustomQuantizationBuilder &operator=(const CustomQuantizationBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateCustomQuantization( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> custom = 0) { + CustomQuantizationBuilder builder_(_fbb); + builder_.add_custom(custom); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateCustomQuantizationDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *custom = nullptr) { + if (custom) { _fbb.ForceVectorAlignment(custom->size(), sizeof(uint8_t), 16); } + auto custom__ = custom ? _fbb.CreateVector(*custom) : 0; + return tflite::CreateCustomQuantization( + _fbb, + custom__); +} + +flatbuffers::Offset CreateCustomQuantization(flatbuffers::FlatBufferBuilder &_fbb, const CustomQuantizationT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct QuantizationParametersT : public flatbuffers::NativeTable { + typedef QuantizationParameters TableType; + std::vector min; + std::vector max; + std::vector scale; + std::vector zero_point; + tflite::QuantizationDetailsUnion details; + int32_t quantized_dimension; + QuantizationParametersT() + : quantized_dimension(0) { + } +}; + +struct QuantizationParameters FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef QuantizationParametersT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_MIN = 4, + VT_MAX = 6, + VT_SCALE = 8, + VT_ZERO_POINT = 10, + VT_DETAILS_TYPE = 12, + VT_DETAILS = 14, + VT_QUANTIZED_DIMENSION = 16 + }; + const flatbuffers::Vector *min() const { + return GetPointer *>(VT_MIN); + } + const flatbuffers::Vector *max() const { + return GetPointer *>(VT_MAX); + } + const flatbuffers::Vector *scale() const { + return GetPointer *>(VT_SCALE); + } + const flatbuffers::Vector *zero_point() const { + return GetPointer *>(VT_ZERO_POINT); + } + tflite::QuantizationDetails details_type() const { + return static_cast(GetField(VT_DETAILS_TYPE, 0)); + } + const void *details() const { + return GetPointer(VT_DETAILS); + } + template const T *details_as() const; + const tflite::CustomQuantization *details_as_CustomQuantization() const { + return details_type() == tflite::QuantizationDetails_CustomQuantization ? static_cast(details()) : nullptr; + } + int32_t quantized_dimension() const { + return GetField(VT_QUANTIZED_DIMENSION, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_MIN) && + verifier.VerifyVector(min()) && + VerifyOffset(verifier, VT_MAX) && + verifier.VerifyVector(max()) && + VerifyOffset(verifier, VT_SCALE) && + verifier.VerifyVector(scale()) && + VerifyOffset(verifier, VT_ZERO_POINT) && + verifier.VerifyVector(zero_point()) && + VerifyField(verifier, VT_DETAILS_TYPE) && + VerifyOffset(verifier, VT_DETAILS) && + VerifyQuantizationDetails(verifier, details(), details_type()) && + VerifyField(verifier, VT_QUANTIZED_DIMENSION) && + verifier.EndTable(); + } + QuantizationParametersT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(QuantizationParametersT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const QuantizationParametersT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +template<> inline const tflite::CustomQuantization *QuantizationParameters::details_as() const { + return details_as_CustomQuantization(); +} + +struct QuantizationParametersBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_min(flatbuffers::Offset> min) { + fbb_.AddOffset(QuantizationParameters::VT_MIN, min); + } + void add_max(flatbuffers::Offset> max) { + fbb_.AddOffset(QuantizationParameters::VT_MAX, max); + } + void add_scale(flatbuffers::Offset> scale) { + fbb_.AddOffset(QuantizationParameters::VT_SCALE, scale); + } + void add_zero_point(flatbuffers::Offset> zero_point) { + fbb_.AddOffset(QuantizationParameters::VT_ZERO_POINT, zero_point); + } + void add_details_type(tflite::QuantizationDetails details_type) { + fbb_.AddElement(QuantizationParameters::VT_DETAILS_TYPE, static_cast(details_type), 0); + } + void add_details(flatbuffers::Offset details) { + fbb_.AddOffset(QuantizationParameters::VT_DETAILS, details); + } + void add_quantized_dimension(int32_t quantized_dimension) { + fbb_.AddElement(QuantizationParameters::VT_QUANTIZED_DIMENSION, quantized_dimension, 0); + } + explicit QuantizationParametersBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + QuantizationParametersBuilder &operator=(const QuantizationParametersBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateQuantizationParameters( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> min = 0, + flatbuffers::Offset> max = 0, + flatbuffers::Offset> scale = 0, + flatbuffers::Offset> zero_point = 0, + tflite::QuantizationDetails details_type = tflite::QuantizationDetails_NONE, + flatbuffers::Offset details = 0, + int32_t quantized_dimension = 0) { + QuantizationParametersBuilder builder_(_fbb); + builder_.add_quantized_dimension(quantized_dimension); + builder_.add_details(details); + builder_.add_zero_point(zero_point); + builder_.add_scale(scale); + builder_.add_max(max); + builder_.add_min(min); + builder_.add_details_type(details_type); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateQuantizationParametersDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *min = nullptr, + const std::vector *max = nullptr, + const std::vector *scale = nullptr, + const std::vector *zero_point = nullptr, + tflite::QuantizationDetails details_type = tflite::QuantizationDetails_NONE, + flatbuffers::Offset details = 0, + int32_t quantized_dimension = 0) { + auto min__ = min ? _fbb.CreateVector(*min) : 0; + auto max__ = max ? _fbb.CreateVector(*max) : 0; + auto scale__ = scale ? _fbb.CreateVector(*scale) : 0; + auto zero_point__ = zero_point ? _fbb.CreateVector(*zero_point) : 0; + return tflite::CreateQuantizationParameters( + _fbb, + min__, + max__, + scale__, + zero_point__, + details_type, + details, + quantized_dimension); +} + +flatbuffers::Offset CreateQuantizationParameters(flatbuffers::FlatBufferBuilder &_fbb, const QuantizationParametersT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct Int32VectorT : public flatbuffers::NativeTable { + typedef Int32Vector TableType; + std::vector values; + Int32VectorT() { + } +}; + +struct Int32Vector FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef Int32VectorT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_VALUES = 4 + }; + const flatbuffers::Vector *values() const { + return GetPointer *>(VT_VALUES); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_VALUES) && + verifier.VerifyVector(values()) && + verifier.EndTable(); + } + Int32VectorT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(Int32VectorT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const Int32VectorT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct Int32VectorBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_values(flatbuffers::Offset> values) { + fbb_.AddOffset(Int32Vector::VT_VALUES, values); + } + explicit Int32VectorBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + Int32VectorBuilder &operator=(const Int32VectorBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateInt32Vector( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> values = 0) { + Int32VectorBuilder builder_(_fbb); + builder_.add_values(values); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateInt32VectorDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *values = nullptr) { + auto values__ = values ? _fbb.CreateVector(*values) : 0; + return tflite::CreateInt32Vector( + _fbb, + values__); +} + +flatbuffers::Offset CreateInt32Vector(flatbuffers::FlatBufferBuilder &_fbb, const Int32VectorT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct Uint16VectorT : public flatbuffers::NativeTable { + typedef Uint16Vector TableType; + std::vector values; + Uint16VectorT() { + } +}; + +struct Uint16Vector FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef Uint16VectorT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_VALUES = 4 + }; + const flatbuffers::Vector *values() const { + return GetPointer *>(VT_VALUES); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_VALUES) && + verifier.VerifyVector(values()) && + verifier.EndTable(); + } + Uint16VectorT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(Uint16VectorT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const Uint16VectorT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct Uint16VectorBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_values(flatbuffers::Offset> values) { + fbb_.AddOffset(Uint16Vector::VT_VALUES, values); + } + explicit Uint16VectorBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + Uint16VectorBuilder &operator=(const Uint16VectorBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateUint16Vector( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> values = 0) { + Uint16VectorBuilder builder_(_fbb); + builder_.add_values(values); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateUint16VectorDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *values = nullptr) { + if (values) { _fbb.ForceVectorAlignment(values->size(), sizeof(uint16_t), 4); } + auto values__ = values ? _fbb.CreateVector(*values) : 0; + return tflite::CreateUint16Vector( + _fbb, + values__); +} + +flatbuffers::Offset CreateUint16Vector(flatbuffers::FlatBufferBuilder &_fbb, const Uint16VectorT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct Uint8VectorT : public flatbuffers::NativeTable { + typedef Uint8Vector TableType; + std::vector values; + Uint8VectorT() { + } +}; + +struct Uint8Vector FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef Uint8VectorT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_VALUES = 4 + }; + const flatbuffers::Vector *values() const { + return GetPointer *>(VT_VALUES); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_VALUES) && + verifier.VerifyVector(values()) && + verifier.EndTable(); + } + Uint8VectorT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(Uint8VectorT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const Uint8VectorT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct Uint8VectorBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_values(flatbuffers::Offset> values) { + fbb_.AddOffset(Uint8Vector::VT_VALUES, values); + } + explicit Uint8VectorBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + Uint8VectorBuilder &operator=(const Uint8VectorBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateUint8Vector( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> values = 0) { + Uint8VectorBuilder builder_(_fbb); + builder_.add_values(values); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateUint8VectorDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *values = nullptr) { + if (values) { _fbb.ForceVectorAlignment(values->size(), sizeof(uint8_t), 4); } + auto values__ = values ? _fbb.CreateVector(*values) : 0; + return tflite::CreateUint8Vector( + _fbb, + values__); +} + +flatbuffers::Offset CreateUint8Vector(flatbuffers::FlatBufferBuilder &_fbb, const Uint8VectorT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct DimensionMetadataT : public flatbuffers::NativeTable { + typedef DimensionMetadata TableType; + tflite::DimensionType format; + int32_t dense_size; + tflite::SparseIndexVectorUnion array_segments; + tflite::SparseIndexVectorUnion array_indices; + DimensionMetadataT() + : format(tflite::DimensionType_DENSE), + dense_size(0) { + } +}; + +struct DimensionMetadata FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef DimensionMetadataT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FORMAT = 4, + VT_DENSE_SIZE = 6, + VT_ARRAY_SEGMENTS_TYPE = 8, + VT_ARRAY_SEGMENTS = 10, + VT_ARRAY_INDICES_TYPE = 12, + VT_ARRAY_INDICES = 14 + }; + tflite::DimensionType format() const { + return static_cast(GetField(VT_FORMAT, 0)); + } + int32_t dense_size() const { + return GetField(VT_DENSE_SIZE, 0); + } + tflite::SparseIndexVector array_segments_type() const { + return static_cast(GetField(VT_ARRAY_SEGMENTS_TYPE, 0)); + } + const void *array_segments() const { + return GetPointer(VT_ARRAY_SEGMENTS); + } + template const T *array_segments_as() const; + const tflite::Int32Vector *array_segments_as_Int32Vector() const { + return array_segments_type() == tflite::SparseIndexVector_Int32Vector ? static_cast(array_segments()) : nullptr; + } + const tflite::Uint16Vector *array_segments_as_Uint16Vector() const { + return array_segments_type() == tflite::SparseIndexVector_Uint16Vector ? static_cast(array_segments()) : nullptr; + } + const tflite::Uint8Vector *array_segments_as_Uint8Vector() const { + return array_segments_type() == tflite::SparseIndexVector_Uint8Vector ? static_cast(array_segments()) : nullptr; + } + tflite::SparseIndexVector array_indices_type() const { + return static_cast(GetField(VT_ARRAY_INDICES_TYPE, 0)); + } + const void *array_indices() const { + return GetPointer(VT_ARRAY_INDICES); + } + template const T *array_indices_as() const; + const tflite::Int32Vector *array_indices_as_Int32Vector() const { + return array_indices_type() == tflite::SparseIndexVector_Int32Vector ? static_cast(array_indices()) : nullptr; + } + const tflite::Uint16Vector *array_indices_as_Uint16Vector() const { + return array_indices_type() == tflite::SparseIndexVector_Uint16Vector ? static_cast(array_indices()) : nullptr; + } + const tflite::Uint8Vector *array_indices_as_Uint8Vector() const { + return array_indices_type() == tflite::SparseIndexVector_Uint8Vector ? static_cast(array_indices()) : nullptr; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FORMAT) && + VerifyField(verifier, VT_DENSE_SIZE) && + VerifyField(verifier, VT_ARRAY_SEGMENTS_TYPE) && + VerifyOffset(verifier, VT_ARRAY_SEGMENTS) && + VerifySparseIndexVector(verifier, array_segments(), array_segments_type()) && + VerifyField(verifier, VT_ARRAY_INDICES_TYPE) && + VerifyOffset(verifier, VT_ARRAY_INDICES) && + VerifySparseIndexVector(verifier, array_indices(), array_indices_type()) && + verifier.EndTable(); + } + DimensionMetadataT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(DimensionMetadataT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +template<> inline const tflite::Int32Vector *DimensionMetadata::array_segments_as() const { + return array_segments_as_Int32Vector(); +} + +template<> inline const tflite::Uint16Vector *DimensionMetadata::array_segments_as() const { + return array_segments_as_Uint16Vector(); +} + +template<> inline const tflite::Uint8Vector *DimensionMetadata::array_segments_as() const { + return array_segments_as_Uint8Vector(); +} + +template<> inline const tflite::Int32Vector *DimensionMetadata::array_indices_as() const { + return array_indices_as_Int32Vector(); +} + +template<> inline const tflite::Uint16Vector *DimensionMetadata::array_indices_as() const { + return array_indices_as_Uint16Vector(); +} + +template<> inline const tflite::Uint8Vector *DimensionMetadata::array_indices_as() const { + return array_indices_as_Uint8Vector(); +} + +struct DimensionMetadataBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_format(tflite::DimensionType format) { + fbb_.AddElement(DimensionMetadata::VT_FORMAT, static_cast(format), 0); + } + void add_dense_size(int32_t dense_size) { + fbb_.AddElement(DimensionMetadata::VT_DENSE_SIZE, dense_size, 0); + } + void add_array_segments_type(tflite::SparseIndexVector array_segments_type) { + fbb_.AddElement(DimensionMetadata::VT_ARRAY_SEGMENTS_TYPE, static_cast(array_segments_type), 0); + } + void add_array_segments(flatbuffers::Offset array_segments) { + fbb_.AddOffset(DimensionMetadata::VT_ARRAY_SEGMENTS, array_segments); + } + void add_array_indices_type(tflite::SparseIndexVector array_indices_type) { + fbb_.AddElement(DimensionMetadata::VT_ARRAY_INDICES_TYPE, static_cast(array_indices_type), 0); + } + void add_array_indices(flatbuffers::Offset array_indices) { + fbb_.AddOffset(DimensionMetadata::VT_ARRAY_INDICES, array_indices); + } + explicit DimensionMetadataBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + DimensionMetadataBuilder &operator=(const DimensionMetadataBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateDimensionMetadata( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::DimensionType format = tflite::DimensionType_DENSE, + int32_t dense_size = 0, + tflite::SparseIndexVector array_segments_type = tflite::SparseIndexVector_NONE, + flatbuffers::Offset array_segments = 0, + tflite::SparseIndexVector array_indices_type = tflite::SparseIndexVector_NONE, + flatbuffers::Offset array_indices = 0) { + DimensionMetadataBuilder builder_(_fbb); + builder_.add_array_indices(array_indices); + builder_.add_array_segments(array_segments); + builder_.add_dense_size(dense_size); + builder_.add_array_indices_type(array_indices_type); + builder_.add_array_segments_type(array_segments_type); + builder_.add_format(format); + return builder_.Finish(); +} + +flatbuffers::Offset CreateDimensionMetadata(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SparsityParametersT : public flatbuffers::NativeTable { + typedef SparsityParameters TableType; + std::vector traversal_order; + std::vector block_map; + std::vector> dim_metadata; + SparsityParametersT() { + } +}; + +struct SparsityParameters FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SparsityParametersT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_TRAVERSAL_ORDER = 4, + VT_BLOCK_MAP = 6, + VT_DIM_METADATA = 8 + }; + const flatbuffers::Vector *traversal_order() const { + return GetPointer *>(VT_TRAVERSAL_ORDER); + } + const flatbuffers::Vector *block_map() const { + return GetPointer *>(VT_BLOCK_MAP); + } + const flatbuffers::Vector> *dim_metadata() const { + return GetPointer> *>(VT_DIM_METADATA); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_TRAVERSAL_ORDER) && + verifier.VerifyVector(traversal_order()) && + VerifyOffset(verifier, VT_BLOCK_MAP) && + verifier.VerifyVector(block_map()) && + VerifyOffset(verifier, VT_DIM_METADATA) && + verifier.VerifyVector(dim_metadata()) && + verifier.VerifyVectorOfTables(dim_metadata()) && + verifier.EndTable(); + } + SparsityParametersT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SparsityParametersT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SparsityParametersBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_traversal_order(flatbuffers::Offset> traversal_order) { + fbb_.AddOffset(SparsityParameters::VT_TRAVERSAL_ORDER, traversal_order); + } + void add_block_map(flatbuffers::Offset> block_map) { + fbb_.AddOffset(SparsityParameters::VT_BLOCK_MAP, block_map); + } + void add_dim_metadata(flatbuffers::Offset>> dim_metadata) { + fbb_.AddOffset(SparsityParameters::VT_DIM_METADATA, dim_metadata); + } + explicit SparsityParametersBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SparsityParametersBuilder &operator=(const SparsityParametersBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSparsityParameters( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> traversal_order = 0, + flatbuffers::Offset> block_map = 0, + flatbuffers::Offset>> dim_metadata = 0) { + SparsityParametersBuilder builder_(_fbb); + builder_.add_dim_metadata(dim_metadata); + builder_.add_block_map(block_map); + builder_.add_traversal_order(traversal_order); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateSparsityParametersDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *traversal_order = nullptr, + const std::vector *block_map = nullptr, + const std::vector> *dim_metadata = nullptr) { + auto traversal_order__ = traversal_order ? _fbb.CreateVector(*traversal_order) : 0; + auto block_map__ = block_map ? _fbb.CreateVector(*block_map) : 0; + auto dim_metadata__ = dim_metadata ? _fbb.CreateVector>(*dim_metadata) : 0; + return tflite::CreateSparsityParameters( + _fbb, + traversal_order__, + block_map__, + dim_metadata__); +} + +flatbuffers::Offset CreateSparsityParameters(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct TensorT : public flatbuffers::NativeTable { + typedef Tensor TableType; + std::vector shape; + tflite::TensorType type; + uint32_t buffer; + std::string name; + std::unique_ptr quantization; + bool is_variable; + std::unique_ptr sparsity; + std::vector shape_signature; + TensorT() + : type(tflite::TensorType_FLOAT32), + buffer(0), + is_variable(false) { + } +}; + +struct Tensor FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef TensorT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_SHAPE = 4, + VT_TYPE = 6, + VT_BUFFER = 8, + VT_NAME = 10, + VT_QUANTIZATION = 12, + VT_IS_VARIABLE = 14, + VT_SPARSITY = 16, + VT_SHAPE_SIGNATURE = 18 + }; + const flatbuffers::Vector *shape() const { + return GetPointer *>(VT_SHAPE); + } + tflite::TensorType type() const { + return static_cast(GetField(VT_TYPE, 0)); + } + uint32_t buffer() const { + return GetField(VT_BUFFER, 0); + } + const flatbuffers::String *name() const { + return GetPointer(VT_NAME); + } + const tflite::QuantizationParameters *quantization() const { + return GetPointer(VT_QUANTIZATION); + } + bool is_variable() const { + return GetField(VT_IS_VARIABLE, 0) != 0; + } + const tflite::SparsityParameters *sparsity() const { + return GetPointer(VT_SPARSITY); + } + const flatbuffers::Vector *shape_signature() const { + return GetPointer *>(VT_SHAPE_SIGNATURE); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_SHAPE) && + verifier.VerifyVector(shape()) && + VerifyField(verifier, VT_TYPE) && + VerifyField(verifier, VT_BUFFER) && + VerifyOffset(verifier, VT_NAME) && + verifier.VerifyString(name()) && + VerifyOffset(verifier, VT_QUANTIZATION) && + verifier.VerifyTable(quantization()) && + VerifyField(verifier, VT_IS_VARIABLE) && + VerifyOffset(verifier, VT_SPARSITY) && + verifier.VerifyTable(sparsity()) && + VerifyOffset(verifier, VT_SHAPE_SIGNATURE) && + verifier.VerifyVector(shape_signature()) && + verifier.EndTable(); + } + TensorT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(TensorT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const TensorT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct TensorBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_shape(flatbuffers::Offset> shape) { + fbb_.AddOffset(Tensor::VT_SHAPE, shape); + } + void add_type(tflite::TensorType type) { + fbb_.AddElement(Tensor::VT_TYPE, static_cast(type), 0); + } + void add_buffer(uint32_t buffer) { + fbb_.AddElement(Tensor::VT_BUFFER, buffer, 0); + } + void add_name(flatbuffers::Offset name) { + fbb_.AddOffset(Tensor::VT_NAME, name); + } + void add_quantization(flatbuffers::Offset quantization) { + fbb_.AddOffset(Tensor::VT_QUANTIZATION, quantization); + } + void add_is_variable(bool is_variable) { + fbb_.AddElement(Tensor::VT_IS_VARIABLE, static_cast(is_variable), 0); + } + void add_sparsity(flatbuffers::Offset sparsity) { + fbb_.AddOffset(Tensor::VT_SPARSITY, sparsity); + } + void add_shape_signature(flatbuffers::Offset> shape_signature) { + fbb_.AddOffset(Tensor::VT_SHAPE_SIGNATURE, shape_signature); + } + explicit TensorBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + TensorBuilder &operator=(const TensorBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateTensor( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> shape = 0, + tflite::TensorType type = tflite::TensorType_FLOAT32, + uint32_t buffer = 0, + flatbuffers::Offset name = 0, + flatbuffers::Offset quantization = 0, + bool is_variable = false, + flatbuffers::Offset sparsity = 0, + flatbuffers::Offset> shape_signature = 0) { + TensorBuilder builder_(_fbb); + builder_.add_shape_signature(shape_signature); + builder_.add_sparsity(sparsity); + builder_.add_quantization(quantization); + builder_.add_name(name); + builder_.add_buffer(buffer); + builder_.add_shape(shape); + builder_.add_is_variable(is_variable); + builder_.add_type(type); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateTensorDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *shape = nullptr, + tflite::TensorType type = tflite::TensorType_FLOAT32, + uint32_t buffer = 0, + const char *name = nullptr, + flatbuffers::Offset quantization = 0, + bool is_variable = false, + flatbuffers::Offset sparsity = 0, + const std::vector *shape_signature = nullptr) { + auto shape__ = shape ? _fbb.CreateVector(*shape) : 0; + auto name__ = name ? _fbb.CreateString(name) : 0; + auto shape_signature__ = shape_signature ? _fbb.CreateVector(*shape_signature) : 0; + return tflite::CreateTensor( + _fbb, + shape__, + type, + buffer, + name__, + quantization, + is_variable, + sparsity, + shape_signature__); +} + +flatbuffers::Offset CreateTensor(flatbuffers::FlatBufferBuilder &_fbb, const TensorT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct Conv2DOptionsT : public flatbuffers::NativeTable { + typedef Conv2DOptions TableType; + tflite::Padding padding; + int32_t stride_w; + int32_t stride_h; + tflite::ActivationFunctionType fused_activation_function; + int32_t dilation_w_factor; + int32_t dilation_h_factor; + Conv2DOptionsT() + : padding(tflite::Padding_SAME), + stride_w(0), + stride_h(0), + fused_activation_function(tflite::ActivationFunctionType_NONE), + dilation_w_factor(1), + dilation_h_factor(1) { + } +}; + +struct Conv2DOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef Conv2DOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_PADDING = 4, + VT_STRIDE_W = 6, + VT_STRIDE_H = 8, + VT_FUSED_ACTIVATION_FUNCTION = 10, + VT_DILATION_W_FACTOR = 12, + VT_DILATION_H_FACTOR = 14 + }; + tflite::Padding padding() const { + return static_cast(GetField(VT_PADDING, 0)); + } + int32_t stride_w() const { + return GetField(VT_STRIDE_W, 0); + } + int32_t stride_h() const { + return GetField(VT_STRIDE_H, 0); + } + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + int32_t dilation_w_factor() const { + return GetField(VT_DILATION_W_FACTOR, 1); + } + int32_t dilation_h_factor() const { + return GetField(VT_DILATION_H_FACTOR, 1); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_PADDING) && + VerifyField(verifier, VT_STRIDE_W) && + VerifyField(verifier, VT_STRIDE_H) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_DILATION_W_FACTOR) && + VerifyField(verifier, VT_DILATION_H_FACTOR) && + verifier.EndTable(); + } + Conv2DOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(Conv2DOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const Conv2DOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct Conv2DOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_padding(tflite::Padding padding) { + fbb_.AddElement(Conv2DOptions::VT_PADDING, static_cast(padding), 0); + } + void add_stride_w(int32_t stride_w) { + fbb_.AddElement(Conv2DOptions::VT_STRIDE_W, stride_w, 0); + } + void add_stride_h(int32_t stride_h) { + fbb_.AddElement(Conv2DOptions::VT_STRIDE_H, stride_h, 0); + } + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(Conv2DOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_dilation_w_factor(int32_t dilation_w_factor) { + fbb_.AddElement(Conv2DOptions::VT_DILATION_W_FACTOR, dilation_w_factor, 1); + } + void add_dilation_h_factor(int32_t dilation_h_factor) { + fbb_.AddElement(Conv2DOptions::VT_DILATION_H_FACTOR, dilation_h_factor, 1); + } + explicit Conv2DOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + Conv2DOptionsBuilder &operator=(const Conv2DOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateConv2DOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::Padding padding = tflite::Padding_SAME, + int32_t stride_w = 0, + int32_t stride_h = 0, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + int32_t dilation_w_factor = 1, + int32_t dilation_h_factor = 1) { + Conv2DOptionsBuilder builder_(_fbb); + builder_.add_dilation_h_factor(dilation_h_factor); + builder_.add_dilation_w_factor(dilation_w_factor); + builder_.add_stride_h(stride_h); + builder_.add_stride_w(stride_w); + builder_.add_fused_activation_function(fused_activation_function); + builder_.add_padding(padding); + return builder_.Finish(); +} + +flatbuffers::Offset CreateConv2DOptions(flatbuffers::FlatBufferBuilder &_fbb, const Conv2DOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct Pool2DOptionsT : public flatbuffers::NativeTable { + typedef Pool2DOptions TableType; + tflite::Padding padding; + int32_t stride_w; + int32_t stride_h; + int32_t filter_width; + int32_t filter_height; + tflite::ActivationFunctionType fused_activation_function; + Pool2DOptionsT() + : padding(tflite::Padding_SAME), + stride_w(0), + stride_h(0), + filter_width(0), + filter_height(0), + fused_activation_function(tflite::ActivationFunctionType_NONE) { + } +}; + +struct Pool2DOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef Pool2DOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_PADDING = 4, + VT_STRIDE_W = 6, + VT_STRIDE_H = 8, + VT_FILTER_WIDTH = 10, + VT_FILTER_HEIGHT = 12, + VT_FUSED_ACTIVATION_FUNCTION = 14 + }; + tflite::Padding padding() const { + return static_cast(GetField(VT_PADDING, 0)); + } + int32_t stride_w() const { + return GetField(VT_STRIDE_W, 0); + } + int32_t stride_h() const { + return GetField(VT_STRIDE_H, 0); + } + int32_t filter_width() const { + return GetField(VT_FILTER_WIDTH, 0); + } + int32_t filter_height() const { + return GetField(VT_FILTER_HEIGHT, 0); + } + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_PADDING) && + VerifyField(verifier, VT_STRIDE_W) && + VerifyField(verifier, VT_STRIDE_H) && + VerifyField(verifier, VT_FILTER_WIDTH) && + VerifyField(verifier, VT_FILTER_HEIGHT) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + verifier.EndTable(); + } + Pool2DOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(Pool2DOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const Pool2DOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct Pool2DOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_padding(tflite::Padding padding) { + fbb_.AddElement(Pool2DOptions::VT_PADDING, static_cast(padding), 0); + } + void add_stride_w(int32_t stride_w) { + fbb_.AddElement(Pool2DOptions::VT_STRIDE_W, stride_w, 0); + } + void add_stride_h(int32_t stride_h) { + fbb_.AddElement(Pool2DOptions::VT_STRIDE_H, stride_h, 0); + } + void add_filter_width(int32_t filter_width) { + fbb_.AddElement(Pool2DOptions::VT_FILTER_WIDTH, filter_width, 0); + } + void add_filter_height(int32_t filter_height) { + fbb_.AddElement(Pool2DOptions::VT_FILTER_HEIGHT, filter_height, 0); + } + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(Pool2DOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + explicit Pool2DOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + Pool2DOptionsBuilder &operator=(const Pool2DOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreatePool2DOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::Padding padding = tflite::Padding_SAME, + int32_t stride_w = 0, + int32_t stride_h = 0, + int32_t filter_width = 0, + int32_t filter_height = 0, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE) { + Pool2DOptionsBuilder builder_(_fbb); + builder_.add_filter_height(filter_height); + builder_.add_filter_width(filter_width); + builder_.add_stride_h(stride_h); + builder_.add_stride_w(stride_w); + builder_.add_fused_activation_function(fused_activation_function); + builder_.add_padding(padding); + return builder_.Finish(); +} + +flatbuffers::Offset CreatePool2DOptions(flatbuffers::FlatBufferBuilder &_fbb, const Pool2DOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct DepthwiseConv2DOptionsT : public flatbuffers::NativeTable { + typedef DepthwiseConv2DOptions TableType; + tflite::Padding padding; + int32_t stride_w; + int32_t stride_h; + int32_t depth_multiplier; + tflite::ActivationFunctionType fused_activation_function; + int32_t dilation_w_factor; + int32_t dilation_h_factor; + DepthwiseConv2DOptionsT() + : padding(tflite::Padding_SAME), + stride_w(0), + stride_h(0), + depth_multiplier(0), + fused_activation_function(tflite::ActivationFunctionType_NONE), + dilation_w_factor(1), + dilation_h_factor(1) { + } +}; + +struct DepthwiseConv2DOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef DepthwiseConv2DOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_PADDING = 4, + VT_STRIDE_W = 6, + VT_STRIDE_H = 8, + VT_DEPTH_MULTIPLIER = 10, + VT_FUSED_ACTIVATION_FUNCTION = 12, + VT_DILATION_W_FACTOR = 14, + VT_DILATION_H_FACTOR = 16 + }; + tflite::Padding padding() const { + return static_cast(GetField(VT_PADDING, 0)); + } + int32_t stride_w() const { + return GetField(VT_STRIDE_W, 0); + } + int32_t stride_h() const { + return GetField(VT_STRIDE_H, 0); + } + int32_t depth_multiplier() const { + return GetField(VT_DEPTH_MULTIPLIER, 0); + } + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + int32_t dilation_w_factor() const { + return GetField(VT_DILATION_W_FACTOR, 1); + } + int32_t dilation_h_factor() const { + return GetField(VT_DILATION_H_FACTOR, 1); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_PADDING) && + VerifyField(verifier, VT_STRIDE_W) && + VerifyField(verifier, VT_STRIDE_H) && + VerifyField(verifier, VT_DEPTH_MULTIPLIER) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_DILATION_W_FACTOR) && + VerifyField(verifier, VT_DILATION_H_FACTOR) && + verifier.EndTable(); + } + DepthwiseConv2DOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(DepthwiseConv2DOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DepthwiseConv2DOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct DepthwiseConv2DOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_padding(tflite::Padding padding) { + fbb_.AddElement(DepthwiseConv2DOptions::VT_PADDING, static_cast(padding), 0); + } + void add_stride_w(int32_t stride_w) { + fbb_.AddElement(DepthwiseConv2DOptions::VT_STRIDE_W, stride_w, 0); + } + void add_stride_h(int32_t stride_h) { + fbb_.AddElement(DepthwiseConv2DOptions::VT_STRIDE_H, stride_h, 0); + } + void add_depth_multiplier(int32_t depth_multiplier) { + fbb_.AddElement(DepthwiseConv2DOptions::VT_DEPTH_MULTIPLIER, depth_multiplier, 0); + } + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(DepthwiseConv2DOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_dilation_w_factor(int32_t dilation_w_factor) { + fbb_.AddElement(DepthwiseConv2DOptions::VT_DILATION_W_FACTOR, dilation_w_factor, 1); + } + void add_dilation_h_factor(int32_t dilation_h_factor) { + fbb_.AddElement(DepthwiseConv2DOptions::VT_DILATION_H_FACTOR, dilation_h_factor, 1); + } + explicit DepthwiseConv2DOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + DepthwiseConv2DOptionsBuilder &operator=(const DepthwiseConv2DOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateDepthwiseConv2DOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::Padding padding = tflite::Padding_SAME, + int32_t stride_w = 0, + int32_t stride_h = 0, + int32_t depth_multiplier = 0, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + int32_t dilation_w_factor = 1, + int32_t dilation_h_factor = 1) { + DepthwiseConv2DOptionsBuilder builder_(_fbb); + builder_.add_dilation_h_factor(dilation_h_factor); + builder_.add_dilation_w_factor(dilation_w_factor); + builder_.add_depth_multiplier(depth_multiplier); + builder_.add_stride_h(stride_h); + builder_.add_stride_w(stride_w); + builder_.add_fused_activation_function(fused_activation_function); + builder_.add_padding(padding); + return builder_.Finish(); +} + +flatbuffers::Offset CreateDepthwiseConv2DOptions(flatbuffers::FlatBufferBuilder &_fbb, const DepthwiseConv2DOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ConcatEmbeddingsOptionsT : public flatbuffers::NativeTable { + typedef ConcatEmbeddingsOptions TableType; + int32_t num_channels; + std::vector num_columns_per_channel; + std::vector embedding_dim_per_channel; + ConcatEmbeddingsOptionsT() + : num_channels(0) { + } +}; + +struct ConcatEmbeddingsOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ConcatEmbeddingsOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NUM_CHANNELS = 4, + VT_NUM_COLUMNS_PER_CHANNEL = 6, + VT_EMBEDDING_DIM_PER_CHANNEL = 8 + }; + int32_t num_channels() const { + return GetField(VT_NUM_CHANNELS, 0); + } + const flatbuffers::Vector *num_columns_per_channel() const { + return GetPointer *>(VT_NUM_COLUMNS_PER_CHANNEL); + } + const flatbuffers::Vector *embedding_dim_per_channel() const { + return GetPointer *>(VT_EMBEDDING_DIM_PER_CHANNEL); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_NUM_CHANNELS) && + VerifyOffset(verifier, VT_NUM_COLUMNS_PER_CHANNEL) && + verifier.VerifyVector(num_columns_per_channel()) && + VerifyOffset(verifier, VT_EMBEDDING_DIM_PER_CHANNEL) && + verifier.VerifyVector(embedding_dim_per_channel()) && + verifier.EndTable(); + } + ConcatEmbeddingsOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ConcatEmbeddingsOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ConcatEmbeddingsOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ConcatEmbeddingsOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_num_channels(int32_t num_channels) { + fbb_.AddElement(ConcatEmbeddingsOptions::VT_NUM_CHANNELS, num_channels, 0); + } + void add_num_columns_per_channel(flatbuffers::Offset> num_columns_per_channel) { + fbb_.AddOffset(ConcatEmbeddingsOptions::VT_NUM_COLUMNS_PER_CHANNEL, num_columns_per_channel); + } + void add_embedding_dim_per_channel(flatbuffers::Offset> embedding_dim_per_channel) { + fbb_.AddOffset(ConcatEmbeddingsOptions::VT_EMBEDDING_DIM_PER_CHANNEL, embedding_dim_per_channel); + } + explicit ConcatEmbeddingsOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ConcatEmbeddingsOptionsBuilder &operator=(const ConcatEmbeddingsOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateConcatEmbeddingsOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t num_channels = 0, + flatbuffers::Offset> num_columns_per_channel = 0, + flatbuffers::Offset> embedding_dim_per_channel = 0) { + ConcatEmbeddingsOptionsBuilder builder_(_fbb); + builder_.add_embedding_dim_per_channel(embedding_dim_per_channel); + builder_.add_num_columns_per_channel(num_columns_per_channel); + builder_.add_num_channels(num_channels); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateConcatEmbeddingsOptionsDirect( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t num_channels = 0, + const std::vector *num_columns_per_channel = nullptr, + const std::vector *embedding_dim_per_channel = nullptr) { + auto num_columns_per_channel__ = num_columns_per_channel ? _fbb.CreateVector(*num_columns_per_channel) : 0; + auto embedding_dim_per_channel__ = embedding_dim_per_channel ? _fbb.CreateVector(*embedding_dim_per_channel) : 0; + return tflite::CreateConcatEmbeddingsOptions( + _fbb, + num_channels, + num_columns_per_channel__, + embedding_dim_per_channel__); +} + +flatbuffers::Offset CreateConcatEmbeddingsOptions(flatbuffers::FlatBufferBuilder &_fbb, const ConcatEmbeddingsOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LSHProjectionOptionsT : public flatbuffers::NativeTable { + typedef LSHProjectionOptions TableType; + tflite::LSHProjectionType type; + LSHProjectionOptionsT() + : type(tflite::LSHProjectionType_UNKNOWN) { + } +}; + +struct LSHProjectionOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LSHProjectionOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_TYPE = 4 + }; + tflite::LSHProjectionType type() const { + return static_cast(GetField(VT_TYPE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_TYPE) && + verifier.EndTable(); + } + LSHProjectionOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LSHProjectionOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LSHProjectionOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LSHProjectionOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_type(tflite::LSHProjectionType type) { + fbb_.AddElement(LSHProjectionOptions::VT_TYPE, static_cast(type), 0); + } + explicit LSHProjectionOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LSHProjectionOptionsBuilder &operator=(const LSHProjectionOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLSHProjectionOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::LSHProjectionType type = tflite::LSHProjectionType_UNKNOWN) { + LSHProjectionOptionsBuilder builder_(_fbb); + builder_.add_type(type); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLSHProjectionOptions(flatbuffers::FlatBufferBuilder &_fbb, const LSHProjectionOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SVDFOptionsT : public flatbuffers::NativeTable { + typedef SVDFOptions TableType; + int32_t rank; + tflite::ActivationFunctionType fused_activation_function; + bool asymmetric_quantize_inputs; + SVDFOptionsT() + : rank(0), + fused_activation_function(tflite::ActivationFunctionType_NONE), + asymmetric_quantize_inputs(false) { + } +}; + +struct SVDFOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SVDFOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_RANK = 4, + VT_FUSED_ACTIVATION_FUNCTION = 6, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 8 + }; + int32_t rank() const { + return GetField(VT_RANK, 0); + } + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_RANK) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + SVDFOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SVDFOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SVDFOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SVDFOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_rank(int32_t rank) { + fbb_.AddElement(SVDFOptions::VT_RANK, rank, 0); + } + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(SVDFOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(SVDFOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit SVDFOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SVDFOptionsBuilder &operator=(const SVDFOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSVDFOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t rank = 0, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + bool asymmetric_quantize_inputs = false) { + SVDFOptionsBuilder builder_(_fbb); + builder_.add_rank(rank); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSVDFOptions(flatbuffers::FlatBufferBuilder &_fbb, const SVDFOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct RNNOptionsT : public flatbuffers::NativeTable { + typedef RNNOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + bool asymmetric_quantize_inputs; + RNNOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE), + asymmetric_quantize_inputs(false) { + } +}; + +struct RNNOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef RNNOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 6 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + RNNOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(RNNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const RNNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct RNNOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(RNNOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(RNNOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit RNNOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + RNNOptionsBuilder &operator=(const RNNOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateRNNOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + bool asymmetric_quantize_inputs = false) { + RNNOptionsBuilder builder_(_fbb); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateRNNOptions(flatbuffers::FlatBufferBuilder &_fbb, const RNNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SequenceRNNOptionsT : public flatbuffers::NativeTable { + typedef SequenceRNNOptions TableType; + bool time_major; + tflite::ActivationFunctionType fused_activation_function; + bool asymmetric_quantize_inputs; + SequenceRNNOptionsT() + : time_major(false), + fused_activation_function(tflite::ActivationFunctionType_NONE), + asymmetric_quantize_inputs(false) { + } +}; + +struct SequenceRNNOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SequenceRNNOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_TIME_MAJOR = 4, + VT_FUSED_ACTIVATION_FUNCTION = 6, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 8 + }; + bool time_major() const { + return GetField(VT_TIME_MAJOR, 0) != 0; + } + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_TIME_MAJOR) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + SequenceRNNOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SequenceRNNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SequenceRNNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SequenceRNNOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_time_major(bool time_major) { + fbb_.AddElement(SequenceRNNOptions::VT_TIME_MAJOR, static_cast(time_major), 0); + } + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(SequenceRNNOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(SequenceRNNOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit SequenceRNNOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SequenceRNNOptionsBuilder &operator=(const SequenceRNNOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSequenceRNNOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool time_major = false, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + bool asymmetric_quantize_inputs = false) { + SequenceRNNOptionsBuilder builder_(_fbb); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_fused_activation_function(fused_activation_function); + builder_.add_time_major(time_major); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSequenceRNNOptions(flatbuffers::FlatBufferBuilder &_fbb, const SequenceRNNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct BidirectionalSequenceRNNOptionsT : public flatbuffers::NativeTable { + typedef BidirectionalSequenceRNNOptions TableType; + bool time_major; + tflite::ActivationFunctionType fused_activation_function; + bool merge_outputs; + bool asymmetric_quantize_inputs; + BidirectionalSequenceRNNOptionsT() + : time_major(false), + fused_activation_function(tflite::ActivationFunctionType_NONE), + merge_outputs(false), + asymmetric_quantize_inputs(false) { + } +}; + +struct BidirectionalSequenceRNNOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef BidirectionalSequenceRNNOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_TIME_MAJOR = 4, + VT_FUSED_ACTIVATION_FUNCTION = 6, + VT_MERGE_OUTPUTS = 8, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 10 + }; + bool time_major() const { + return GetField(VT_TIME_MAJOR, 0) != 0; + } + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool merge_outputs() const { + return GetField(VT_MERGE_OUTPUTS, 0) != 0; + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_TIME_MAJOR) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_MERGE_OUTPUTS) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + BidirectionalSequenceRNNOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(BidirectionalSequenceRNNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceRNNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct BidirectionalSequenceRNNOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_time_major(bool time_major) { + fbb_.AddElement(BidirectionalSequenceRNNOptions::VT_TIME_MAJOR, static_cast(time_major), 0); + } + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(BidirectionalSequenceRNNOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_merge_outputs(bool merge_outputs) { + fbb_.AddElement(BidirectionalSequenceRNNOptions::VT_MERGE_OUTPUTS, static_cast(merge_outputs), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(BidirectionalSequenceRNNOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit BidirectionalSequenceRNNOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + BidirectionalSequenceRNNOptionsBuilder &operator=(const BidirectionalSequenceRNNOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateBidirectionalSequenceRNNOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool time_major = false, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + bool merge_outputs = false, + bool asymmetric_quantize_inputs = false) { + BidirectionalSequenceRNNOptionsBuilder builder_(_fbb); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_merge_outputs(merge_outputs); + builder_.add_fused_activation_function(fused_activation_function); + builder_.add_time_major(time_major); + return builder_.Finish(); +} + +flatbuffers::Offset CreateBidirectionalSequenceRNNOptions(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceRNNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct FullyConnectedOptionsT : public flatbuffers::NativeTable { + typedef FullyConnectedOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + tflite::FullyConnectedOptionsWeightsFormat weights_format; + bool keep_num_dims; + bool asymmetric_quantize_inputs; + FullyConnectedOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE), + weights_format(tflite::FullyConnectedOptionsWeightsFormat_DEFAULT), + keep_num_dims(false), + asymmetric_quantize_inputs(false) { + } +}; + +struct FullyConnectedOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef FullyConnectedOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_WEIGHTS_FORMAT = 6, + VT_KEEP_NUM_DIMS = 8, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 10 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + tflite::FullyConnectedOptionsWeightsFormat weights_format() const { + return static_cast(GetField(VT_WEIGHTS_FORMAT, 0)); + } + bool keep_num_dims() const { + return GetField(VT_KEEP_NUM_DIMS, 0) != 0; + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_WEIGHTS_FORMAT) && + VerifyField(verifier, VT_KEEP_NUM_DIMS) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + FullyConnectedOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(FullyConnectedOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const FullyConnectedOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct FullyConnectedOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(FullyConnectedOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_weights_format(tflite::FullyConnectedOptionsWeightsFormat weights_format) { + fbb_.AddElement(FullyConnectedOptions::VT_WEIGHTS_FORMAT, static_cast(weights_format), 0); + } + void add_keep_num_dims(bool keep_num_dims) { + fbb_.AddElement(FullyConnectedOptions::VT_KEEP_NUM_DIMS, static_cast(keep_num_dims), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(FullyConnectedOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit FullyConnectedOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + FullyConnectedOptionsBuilder &operator=(const FullyConnectedOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateFullyConnectedOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + tflite::FullyConnectedOptionsWeightsFormat weights_format = tflite::FullyConnectedOptionsWeightsFormat_DEFAULT, + bool keep_num_dims = false, + bool asymmetric_quantize_inputs = false) { + FullyConnectedOptionsBuilder builder_(_fbb); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_keep_num_dims(keep_num_dims); + builder_.add_weights_format(weights_format); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateFullyConnectedOptions(flatbuffers::FlatBufferBuilder &_fbb, const FullyConnectedOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SoftmaxOptionsT : public flatbuffers::NativeTable { + typedef SoftmaxOptions TableType; + float beta; + SoftmaxOptionsT() + : beta(0.0f) { + } +}; + +struct SoftmaxOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SoftmaxOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_BETA = 4 + }; + float beta() const { + return GetField(VT_BETA, 0.0f); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_BETA) && + verifier.EndTable(); + } + SoftmaxOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SoftmaxOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SoftmaxOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SoftmaxOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_beta(float beta) { + fbb_.AddElement(SoftmaxOptions::VT_BETA, beta, 0.0f); + } + explicit SoftmaxOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SoftmaxOptionsBuilder &operator=(const SoftmaxOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSoftmaxOptions( + flatbuffers::FlatBufferBuilder &_fbb, + float beta = 0.0f) { + SoftmaxOptionsBuilder builder_(_fbb); + builder_.add_beta(beta); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSoftmaxOptions(flatbuffers::FlatBufferBuilder &_fbb, const SoftmaxOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ConcatenationOptionsT : public flatbuffers::NativeTable { + typedef ConcatenationOptions TableType; + int32_t axis; + tflite::ActivationFunctionType fused_activation_function; + ConcatenationOptionsT() + : axis(0), + fused_activation_function(tflite::ActivationFunctionType_NONE) { + } +}; + +struct ConcatenationOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ConcatenationOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_AXIS = 4, + VT_FUSED_ACTIVATION_FUNCTION = 6 + }; + int32_t axis() const { + return GetField(VT_AXIS, 0); + } + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_AXIS) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + verifier.EndTable(); + } + ConcatenationOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ConcatenationOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ConcatenationOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ConcatenationOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_axis(int32_t axis) { + fbb_.AddElement(ConcatenationOptions::VT_AXIS, axis, 0); + } + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(ConcatenationOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + explicit ConcatenationOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ConcatenationOptionsBuilder &operator=(const ConcatenationOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateConcatenationOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t axis = 0, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE) { + ConcatenationOptionsBuilder builder_(_fbb); + builder_.add_axis(axis); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateConcatenationOptions(flatbuffers::FlatBufferBuilder &_fbb, const ConcatenationOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct AddOptionsT : public flatbuffers::NativeTable { + typedef AddOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + bool pot_scale_int16; + AddOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE), + pot_scale_int16(true) { + } +}; + +struct AddOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef AddOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_POT_SCALE_INT16 = 6 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool pot_scale_int16() const { + return GetField(VT_POT_SCALE_INT16, 1) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_POT_SCALE_INT16) && + verifier.EndTable(); + } + AddOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(AddOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const AddOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct AddOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(AddOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_pot_scale_int16(bool pot_scale_int16) { + fbb_.AddElement(AddOptions::VT_POT_SCALE_INT16, static_cast(pot_scale_int16), 1); + } + explicit AddOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + AddOptionsBuilder &operator=(const AddOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateAddOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + bool pot_scale_int16 = true) { + AddOptionsBuilder builder_(_fbb); + builder_.add_pot_scale_int16(pot_scale_int16); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateAddOptions(flatbuffers::FlatBufferBuilder &_fbb, const AddOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct MulOptionsT : public flatbuffers::NativeTable { + typedef MulOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + MulOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE) { + } +}; + +struct MulOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef MulOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + verifier.EndTable(); + } + MulOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(MulOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const MulOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct MulOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(MulOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + explicit MulOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + MulOptionsBuilder &operator=(const MulOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateMulOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE) { + MulOptionsBuilder builder_(_fbb); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateMulOptions(flatbuffers::FlatBufferBuilder &_fbb, const MulOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct L2NormOptionsT : public flatbuffers::NativeTable { + typedef L2NormOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + L2NormOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE) { + } +}; + +struct L2NormOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef L2NormOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + verifier.EndTable(); + } + L2NormOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(L2NormOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const L2NormOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct L2NormOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(L2NormOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + explicit L2NormOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + L2NormOptionsBuilder &operator=(const L2NormOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateL2NormOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE) { + L2NormOptionsBuilder builder_(_fbb); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateL2NormOptions(flatbuffers::FlatBufferBuilder &_fbb, const L2NormOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LocalResponseNormalizationOptionsT : public flatbuffers::NativeTable { + typedef LocalResponseNormalizationOptions TableType; + int32_t radius; + float bias; + float alpha; + float beta; + LocalResponseNormalizationOptionsT() + : radius(0), + bias(0.0f), + alpha(0.0f), + beta(0.0f) { + } +}; + +struct LocalResponseNormalizationOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LocalResponseNormalizationOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_RADIUS = 4, + VT_BIAS = 6, + VT_ALPHA = 8, + VT_BETA = 10 + }; + int32_t radius() const { + return GetField(VT_RADIUS, 0); + } + float bias() const { + return GetField(VT_BIAS, 0.0f); + } + float alpha() const { + return GetField(VT_ALPHA, 0.0f); + } + float beta() const { + return GetField(VT_BETA, 0.0f); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_RADIUS) && + VerifyField(verifier, VT_BIAS) && + VerifyField(verifier, VT_ALPHA) && + VerifyField(verifier, VT_BETA) && + verifier.EndTable(); + } + LocalResponseNormalizationOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LocalResponseNormalizationOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LocalResponseNormalizationOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LocalResponseNormalizationOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_radius(int32_t radius) { + fbb_.AddElement(LocalResponseNormalizationOptions::VT_RADIUS, radius, 0); + } + void add_bias(float bias) { + fbb_.AddElement(LocalResponseNormalizationOptions::VT_BIAS, bias, 0.0f); + } + void add_alpha(float alpha) { + fbb_.AddElement(LocalResponseNormalizationOptions::VT_ALPHA, alpha, 0.0f); + } + void add_beta(float beta) { + fbb_.AddElement(LocalResponseNormalizationOptions::VT_BETA, beta, 0.0f); + } + explicit LocalResponseNormalizationOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LocalResponseNormalizationOptionsBuilder &operator=(const LocalResponseNormalizationOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLocalResponseNormalizationOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t radius = 0, + float bias = 0.0f, + float alpha = 0.0f, + float beta = 0.0f) { + LocalResponseNormalizationOptionsBuilder builder_(_fbb); + builder_.add_beta(beta); + builder_.add_alpha(alpha); + builder_.add_bias(bias); + builder_.add_radius(radius); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLocalResponseNormalizationOptions(flatbuffers::FlatBufferBuilder &_fbb, const LocalResponseNormalizationOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LSTMOptionsT : public flatbuffers::NativeTable { + typedef LSTMOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + float cell_clip; + float proj_clip; + tflite::LSTMKernelType kernel_type; + bool asymmetric_quantize_inputs; + LSTMOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE), + cell_clip(0.0f), + proj_clip(0.0f), + kernel_type(tflite::LSTMKernelType_FULL), + asymmetric_quantize_inputs(false) { + } +}; + +struct LSTMOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LSTMOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_CELL_CLIP = 6, + VT_PROJ_CLIP = 8, + VT_KERNEL_TYPE = 10, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 12 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + float cell_clip() const { + return GetField(VT_CELL_CLIP, 0.0f); + } + float proj_clip() const { + return GetField(VT_PROJ_CLIP, 0.0f); + } + tflite::LSTMKernelType kernel_type() const { + return static_cast(GetField(VT_KERNEL_TYPE, 0)); + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_CELL_CLIP) && + VerifyField(verifier, VT_PROJ_CLIP) && + VerifyField(verifier, VT_KERNEL_TYPE) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + LSTMOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LSTMOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LSTMOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LSTMOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(LSTMOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_cell_clip(float cell_clip) { + fbb_.AddElement(LSTMOptions::VT_CELL_CLIP, cell_clip, 0.0f); + } + void add_proj_clip(float proj_clip) { + fbb_.AddElement(LSTMOptions::VT_PROJ_CLIP, proj_clip, 0.0f); + } + void add_kernel_type(tflite::LSTMKernelType kernel_type) { + fbb_.AddElement(LSTMOptions::VT_KERNEL_TYPE, static_cast(kernel_type), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(LSTMOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit LSTMOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LSTMOptionsBuilder &operator=(const LSTMOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLSTMOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + float cell_clip = 0.0f, + float proj_clip = 0.0f, + tflite::LSTMKernelType kernel_type = tflite::LSTMKernelType_FULL, + bool asymmetric_quantize_inputs = false) { + LSTMOptionsBuilder builder_(_fbb); + builder_.add_proj_clip(proj_clip); + builder_.add_cell_clip(cell_clip); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_kernel_type(kernel_type); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLSTMOptions(flatbuffers::FlatBufferBuilder &_fbb, const LSTMOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct UnidirectionalSequenceLSTMOptionsT : public flatbuffers::NativeTable { + typedef UnidirectionalSequenceLSTMOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + float cell_clip; + float proj_clip; + bool time_major; + bool asymmetric_quantize_inputs; + UnidirectionalSequenceLSTMOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE), + cell_clip(0.0f), + proj_clip(0.0f), + time_major(false), + asymmetric_quantize_inputs(false) { + } +}; + +struct UnidirectionalSequenceLSTMOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef UnidirectionalSequenceLSTMOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_CELL_CLIP = 6, + VT_PROJ_CLIP = 8, + VT_TIME_MAJOR = 10, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 12 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + float cell_clip() const { + return GetField(VT_CELL_CLIP, 0.0f); + } + float proj_clip() const { + return GetField(VT_PROJ_CLIP, 0.0f); + } + bool time_major() const { + return GetField(VT_TIME_MAJOR, 0) != 0; + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_CELL_CLIP) && + VerifyField(verifier, VT_PROJ_CLIP) && + VerifyField(verifier, VT_TIME_MAJOR) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + UnidirectionalSequenceLSTMOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(UnidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const UnidirectionalSequenceLSTMOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct UnidirectionalSequenceLSTMOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(UnidirectionalSequenceLSTMOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_cell_clip(float cell_clip) { + fbb_.AddElement(UnidirectionalSequenceLSTMOptions::VT_CELL_CLIP, cell_clip, 0.0f); + } + void add_proj_clip(float proj_clip) { + fbb_.AddElement(UnidirectionalSequenceLSTMOptions::VT_PROJ_CLIP, proj_clip, 0.0f); + } + void add_time_major(bool time_major) { + fbb_.AddElement(UnidirectionalSequenceLSTMOptions::VT_TIME_MAJOR, static_cast(time_major), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(UnidirectionalSequenceLSTMOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit UnidirectionalSequenceLSTMOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + UnidirectionalSequenceLSTMOptionsBuilder &operator=(const UnidirectionalSequenceLSTMOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateUnidirectionalSequenceLSTMOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + float cell_clip = 0.0f, + float proj_clip = 0.0f, + bool time_major = false, + bool asymmetric_quantize_inputs = false) { + UnidirectionalSequenceLSTMOptionsBuilder builder_(_fbb); + builder_.add_proj_clip(proj_clip); + builder_.add_cell_clip(cell_clip); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_time_major(time_major); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateUnidirectionalSequenceLSTMOptions(flatbuffers::FlatBufferBuilder &_fbb, const UnidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct BidirectionalSequenceLSTMOptionsT : public flatbuffers::NativeTable { + typedef BidirectionalSequenceLSTMOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + float cell_clip; + float proj_clip; + bool merge_outputs; + bool time_major; + bool asymmetric_quantize_inputs; + BidirectionalSequenceLSTMOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE), + cell_clip(0.0f), + proj_clip(0.0f), + merge_outputs(false), + time_major(true), + asymmetric_quantize_inputs(false) { + } +}; + +struct BidirectionalSequenceLSTMOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef BidirectionalSequenceLSTMOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_CELL_CLIP = 6, + VT_PROJ_CLIP = 8, + VT_MERGE_OUTPUTS = 10, + VT_TIME_MAJOR = 12, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 14 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + float cell_clip() const { + return GetField(VT_CELL_CLIP, 0.0f); + } + float proj_clip() const { + return GetField(VT_PROJ_CLIP, 0.0f); + } + bool merge_outputs() const { + return GetField(VT_MERGE_OUTPUTS, 0) != 0; + } + bool time_major() const { + return GetField(VT_TIME_MAJOR, 1) != 0; + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_CELL_CLIP) && + VerifyField(verifier, VT_PROJ_CLIP) && + VerifyField(verifier, VT_MERGE_OUTPUTS) && + VerifyField(verifier, VT_TIME_MAJOR) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + BidirectionalSequenceLSTMOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(BidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceLSTMOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct BidirectionalSequenceLSTMOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(BidirectionalSequenceLSTMOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_cell_clip(float cell_clip) { + fbb_.AddElement(BidirectionalSequenceLSTMOptions::VT_CELL_CLIP, cell_clip, 0.0f); + } + void add_proj_clip(float proj_clip) { + fbb_.AddElement(BidirectionalSequenceLSTMOptions::VT_PROJ_CLIP, proj_clip, 0.0f); + } + void add_merge_outputs(bool merge_outputs) { + fbb_.AddElement(BidirectionalSequenceLSTMOptions::VT_MERGE_OUTPUTS, static_cast(merge_outputs), 0); + } + void add_time_major(bool time_major) { + fbb_.AddElement(BidirectionalSequenceLSTMOptions::VT_TIME_MAJOR, static_cast(time_major), 1); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(BidirectionalSequenceLSTMOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit BidirectionalSequenceLSTMOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + BidirectionalSequenceLSTMOptionsBuilder &operator=(const BidirectionalSequenceLSTMOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateBidirectionalSequenceLSTMOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + float cell_clip = 0.0f, + float proj_clip = 0.0f, + bool merge_outputs = false, + bool time_major = true, + bool asymmetric_quantize_inputs = false) { + BidirectionalSequenceLSTMOptionsBuilder builder_(_fbb); + builder_.add_proj_clip(proj_clip); + builder_.add_cell_clip(cell_clip); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_time_major(time_major); + builder_.add_merge_outputs(merge_outputs); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateBidirectionalSequenceLSTMOptions(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ResizeBilinearOptionsT : public flatbuffers::NativeTable { + typedef ResizeBilinearOptions TableType; + bool align_corners; + bool half_pixel_centers; + ResizeBilinearOptionsT() + : align_corners(false), + half_pixel_centers(false) { + } +}; + +struct ResizeBilinearOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ResizeBilinearOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_ALIGN_CORNERS = 8, + VT_HALF_PIXEL_CENTERS = 10 + }; + bool align_corners() const { + return GetField(VT_ALIGN_CORNERS, 0) != 0; + } + bool half_pixel_centers() const { + return GetField(VT_HALF_PIXEL_CENTERS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_ALIGN_CORNERS) && + VerifyField(verifier, VT_HALF_PIXEL_CENTERS) && + verifier.EndTable(); + } + ResizeBilinearOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ResizeBilinearOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ResizeBilinearOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ResizeBilinearOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_align_corners(bool align_corners) { + fbb_.AddElement(ResizeBilinearOptions::VT_ALIGN_CORNERS, static_cast(align_corners), 0); + } + void add_half_pixel_centers(bool half_pixel_centers) { + fbb_.AddElement(ResizeBilinearOptions::VT_HALF_PIXEL_CENTERS, static_cast(half_pixel_centers), 0); + } + explicit ResizeBilinearOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ResizeBilinearOptionsBuilder &operator=(const ResizeBilinearOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateResizeBilinearOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool align_corners = false, + bool half_pixel_centers = false) { + ResizeBilinearOptionsBuilder builder_(_fbb); + builder_.add_half_pixel_centers(half_pixel_centers); + builder_.add_align_corners(align_corners); + return builder_.Finish(); +} + +flatbuffers::Offset CreateResizeBilinearOptions(flatbuffers::FlatBufferBuilder &_fbb, const ResizeBilinearOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ResizeNearestNeighborOptionsT : public flatbuffers::NativeTable { + typedef ResizeNearestNeighborOptions TableType; + bool align_corners; + bool half_pixel_centers; + ResizeNearestNeighborOptionsT() + : align_corners(false), + half_pixel_centers(false) { + } +}; + +struct ResizeNearestNeighborOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ResizeNearestNeighborOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_ALIGN_CORNERS = 4, + VT_HALF_PIXEL_CENTERS = 6 + }; + bool align_corners() const { + return GetField(VT_ALIGN_CORNERS, 0) != 0; + } + bool half_pixel_centers() const { + return GetField(VT_HALF_PIXEL_CENTERS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_ALIGN_CORNERS) && + VerifyField(verifier, VT_HALF_PIXEL_CENTERS) && + verifier.EndTable(); + } + ResizeNearestNeighborOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ResizeNearestNeighborOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ResizeNearestNeighborOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ResizeNearestNeighborOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_align_corners(bool align_corners) { + fbb_.AddElement(ResizeNearestNeighborOptions::VT_ALIGN_CORNERS, static_cast(align_corners), 0); + } + void add_half_pixel_centers(bool half_pixel_centers) { + fbb_.AddElement(ResizeNearestNeighborOptions::VT_HALF_PIXEL_CENTERS, static_cast(half_pixel_centers), 0); + } + explicit ResizeNearestNeighborOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ResizeNearestNeighborOptionsBuilder &operator=(const ResizeNearestNeighborOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateResizeNearestNeighborOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool align_corners = false, + bool half_pixel_centers = false) { + ResizeNearestNeighborOptionsBuilder builder_(_fbb); + builder_.add_half_pixel_centers(half_pixel_centers); + builder_.add_align_corners(align_corners); + return builder_.Finish(); +} + +flatbuffers::Offset CreateResizeNearestNeighborOptions(flatbuffers::FlatBufferBuilder &_fbb, const ResizeNearestNeighborOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct CallOptionsT : public flatbuffers::NativeTable { + typedef CallOptions TableType; + uint32_t subgraph; + CallOptionsT() + : subgraph(0) { + } +}; + +struct CallOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef CallOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_SUBGRAPH = 4 + }; + uint32_t subgraph() const { + return GetField(VT_SUBGRAPH, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_SUBGRAPH) && + verifier.EndTable(); + } + CallOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(CallOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const CallOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct CallOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_subgraph(uint32_t subgraph) { + fbb_.AddElement(CallOptions::VT_SUBGRAPH, subgraph, 0); + } + explicit CallOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + CallOptionsBuilder &operator=(const CallOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateCallOptions( + flatbuffers::FlatBufferBuilder &_fbb, + uint32_t subgraph = 0) { + CallOptionsBuilder builder_(_fbb); + builder_.add_subgraph(subgraph); + return builder_.Finish(); +} + +flatbuffers::Offset CreateCallOptions(flatbuffers::FlatBufferBuilder &_fbb, const CallOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct PadOptionsT : public flatbuffers::NativeTable { + typedef PadOptions TableType; + PadOptionsT() { + } +}; + +struct PadOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef PadOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + PadOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(PadOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const PadOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct PadOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit PadOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + PadOptionsBuilder &operator=(const PadOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreatePadOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + PadOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreatePadOptions(flatbuffers::FlatBufferBuilder &_fbb, const PadOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct PadV2OptionsT : public flatbuffers::NativeTable { + typedef PadV2Options TableType; + PadV2OptionsT() { + } +}; + +struct PadV2Options FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef PadV2OptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + PadV2OptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(PadV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const PadV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct PadV2OptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit PadV2OptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + PadV2OptionsBuilder &operator=(const PadV2OptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreatePadV2Options( + flatbuffers::FlatBufferBuilder &_fbb) { + PadV2OptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreatePadV2Options(flatbuffers::FlatBufferBuilder &_fbb, const PadV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ReshapeOptionsT : public flatbuffers::NativeTable { + typedef ReshapeOptions TableType; + std::vector new_shape; + ReshapeOptionsT() { + } +}; + +struct ReshapeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ReshapeOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NEW_SHAPE = 4 + }; + const flatbuffers::Vector *new_shape() const { + return GetPointer *>(VT_NEW_SHAPE); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_NEW_SHAPE) && + verifier.VerifyVector(new_shape()) && + verifier.EndTable(); + } + ReshapeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ReshapeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReshapeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ReshapeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_new_shape(flatbuffers::Offset> new_shape) { + fbb_.AddOffset(ReshapeOptions::VT_NEW_SHAPE, new_shape); + } + explicit ReshapeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ReshapeOptionsBuilder &operator=(const ReshapeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateReshapeOptions( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> new_shape = 0) { + ReshapeOptionsBuilder builder_(_fbb); + builder_.add_new_shape(new_shape); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateReshapeOptionsDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *new_shape = nullptr) { + auto new_shape__ = new_shape ? _fbb.CreateVector(*new_shape) : 0; + return tflite::CreateReshapeOptions( + _fbb, + new_shape__); +} + +flatbuffers::Offset CreateReshapeOptions(flatbuffers::FlatBufferBuilder &_fbb, const ReshapeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SpaceToBatchNDOptionsT : public flatbuffers::NativeTable { + typedef SpaceToBatchNDOptions TableType; + SpaceToBatchNDOptionsT() { + } +}; + +struct SpaceToBatchNDOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SpaceToBatchNDOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + SpaceToBatchNDOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SpaceToBatchNDOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToBatchNDOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SpaceToBatchNDOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit SpaceToBatchNDOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SpaceToBatchNDOptionsBuilder &operator=(const SpaceToBatchNDOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSpaceToBatchNDOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + SpaceToBatchNDOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSpaceToBatchNDOptions(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToBatchNDOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct BatchToSpaceNDOptionsT : public flatbuffers::NativeTable { + typedef BatchToSpaceNDOptions TableType; + BatchToSpaceNDOptionsT() { + } +}; + +struct BatchToSpaceNDOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef BatchToSpaceNDOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + BatchToSpaceNDOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(BatchToSpaceNDOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const BatchToSpaceNDOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct BatchToSpaceNDOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit BatchToSpaceNDOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + BatchToSpaceNDOptionsBuilder &operator=(const BatchToSpaceNDOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateBatchToSpaceNDOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + BatchToSpaceNDOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateBatchToSpaceNDOptions(flatbuffers::FlatBufferBuilder &_fbb, const BatchToSpaceNDOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SkipGramOptionsT : public flatbuffers::NativeTable { + typedef SkipGramOptions TableType; + int32_t ngram_size; + int32_t max_skip_size; + bool include_all_ngrams; + SkipGramOptionsT() + : ngram_size(0), + max_skip_size(0), + include_all_ngrams(false) { + } +}; + +struct SkipGramOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SkipGramOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NGRAM_SIZE = 4, + VT_MAX_SKIP_SIZE = 6, + VT_INCLUDE_ALL_NGRAMS = 8 + }; + int32_t ngram_size() const { + return GetField(VT_NGRAM_SIZE, 0); + } + int32_t max_skip_size() const { + return GetField(VT_MAX_SKIP_SIZE, 0); + } + bool include_all_ngrams() const { + return GetField(VT_INCLUDE_ALL_NGRAMS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_NGRAM_SIZE) && + VerifyField(verifier, VT_MAX_SKIP_SIZE) && + VerifyField(verifier, VT_INCLUDE_ALL_NGRAMS) && + verifier.EndTable(); + } + SkipGramOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SkipGramOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SkipGramOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SkipGramOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_ngram_size(int32_t ngram_size) { + fbb_.AddElement(SkipGramOptions::VT_NGRAM_SIZE, ngram_size, 0); + } + void add_max_skip_size(int32_t max_skip_size) { + fbb_.AddElement(SkipGramOptions::VT_MAX_SKIP_SIZE, max_skip_size, 0); + } + void add_include_all_ngrams(bool include_all_ngrams) { + fbb_.AddElement(SkipGramOptions::VT_INCLUDE_ALL_NGRAMS, static_cast(include_all_ngrams), 0); + } + explicit SkipGramOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SkipGramOptionsBuilder &operator=(const SkipGramOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSkipGramOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t ngram_size = 0, + int32_t max_skip_size = 0, + bool include_all_ngrams = false) { + SkipGramOptionsBuilder builder_(_fbb); + builder_.add_max_skip_size(max_skip_size); + builder_.add_ngram_size(ngram_size); + builder_.add_include_all_ngrams(include_all_ngrams); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSkipGramOptions(flatbuffers::FlatBufferBuilder &_fbb, const SkipGramOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SpaceToDepthOptionsT : public flatbuffers::NativeTable { + typedef SpaceToDepthOptions TableType; + int32_t block_size; + SpaceToDepthOptionsT() + : block_size(0) { + } +}; + +struct SpaceToDepthOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SpaceToDepthOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_BLOCK_SIZE = 4 + }; + int32_t block_size() const { + return GetField(VT_BLOCK_SIZE, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_BLOCK_SIZE) && + verifier.EndTable(); + } + SpaceToDepthOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SpaceToDepthOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToDepthOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SpaceToDepthOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_block_size(int32_t block_size) { + fbb_.AddElement(SpaceToDepthOptions::VT_BLOCK_SIZE, block_size, 0); + } + explicit SpaceToDepthOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SpaceToDepthOptionsBuilder &operator=(const SpaceToDepthOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSpaceToDepthOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t block_size = 0) { + SpaceToDepthOptionsBuilder builder_(_fbb); + builder_.add_block_size(block_size); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSpaceToDepthOptions(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToDepthOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct DepthToSpaceOptionsT : public flatbuffers::NativeTable { + typedef DepthToSpaceOptions TableType; + int32_t block_size; + DepthToSpaceOptionsT() + : block_size(0) { + } +}; + +struct DepthToSpaceOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef DepthToSpaceOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_BLOCK_SIZE = 4 + }; + int32_t block_size() const { + return GetField(VT_BLOCK_SIZE, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_BLOCK_SIZE) && + verifier.EndTable(); + } + DepthToSpaceOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(DepthToSpaceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DepthToSpaceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct DepthToSpaceOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_block_size(int32_t block_size) { + fbb_.AddElement(DepthToSpaceOptions::VT_BLOCK_SIZE, block_size, 0); + } + explicit DepthToSpaceOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + DepthToSpaceOptionsBuilder &operator=(const DepthToSpaceOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateDepthToSpaceOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t block_size = 0) { + DepthToSpaceOptionsBuilder builder_(_fbb); + builder_.add_block_size(block_size); + return builder_.Finish(); +} + +flatbuffers::Offset CreateDepthToSpaceOptions(flatbuffers::FlatBufferBuilder &_fbb, const DepthToSpaceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SubOptionsT : public flatbuffers::NativeTable { + typedef SubOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + bool pot_scale_int16; + SubOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE), + pot_scale_int16(true) { + } +}; + +struct SubOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SubOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_POT_SCALE_INT16 = 6 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool pot_scale_int16() const { + return GetField(VT_POT_SCALE_INT16, 1) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_POT_SCALE_INT16) && + verifier.EndTable(); + } + SubOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SubOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SubOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SubOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(SubOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + void add_pot_scale_int16(bool pot_scale_int16) { + fbb_.AddElement(SubOptions::VT_POT_SCALE_INT16, static_cast(pot_scale_int16), 1); + } + explicit SubOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SubOptionsBuilder &operator=(const SubOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSubOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE, + bool pot_scale_int16 = true) { + SubOptionsBuilder builder_(_fbb); + builder_.add_pot_scale_int16(pot_scale_int16); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSubOptions(flatbuffers::FlatBufferBuilder &_fbb, const SubOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct DivOptionsT : public flatbuffers::NativeTable { + typedef DivOptions TableType; + tflite::ActivationFunctionType fused_activation_function; + DivOptionsT() + : fused_activation_function(tflite::ActivationFunctionType_NONE) { + } +}; + +struct DivOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef DivOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_FUSED_ACTIVATION_FUNCTION = 4 + }; + tflite::ActivationFunctionType fused_activation_function() const { + return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + verifier.EndTable(); + } + DivOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(DivOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DivOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct DivOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_fused_activation_function(tflite::ActivationFunctionType fused_activation_function) { + fbb_.AddElement(DivOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); + } + explicit DivOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + DivOptionsBuilder &operator=(const DivOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateDivOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::ActivationFunctionType fused_activation_function = tflite::ActivationFunctionType_NONE) { + DivOptionsBuilder builder_(_fbb); + builder_.add_fused_activation_function(fused_activation_function); + return builder_.Finish(); +} + +flatbuffers::Offset CreateDivOptions(flatbuffers::FlatBufferBuilder &_fbb, const DivOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct TopKV2OptionsT : public flatbuffers::NativeTable { + typedef TopKV2Options TableType; + TopKV2OptionsT() { + } +}; + +struct TopKV2Options FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef TopKV2OptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + TopKV2OptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(TopKV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const TopKV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct TopKV2OptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit TopKV2OptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + TopKV2OptionsBuilder &operator=(const TopKV2OptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateTopKV2Options( + flatbuffers::FlatBufferBuilder &_fbb) { + TopKV2OptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateTopKV2Options(flatbuffers::FlatBufferBuilder &_fbb, const TopKV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct EmbeddingLookupSparseOptionsT : public flatbuffers::NativeTable { + typedef EmbeddingLookupSparseOptions TableType; + tflite::CombinerType combiner; + EmbeddingLookupSparseOptionsT() + : combiner(tflite::CombinerType_SUM) { + } +}; + +struct EmbeddingLookupSparseOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef EmbeddingLookupSparseOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_COMBINER = 4 + }; + tflite::CombinerType combiner() const { + return static_cast(GetField(VT_COMBINER, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_COMBINER) && + verifier.EndTable(); + } + EmbeddingLookupSparseOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(EmbeddingLookupSparseOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const EmbeddingLookupSparseOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct EmbeddingLookupSparseOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_combiner(tflite::CombinerType combiner) { + fbb_.AddElement(EmbeddingLookupSparseOptions::VT_COMBINER, static_cast(combiner), 0); + } + explicit EmbeddingLookupSparseOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + EmbeddingLookupSparseOptionsBuilder &operator=(const EmbeddingLookupSparseOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateEmbeddingLookupSparseOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::CombinerType combiner = tflite::CombinerType_SUM) { + EmbeddingLookupSparseOptionsBuilder builder_(_fbb); + builder_.add_combiner(combiner); + return builder_.Finish(); +} + +flatbuffers::Offset CreateEmbeddingLookupSparseOptions(flatbuffers::FlatBufferBuilder &_fbb, const EmbeddingLookupSparseOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct GatherOptionsT : public flatbuffers::NativeTable { + typedef GatherOptions TableType; + int32_t axis; + GatherOptionsT() + : axis(0) { + } +}; + +struct GatherOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef GatherOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_AXIS = 4 + }; + int32_t axis() const { + return GetField(VT_AXIS, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_AXIS) && + verifier.EndTable(); + } + GatherOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(GatherOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const GatherOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct GatherOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_axis(int32_t axis) { + fbb_.AddElement(GatherOptions::VT_AXIS, axis, 0); + } + explicit GatherOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + GatherOptionsBuilder &operator=(const GatherOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateGatherOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t axis = 0) { + GatherOptionsBuilder builder_(_fbb); + builder_.add_axis(axis); + return builder_.Finish(); +} + +flatbuffers::Offset CreateGatherOptions(flatbuffers::FlatBufferBuilder &_fbb, const GatherOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct TransposeOptionsT : public flatbuffers::NativeTable { + typedef TransposeOptions TableType; + TransposeOptionsT() { + } +}; + +struct TransposeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef TransposeOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + TransposeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(TransposeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const TransposeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct TransposeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit TransposeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + TransposeOptionsBuilder &operator=(const TransposeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateTransposeOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + TransposeOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateTransposeOptions(flatbuffers::FlatBufferBuilder &_fbb, const TransposeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ExpOptionsT : public flatbuffers::NativeTable { + typedef ExpOptions TableType; + ExpOptionsT() { + } +}; + +struct ExpOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ExpOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + ExpOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ExpOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ExpOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ExpOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit ExpOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ExpOptionsBuilder &operator=(const ExpOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateExpOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + ExpOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateExpOptions(flatbuffers::FlatBufferBuilder &_fbb, const ExpOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct CosOptionsT : public flatbuffers::NativeTable { + typedef CosOptions TableType; + CosOptionsT() { + } +}; + +struct CosOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef CosOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + CosOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(CosOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const CosOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct CosOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit CosOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + CosOptionsBuilder &operator=(const CosOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateCosOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + CosOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateCosOptions(flatbuffers::FlatBufferBuilder &_fbb, const CosOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ReducerOptionsT : public flatbuffers::NativeTable { + typedef ReducerOptions TableType; + bool keep_dims; + ReducerOptionsT() + : keep_dims(false) { + } +}; + +struct ReducerOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ReducerOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_KEEP_DIMS = 4 + }; + bool keep_dims() const { + return GetField(VT_KEEP_DIMS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_KEEP_DIMS) && + verifier.EndTable(); + } + ReducerOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ReducerOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReducerOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ReducerOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_keep_dims(bool keep_dims) { + fbb_.AddElement(ReducerOptions::VT_KEEP_DIMS, static_cast(keep_dims), 0); + } + explicit ReducerOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ReducerOptionsBuilder &operator=(const ReducerOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateReducerOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool keep_dims = false) { + ReducerOptionsBuilder builder_(_fbb); + builder_.add_keep_dims(keep_dims); + return builder_.Finish(); +} + +flatbuffers::Offset CreateReducerOptions(flatbuffers::FlatBufferBuilder &_fbb, const ReducerOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SqueezeOptionsT : public flatbuffers::NativeTable { + typedef SqueezeOptions TableType; + std::vector squeeze_dims; + SqueezeOptionsT() { + } +}; + +struct SqueezeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SqueezeOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_SQUEEZE_DIMS = 4 + }; + const flatbuffers::Vector *squeeze_dims() const { + return GetPointer *>(VT_SQUEEZE_DIMS); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_SQUEEZE_DIMS) && + verifier.VerifyVector(squeeze_dims()) && + verifier.EndTable(); + } + SqueezeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SqueezeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SqueezeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SqueezeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_squeeze_dims(flatbuffers::Offset> squeeze_dims) { + fbb_.AddOffset(SqueezeOptions::VT_SQUEEZE_DIMS, squeeze_dims); + } + explicit SqueezeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SqueezeOptionsBuilder &operator=(const SqueezeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSqueezeOptions( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> squeeze_dims = 0) { + SqueezeOptionsBuilder builder_(_fbb); + builder_.add_squeeze_dims(squeeze_dims); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateSqueezeOptionsDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *squeeze_dims = nullptr) { + auto squeeze_dims__ = squeeze_dims ? _fbb.CreateVector(*squeeze_dims) : 0; + return tflite::CreateSqueezeOptions( + _fbb, + squeeze_dims__); +} + +flatbuffers::Offset CreateSqueezeOptions(flatbuffers::FlatBufferBuilder &_fbb, const SqueezeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SplitOptionsT : public flatbuffers::NativeTable { + typedef SplitOptions TableType; + int32_t num_splits; + SplitOptionsT() + : num_splits(0) { + } +}; + +struct SplitOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SplitOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NUM_SPLITS = 4 + }; + int32_t num_splits() const { + return GetField(VT_NUM_SPLITS, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_NUM_SPLITS) && + verifier.EndTable(); + } + SplitOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SplitOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SplitOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SplitOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_num_splits(int32_t num_splits) { + fbb_.AddElement(SplitOptions::VT_NUM_SPLITS, num_splits, 0); + } + explicit SplitOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SplitOptionsBuilder &operator=(const SplitOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSplitOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t num_splits = 0) { + SplitOptionsBuilder builder_(_fbb); + builder_.add_num_splits(num_splits); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSplitOptions(flatbuffers::FlatBufferBuilder &_fbb, const SplitOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SplitVOptionsT : public flatbuffers::NativeTable { + typedef SplitVOptions TableType; + int32_t num_splits; + SplitVOptionsT() + : num_splits(0) { + } +}; + +struct SplitVOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SplitVOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NUM_SPLITS = 4 + }; + int32_t num_splits() const { + return GetField(VT_NUM_SPLITS, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_NUM_SPLITS) && + verifier.EndTable(); + } + SplitVOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SplitVOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SplitVOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SplitVOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_num_splits(int32_t num_splits) { + fbb_.AddElement(SplitVOptions::VT_NUM_SPLITS, num_splits, 0); + } + explicit SplitVOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SplitVOptionsBuilder &operator=(const SplitVOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSplitVOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t num_splits = 0) { + SplitVOptionsBuilder builder_(_fbb); + builder_.add_num_splits(num_splits); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSplitVOptions(flatbuffers::FlatBufferBuilder &_fbb, const SplitVOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct StridedSliceOptionsT : public flatbuffers::NativeTable { + typedef StridedSliceOptions TableType; + int32_t begin_mask; + int32_t end_mask; + int32_t ellipsis_mask; + int32_t new_axis_mask; + int32_t shrink_axis_mask; + StridedSliceOptionsT() + : begin_mask(0), + end_mask(0), + ellipsis_mask(0), + new_axis_mask(0), + shrink_axis_mask(0) { + } +}; + +struct StridedSliceOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef StridedSliceOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_BEGIN_MASK = 4, + VT_END_MASK = 6, + VT_ELLIPSIS_MASK = 8, + VT_NEW_AXIS_MASK = 10, + VT_SHRINK_AXIS_MASK = 12 + }; + int32_t begin_mask() const { + return GetField(VT_BEGIN_MASK, 0); + } + int32_t end_mask() const { + return GetField(VT_END_MASK, 0); + } + int32_t ellipsis_mask() const { + return GetField(VT_ELLIPSIS_MASK, 0); + } + int32_t new_axis_mask() const { + return GetField(VT_NEW_AXIS_MASK, 0); + } + int32_t shrink_axis_mask() const { + return GetField(VT_SHRINK_AXIS_MASK, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_BEGIN_MASK) && + VerifyField(verifier, VT_END_MASK) && + VerifyField(verifier, VT_ELLIPSIS_MASK) && + VerifyField(verifier, VT_NEW_AXIS_MASK) && + VerifyField(verifier, VT_SHRINK_AXIS_MASK) && + verifier.EndTable(); + } + StridedSliceOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(StridedSliceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const StridedSliceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct StridedSliceOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_begin_mask(int32_t begin_mask) { + fbb_.AddElement(StridedSliceOptions::VT_BEGIN_MASK, begin_mask, 0); + } + void add_end_mask(int32_t end_mask) { + fbb_.AddElement(StridedSliceOptions::VT_END_MASK, end_mask, 0); + } + void add_ellipsis_mask(int32_t ellipsis_mask) { + fbb_.AddElement(StridedSliceOptions::VT_ELLIPSIS_MASK, ellipsis_mask, 0); + } + void add_new_axis_mask(int32_t new_axis_mask) { + fbb_.AddElement(StridedSliceOptions::VT_NEW_AXIS_MASK, new_axis_mask, 0); + } + void add_shrink_axis_mask(int32_t shrink_axis_mask) { + fbb_.AddElement(StridedSliceOptions::VT_SHRINK_AXIS_MASK, shrink_axis_mask, 0); + } + explicit StridedSliceOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + StridedSliceOptionsBuilder &operator=(const StridedSliceOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateStridedSliceOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t begin_mask = 0, + int32_t end_mask = 0, + int32_t ellipsis_mask = 0, + int32_t new_axis_mask = 0, + int32_t shrink_axis_mask = 0) { + StridedSliceOptionsBuilder builder_(_fbb); + builder_.add_shrink_axis_mask(shrink_axis_mask); + builder_.add_new_axis_mask(new_axis_mask); + builder_.add_ellipsis_mask(ellipsis_mask); + builder_.add_end_mask(end_mask); + builder_.add_begin_mask(begin_mask); + return builder_.Finish(); +} + +flatbuffers::Offset CreateStridedSliceOptions(flatbuffers::FlatBufferBuilder &_fbb, const StridedSliceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LogSoftmaxOptionsT : public flatbuffers::NativeTable { + typedef LogSoftmaxOptions TableType; + LogSoftmaxOptionsT() { + } +}; + +struct LogSoftmaxOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LogSoftmaxOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + LogSoftmaxOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LogSoftmaxOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogSoftmaxOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LogSoftmaxOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit LogSoftmaxOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LogSoftmaxOptionsBuilder &operator=(const LogSoftmaxOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLogSoftmaxOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + LogSoftmaxOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLogSoftmaxOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogSoftmaxOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct CastOptionsT : public flatbuffers::NativeTable { + typedef CastOptions TableType; + tflite::TensorType in_data_type; + tflite::TensorType out_data_type; + CastOptionsT() + : in_data_type(tflite::TensorType_FLOAT32), + out_data_type(tflite::TensorType_FLOAT32) { + } +}; + +struct CastOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef CastOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_IN_DATA_TYPE = 4, + VT_OUT_DATA_TYPE = 6 + }; + tflite::TensorType in_data_type() const { + return static_cast(GetField(VT_IN_DATA_TYPE, 0)); + } + tflite::TensorType out_data_type() const { + return static_cast(GetField(VT_OUT_DATA_TYPE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_IN_DATA_TYPE) && + VerifyField(verifier, VT_OUT_DATA_TYPE) && + verifier.EndTable(); + } + CastOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(CastOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const CastOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct CastOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_in_data_type(tflite::TensorType in_data_type) { + fbb_.AddElement(CastOptions::VT_IN_DATA_TYPE, static_cast(in_data_type), 0); + } + void add_out_data_type(tflite::TensorType out_data_type) { + fbb_.AddElement(CastOptions::VT_OUT_DATA_TYPE, static_cast(out_data_type), 0); + } + explicit CastOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + CastOptionsBuilder &operator=(const CastOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateCastOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::TensorType in_data_type = tflite::TensorType_FLOAT32, + tflite::TensorType out_data_type = tflite::TensorType_FLOAT32) { + CastOptionsBuilder builder_(_fbb); + builder_.add_out_data_type(out_data_type); + builder_.add_in_data_type(in_data_type); + return builder_.Finish(); +} + +flatbuffers::Offset CreateCastOptions(flatbuffers::FlatBufferBuilder &_fbb, const CastOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct DequantizeOptionsT : public flatbuffers::NativeTable { + typedef DequantizeOptions TableType; + DequantizeOptionsT() { + } +}; + +struct DequantizeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef DequantizeOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + DequantizeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(DequantizeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DequantizeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct DequantizeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit DequantizeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + DequantizeOptionsBuilder &operator=(const DequantizeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateDequantizeOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + DequantizeOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateDequantizeOptions(flatbuffers::FlatBufferBuilder &_fbb, const DequantizeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct MaximumMinimumOptionsT : public flatbuffers::NativeTable { + typedef MaximumMinimumOptions TableType; + MaximumMinimumOptionsT() { + } +}; + +struct MaximumMinimumOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef MaximumMinimumOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + MaximumMinimumOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(MaximumMinimumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const MaximumMinimumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct MaximumMinimumOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit MaximumMinimumOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + MaximumMinimumOptionsBuilder &operator=(const MaximumMinimumOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateMaximumMinimumOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + MaximumMinimumOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateMaximumMinimumOptions(flatbuffers::FlatBufferBuilder &_fbb, const MaximumMinimumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct TileOptionsT : public flatbuffers::NativeTable { + typedef TileOptions TableType; + TileOptionsT() { + } +}; + +struct TileOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef TileOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + TileOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(TileOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const TileOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct TileOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit TileOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + TileOptionsBuilder &operator=(const TileOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateTileOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + TileOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateTileOptions(flatbuffers::FlatBufferBuilder &_fbb, const TileOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ArgMaxOptionsT : public flatbuffers::NativeTable { + typedef ArgMaxOptions TableType; + tflite::TensorType output_type; + ArgMaxOptionsT() + : output_type(tflite::TensorType_FLOAT32) { + } +}; + +struct ArgMaxOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ArgMaxOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_OUTPUT_TYPE = 4 + }; + tflite::TensorType output_type() const { + return static_cast(GetField(VT_OUTPUT_TYPE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_OUTPUT_TYPE) && + verifier.EndTable(); + } + ArgMaxOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ArgMaxOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ArgMaxOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ArgMaxOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_output_type(tflite::TensorType output_type) { + fbb_.AddElement(ArgMaxOptions::VT_OUTPUT_TYPE, static_cast(output_type), 0); + } + explicit ArgMaxOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ArgMaxOptionsBuilder &operator=(const ArgMaxOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateArgMaxOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::TensorType output_type = tflite::TensorType_FLOAT32) { + ArgMaxOptionsBuilder builder_(_fbb); + builder_.add_output_type(output_type); + return builder_.Finish(); +} + +flatbuffers::Offset CreateArgMaxOptions(flatbuffers::FlatBufferBuilder &_fbb, const ArgMaxOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ArgMinOptionsT : public flatbuffers::NativeTable { + typedef ArgMinOptions TableType; + tflite::TensorType output_type; + ArgMinOptionsT() + : output_type(tflite::TensorType_FLOAT32) { + } +}; + +struct ArgMinOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ArgMinOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_OUTPUT_TYPE = 4 + }; + tflite::TensorType output_type() const { + return static_cast(GetField(VT_OUTPUT_TYPE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_OUTPUT_TYPE) && + verifier.EndTable(); + } + ArgMinOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ArgMinOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ArgMinOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_output_type(tflite::TensorType output_type) { + fbb_.AddElement(ArgMinOptions::VT_OUTPUT_TYPE, static_cast(output_type), 0); + } + explicit ArgMinOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ArgMinOptionsBuilder &operator=(const ArgMinOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateArgMinOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::TensorType output_type = tflite::TensorType_FLOAT32) { + ArgMinOptionsBuilder builder_(_fbb); + builder_.add_output_type(output_type); + return builder_.Finish(); +} + +flatbuffers::Offset CreateArgMinOptions(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct GreaterOptionsT : public flatbuffers::NativeTable { + typedef GreaterOptions TableType; + GreaterOptionsT() { + } +}; + +struct GreaterOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef GreaterOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + GreaterOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(GreaterOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const GreaterOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct GreaterOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit GreaterOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + GreaterOptionsBuilder &operator=(const GreaterOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateGreaterOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + GreaterOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateGreaterOptions(flatbuffers::FlatBufferBuilder &_fbb, const GreaterOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct GreaterEqualOptionsT : public flatbuffers::NativeTable { + typedef GreaterEqualOptions TableType; + GreaterEqualOptionsT() { + } +}; + +struct GreaterEqualOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef GreaterEqualOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + GreaterEqualOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(GreaterEqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const GreaterEqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct GreaterEqualOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit GreaterEqualOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + GreaterEqualOptionsBuilder &operator=(const GreaterEqualOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateGreaterEqualOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + GreaterEqualOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateGreaterEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const GreaterEqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LessOptionsT : public flatbuffers::NativeTable { + typedef LessOptions TableType; + LessOptionsT() { + } +}; + +struct LessOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LessOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + LessOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LessOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LessOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LessOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit LessOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LessOptionsBuilder &operator=(const LessOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLessOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + LessOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLessOptions(flatbuffers::FlatBufferBuilder &_fbb, const LessOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LessEqualOptionsT : public flatbuffers::NativeTable { + typedef LessEqualOptions TableType; + LessEqualOptionsT() { + } +}; + +struct LessEqualOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LessEqualOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + LessEqualOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LessEqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LessEqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LessEqualOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit LessEqualOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LessEqualOptionsBuilder &operator=(const LessEqualOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLessEqualOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + LessEqualOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLessEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const LessEqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct NegOptionsT : public flatbuffers::NativeTable { + typedef NegOptions TableType; + NegOptionsT() { + } +}; + +struct NegOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef NegOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + NegOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(NegOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const NegOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct NegOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit NegOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + NegOptionsBuilder &operator=(const NegOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateNegOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + NegOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateNegOptions(flatbuffers::FlatBufferBuilder &_fbb, const NegOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SelectOptionsT : public flatbuffers::NativeTable { + typedef SelectOptions TableType; + SelectOptionsT() { + } +}; + +struct SelectOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SelectOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + SelectOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SelectOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SelectOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SelectOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit SelectOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SelectOptionsBuilder &operator=(const SelectOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSelectOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + SelectOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSelectOptions(flatbuffers::FlatBufferBuilder &_fbb, const SelectOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SliceOptionsT : public flatbuffers::NativeTable { + typedef SliceOptions TableType; + SliceOptionsT() { + } +}; + +struct SliceOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SliceOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + SliceOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SliceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SliceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SliceOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit SliceOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SliceOptionsBuilder &operator=(const SliceOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSliceOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + SliceOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSliceOptions(flatbuffers::FlatBufferBuilder &_fbb, const SliceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct TransposeConvOptionsT : public flatbuffers::NativeTable { + typedef TransposeConvOptions TableType; + tflite::Padding padding; + int32_t stride_w; + int32_t stride_h; + TransposeConvOptionsT() + : padding(tflite::Padding_SAME), + stride_w(0), + stride_h(0) { + } +}; + +struct TransposeConvOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef TransposeConvOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_PADDING = 4, + VT_STRIDE_W = 6, + VT_STRIDE_H = 8 + }; + tflite::Padding padding() const { + return static_cast(GetField(VT_PADDING, 0)); + } + int32_t stride_w() const { + return GetField(VT_STRIDE_W, 0); + } + int32_t stride_h() const { + return GetField(VT_STRIDE_H, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_PADDING) && + VerifyField(verifier, VT_STRIDE_W) && + VerifyField(verifier, VT_STRIDE_H) && + verifier.EndTable(); + } + TransposeConvOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(TransposeConvOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const TransposeConvOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct TransposeConvOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_padding(tflite::Padding padding) { + fbb_.AddElement(TransposeConvOptions::VT_PADDING, static_cast(padding), 0); + } + void add_stride_w(int32_t stride_w) { + fbb_.AddElement(TransposeConvOptions::VT_STRIDE_W, stride_w, 0); + } + void add_stride_h(int32_t stride_h) { + fbb_.AddElement(TransposeConvOptions::VT_STRIDE_H, stride_h, 0); + } + explicit TransposeConvOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + TransposeConvOptionsBuilder &operator=(const TransposeConvOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateTransposeConvOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::Padding padding = tflite::Padding_SAME, + int32_t stride_w = 0, + int32_t stride_h = 0) { + TransposeConvOptionsBuilder builder_(_fbb); + builder_.add_stride_h(stride_h); + builder_.add_stride_w(stride_w); + builder_.add_padding(padding); + return builder_.Finish(); +} + +flatbuffers::Offset CreateTransposeConvOptions(flatbuffers::FlatBufferBuilder &_fbb, const TransposeConvOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ExpandDimsOptionsT : public flatbuffers::NativeTable { + typedef ExpandDimsOptions TableType; + ExpandDimsOptionsT() { + } +}; + +struct ExpandDimsOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ExpandDimsOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + ExpandDimsOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ExpandDimsOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ExpandDimsOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ExpandDimsOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit ExpandDimsOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ExpandDimsOptionsBuilder &operator=(const ExpandDimsOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateExpandDimsOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + ExpandDimsOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateExpandDimsOptions(flatbuffers::FlatBufferBuilder &_fbb, const ExpandDimsOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SparseToDenseOptionsT : public flatbuffers::NativeTable { + typedef SparseToDenseOptions TableType; + bool validate_indices; + SparseToDenseOptionsT() + : validate_indices(false) { + } +}; + +struct SparseToDenseOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SparseToDenseOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_VALIDATE_INDICES = 4 + }; + bool validate_indices() const { + return GetField(VT_VALIDATE_INDICES, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_VALIDATE_INDICES) && + verifier.EndTable(); + } + SparseToDenseOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SparseToDenseOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SparseToDenseOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SparseToDenseOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_validate_indices(bool validate_indices) { + fbb_.AddElement(SparseToDenseOptions::VT_VALIDATE_INDICES, static_cast(validate_indices), 0); + } + explicit SparseToDenseOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SparseToDenseOptionsBuilder &operator=(const SparseToDenseOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSparseToDenseOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool validate_indices = false) { + SparseToDenseOptionsBuilder builder_(_fbb); + builder_.add_validate_indices(validate_indices); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSparseToDenseOptions(flatbuffers::FlatBufferBuilder &_fbb, const SparseToDenseOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct EqualOptionsT : public flatbuffers::NativeTable { + typedef EqualOptions TableType; + EqualOptionsT() { + } +}; + +struct EqualOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef EqualOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + EqualOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(EqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const EqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct EqualOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit EqualOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + EqualOptionsBuilder &operator=(const EqualOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateEqualOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + EqualOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const EqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct NotEqualOptionsT : public flatbuffers::NativeTable { + typedef NotEqualOptions TableType; + NotEqualOptionsT() { + } +}; + +struct NotEqualOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef NotEqualOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + NotEqualOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(NotEqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const NotEqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct NotEqualOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit NotEqualOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + NotEqualOptionsBuilder &operator=(const NotEqualOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateNotEqualOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + NotEqualOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateNotEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const NotEqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ShapeOptionsT : public flatbuffers::NativeTable { + typedef ShapeOptions TableType; + tflite::TensorType out_type; + ShapeOptionsT() + : out_type(tflite::TensorType_FLOAT32) { + } +}; + +struct ShapeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ShapeOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_OUT_TYPE = 4 + }; + tflite::TensorType out_type() const { + return static_cast(GetField(VT_OUT_TYPE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_OUT_TYPE) && + verifier.EndTable(); + } + ShapeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ShapeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ShapeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ShapeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_out_type(tflite::TensorType out_type) { + fbb_.AddElement(ShapeOptions::VT_OUT_TYPE, static_cast(out_type), 0); + } + explicit ShapeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ShapeOptionsBuilder &operator=(const ShapeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateShapeOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::TensorType out_type = tflite::TensorType_FLOAT32) { + ShapeOptionsBuilder builder_(_fbb); + builder_.add_out_type(out_type); + return builder_.Finish(); +} + +flatbuffers::Offset CreateShapeOptions(flatbuffers::FlatBufferBuilder &_fbb, const ShapeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct RankOptionsT : public flatbuffers::NativeTable { + typedef RankOptions TableType; + RankOptionsT() { + } +}; + +struct RankOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef RankOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + RankOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(RankOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const RankOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct RankOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit RankOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + RankOptionsBuilder &operator=(const RankOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateRankOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + RankOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateRankOptions(flatbuffers::FlatBufferBuilder &_fbb, const RankOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct PowOptionsT : public flatbuffers::NativeTable { + typedef PowOptions TableType; + PowOptionsT() { + } +}; + +struct PowOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef PowOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + PowOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(PowOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct PowOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit PowOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + PowOptionsBuilder &operator=(const PowOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreatePowOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + PowOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreatePowOptions(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct FakeQuantOptionsT : public flatbuffers::NativeTable { + typedef FakeQuantOptions TableType; + float min; + float max; + int32_t num_bits; + bool narrow_range; + FakeQuantOptionsT() + : min(0.0f), + max(0.0f), + num_bits(0), + narrow_range(false) { + } +}; + +struct FakeQuantOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef FakeQuantOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_MIN = 4, + VT_MAX = 6, + VT_NUM_BITS = 8, + VT_NARROW_RANGE = 10 + }; + float min() const { + return GetField(VT_MIN, 0.0f); + } + float max() const { + return GetField(VT_MAX, 0.0f); + } + int32_t num_bits() const { + return GetField(VT_NUM_BITS, 0); + } + bool narrow_range() const { + return GetField(VT_NARROW_RANGE, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_MIN) && + VerifyField(verifier, VT_MAX) && + VerifyField(verifier, VT_NUM_BITS) && + VerifyField(verifier, VT_NARROW_RANGE) && + verifier.EndTable(); + } + FakeQuantOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(FakeQuantOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct FakeQuantOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_min(float min) { + fbb_.AddElement(FakeQuantOptions::VT_MIN, min, 0.0f); + } + void add_max(float max) { + fbb_.AddElement(FakeQuantOptions::VT_MAX, max, 0.0f); + } + void add_num_bits(int32_t num_bits) { + fbb_.AddElement(FakeQuantOptions::VT_NUM_BITS, num_bits, 0); + } + void add_narrow_range(bool narrow_range) { + fbb_.AddElement(FakeQuantOptions::VT_NARROW_RANGE, static_cast(narrow_range), 0); + } + explicit FakeQuantOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + FakeQuantOptionsBuilder &operator=(const FakeQuantOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateFakeQuantOptions( + flatbuffers::FlatBufferBuilder &_fbb, + float min = 0.0f, + float max = 0.0f, + int32_t num_bits = 0, + bool narrow_range = false) { + FakeQuantOptionsBuilder builder_(_fbb); + builder_.add_num_bits(num_bits); + builder_.add_max(max); + builder_.add_min(min); + builder_.add_narrow_range(narrow_range); + return builder_.Finish(); +} + +flatbuffers::Offset CreateFakeQuantOptions(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct PackOptionsT : public flatbuffers::NativeTable { + typedef PackOptions TableType; + int32_t values_count; + int32_t axis; + PackOptionsT() + : values_count(0), + axis(0) { + } +}; + +struct PackOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef PackOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_VALUES_COUNT = 4, + VT_AXIS = 6 + }; + int32_t values_count() const { + return GetField(VT_VALUES_COUNT, 0); + } + int32_t axis() const { + return GetField(VT_AXIS, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_VALUES_COUNT) && + VerifyField(verifier, VT_AXIS) && + verifier.EndTable(); + } + PackOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(PackOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const PackOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct PackOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_values_count(int32_t values_count) { + fbb_.AddElement(PackOptions::VT_VALUES_COUNT, values_count, 0); + } + void add_axis(int32_t axis) { + fbb_.AddElement(PackOptions::VT_AXIS, axis, 0); + } + explicit PackOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + PackOptionsBuilder &operator=(const PackOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreatePackOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t values_count = 0, + int32_t axis = 0) { + PackOptionsBuilder builder_(_fbb); + builder_.add_axis(axis); + builder_.add_values_count(values_count); + return builder_.Finish(); +} + +flatbuffers::Offset CreatePackOptions(flatbuffers::FlatBufferBuilder &_fbb, const PackOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LogicalOrOptionsT : public flatbuffers::NativeTable { + typedef LogicalOrOptions TableType; + LogicalOrOptionsT() { + } +}; + +struct LogicalOrOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LogicalOrOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + LogicalOrOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LogicalOrOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogicalOrOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LogicalOrOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit LogicalOrOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LogicalOrOptionsBuilder &operator=(const LogicalOrOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLogicalOrOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + LogicalOrOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLogicalOrOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogicalOrOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct OneHotOptionsT : public flatbuffers::NativeTable { + typedef OneHotOptions TableType; + int32_t axis; + OneHotOptionsT() + : axis(0) { + } +}; + +struct OneHotOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef OneHotOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_AXIS = 4 + }; + int32_t axis() const { + return GetField(VT_AXIS, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_AXIS) && + verifier.EndTable(); + } + OneHotOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(OneHotOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const OneHotOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct OneHotOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_axis(int32_t axis) { + fbb_.AddElement(OneHotOptions::VT_AXIS, axis, 0); + } + explicit OneHotOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + OneHotOptionsBuilder &operator=(const OneHotOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateOneHotOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t axis = 0) { + OneHotOptionsBuilder builder_(_fbb); + builder_.add_axis(axis); + return builder_.Finish(); +} + +flatbuffers::Offset CreateOneHotOptions(flatbuffers::FlatBufferBuilder &_fbb, const OneHotOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct AbsOptionsT : public flatbuffers::NativeTable { + typedef AbsOptions TableType; + AbsOptionsT() { + } +}; + +struct AbsOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef AbsOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + AbsOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(AbsOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const AbsOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct AbsOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit AbsOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + AbsOptionsBuilder &operator=(const AbsOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateAbsOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + AbsOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateAbsOptions(flatbuffers::FlatBufferBuilder &_fbb, const AbsOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct HardSwishOptionsT : public flatbuffers::NativeTable { + typedef HardSwishOptions TableType; + HardSwishOptionsT() { + } +}; + +struct HardSwishOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef HardSwishOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + HardSwishOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(HardSwishOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const HardSwishOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct HardSwishOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit HardSwishOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + HardSwishOptionsBuilder &operator=(const HardSwishOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateHardSwishOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + HardSwishOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateHardSwishOptions(flatbuffers::FlatBufferBuilder &_fbb, const HardSwishOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LogicalAndOptionsT : public flatbuffers::NativeTable { + typedef LogicalAndOptions TableType; + LogicalAndOptionsT() { + } +}; + +struct LogicalAndOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LogicalAndOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + LogicalAndOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LogicalAndOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogicalAndOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LogicalAndOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit LogicalAndOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LogicalAndOptionsBuilder &operator=(const LogicalAndOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLogicalAndOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + LogicalAndOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLogicalAndOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogicalAndOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LogicalNotOptionsT : public flatbuffers::NativeTable { + typedef LogicalNotOptions TableType; + LogicalNotOptionsT() { + } +}; + +struct LogicalNotOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LogicalNotOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + LogicalNotOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LogicalNotOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogicalNotOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LogicalNotOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit LogicalNotOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LogicalNotOptionsBuilder &operator=(const LogicalNotOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLogicalNotOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + LogicalNotOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLogicalNotOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogicalNotOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct UnpackOptionsT : public flatbuffers::NativeTable { + typedef UnpackOptions TableType; + int32_t num; + int32_t axis; + UnpackOptionsT() + : num(0), + axis(0) { + } +}; + +struct UnpackOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef UnpackOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NUM = 4, + VT_AXIS = 6 + }; + int32_t num() const { + return GetField(VT_NUM, 0); + } + int32_t axis() const { + return GetField(VT_AXIS, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_NUM) && + VerifyField(verifier, VT_AXIS) && + verifier.EndTable(); + } + UnpackOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(UnpackOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const UnpackOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct UnpackOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_num(int32_t num) { + fbb_.AddElement(UnpackOptions::VT_NUM, num, 0); + } + void add_axis(int32_t axis) { + fbb_.AddElement(UnpackOptions::VT_AXIS, axis, 0); + } + explicit UnpackOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + UnpackOptionsBuilder &operator=(const UnpackOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateUnpackOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t num = 0, + int32_t axis = 0) { + UnpackOptionsBuilder builder_(_fbb); + builder_.add_axis(axis); + builder_.add_num(num); + return builder_.Finish(); +} + +flatbuffers::Offset CreateUnpackOptions(flatbuffers::FlatBufferBuilder &_fbb, const UnpackOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct FloorDivOptionsT : public flatbuffers::NativeTable { + typedef FloorDivOptions TableType; + FloorDivOptionsT() { + } +}; + +struct FloorDivOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef FloorDivOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + FloorDivOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(FloorDivOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const FloorDivOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct FloorDivOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit FloorDivOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + FloorDivOptionsBuilder &operator=(const FloorDivOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateFloorDivOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + FloorDivOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateFloorDivOptions(flatbuffers::FlatBufferBuilder &_fbb, const FloorDivOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SquareOptionsT : public flatbuffers::NativeTable { + typedef SquareOptions TableType; + SquareOptionsT() { + } +}; + +struct SquareOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SquareOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + SquareOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SquareOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SquareOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SquareOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit SquareOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SquareOptionsBuilder &operator=(const SquareOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSquareOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + SquareOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSquareOptions(flatbuffers::FlatBufferBuilder &_fbb, const SquareOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ZerosLikeOptionsT : public flatbuffers::NativeTable { + typedef ZerosLikeOptions TableType; + ZerosLikeOptionsT() { + } +}; + +struct ZerosLikeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ZerosLikeOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + ZerosLikeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ZerosLikeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ZerosLikeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ZerosLikeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit ZerosLikeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ZerosLikeOptionsBuilder &operator=(const ZerosLikeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateZerosLikeOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + ZerosLikeOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateZerosLikeOptions(flatbuffers::FlatBufferBuilder &_fbb, const ZerosLikeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct FillOptionsT : public flatbuffers::NativeTable { + typedef FillOptions TableType; + FillOptionsT() { + } +}; + +struct FillOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef FillOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + FillOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(FillOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const FillOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct FillOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit FillOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + FillOptionsBuilder &operator=(const FillOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateFillOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + FillOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateFillOptions(flatbuffers::FlatBufferBuilder &_fbb, const FillOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct FloorModOptionsT : public flatbuffers::NativeTable { + typedef FloorModOptions TableType; + FloorModOptionsT() { + } +}; + +struct FloorModOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef FloorModOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + FloorModOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(FloorModOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const FloorModOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct FloorModOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit FloorModOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + FloorModOptionsBuilder &operator=(const FloorModOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateFloorModOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + FloorModOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateFloorModOptions(flatbuffers::FlatBufferBuilder &_fbb, const FloorModOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct RangeOptionsT : public flatbuffers::NativeTable { + typedef RangeOptions TableType; + RangeOptionsT() { + } +}; + +struct RangeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef RangeOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + RangeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(RangeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const RangeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct RangeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit RangeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + RangeOptionsBuilder &operator=(const RangeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateRangeOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + RangeOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateRangeOptions(flatbuffers::FlatBufferBuilder &_fbb, const RangeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct LeakyReluOptionsT : public flatbuffers::NativeTable { + typedef LeakyReluOptions TableType; + float alpha; + LeakyReluOptionsT() + : alpha(0.0f) { + } +}; + +struct LeakyReluOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef LeakyReluOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_ALPHA = 4 + }; + float alpha() const { + return GetField(VT_ALPHA, 0.0f); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_ALPHA) && + verifier.EndTable(); + } + LeakyReluOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(LeakyReluOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const LeakyReluOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct LeakyReluOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_alpha(float alpha) { + fbb_.AddElement(LeakyReluOptions::VT_ALPHA, alpha, 0.0f); + } + explicit LeakyReluOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + LeakyReluOptionsBuilder &operator=(const LeakyReluOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateLeakyReluOptions( + flatbuffers::FlatBufferBuilder &_fbb, + float alpha = 0.0f) { + LeakyReluOptionsBuilder builder_(_fbb); + builder_.add_alpha(alpha); + return builder_.Finish(); +} + +flatbuffers::Offset CreateLeakyReluOptions(flatbuffers::FlatBufferBuilder &_fbb, const LeakyReluOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SquaredDifferenceOptionsT : public flatbuffers::NativeTable { + typedef SquaredDifferenceOptions TableType; + SquaredDifferenceOptionsT() { + } +}; + +struct SquaredDifferenceOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SquaredDifferenceOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + SquaredDifferenceOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SquaredDifferenceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SquaredDifferenceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SquaredDifferenceOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit SquaredDifferenceOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SquaredDifferenceOptionsBuilder &operator=(const SquaredDifferenceOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSquaredDifferenceOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + SquaredDifferenceOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSquaredDifferenceOptions(flatbuffers::FlatBufferBuilder &_fbb, const SquaredDifferenceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct MirrorPadOptionsT : public flatbuffers::NativeTable { + typedef MirrorPadOptions TableType; + tflite::MirrorPadMode mode; + MirrorPadOptionsT() + : mode(tflite::MirrorPadMode_REFLECT) { + } +}; + +struct MirrorPadOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef MirrorPadOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_MODE = 4 + }; + tflite::MirrorPadMode mode() const { + return static_cast(GetField(VT_MODE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_MODE) && + verifier.EndTable(); + } + MirrorPadOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(MirrorPadOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const MirrorPadOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct MirrorPadOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_mode(tflite::MirrorPadMode mode) { + fbb_.AddElement(MirrorPadOptions::VT_MODE, static_cast(mode), 0); + } + explicit MirrorPadOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + MirrorPadOptionsBuilder &operator=(const MirrorPadOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateMirrorPadOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::MirrorPadMode mode = tflite::MirrorPadMode_REFLECT) { + MirrorPadOptionsBuilder builder_(_fbb); + builder_.add_mode(mode); + return builder_.Finish(); +} + +flatbuffers::Offset CreateMirrorPadOptions(flatbuffers::FlatBufferBuilder &_fbb, const MirrorPadOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct UniqueOptionsT : public flatbuffers::NativeTable { + typedef UniqueOptions TableType; + tflite::TensorType idx_out_type; + UniqueOptionsT() + : idx_out_type(tflite::TensorType_INT32) { + } +}; + +struct UniqueOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef UniqueOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_IDX_OUT_TYPE = 4 + }; + tflite::TensorType idx_out_type() const { + return static_cast(GetField(VT_IDX_OUT_TYPE, 2)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_IDX_OUT_TYPE) && + verifier.EndTable(); + } + UniqueOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(UniqueOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const UniqueOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct UniqueOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_idx_out_type(tflite::TensorType idx_out_type) { + fbb_.AddElement(UniqueOptions::VT_IDX_OUT_TYPE, static_cast(idx_out_type), 2); + } + explicit UniqueOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + UniqueOptionsBuilder &operator=(const UniqueOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateUniqueOptions( + flatbuffers::FlatBufferBuilder &_fbb, + tflite::TensorType idx_out_type = tflite::TensorType_INT32) { + UniqueOptionsBuilder builder_(_fbb); + builder_.add_idx_out_type(idx_out_type); + return builder_.Finish(); +} + +flatbuffers::Offset CreateUniqueOptions(flatbuffers::FlatBufferBuilder &_fbb, const UniqueOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ReverseV2OptionsT : public flatbuffers::NativeTable { + typedef ReverseV2Options TableType; + ReverseV2OptionsT() { + } +}; + +struct ReverseV2Options FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ReverseV2OptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + ReverseV2OptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ReverseV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReverseV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ReverseV2OptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit ReverseV2OptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ReverseV2OptionsBuilder &operator=(const ReverseV2OptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateReverseV2Options( + flatbuffers::FlatBufferBuilder &_fbb) { + ReverseV2OptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateReverseV2Options(flatbuffers::FlatBufferBuilder &_fbb, const ReverseV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct AddNOptionsT : public flatbuffers::NativeTable { + typedef AddNOptions TableType; + AddNOptionsT() { + } +}; + +struct AddNOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef AddNOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + AddNOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(AddNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const AddNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct AddNOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit AddNOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + AddNOptionsBuilder &operator=(const AddNOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateAddNOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + AddNOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateAddNOptions(flatbuffers::FlatBufferBuilder &_fbb, const AddNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct GatherNdOptionsT : public flatbuffers::NativeTable { + typedef GatherNdOptions TableType; + GatherNdOptionsT() { + } +}; + +struct GatherNdOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef GatherNdOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + GatherNdOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(GatherNdOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const GatherNdOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct GatherNdOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit GatherNdOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + GatherNdOptionsBuilder &operator=(const GatherNdOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateGatherNdOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + GatherNdOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateGatherNdOptions(flatbuffers::FlatBufferBuilder &_fbb, const GatherNdOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct WhereOptionsT : public flatbuffers::NativeTable { + typedef WhereOptions TableType; + WhereOptionsT() { + } +}; + +struct WhereOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef WhereOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + WhereOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(WhereOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const WhereOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct WhereOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit WhereOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + WhereOptionsBuilder &operator=(const WhereOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateWhereOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + WhereOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateWhereOptions(flatbuffers::FlatBufferBuilder &_fbb, const WhereOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ReverseSequenceOptionsT : public flatbuffers::NativeTable { + typedef ReverseSequenceOptions TableType; + int32_t seq_dim; + int32_t batch_dim; + ReverseSequenceOptionsT() + : seq_dim(0), + batch_dim(0) { + } +}; + +struct ReverseSequenceOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ReverseSequenceOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_SEQ_DIM = 4, + VT_BATCH_DIM = 6 + }; + int32_t seq_dim() const { + return GetField(VT_SEQ_DIM, 0); + } + int32_t batch_dim() const { + return GetField(VT_BATCH_DIM, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_SEQ_DIM) && + VerifyField(verifier, VT_BATCH_DIM) && + verifier.EndTable(); + } + ReverseSequenceOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ReverseSequenceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReverseSequenceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ReverseSequenceOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_seq_dim(int32_t seq_dim) { + fbb_.AddElement(ReverseSequenceOptions::VT_SEQ_DIM, seq_dim, 0); + } + void add_batch_dim(int32_t batch_dim) { + fbb_.AddElement(ReverseSequenceOptions::VT_BATCH_DIM, batch_dim, 0); + } + explicit ReverseSequenceOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ReverseSequenceOptionsBuilder &operator=(const ReverseSequenceOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateReverseSequenceOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t seq_dim = 0, + int32_t batch_dim = 0) { + ReverseSequenceOptionsBuilder builder_(_fbb); + builder_.add_batch_dim(batch_dim); + builder_.add_seq_dim(seq_dim); + return builder_.Finish(); +} + +flatbuffers::Offset CreateReverseSequenceOptions(flatbuffers::FlatBufferBuilder &_fbb, const ReverseSequenceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct MatrixDiagOptionsT : public flatbuffers::NativeTable { + typedef MatrixDiagOptions TableType; + MatrixDiagOptionsT() { + } +}; + +struct MatrixDiagOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef MatrixDiagOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + MatrixDiagOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(MatrixDiagOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const MatrixDiagOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct MatrixDiagOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit MatrixDiagOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + MatrixDiagOptionsBuilder &operator=(const MatrixDiagOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateMatrixDiagOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + MatrixDiagOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateMatrixDiagOptions(flatbuffers::FlatBufferBuilder &_fbb, const MatrixDiagOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct QuantizeOptionsT : public flatbuffers::NativeTable { + typedef QuantizeOptions TableType; + QuantizeOptionsT() { + } +}; + +struct QuantizeOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef QuantizeOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + QuantizeOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(QuantizeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const QuantizeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct QuantizeOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit QuantizeOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + QuantizeOptionsBuilder &operator=(const QuantizeOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateQuantizeOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + QuantizeOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateQuantizeOptions(flatbuffers::FlatBufferBuilder &_fbb, const QuantizeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct MatrixSetDiagOptionsT : public flatbuffers::NativeTable { + typedef MatrixSetDiagOptions TableType; + MatrixSetDiagOptionsT() { + } +}; + +struct MatrixSetDiagOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef MatrixSetDiagOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + MatrixSetDiagOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(MatrixSetDiagOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const MatrixSetDiagOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct MatrixSetDiagOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit MatrixSetDiagOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + MatrixSetDiagOptionsBuilder &operator=(const MatrixSetDiagOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateMatrixSetDiagOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + MatrixSetDiagOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateMatrixSetDiagOptions(flatbuffers::FlatBufferBuilder &_fbb, const MatrixSetDiagOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct IfOptionsT : public flatbuffers::NativeTable { + typedef IfOptions TableType; + int32_t then_subgraph_index; + int32_t else_subgraph_index; + IfOptionsT() + : then_subgraph_index(0), + else_subgraph_index(0) { + } +}; + +struct IfOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef IfOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_THEN_SUBGRAPH_INDEX = 4, + VT_ELSE_SUBGRAPH_INDEX = 6 + }; + int32_t then_subgraph_index() const { + return GetField(VT_THEN_SUBGRAPH_INDEX, 0); + } + int32_t else_subgraph_index() const { + return GetField(VT_ELSE_SUBGRAPH_INDEX, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_THEN_SUBGRAPH_INDEX) && + VerifyField(verifier, VT_ELSE_SUBGRAPH_INDEX) && + verifier.EndTable(); + } + IfOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(IfOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const IfOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct IfOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_then_subgraph_index(int32_t then_subgraph_index) { + fbb_.AddElement(IfOptions::VT_THEN_SUBGRAPH_INDEX, then_subgraph_index, 0); + } + void add_else_subgraph_index(int32_t else_subgraph_index) { + fbb_.AddElement(IfOptions::VT_ELSE_SUBGRAPH_INDEX, else_subgraph_index, 0); + } + explicit IfOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + IfOptionsBuilder &operator=(const IfOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateIfOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t then_subgraph_index = 0, + int32_t else_subgraph_index = 0) { + IfOptionsBuilder builder_(_fbb); + builder_.add_else_subgraph_index(else_subgraph_index); + builder_.add_then_subgraph_index(then_subgraph_index); + return builder_.Finish(); +} + +flatbuffers::Offset CreateIfOptions(flatbuffers::FlatBufferBuilder &_fbb, const IfOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct CallOnceOptionsT : public flatbuffers::NativeTable { + typedef CallOnceOptions TableType; + int32_t init_subgraph_index; + CallOnceOptionsT() + : init_subgraph_index(0) { + } +}; + +struct CallOnceOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef CallOnceOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_INIT_SUBGRAPH_INDEX = 4 + }; + int32_t init_subgraph_index() const { + return GetField(VT_INIT_SUBGRAPH_INDEX, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_INIT_SUBGRAPH_INDEX) && + verifier.EndTable(); + } + CallOnceOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(CallOnceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const CallOnceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct CallOnceOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_init_subgraph_index(int32_t init_subgraph_index) { + fbb_.AddElement(CallOnceOptions::VT_INIT_SUBGRAPH_INDEX, init_subgraph_index, 0); + } + explicit CallOnceOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + CallOnceOptionsBuilder &operator=(const CallOnceOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateCallOnceOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t init_subgraph_index = 0) { + CallOnceOptionsBuilder builder_(_fbb); + builder_.add_init_subgraph_index(init_subgraph_index); + return builder_.Finish(); +} + +flatbuffers::Offset CreateCallOnceOptions(flatbuffers::FlatBufferBuilder &_fbb, const CallOnceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct WhileOptionsT : public flatbuffers::NativeTable { + typedef WhileOptions TableType; + int32_t cond_subgraph_index; + int32_t body_subgraph_index; + WhileOptionsT() + : cond_subgraph_index(0), + body_subgraph_index(0) { + } +}; + +struct WhileOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef WhileOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_COND_SUBGRAPH_INDEX = 4, + VT_BODY_SUBGRAPH_INDEX = 6 + }; + int32_t cond_subgraph_index() const { + return GetField(VT_COND_SUBGRAPH_INDEX, 0); + } + int32_t body_subgraph_index() const { + return GetField(VT_BODY_SUBGRAPH_INDEX, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_COND_SUBGRAPH_INDEX) && + VerifyField(verifier, VT_BODY_SUBGRAPH_INDEX) && + verifier.EndTable(); + } + WhileOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(WhileOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const WhileOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct WhileOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_cond_subgraph_index(int32_t cond_subgraph_index) { + fbb_.AddElement(WhileOptions::VT_COND_SUBGRAPH_INDEX, cond_subgraph_index, 0); + } + void add_body_subgraph_index(int32_t body_subgraph_index) { + fbb_.AddElement(WhileOptions::VT_BODY_SUBGRAPH_INDEX, body_subgraph_index, 0); + } + explicit WhileOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + WhileOptionsBuilder &operator=(const WhileOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateWhileOptions( + flatbuffers::FlatBufferBuilder &_fbb, + int32_t cond_subgraph_index = 0, + int32_t body_subgraph_index = 0) { + WhileOptionsBuilder builder_(_fbb); + builder_.add_body_subgraph_index(body_subgraph_index); + builder_.add_cond_subgraph_index(cond_subgraph_index); + return builder_.Finish(); +} + +flatbuffers::Offset CreateWhileOptions(flatbuffers::FlatBufferBuilder &_fbb, const WhileOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct NonMaxSuppressionV4OptionsT : public flatbuffers::NativeTable { + typedef NonMaxSuppressionV4Options TableType; + NonMaxSuppressionV4OptionsT() { + } +}; + +struct NonMaxSuppressionV4Options FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef NonMaxSuppressionV4OptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + NonMaxSuppressionV4OptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(NonMaxSuppressionV4OptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV4OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct NonMaxSuppressionV4OptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit NonMaxSuppressionV4OptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + NonMaxSuppressionV4OptionsBuilder &operator=(const NonMaxSuppressionV4OptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateNonMaxSuppressionV4Options( + flatbuffers::FlatBufferBuilder &_fbb) { + NonMaxSuppressionV4OptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateNonMaxSuppressionV4Options(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV4OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct NonMaxSuppressionV5OptionsT : public flatbuffers::NativeTable { + typedef NonMaxSuppressionV5Options TableType; + NonMaxSuppressionV5OptionsT() { + } +}; + +struct NonMaxSuppressionV5Options FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef NonMaxSuppressionV5OptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + NonMaxSuppressionV5OptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(NonMaxSuppressionV5OptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV5OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct NonMaxSuppressionV5OptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit NonMaxSuppressionV5OptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + NonMaxSuppressionV5OptionsBuilder &operator=(const NonMaxSuppressionV5OptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateNonMaxSuppressionV5Options( + flatbuffers::FlatBufferBuilder &_fbb) { + NonMaxSuppressionV5OptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateNonMaxSuppressionV5Options(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV5OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ScatterNdOptionsT : public flatbuffers::NativeTable { + typedef ScatterNdOptions TableType; + ScatterNdOptionsT() { + } +}; + +struct ScatterNdOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ScatterNdOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + ScatterNdOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ScatterNdOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ScatterNdOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ScatterNdOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit ScatterNdOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ScatterNdOptionsBuilder &operator=(const ScatterNdOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateScatterNdOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + ScatterNdOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateScatterNdOptions(flatbuffers::FlatBufferBuilder &_fbb, const ScatterNdOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SelectV2OptionsT : public flatbuffers::NativeTable { + typedef SelectV2Options TableType; + SelectV2OptionsT() { + } +}; + +struct SelectV2Options FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SelectV2OptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + SelectV2OptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SelectV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SelectV2OptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit SelectV2OptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SelectV2OptionsBuilder &operator=(const SelectV2OptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSelectV2Options( + flatbuffers::FlatBufferBuilder &_fbb) { + SelectV2OptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSelectV2Options(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct DensifyOptionsT : public flatbuffers::NativeTable { + typedef DensifyOptions TableType; + DensifyOptionsT() { + } +}; + +struct DensifyOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef DensifyOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + DensifyOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(DensifyOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct DensifyOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit DensifyOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + DensifyOptionsBuilder &operator=(const DensifyOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateDensifyOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + DensifyOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateDensifyOptions(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SegmentSumOptionsT : public flatbuffers::NativeTable { + typedef SegmentSumOptions TableType; + SegmentSumOptionsT() { + } +}; + +struct SegmentSumOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SegmentSumOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + SegmentSumOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SegmentSumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SegmentSumOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit SegmentSumOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SegmentSumOptionsBuilder &operator=(const SegmentSumOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSegmentSumOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + SegmentSumOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateSegmentSumOptions(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct BatchMatMulOptionsT : public flatbuffers::NativeTable { + typedef BatchMatMulOptions TableType; + bool adj_x; + bool adj_y; + bool asymmetric_quantize_inputs; + BatchMatMulOptionsT() + : adj_x(false), + adj_y(false), + asymmetric_quantize_inputs(false) { + } +}; + +struct BatchMatMulOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef BatchMatMulOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_ADJ_X = 4, + VT_ADJ_Y = 6, + VT_ASYMMETRIC_QUANTIZE_INPUTS = 8 + }; + bool adj_x() const { + return GetField(VT_ADJ_X, 0) != 0; + } + bool adj_y() const { + return GetField(VT_ADJ_Y, 0) != 0; + } + bool asymmetric_quantize_inputs() const { + return GetField(VT_ASYMMETRIC_QUANTIZE_INPUTS, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_ADJ_X) && + VerifyField(verifier, VT_ADJ_Y) && + VerifyField(verifier, VT_ASYMMETRIC_QUANTIZE_INPUTS) && + verifier.EndTable(); + } + BatchMatMulOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(BatchMatMulOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const BatchMatMulOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct BatchMatMulOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_adj_x(bool adj_x) { + fbb_.AddElement(BatchMatMulOptions::VT_ADJ_X, static_cast(adj_x), 0); + } + void add_adj_y(bool adj_y) { + fbb_.AddElement(BatchMatMulOptions::VT_ADJ_Y, static_cast(adj_y), 0); + } + void add_asymmetric_quantize_inputs(bool asymmetric_quantize_inputs) { + fbb_.AddElement(BatchMatMulOptions::VT_ASYMMETRIC_QUANTIZE_INPUTS, static_cast(asymmetric_quantize_inputs), 0); + } + explicit BatchMatMulOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + BatchMatMulOptionsBuilder &operator=(const BatchMatMulOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateBatchMatMulOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool adj_x = false, + bool adj_y = false, + bool asymmetric_quantize_inputs = false) { + BatchMatMulOptionsBuilder builder_(_fbb); + builder_.add_asymmetric_quantize_inputs(asymmetric_quantize_inputs); + builder_.add_adj_y(adj_y); + builder_.add_adj_x(adj_x); + return builder_.Finish(); +} + +flatbuffers::Offset CreateBatchMatMulOptions(flatbuffers::FlatBufferBuilder &_fbb, const BatchMatMulOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct CumsumOptionsT : public flatbuffers::NativeTable { + typedef CumsumOptions TableType; + bool exclusive; + bool reverse; + CumsumOptionsT() + : exclusive(false), + reverse(false) { + } +}; + +struct CumsumOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef CumsumOptionsT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_EXCLUSIVE = 4, + VT_REVERSE = 6 + }; + bool exclusive() const { + return GetField(VT_EXCLUSIVE, 0) != 0; + } + bool reverse() const { + return GetField(VT_REVERSE, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_EXCLUSIVE) && + VerifyField(verifier, VT_REVERSE) && + verifier.EndTable(); + } + CumsumOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(CumsumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const CumsumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct CumsumOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_exclusive(bool exclusive) { + fbb_.AddElement(CumsumOptions::VT_EXCLUSIVE, static_cast(exclusive), 0); + } + void add_reverse(bool reverse) { + fbb_.AddElement(CumsumOptions::VT_REVERSE, static_cast(reverse), 0); + } + explicit CumsumOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + CumsumOptionsBuilder &operator=(const CumsumOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateCumsumOptions( + flatbuffers::FlatBufferBuilder &_fbb, + bool exclusive = false, + bool reverse = false) { + CumsumOptionsBuilder builder_(_fbb); + builder_.add_reverse(reverse); + builder_.add_exclusive(exclusive); + return builder_.Finish(); +} + +flatbuffers::Offset CreateCumsumOptions(flatbuffers::FlatBufferBuilder &_fbb, const CumsumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct BroadcastToOptionsT : public flatbuffers::NativeTable { + typedef BroadcastToOptions TableType; + BroadcastToOptionsT() { + } +}; + +struct BroadcastToOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef BroadcastToOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + BroadcastToOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(BroadcastToOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const BroadcastToOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct BroadcastToOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit BroadcastToOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + BroadcastToOptionsBuilder &operator=(const BroadcastToOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateBroadcastToOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + BroadcastToOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateBroadcastToOptions(flatbuffers::FlatBufferBuilder &_fbb, const BroadcastToOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct Rfft2dOptionsT : public flatbuffers::NativeTable { + typedef Rfft2dOptions TableType; + Rfft2dOptionsT() {} +}; + +struct Rfft2dOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef Rfft2dOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && verifier.EndTable(); + } + Rfft2dOptionsT *UnPack( + const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo( + Rfft2dOptionsT *_o, + const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack( + flatbuffers::FlatBufferBuilder &_fbb, const Rfft2dOptionsT *_o, + const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct Rfft2dOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit Rfft2dOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + Rfft2dOptionsBuilder &operator=(const Rfft2dOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateRfft2dOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + Rfft2dOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreateRfft2dOptions( + flatbuffers::FlatBufferBuilder &_fbb, const Rfft2dOptionsT *_o, + const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct OperatorCodeT : public flatbuffers::NativeTable { + typedef OperatorCode TableType; + int8_t deprecated_builtin_code; + std::string custom_code; + int32_t version; + tflite::BuiltinOperator builtin_code; + OperatorCodeT() + : deprecated_builtin_code(0), + version(1), + builtin_code(tflite::BuiltinOperator_ADD) { + } +}; + +struct OperatorCode FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef OperatorCodeT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_DEPRECATED_BUILTIN_CODE = 4, + VT_CUSTOM_CODE = 6, + VT_VERSION = 8, + VT_BUILTIN_CODE = 10 + }; + int8_t deprecated_builtin_code() const { + return GetField(VT_DEPRECATED_BUILTIN_CODE, 0); + } + const flatbuffers::String *custom_code() const { + return GetPointer(VT_CUSTOM_CODE); + } + int32_t version() const { + return GetField(VT_VERSION, 1); + } + tflite::BuiltinOperator builtin_code() const { + return static_cast(GetField(VT_BUILTIN_CODE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_DEPRECATED_BUILTIN_CODE) && + VerifyOffset(verifier, VT_CUSTOM_CODE) && + verifier.VerifyString(custom_code()) && + VerifyField(verifier, VT_VERSION) && + VerifyField(verifier, VT_BUILTIN_CODE) && + verifier.EndTable(); + } + OperatorCodeT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(OperatorCodeT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const OperatorCodeT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct OperatorCodeBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_deprecated_builtin_code(int8_t deprecated_builtin_code) { + fbb_.AddElement(OperatorCode::VT_DEPRECATED_BUILTIN_CODE, deprecated_builtin_code, 0); + } + void add_custom_code(flatbuffers::Offset custom_code) { + fbb_.AddOffset(OperatorCode::VT_CUSTOM_CODE, custom_code); + } + void add_version(int32_t version) { + fbb_.AddElement(OperatorCode::VT_VERSION, version, 1); + } + void add_builtin_code(tflite::BuiltinOperator builtin_code) { + fbb_.AddElement(OperatorCode::VT_BUILTIN_CODE, static_cast(builtin_code), 0); + } + explicit OperatorCodeBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + OperatorCodeBuilder &operator=(const OperatorCodeBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateOperatorCode( + flatbuffers::FlatBufferBuilder &_fbb, + int8_t deprecated_builtin_code = 0, + flatbuffers::Offset custom_code = 0, + int32_t version = 1, + tflite::BuiltinOperator builtin_code = tflite::BuiltinOperator_ADD) { + OperatorCodeBuilder builder_(_fbb); + builder_.add_builtin_code(builtin_code); + builder_.add_version(version); + builder_.add_custom_code(custom_code); + builder_.add_deprecated_builtin_code(deprecated_builtin_code); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateOperatorCodeDirect( + flatbuffers::FlatBufferBuilder &_fbb, + int8_t deprecated_builtin_code = 0, + const char *custom_code = nullptr, + int32_t version = 1, + tflite::BuiltinOperator builtin_code = tflite::BuiltinOperator_ADD) { + auto custom_code__ = custom_code ? _fbb.CreateString(custom_code) : 0; + return tflite::CreateOperatorCode( + _fbb, + deprecated_builtin_code, + custom_code__, + version, + builtin_code); +} + +flatbuffers::Offset CreateOperatorCode(flatbuffers::FlatBufferBuilder &_fbb, const OperatorCodeT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct OperatorT : public flatbuffers::NativeTable { + typedef Operator TableType; + uint32_t opcode_index; + std::vector inputs; + std::vector outputs; + tflite::BuiltinOptionsUnion builtin_options; + std::vector custom_options; + tflite::CustomOptionsFormat custom_options_format; + std::vector mutating_variable_inputs; + std::vector intermediates; + OperatorT() + : opcode_index(0), + custom_options_format(tflite::CustomOptionsFormat_FLEXBUFFERS) { + } +}; + +struct Operator FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef OperatorT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_OPCODE_INDEX = 4, + VT_INPUTS = 6, + VT_OUTPUTS = 8, + VT_BUILTIN_OPTIONS_TYPE = 10, + VT_BUILTIN_OPTIONS = 12, + VT_CUSTOM_OPTIONS = 14, + VT_CUSTOM_OPTIONS_FORMAT = 16, + VT_MUTATING_VARIABLE_INPUTS = 18, + VT_INTERMEDIATES = 20 + }; + uint32_t opcode_index() const { + return GetField(VT_OPCODE_INDEX, 0); + } + const flatbuffers::Vector *inputs() const { + return GetPointer *>(VT_INPUTS); + } + const flatbuffers::Vector *outputs() const { + return GetPointer *>(VT_OUTPUTS); + } + tflite::BuiltinOptions builtin_options_type() const { + return static_cast(GetField(VT_BUILTIN_OPTIONS_TYPE, 0)); + } + const void *builtin_options() const { + return GetPointer(VT_BUILTIN_OPTIONS); + } + template const T *builtin_options_as() const; + const tflite::Conv2DOptions *builtin_options_as_Conv2DOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_Conv2DOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::DepthwiseConv2DOptions *builtin_options_as_DepthwiseConv2DOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_DepthwiseConv2DOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ConcatEmbeddingsOptions *builtin_options_as_ConcatEmbeddingsOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ConcatEmbeddingsOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LSHProjectionOptions *builtin_options_as_LSHProjectionOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LSHProjectionOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::Pool2DOptions *builtin_options_as_Pool2DOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_Pool2DOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SVDFOptions *builtin_options_as_SVDFOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SVDFOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::RNNOptions *builtin_options_as_RNNOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_RNNOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::FullyConnectedOptions *builtin_options_as_FullyConnectedOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_FullyConnectedOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SoftmaxOptions *builtin_options_as_SoftmaxOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SoftmaxOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ConcatenationOptions *builtin_options_as_ConcatenationOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ConcatenationOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::AddOptions *builtin_options_as_AddOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_AddOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::L2NormOptions *builtin_options_as_L2NormOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_L2NormOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LocalResponseNormalizationOptions *builtin_options_as_LocalResponseNormalizationOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LocalResponseNormalizationOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LSTMOptions *builtin_options_as_LSTMOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LSTMOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ResizeBilinearOptions *builtin_options_as_ResizeBilinearOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ResizeBilinearOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::CallOptions *builtin_options_as_CallOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_CallOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ReshapeOptions *builtin_options_as_ReshapeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ReshapeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SkipGramOptions *builtin_options_as_SkipGramOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SkipGramOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SpaceToDepthOptions *builtin_options_as_SpaceToDepthOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SpaceToDepthOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::EmbeddingLookupSparseOptions *builtin_options_as_EmbeddingLookupSparseOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_EmbeddingLookupSparseOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::MulOptions *builtin_options_as_MulOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_MulOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::PadOptions *builtin_options_as_PadOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_PadOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::GatherOptions *builtin_options_as_GatherOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_GatherOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::BatchToSpaceNDOptions *builtin_options_as_BatchToSpaceNDOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_BatchToSpaceNDOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SpaceToBatchNDOptions *builtin_options_as_SpaceToBatchNDOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SpaceToBatchNDOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::TransposeOptions *builtin_options_as_TransposeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_TransposeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ReducerOptions *builtin_options_as_ReducerOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ReducerOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SubOptions *builtin_options_as_SubOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SubOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::DivOptions *builtin_options_as_DivOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_DivOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SqueezeOptions *builtin_options_as_SqueezeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SqueezeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SequenceRNNOptions *builtin_options_as_SequenceRNNOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SequenceRNNOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::StridedSliceOptions *builtin_options_as_StridedSliceOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_StridedSliceOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ExpOptions *builtin_options_as_ExpOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ExpOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::TopKV2Options *builtin_options_as_TopKV2Options() const { + return builtin_options_type() == tflite::BuiltinOptions_TopKV2Options ? static_cast(builtin_options()) : nullptr; + } + const tflite::SplitOptions *builtin_options_as_SplitOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SplitOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LogSoftmaxOptions *builtin_options_as_LogSoftmaxOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LogSoftmaxOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::CastOptions *builtin_options_as_CastOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_CastOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::DequantizeOptions *builtin_options_as_DequantizeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_DequantizeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::MaximumMinimumOptions *builtin_options_as_MaximumMinimumOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_MaximumMinimumOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ArgMaxOptions *builtin_options_as_ArgMaxOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ArgMaxOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LessOptions *builtin_options_as_LessOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LessOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::NegOptions *builtin_options_as_NegOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_NegOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::PadV2Options *builtin_options_as_PadV2Options() const { + return builtin_options_type() == tflite::BuiltinOptions_PadV2Options ? static_cast(builtin_options()) : nullptr; + } + const tflite::GreaterOptions *builtin_options_as_GreaterOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_GreaterOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::GreaterEqualOptions *builtin_options_as_GreaterEqualOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_GreaterEqualOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LessEqualOptions *builtin_options_as_LessEqualOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LessEqualOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SelectOptions *builtin_options_as_SelectOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SelectOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SliceOptions *builtin_options_as_SliceOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SliceOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::TransposeConvOptions *builtin_options_as_TransposeConvOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_TransposeConvOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SparseToDenseOptions *builtin_options_as_SparseToDenseOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SparseToDenseOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::TileOptions *builtin_options_as_TileOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_TileOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ExpandDimsOptions *builtin_options_as_ExpandDimsOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ExpandDimsOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::EqualOptions *builtin_options_as_EqualOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_EqualOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::NotEqualOptions *builtin_options_as_NotEqualOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_NotEqualOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ShapeOptions *builtin_options_as_ShapeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ShapeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::PowOptions *builtin_options_as_PowOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_PowOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ArgMinOptions *builtin_options_as_ArgMinOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ArgMinOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::FakeQuantOptions *builtin_options_as_FakeQuantOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_FakeQuantOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::PackOptions *builtin_options_as_PackOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_PackOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LogicalOrOptions *builtin_options_as_LogicalOrOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LogicalOrOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::OneHotOptions *builtin_options_as_OneHotOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_OneHotOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LogicalAndOptions *builtin_options_as_LogicalAndOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LogicalAndOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LogicalNotOptions *builtin_options_as_LogicalNotOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LogicalNotOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::UnpackOptions *builtin_options_as_UnpackOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_UnpackOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::FloorDivOptions *builtin_options_as_FloorDivOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_FloorDivOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SquareOptions *builtin_options_as_SquareOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SquareOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ZerosLikeOptions *builtin_options_as_ZerosLikeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ZerosLikeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::FillOptions *builtin_options_as_FillOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_FillOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::BidirectionalSequenceLSTMOptions *builtin_options_as_BidirectionalSequenceLSTMOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_BidirectionalSequenceLSTMOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::BidirectionalSequenceRNNOptions *builtin_options_as_BidirectionalSequenceRNNOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_BidirectionalSequenceRNNOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::UnidirectionalSequenceLSTMOptions *builtin_options_as_UnidirectionalSequenceLSTMOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_UnidirectionalSequenceLSTMOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::FloorModOptions *builtin_options_as_FloorModOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_FloorModOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::RangeOptions *builtin_options_as_RangeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_RangeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ResizeNearestNeighborOptions *builtin_options_as_ResizeNearestNeighborOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ResizeNearestNeighborOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::LeakyReluOptions *builtin_options_as_LeakyReluOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_LeakyReluOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SquaredDifferenceOptions *builtin_options_as_SquaredDifferenceOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SquaredDifferenceOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::MirrorPadOptions *builtin_options_as_MirrorPadOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_MirrorPadOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::AbsOptions *builtin_options_as_AbsOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_AbsOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SplitVOptions *builtin_options_as_SplitVOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SplitVOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::UniqueOptions *builtin_options_as_UniqueOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_UniqueOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ReverseV2Options *builtin_options_as_ReverseV2Options() const { + return builtin_options_type() == tflite::BuiltinOptions_ReverseV2Options ? static_cast(builtin_options()) : nullptr; + } + const tflite::AddNOptions *builtin_options_as_AddNOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_AddNOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::GatherNdOptions *builtin_options_as_GatherNdOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_GatherNdOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::CosOptions *builtin_options_as_CosOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_CosOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::WhereOptions *builtin_options_as_WhereOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_WhereOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::RankOptions *builtin_options_as_RankOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_RankOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::ReverseSequenceOptions *builtin_options_as_ReverseSequenceOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ReverseSequenceOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::MatrixDiagOptions *builtin_options_as_MatrixDiagOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_MatrixDiagOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::QuantizeOptions *builtin_options_as_QuantizeOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_QuantizeOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::MatrixSetDiagOptions *builtin_options_as_MatrixSetDiagOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_MatrixSetDiagOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::HardSwishOptions *builtin_options_as_HardSwishOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_HardSwishOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::IfOptions *builtin_options_as_IfOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_IfOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::WhileOptions *builtin_options_as_WhileOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_WhileOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::DepthToSpaceOptions *builtin_options_as_DepthToSpaceOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_DepthToSpaceOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::NonMaxSuppressionV4Options *builtin_options_as_NonMaxSuppressionV4Options() const { + return builtin_options_type() == tflite::BuiltinOptions_NonMaxSuppressionV4Options ? static_cast(builtin_options()) : nullptr; + } + const tflite::NonMaxSuppressionV5Options *builtin_options_as_NonMaxSuppressionV5Options() const { + return builtin_options_type() == tflite::BuiltinOptions_NonMaxSuppressionV5Options ? static_cast(builtin_options()) : nullptr; + } + const tflite::ScatterNdOptions *builtin_options_as_ScatterNdOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_ScatterNdOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SelectV2Options *builtin_options_as_SelectV2Options() const { + return builtin_options_type() == tflite::BuiltinOptions_SelectV2Options ? static_cast(builtin_options()) : nullptr; + } + const tflite::DensifyOptions *builtin_options_as_DensifyOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_DensifyOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::SegmentSumOptions *builtin_options_as_SegmentSumOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_SegmentSumOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::BatchMatMulOptions *builtin_options_as_BatchMatMulOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_BatchMatMulOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::CumsumOptions *builtin_options_as_CumsumOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_CumsumOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::CallOnceOptions *builtin_options_as_CallOnceOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_CallOnceOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::BroadcastToOptions *builtin_options_as_BroadcastToOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_BroadcastToOptions ? static_cast(builtin_options()) : nullptr; + } + const tflite::Rfft2dOptions *builtin_options_as_Rfft2dOptions() const { + return builtin_options_type() == tflite::BuiltinOptions_Rfft2dOptions + ? static_cast(builtin_options()) + : nullptr; + } + const flatbuffers::Vector *custom_options() const { + return GetPointer *>(VT_CUSTOM_OPTIONS); + } + tflite::CustomOptionsFormat custom_options_format() const { + return static_cast(GetField(VT_CUSTOM_OPTIONS_FORMAT, 0)); + } + const flatbuffers::Vector *mutating_variable_inputs() const { + return GetPointer *>(VT_MUTATING_VARIABLE_INPUTS); + } + const flatbuffers::Vector *intermediates() const { + return GetPointer *>(VT_INTERMEDIATES); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_OPCODE_INDEX) && + VerifyOffset(verifier, VT_INPUTS) && + verifier.VerifyVector(inputs()) && + VerifyOffset(verifier, VT_OUTPUTS) && + verifier.VerifyVector(outputs()) && + VerifyField(verifier, VT_BUILTIN_OPTIONS_TYPE) && + VerifyOffset(verifier, VT_BUILTIN_OPTIONS) && + VerifyBuiltinOptions(verifier, builtin_options(), builtin_options_type()) && + VerifyOffset(verifier, VT_CUSTOM_OPTIONS) && + verifier.VerifyVector(custom_options()) && + VerifyField(verifier, VT_CUSTOM_OPTIONS_FORMAT) && + VerifyOffset(verifier, VT_MUTATING_VARIABLE_INPUTS) && + verifier.VerifyVector(mutating_variable_inputs()) && + VerifyOffset(verifier, VT_INTERMEDIATES) && + verifier.VerifyVector(intermediates()) && + verifier.EndTable(); + } + OperatorT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(OperatorT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const OperatorT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +template<> inline const tflite::Conv2DOptions *Operator::builtin_options_as() const { + return builtin_options_as_Conv2DOptions(); +} + +template<> inline const tflite::DepthwiseConv2DOptions *Operator::builtin_options_as() const { + return builtin_options_as_DepthwiseConv2DOptions(); +} + +template<> inline const tflite::ConcatEmbeddingsOptions *Operator::builtin_options_as() const { + return builtin_options_as_ConcatEmbeddingsOptions(); +} + +template<> inline const tflite::LSHProjectionOptions *Operator::builtin_options_as() const { + return builtin_options_as_LSHProjectionOptions(); +} + +template<> inline const tflite::Pool2DOptions *Operator::builtin_options_as() const { + return builtin_options_as_Pool2DOptions(); +} + +template<> inline const tflite::SVDFOptions *Operator::builtin_options_as() const { + return builtin_options_as_SVDFOptions(); +} + +template<> inline const tflite::RNNOptions *Operator::builtin_options_as() const { + return builtin_options_as_RNNOptions(); +} + +template<> inline const tflite::FullyConnectedOptions *Operator::builtin_options_as() const { + return builtin_options_as_FullyConnectedOptions(); +} + +template<> inline const tflite::SoftmaxOptions *Operator::builtin_options_as() const { + return builtin_options_as_SoftmaxOptions(); +} + +template<> inline const tflite::ConcatenationOptions *Operator::builtin_options_as() const { + return builtin_options_as_ConcatenationOptions(); +} + +template<> inline const tflite::AddOptions *Operator::builtin_options_as() const { + return builtin_options_as_AddOptions(); +} + +template<> inline const tflite::L2NormOptions *Operator::builtin_options_as() const { + return builtin_options_as_L2NormOptions(); +} + +template<> inline const tflite::LocalResponseNormalizationOptions *Operator::builtin_options_as() const { + return builtin_options_as_LocalResponseNormalizationOptions(); +} + +template<> inline const tflite::LSTMOptions *Operator::builtin_options_as() const { + return builtin_options_as_LSTMOptions(); +} + +template<> inline const tflite::ResizeBilinearOptions *Operator::builtin_options_as() const { + return builtin_options_as_ResizeBilinearOptions(); +} + +template<> inline const tflite::CallOptions *Operator::builtin_options_as() const { + return builtin_options_as_CallOptions(); +} + +template<> inline const tflite::ReshapeOptions *Operator::builtin_options_as() const { + return builtin_options_as_ReshapeOptions(); +} + +template<> inline const tflite::SkipGramOptions *Operator::builtin_options_as() const { + return builtin_options_as_SkipGramOptions(); +} + +template<> inline const tflite::SpaceToDepthOptions *Operator::builtin_options_as() const { + return builtin_options_as_SpaceToDepthOptions(); +} + +template<> inline const tflite::EmbeddingLookupSparseOptions *Operator::builtin_options_as() const { + return builtin_options_as_EmbeddingLookupSparseOptions(); +} + +template<> inline const tflite::MulOptions *Operator::builtin_options_as() const { + return builtin_options_as_MulOptions(); +} + +template<> inline const tflite::PadOptions *Operator::builtin_options_as() const { + return builtin_options_as_PadOptions(); +} + +template<> inline const tflite::GatherOptions *Operator::builtin_options_as() const { + return builtin_options_as_GatherOptions(); +} + +template<> inline const tflite::BatchToSpaceNDOptions *Operator::builtin_options_as() const { + return builtin_options_as_BatchToSpaceNDOptions(); +} + +template<> inline const tflite::SpaceToBatchNDOptions *Operator::builtin_options_as() const { + return builtin_options_as_SpaceToBatchNDOptions(); +} + +template<> inline const tflite::TransposeOptions *Operator::builtin_options_as() const { + return builtin_options_as_TransposeOptions(); +} + +template<> inline const tflite::ReducerOptions *Operator::builtin_options_as() const { + return builtin_options_as_ReducerOptions(); +} + +template<> inline const tflite::SubOptions *Operator::builtin_options_as() const { + return builtin_options_as_SubOptions(); +} + +template<> inline const tflite::DivOptions *Operator::builtin_options_as() const { + return builtin_options_as_DivOptions(); +} + +template<> inline const tflite::SqueezeOptions *Operator::builtin_options_as() const { + return builtin_options_as_SqueezeOptions(); +} + +template<> inline const tflite::SequenceRNNOptions *Operator::builtin_options_as() const { + return builtin_options_as_SequenceRNNOptions(); +} + +template<> inline const tflite::StridedSliceOptions *Operator::builtin_options_as() const { + return builtin_options_as_StridedSliceOptions(); +} + +template<> inline const tflite::ExpOptions *Operator::builtin_options_as() const { + return builtin_options_as_ExpOptions(); +} + +template<> inline const tflite::TopKV2Options *Operator::builtin_options_as() const { + return builtin_options_as_TopKV2Options(); +} + +template<> inline const tflite::SplitOptions *Operator::builtin_options_as() const { + return builtin_options_as_SplitOptions(); +} + +template<> inline const tflite::LogSoftmaxOptions *Operator::builtin_options_as() const { + return builtin_options_as_LogSoftmaxOptions(); +} + +template<> inline const tflite::CastOptions *Operator::builtin_options_as() const { + return builtin_options_as_CastOptions(); +} + +template<> inline const tflite::DequantizeOptions *Operator::builtin_options_as() const { + return builtin_options_as_DequantizeOptions(); +} + +template<> inline const tflite::MaximumMinimumOptions *Operator::builtin_options_as() const { + return builtin_options_as_MaximumMinimumOptions(); +} + +template<> inline const tflite::ArgMaxOptions *Operator::builtin_options_as() const { + return builtin_options_as_ArgMaxOptions(); +} + +template<> inline const tflite::LessOptions *Operator::builtin_options_as() const { + return builtin_options_as_LessOptions(); +} + +template<> inline const tflite::NegOptions *Operator::builtin_options_as() const { + return builtin_options_as_NegOptions(); +} + +template<> inline const tflite::PadV2Options *Operator::builtin_options_as() const { + return builtin_options_as_PadV2Options(); +} + +template<> inline const tflite::GreaterOptions *Operator::builtin_options_as() const { + return builtin_options_as_GreaterOptions(); +} + +template<> inline const tflite::GreaterEqualOptions *Operator::builtin_options_as() const { + return builtin_options_as_GreaterEqualOptions(); +} + +template<> inline const tflite::LessEqualOptions *Operator::builtin_options_as() const { + return builtin_options_as_LessEqualOptions(); +} + +template<> inline const tflite::SelectOptions *Operator::builtin_options_as() const { + return builtin_options_as_SelectOptions(); +} + +template<> inline const tflite::SliceOptions *Operator::builtin_options_as() const { + return builtin_options_as_SliceOptions(); +} + +template<> inline const tflite::TransposeConvOptions *Operator::builtin_options_as() const { + return builtin_options_as_TransposeConvOptions(); +} + +template<> inline const tflite::SparseToDenseOptions *Operator::builtin_options_as() const { + return builtin_options_as_SparseToDenseOptions(); +} + +template<> inline const tflite::TileOptions *Operator::builtin_options_as() const { + return builtin_options_as_TileOptions(); +} + +template<> inline const tflite::ExpandDimsOptions *Operator::builtin_options_as() const { + return builtin_options_as_ExpandDimsOptions(); +} + +template<> inline const tflite::EqualOptions *Operator::builtin_options_as() const { + return builtin_options_as_EqualOptions(); +} + +template<> inline const tflite::NotEqualOptions *Operator::builtin_options_as() const { + return builtin_options_as_NotEqualOptions(); +} + +template<> inline const tflite::ShapeOptions *Operator::builtin_options_as() const { + return builtin_options_as_ShapeOptions(); +} + +template<> inline const tflite::PowOptions *Operator::builtin_options_as() const { + return builtin_options_as_PowOptions(); +} + +template<> inline const tflite::ArgMinOptions *Operator::builtin_options_as() const { + return builtin_options_as_ArgMinOptions(); +} + +template<> inline const tflite::FakeQuantOptions *Operator::builtin_options_as() const { + return builtin_options_as_FakeQuantOptions(); +} + +template<> inline const tflite::PackOptions *Operator::builtin_options_as() const { + return builtin_options_as_PackOptions(); +} + +template<> inline const tflite::LogicalOrOptions *Operator::builtin_options_as() const { + return builtin_options_as_LogicalOrOptions(); +} + +template<> inline const tflite::OneHotOptions *Operator::builtin_options_as() const { + return builtin_options_as_OneHotOptions(); +} + +template<> inline const tflite::LogicalAndOptions *Operator::builtin_options_as() const { + return builtin_options_as_LogicalAndOptions(); +} + +template<> inline const tflite::LogicalNotOptions *Operator::builtin_options_as() const { + return builtin_options_as_LogicalNotOptions(); +} + +template<> inline const tflite::UnpackOptions *Operator::builtin_options_as() const { + return builtin_options_as_UnpackOptions(); +} + +template<> inline const tflite::FloorDivOptions *Operator::builtin_options_as() const { + return builtin_options_as_FloorDivOptions(); +} + +template<> inline const tflite::SquareOptions *Operator::builtin_options_as() const { + return builtin_options_as_SquareOptions(); +} + +template<> inline const tflite::ZerosLikeOptions *Operator::builtin_options_as() const { + return builtin_options_as_ZerosLikeOptions(); +} + +template<> inline const tflite::FillOptions *Operator::builtin_options_as() const { + return builtin_options_as_FillOptions(); +} + +template<> inline const tflite::BidirectionalSequenceLSTMOptions *Operator::builtin_options_as() const { + return builtin_options_as_BidirectionalSequenceLSTMOptions(); +} + +template<> inline const tflite::BidirectionalSequenceRNNOptions *Operator::builtin_options_as() const { + return builtin_options_as_BidirectionalSequenceRNNOptions(); +} + +template<> inline const tflite::UnidirectionalSequenceLSTMOptions *Operator::builtin_options_as() const { + return builtin_options_as_UnidirectionalSequenceLSTMOptions(); +} + +template<> inline const tflite::FloorModOptions *Operator::builtin_options_as() const { + return builtin_options_as_FloorModOptions(); +} + +template<> inline const tflite::RangeOptions *Operator::builtin_options_as() const { + return builtin_options_as_RangeOptions(); +} + +template<> inline const tflite::ResizeNearestNeighborOptions *Operator::builtin_options_as() const { + return builtin_options_as_ResizeNearestNeighborOptions(); +} + +template<> inline const tflite::LeakyReluOptions *Operator::builtin_options_as() const { + return builtin_options_as_LeakyReluOptions(); +} + +template<> inline const tflite::SquaredDifferenceOptions *Operator::builtin_options_as() const { + return builtin_options_as_SquaredDifferenceOptions(); +} + +template<> inline const tflite::MirrorPadOptions *Operator::builtin_options_as() const { + return builtin_options_as_MirrorPadOptions(); +} + +template<> inline const tflite::AbsOptions *Operator::builtin_options_as() const { + return builtin_options_as_AbsOptions(); +} + +template<> inline const tflite::SplitVOptions *Operator::builtin_options_as() const { + return builtin_options_as_SplitVOptions(); +} + +template<> inline const tflite::UniqueOptions *Operator::builtin_options_as() const { + return builtin_options_as_UniqueOptions(); +} + +template<> inline const tflite::ReverseV2Options *Operator::builtin_options_as() const { + return builtin_options_as_ReverseV2Options(); +} + +template<> inline const tflite::AddNOptions *Operator::builtin_options_as() const { + return builtin_options_as_AddNOptions(); +} + +template<> inline const tflite::GatherNdOptions *Operator::builtin_options_as() const { + return builtin_options_as_GatherNdOptions(); +} + +template<> inline const tflite::CosOptions *Operator::builtin_options_as() const { + return builtin_options_as_CosOptions(); +} + +template<> inline const tflite::WhereOptions *Operator::builtin_options_as() const { + return builtin_options_as_WhereOptions(); +} + +template<> inline const tflite::RankOptions *Operator::builtin_options_as() const { + return builtin_options_as_RankOptions(); +} + +template<> inline const tflite::ReverseSequenceOptions *Operator::builtin_options_as() const { + return builtin_options_as_ReverseSequenceOptions(); +} + +template<> inline const tflite::MatrixDiagOptions *Operator::builtin_options_as() const { + return builtin_options_as_MatrixDiagOptions(); +} + +template<> inline const tflite::QuantizeOptions *Operator::builtin_options_as() const { + return builtin_options_as_QuantizeOptions(); +} + +template<> inline const tflite::MatrixSetDiagOptions *Operator::builtin_options_as() const { + return builtin_options_as_MatrixSetDiagOptions(); +} + +template<> inline const tflite::HardSwishOptions *Operator::builtin_options_as() const { + return builtin_options_as_HardSwishOptions(); +} + +template<> inline const tflite::IfOptions *Operator::builtin_options_as() const { + return builtin_options_as_IfOptions(); +} + +template<> inline const tflite::WhileOptions *Operator::builtin_options_as() const { + return builtin_options_as_WhileOptions(); +} + +template<> inline const tflite::DepthToSpaceOptions *Operator::builtin_options_as() const { + return builtin_options_as_DepthToSpaceOptions(); +} + +template<> inline const tflite::NonMaxSuppressionV4Options *Operator::builtin_options_as() const { + return builtin_options_as_NonMaxSuppressionV4Options(); +} + +template<> inline const tflite::NonMaxSuppressionV5Options *Operator::builtin_options_as() const { + return builtin_options_as_NonMaxSuppressionV5Options(); +} + +template<> inline const tflite::ScatterNdOptions *Operator::builtin_options_as() const { + return builtin_options_as_ScatterNdOptions(); +} + +template<> inline const tflite::SelectV2Options *Operator::builtin_options_as() const { + return builtin_options_as_SelectV2Options(); +} + +template<> inline const tflite::DensifyOptions *Operator::builtin_options_as() const { + return builtin_options_as_DensifyOptions(); +} + +template<> inline const tflite::SegmentSumOptions *Operator::builtin_options_as() const { + return builtin_options_as_SegmentSumOptions(); +} + +template<> inline const tflite::BatchMatMulOptions *Operator::builtin_options_as() const { + return builtin_options_as_BatchMatMulOptions(); +} + +template<> inline const tflite::CumsumOptions *Operator::builtin_options_as() const { + return builtin_options_as_CumsumOptions(); +} + +template<> inline const tflite::CallOnceOptions *Operator::builtin_options_as() const { + return builtin_options_as_CallOnceOptions(); +} + +template<> inline const tflite::BroadcastToOptions *Operator::builtin_options_as() const { + return builtin_options_as_BroadcastToOptions(); +} + +template <> +inline const tflite::Rfft2dOptions * +Operator::builtin_options_as() const { + return builtin_options_as_Rfft2dOptions(); +} + +struct OperatorBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_opcode_index(uint32_t opcode_index) { + fbb_.AddElement(Operator::VT_OPCODE_INDEX, opcode_index, 0); + } + void add_inputs(flatbuffers::Offset> inputs) { + fbb_.AddOffset(Operator::VT_INPUTS, inputs); + } + void add_outputs(flatbuffers::Offset> outputs) { + fbb_.AddOffset(Operator::VT_OUTPUTS, outputs); + } + void add_builtin_options_type(tflite::BuiltinOptions builtin_options_type) { + fbb_.AddElement(Operator::VT_BUILTIN_OPTIONS_TYPE, static_cast(builtin_options_type), 0); + } + void add_builtin_options(flatbuffers::Offset builtin_options) { + fbb_.AddOffset(Operator::VT_BUILTIN_OPTIONS, builtin_options); + } + void add_custom_options(flatbuffers::Offset> custom_options) { + fbb_.AddOffset(Operator::VT_CUSTOM_OPTIONS, custom_options); + } + void add_custom_options_format(tflite::CustomOptionsFormat custom_options_format) { + fbb_.AddElement(Operator::VT_CUSTOM_OPTIONS_FORMAT, static_cast(custom_options_format), 0); + } + void add_mutating_variable_inputs(flatbuffers::Offset> mutating_variable_inputs) { + fbb_.AddOffset(Operator::VT_MUTATING_VARIABLE_INPUTS, mutating_variable_inputs); + } + void add_intermediates(flatbuffers::Offset> intermediates) { + fbb_.AddOffset(Operator::VT_INTERMEDIATES, intermediates); + } + explicit OperatorBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + OperatorBuilder &operator=(const OperatorBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateOperator( + flatbuffers::FlatBufferBuilder &_fbb, + uint32_t opcode_index = 0, + flatbuffers::Offset> inputs = 0, + flatbuffers::Offset> outputs = 0, + tflite::BuiltinOptions builtin_options_type = tflite::BuiltinOptions_NONE, + flatbuffers::Offset builtin_options = 0, + flatbuffers::Offset> custom_options = 0, + tflite::CustomOptionsFormat custom_options_format = tflite::CustomOptionsFormat_FLEXBUFFERS, + flatbuffers::Offset> mutating_variable_inputs = 0, + flatbuffers::Offset> intermediates = 0) { + OperatorBuilder builder_(_fbb); + builder_.add_intermediates(intermediates); + builder_.add_mutating_variable_inputs(mutating_variable_inputs); + builder_.add_custom_options(custom_options); + builder_.add_builtin_options(builtin_options); + builder_.add_outputs(outputs); + builder_.add_inputs(inputs); + builder_.add_opcode_index(opcode_index); + builder_.add_custom_options_format(custom_options_format); + builder_.add_builtin_options_type(builtin_options_type); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateOperatorDirect( + flatbuffers::FlatBufferBuilder &_fbb, + uint32_t opcode_index = 0, + const std::vector *inputs = nullptr, + const std::vector *outputs = nullptr, + tflite::BuiltinOptions builtin_options_type = tflite::BuiltinOptions_NONE, + flatbuffers::Offset builtin_options = 0, + const std::vector *custom_options = nullptr, + tflite::CustomOptionsFormat custom_options_format = tflite::CustomOptionsFormat_FLEXBUFFERS, + const std::vector *mutating_variable_inputs = nullptr, + const std::vector *intermediates = nullptr) { + auto inputs__ = inputs ? _fbb.CreateVector(*inputs) : 0; + auto outputs__ = outputs ? _fbb.CreateVector(*outputs) : 0; + auto custom_options__ = custom_options ? _fbb.CreateVector(*custom_options) : 0; + auto mutating_variable_inputs__ = mutating_variable_inputs ? _fbb.CreateVector(*mutating_variable_inputs) : 0; + auto intermediates__ = intermediates ? _fbb.CreateVector(*intermediates) : 0; + return tflite::CreateOperator( + _fbb, + opcode_index, + inputs__, + outputs__, + builtin_options_type, + builtin_options, + custom_options__, + custom_options_format, + mutating_variable_inputs__, + intermediates__); +} + +flatbuffers::Offset CreateOperator(flatbuffers::FlatBufferBuilder &_fbb, const OperatorT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SubGraphT : public flatbuffers::NativeTable { + typedef SubGraph TableType; + std::vector> tensors; + std::vector inputs; + std::vector outputs; + std::vector> operators; + std::string name; + SubGraphT() { + } +}; + +struct SubGraph FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SubGraphT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_TENSORS = 4, + VT_INPUTS = 6, + VT_OUTPUTS = 8, + VT_OPERATORS = 10, + VT_NAME = 12 + }; + const flatbuffers::Vector> *tensors() const { + return GetPointer> *>(VT_TENSORS); + } + const flatbuffers::Vector *inputs() const { + return GetPointer *>(VT_INPUTS); + } + const flatbuffers::Vector *outputs() const { + return GetPointer *>(VT_OUTPUTS); + } + const flatbuffers::Vector> *operators() const { + return GetPointer> *>(VT_OPERATORS); + } + const flatbuffers::String *name() const { + return GetPointer(VT_NAME); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_TENSORS) && + verifier.VerifyVector(tensors()) && + verifier.VerifyVectorOfTables(tensors()) && + VerifyOffset(verifier, VT_INPUTS) && + verifier.VerifyVector(inputs()) && + VerifyOffset(verifier, VT_OUTPUTS) && + verifier.VerifyVector(outputs()) && + VerifyOffset(verifier, VT_OPERATORS) && + verifier.VerifyVector(operators()) && + verifier.VerifyVectorOfTables(operators()) && + VerifyOffset(verifier, VT_NAME) && + verifier.VerifyString(name()) && + verifier.EndTable(); + } + SubGraphT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SubGraphT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SubGraphT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SubGraphBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_tensors(flatbuffers::Offset>> tensors) { + fbb_.AddOffset(SubGraph::VT_TENSORS, tensors); + } + void add_inputs(flatbuffers::Offset> inputs) { + fbb_.AddOffset(SubGraph::VT_INPUTS, inputs); + } + void add_outputs(flatbuffers::Offset> outputs) { + fbb_.AddOffset(SubGraph::VT_OUTPUTS, outputs); + } + void add_operators(flatbuffers::Offset>> operators) { + fbb_.AddOffset(SubGraph::VT_OPERATORS, operators); + } + void add_name(flatbuffers::Offset name) { + fbb_.AddOffset(SubGraph::VT_NAME, name); + } + explicit SubGraphBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SubGraphBuilder &operator=(const SubGraphBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSubGraph( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset>> tensors = 0, + flatbuffers::Offset> inputs = 0, + flatbuffers::Offset> outputs = 0, + flatbuffers::Offset>> operators = 0, + flatbuffers::Offset name = 0) { + SubGraphBuilder builder_(_fbb); + builder_.add_name(name); + builder_.add_operators(operators); + builder_.add_outputs(outputs); + builder_.add_inputs(inputs); + builder_.add_tensors(tensors); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateSubGraphDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector> *tensors = nullptr, + const std::vector *inputs = nullptr, + const std::vector *outputs = nullptr, + const std::vector> *operators = nullptr, + const char *name = nullptr) { + auto tensors__ = tensors ? _fbb.CreateVector>(*tensors) : 0; + auto inputs__ = inputs ? _fbb.CreateVector(*inputs) : 0; + auto outputs__ = outputs ? _fbb.CreateVector(*outputs) : 0; + auto operators__ = operators ? _fbb.CreateVector>(*operators) : 0; + auto name__ = name ? _fbb.CreateString(name) : 0; + return tflite::CreateSubGraph( + _fbb, + tensors__, + inputs__, + outputs__, + operators__, + name__); +} + +flatbuffers::Offset CreateSubGraph(flatbuffers::FlatBufferBuilder &_fbb, const SubGraphT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct BufferT : public flatbuffers::NativeTable { + typedef Buffer TableType; + std::vector data; + BufferT() { + } +}; + +struct Buffer FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef BufferT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_DATA = 4 + }; + const flatbuffers::Vector *data() const { + return GetPointer *>(VT_DATA); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_DATA) && + verifier.VerifyVector(data()) && + verifier.EndTable(); + } + BufferT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(BufferT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const BufferT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct BufferBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_data(flatbuffers::Offset> data) { + fbb_.AddOffset(Buffer::VT_DATA, data); + } + explicit BufferBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + BufferBuilder &operator=(const BufferBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateBuffer( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset> data = 0) { + BufferBuilder builder_(_fbb); + builder_.add_data(data); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateBufferDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector *data = nullptr) { + if (data) { _fbb.ForceVectorAlignment(data->size(), sizeof(uint8_t), 16); } + auto data__ = data ? _fbb.CreateVector(*data) : 0; + return tflite::CreateBuffer( + _fbb, + data__); +} + +flatbuffers::Offset CreateBuffer(flatbuffers::FlatBufferBuilder &_fbb, const BufferT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct MetadataT : public flatbuffers::NativeTable { + typedef Metadata TableType; + std::string name; + uint32_t buffer; + MetadataT() + : buffer(0) { + } +}; + +struct Metadata FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef MetadataT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NAME = 4, + VT_BUFFER = 6 + }; + const flatbuffers::String *name() const { + return GetPointer(VT_NAME); + } + uint32_t buffer() const { + return GetField(VT_BUFFER, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_NAME) && + verifier.VerifyString(name()) && + VerifyField(verifier, VT_BUFFER) && + verifier.EndTable(); + } + MetadataT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(MetadataT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const MetadataT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct MetadataBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_name(flatbuffers::Offset name) { + fbb_.AddOffset(Metadata::VT_NAME, name); + } + void add_buffer(uint32_t buffer) { + fbb_.AddElement(Metadata::VT_BUFFER, buffer, 0); + } + explicit MetadataBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + MetadataBuilder &operator=(const MetadataBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateMetadata( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset name = 0, + uint32_t buffer = 0) { + MetadataBuilder builder_(_fbb); + builder_.add_buffer(buffer); + builder_.add_name(name); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateMetadataDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const char *name = nullptr, + uint32_t buffer = 0) { + auto name__ = name ? _fbb.CreateString(name) : 0; + return tflite::CreateMetadata( + _fbb, + name__, + buffer); +} + +flatbuffers::Offset CreateMetadata(flatbuffers::FlatBufferBuilder &_fbb, const MetadataT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct TensorMapT : public flatbuffers::NativeTable { + typedef TensorMap TableType; + std::string name; + uint32_t tensor_index; + TensorMapT() + : tensor_index(0) { + } +}; + +struct TensorMap FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef TensorMapT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_NAME = 4, + VT_TENSOR_INDEX = 6 + }; + const flatbuffers::String *name() const { + return GetPointer(VT_NAME); + } + uint32_t tensor_index() const { + return GetField(VT_TENSOR_INDEX, 0); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_NAME) && + verifier.VerifyString(name()) && + VerifyField(verifier, VT_TENSOR_INDEX) && + verifier.EndTable(); + } + TensorMapT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(TensorMapT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const TensorMapT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct TensorMapBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_name(flatbuffers::Offset name) { + fbb_.AddOffset(TensorMap::VT_NAME, name); + } + void add_tensor_index(uint32_t tensor_index) { + fbb_.AddElement(TensorMap::VT_TENSOR_INDEX, tensor_index, 0); + } + explicit TensorMapBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + TensorMapBuilder &operator=(const TensorMapBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateTensorMap( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset name = 0, + uint32_t tensor_index = 0) { + TensorMapBuilder builder_(_fbb); + builder_.add_tensor_index(tensor_index); + builder_.add_name(name); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateTensorMapDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const char *name = nullptr, + uint32_t tensor_index = 0) { + auto name__ = name ? _fbb.CreateString(name) : 0; + return tflite::CreateTensorMap( + _fbb, + name__, + tensor_index); +} + +flatbuffers::Offset CreateTensorMap(flatbuffers::FlatBufferBuilder &_fbb, const TensorMapT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct SignatureDefT : public flatbuffers::NativeTable { + typedef SignatureDef TableType; + std::vector> inputs; + std::vector> outputs; + std::string method_name; + std::string key; + SignatureDefT() { + } +}; + +struct SignatureDef FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef SignatureDefT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_INPUTS = 4, + VT_OUTPUTS = 6, + VT_METHOD_NAME = 8, + VT_KEY = 10 + }; + const flatbuffers::Vector> *inputs() const { + return GetPointer> *>(VT_INPUTS); + } + const flatbuffers::Vector> *outputs() const { + return GetPointer> *>(VT_OUTPUTS); + } + const flatbuffers::String *method_name() const { + return GetPointer(VT_METHOD_NAME); + } + const flatbuffers::String *key() const { + return GetPointer(VT_KEY); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyOffset(verifier, VT_INPUTS) && + verifier.VerifyVector(inputs()) && + verifier.VerifyVectorOfTables(inputs()) && + VerifyOffset(verifier, VT_OUTPUTS) && + verifier.VerifyVector(outputs()) && + verifier.VerifyVectorOfTables(outputs()) && + VerifyOffset(verifier, VT_METHOD_NAME) && + verifier.VerifyString(method_name()) && + VerifyOffset(verifier, VT_KEY) && + verifier.VerifyString(key()) && + verifier.EndTable(); + } + SignatureDefT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(SignatureDefT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SignatureDefT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct SignatureDefBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_inputs(flatbuffers::Offset>> inputs) { + fbb_.AddOffset(SignatureDef::VT_INPUTS, inputs); + } + void add_outputs(flatbuffers::Offset>> outputs) { + fbb_.AddOffset(SignatureDef::VT_OUTPUTS, outputs); + } + void add_method_name(flatbuffers::Offset method_name) { + fbb_.AddOffset(SignatureDef::VT_METHOD_NAME, method_name); + } + void add_key(flatbuffers::Offset key) { + fbb_.AddOffset(SignatureDef::VT_KEY, key); + } + explicit SignatureDefBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + SignatureDefBuilder &operator=(const SignatureDefBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateSignatureDef( + flatbuffers::FlatBufferBuilder &_fbb, + flatbuffers::Offset>> inputs = 0, + flatbuffers::Offset>> outputs = 0, + flatbuffers::Offset method_name = 0, + flatbuffers::Offset key = 0) { + SignatureDefBuilder builder_(_fbb); + builder_.add_key(key); + builder_.add_method_name(method_name); + builder_.add_outputs(outputs); + builder_.add_inputs(inputs); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateSignatureDefDirect( + flatbuffers::FlatBufferBuilder &_fbb, + const std::vector> *inputs = nullptr, + const std::vector> *outputs = nullptr, + const char *method_name = nullptr, + const char *key = nullptr) { + auto inputs__ = inputs ? _fbb.CreateVector>(*inputs) : 0; + auto outputs__ = outputs ? _fbb.CreateVector>(*outputs) : 0; + auto method_name__ = method_name ? _fbb.CreateString(method_name) : 0; + auto key__ = key ? _fbb.CreateString(key) : 0; + return tflite::CreateSignatureDef( + _fbb, + inputs__, + outputs__, + method_name__, + key__); +} + +flatbuffers::Offset CreateSignatureDef(flatbuffers::FlatBufferBuilder &_fbb, const SignatureDefT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct ModelT : public flatbuffers::NativeTable { + typedef Model TableType; + uint32_t version; + std::vector> operator_codes; + std::vector> subgraphs; + std::string description; + std::vector> buffers; + std::vector metadata_buffer; + std::vector> metadata; + std::vector> signature_defs; + ModelT() + : version(0) { + } +}; + +struct Model FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ModelT NativeTableType; + enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { + VT_VERSION = 4, + VT_OPERATOR_CODES = 6, + VT_SUBGRAPHS = 8, + VT_DESCRIPTION = 10, + VT_BUFFERS = 12, + VT_METADATA_BUFFER = 14, + VT_METADATA = 16, + VT_SIGNATURE_DEFS = 18 + }; + uint32_t version() const { + return GetField(VT_VERSION, 0); + } + const flatbuffers::Vector> *operator_codes() const { + return GetPointer> *>(VT_OPERATOR_CODES); + } + const flatbuffers::Vector> *subgraphs() const { + return GetPointer> *>(VT_SUBGRAPHS); + } + const flatbuffers::String *description() const { + return GetPointer(VT_DESCRIPTION); + } + const flatbuffers::Vector> *buffers() const { + return GetPointer> *>(VT_BUFFERS); + } + const flatbuffers::Vector *metadata_buffer() const { + return GetPointer *>(VT_METADATA_BUFFER); + } + const flatbuffers::Vector> *metadata() const { + return GetPointer> *>(VT_METADATA); + } + const flatbuffers::Vector> *signature_defs() const { + return GetPointer> *>(VT_SIGNATURE_DEFS); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_VERSION) && + VerifyOffset(verifier, VT_OPERATOR_CODES) && + verifier.VerifyVector(operator_codes()) && + verifier.VerifyVectorOfTables(operator_codes()) && + VerifyOffset(verifier, VT_SUBGRAPHS) && + verifier.VerifyVector(subgraphs()) && + verifier.VerifyVectorOfTables(subgraphs()) && + VerifyOffset(verifier, VT_DESCRIPTION) && + verifier.VerifyString(description()) && + VerifyOffset(verifier, VT_BUFFERS) && + verifier.VerifyVector(buffers()) && + verifier.VerifyVectorOfTables(buffers()) && + VerifyOffset(verifier, VT_METADATA_BUFFER) && + verifier.VerifyVector(metadata_buffer()) && + VerifyOffset(verifier, VT_METADATA) && + verifier.VerifyVector(metadata()) && + verifier.VerifyVectorOfTables(metadata()) && + VerifyOffset(verifier, VT_SIGNATURE_DEFS) && + verifier.VerifyVector(signature_defs()) && + verifier.VerifyVectorOfTables(signature_defs()) && + verifier.EndTable(); + } + ModelT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ModelT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ModelT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ModelBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_version(uint32_t version) { + fbb_.AddElement(Model::VT_VERSION, version, 0); + } + void add_operator_codes(flatbuffers::Offset>> operator_codes) { + fbb_.AddOffset(Model::VT_OPERATOR_CODES, operator_codes); + } + void add_subgraphs(flatbuffers::Offset>> subgraphs) { + fbb_.AddOffset(Model::VT_SUBGRAPHS, subgraphs); + } + void add_description(flatbuffers::Offset description) { + fbb_.AddOffset(Model::VT_DESCRIPTION, description); + } + void add_buffers(flatbuffers::Offset>> buffers) { + fbb_.AddOffset(Model::VT_BUFFERS, buffers); + } + void add_metadata_buffer(flatbuffers::Offset> metadata_buffer) { + fbb_.AddOffset(Model::VT_METADATA_BUFFER, metadata_buffer); + } + void add_metadata(flatbuffers::Offset>> metadata) { + fbb_.AddOffset(Model::VT_METADATA, metadata); + } + void add_signature_defs(flatbuffers::Offset>> signature_defs) { + fbb_.AddOffset(Model::VT_SIGNATURE_DEFS, signature_defs); + } + explicit ModelBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ModelBuilder &operator=(const ModelBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateModel( + flatbuffers::FlatBufferBuilder &_fbb, + uint32_t version = 0, + flatbuffers::Offset>> operator_codes = 0, + flatbuffers::Offset>> subgraphs = 0, + flatbuffers::Offset description = 0, + flatbuffers::Offset>> buffers = 0, + flatbuffers::Offset> metadata_buffer = 0, + flatbuffers::Offset>> metadata = 0, + flatbuffers::Offset>> signature_defs = 0) { + ModelBuilder builder_(_fbb); + builder_.add_signature_defs(signature_defs); + builder_.add_metadata(metadata); + builder_.add_metadata_buffer(metadata_buffer); + builder_.add_buffers(buffers); + builder_.add_description(description); + builder_.add_subgraphs(subgraphs); + builder_.add_operator_codes(operator_codes); + builder_.add_version(version); + return builder_.Finish(); +} + +inline flatbuffers::Offset CreateModelDirect( + flatbuffers::FlatBufferBuilder &_fbb, + uint32_t version = 0, + const std::vector> *operator_codes = nullptr, + const std::vector> *subgraphs = nullptr, + const char *description = nullptr, + const std::vector> *buffers = nullptr, + const std::vector *metadata_buffer = nullptr, + const std::vector> *metadata = nullptr, + const std::vector> *signature_defs = nullptr) { + auto operator_codes__ = operator_codes ? _fbb.CreateVector>(*operator_codes) : 0; + auto subgraphs__ = subgraphs ? _fbb.CreateVector>(*subgraphs) : 0; + auto description__ = description ? _fbb.CreateString(description) : 0; + auto buffers__ = buffers ? _fbb.CreateVector>(*buffers) : 0; + auto metadata_buffer__ = metadata_buffer ? _fbb.CreateVector(*metadata_buffer) : 0; + auto metadata__ = metadata ? _fbb.CreateVector>(*metadata) : 0; + auto signature_defs__ = signature_defs ? _fbb.CreateVector>(*signature_defs) : 0; + return tflite::CreateModel( + _fbb, + version, + operator_codes__, + subgraphs__, + description__, + buffers__, + metadata_buffer__, + metadata__, + signature_defs__); +} + +flatbuffers::Offset CreateModel(flatbuffers::FlatBufferBuilder &_fbb, const ModelT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +inline CustomQuantizationT *CustomQuantization::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new CustomQuantizationT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void CustomQuantization::UnPackTo(CustomQuantizationT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = custom(); if (_e) { _o->custom.resize(_e->size()); std::copy(_e->begin(), _e->end(), _o->custom.begin()); } } +} + +inline flatbuffers::Offset CustomQuantization::Pack(flatbuffers::FlatBufferBuilder &_fbb, const CustomQuantizationT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateCustomQuantization(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateCustomQuantization(flatbuffers::FlatBufferBuilder &_fbb, const CustomQuantizationT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const CustomQuantizationT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + _fbb.ForceVectorAlignment(_o->custom.size(), sizeof(uint8_t), 16); + auto _custom = _o->custom.size() ? _fbb.CreateVector(_o->custom) : 0; + return tflite::CreateCustomQuantization( + _fbb, + _custom); +} + +inline QuantizationParametersT *QuantizationParameters::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new QuantizationParametersT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void QuantizationParameters::UnPackTo(QuantizationParametersT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = min(); if (_e) { _o->min.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->min[_i] = _e->Get(_i); } } } + { auto _e = max(); if (_e) { _o->max.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->max[_i] = _e->Get(_i); } } } + { auto _e = scale(); if (_e) { _o->scale.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->scale[_i] = _e->Get(_i); } } } + { auto _e = zero_point(); if (_e) { _o->zero_point.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->zero_point[_i] = _e->Get(_i); } } } + { auto _e = details_type(); _o->details.type = _e; } + { auto _e = details(); if (_e) _o->details.value = tflite::QuantizationDetailsUnion::UnPack(_e, details_type(), _resolver); } + { auto _e = quantized_dimension(); _o->quantized_dimension = _e; } +} + +inline flatbuffers::Offset QuantizationParameters::Pack(flatbuffers::FlatBufferBuilder &_fbb, const QuantizationParametersT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateQuantizationParameters(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateQuantizationParameters(flatbuffers::FlatBufferBuilder &_fbb, const QuantizationParametersT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const QuantizationParametersT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _min = _o->min.size() ? _fbb.CreateVector(_o->min) : 0; + auto _max = _o->max.size() ? _fbb.CreateVector(_o->max) : 0; + auto _scale = _o->scale.size() ? _fbb.CreateVector(_o->scale) : 0; + auto _zero_point = _o->zero_point.size() ? _fbb.CreateVector(_o->zero_point) : 0; + auto _details_type = _o->details.type; + auto _details = _o->details.Pack(_fbb); + auto _quantized_dimension = _o->quantized_dimension; + return tflite::CreateQuantizationParameters( + _fbb, + _min, + _max, + _scale, + _zero_point, + _details_type, + _details, + _quantized_dimension); +} + +inline Int32VectorT *Int32Vector::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new Int32VectorT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Int32Vector::UnPackTo(Int32VectorT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = values(); if (_e) { _o->values.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->values[_i] = _e->Get(_i); } } } +} + +inline flatbuffers::Offset Int32Vector::Pack(flatbuffers::FlatBufferBuilder &_fbb, const Int32VectorT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateInt32Vector(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateInt32Vector(flatbuffers::FlatBufferBuilder &_fbb, const Int32VectorT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const Int32VectorT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _values = _o->values.size() ? _fbb.CreateVector(_o->values) : 0; + return tflite::CreateInt32Vector( + _fbb, + _values); +} + +inline Uint16VectorT *Uint16Vector::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new Uint16VectorT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Uint16Vector::UnPackTo(Uint16VectorT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = values(); if (_e) { _o->values.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->values[_i] = _e->Get(_i); } } } +} + +inline flatbuffers::Offset Uint16Vector::Pack(flatbuffers::FlatBufferBuilder &_fbb, const Uint16VectorT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateUint16Vector(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateUint16Vector(flatbuffers::FlatBufferBuilder &_fbb, const Uint16VectorT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const Uint16VectorT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + _fbb.ForceVectorAlignment(_o->values.size(), sizeof(uint16_t), 4); + auto _values = _o->values.size() ? _fbb.CreateVector(_o->values) : 0; + return tflite::CreateUint16Vector( + _fbb, + _values); +} + +inline Uint8VectorT *Uint8Vector::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new Uint8VectorT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Uint8Vector::UnPackTo(Uint8VectorT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = values(); if (_e) { _o->values.resize(_e->size()); std::copy(_e->begin(), _e->end(), _o->values.begin()); } } +} + +inline flatbuffers::Offset Uint8Vector::Pack(flatbuffers::FlatBufferBuilder &_fbb, const Uint8VectorT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateUint8Vector(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateUint8Vector(flatbuffers::FlatBufferBuilder &_fbb, const Uint8VectorT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const Uint8VectorT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + _fbb.ForceVectorAlignment(_o->values.size(), sizeof(uint8_t), 4); + auto _values = _o->values.size() ? _fbb.CreateVector(_o->values) : 0; + return tflite::CreateUint8Vector( + _fbb, + _values); +} + +inline DimensionMetadataT *DimensionMetadata::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new DimensionMetadataT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void DimensionMetadata::UnPackTo(DimensionMetadataT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = format(); _o->format = _e; } + { auto _e = dense_size(); _o->dense_size = _e; } + { auto _e = array_segments_type(); _o->array_segments.type = _e; } + { auto _e = array_segments(); if (_e) _o->array_segments.value = tflite::SparseIndexVectorUnion::UnPack(_e, array_segments_type(), _resolver); } + { auto _e = array_indices_type(); _o->array_indices.type = _e; } + { auto _e = array_indices(); if (_e) _o->array_indices.value = tflite::SparseIndexVectorUnion::UnPack(_e, array_indices_type(), _resolver); } +} + +inline flatbuffers::Offset DimensionMetadata::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateDimensionMetadata(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateDimensionMetadata(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DimensionMetadataT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _format = _o->format; + auto _dense_size = _o->dense_size; + auto _array_segments_type = _o->array_segments.type; + auto _array_segments = _o->array_segments.Pack(_fbb); + auto _array_indices_type = _o->array_indices.type; + auto _array_indices = _o->array_indices.Pack(_fbb); + return tflite::CreateDimensionMetadata( + _fbb, + _format, + _dense_size, + _array_segments_type, + _array_segments, + _array_indices_type, + _array_indices); +} + +inline SparsityParametersT *SparsityParameters::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SparsityParametersT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SparsityParameters::UnPackTo(SparsityParametersT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = traversal_order(); if (_e) { _o->traversal_order.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->traversal_order[_i] = _e->Get(_i); } } } + { auto _e = block_map(); if (_e) { _o->block_map.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->block_map[_i] = _e->Get(_i); } } } + { auto _e = dim_metadata(); if (_e) { _o->dim_metadata.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->dim_metadata[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } +} + +inline flatbuffers::Offset SparsityParameters::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSparsityParameters(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSparsityParameters(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SparsityParametersT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _traversal_order = _o->traversal_order.size() ? _fbb.CreateVector(_o->traversal_order) : 0; + auto _block_map = _o->block_map.size() ? _fbb.CreateVector(_o->block_map) : 0; + auto _dim_metadata = _o->dim_metadata.size() ? _fbb.CreateVector> (_o->dim_metadata.size(), [](size_t i, _VectorArgs *__va) { return CreateDimensionMetadata(*__va->__fbb, __va->__o->dim_metadata[i].get(), __va->__rehasher); }, &_va ) : 0; + return tflite::CreateSparsityParameters( + _fbb, + _traversal_order, + _block_map, + _dim_metadata); +} + +inline TensorT *Tensor::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new TensorT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Tensor::UnPackTo(TensorT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = shape(); if (_e) { _o->shape.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->shape[_i] = _e->Get(_i); } } } + { auto _e = type(); _o->type = _e; } + { auto _e = buffer(); _o->buffer = _e; } + { auto _e = name(); if (_e) _o->name = _e->str(); } + { auto _e = quantization(); if (_e) _o->quantization = std::unique_ptr(_e->UnPack(_resolver)); } + { auto _e = is_variable(); _o->is_variable = _e; } + { auto _e = sparsity(); if (_e) _o->sparsity = std::unique_ptr(_e->UnPack(_resolver)); } + { auto _e = shape_signature(); if (_e) { _o->shape_signature.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->shape_signature[_i] = _e->Get(_i); } } } +} + +inline flatbuffers::Offset Tensor::Pack(flatbuffers::FlatBufferBuilder &_fbb, const TensorT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateTensor(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateTensor(flatbuffers::FlatBufferBuilder &_fbb, const TensorT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const TensorT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _shape = _o->shape.size() ? _fbb.CreateVector(_o->shape) : 0; + auto _type = _o->type; + auto _buffer = _o->buffer; + auto _name = _o->name.empty() ? 0 : _fbb.CreateString(_o->name); + auto _quantization = _o->quantization ? CreateQuantizationParameters(_fbb, _o->quantization.get(), _rehasher) : 0; + auto _is_variable = _o->is_variable; + auto _sparsity = _o->sparsity ? CreateSparsityParameters(_fbb, _o->sparsity.get(), _rehasher) : 0; + auto _shape_signature = _o->shape_signature.size() ? _fbb.CreateVector(_o->shape_signature) : 0; + return tflite::CreateTensor( + _fbb, + _shape, + _type, + _buffer, + _name, + _quantization, + _is_variable, + _sparsity, + _shape_signature); +} + +inline Conv2DOptionsT *Conv2DOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new Conv2DOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Conv2DOptions::UnPackTo(Conv2DOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = padding(); _o->padding = _e; } + { auto _e = stride_w(); _o->stride_w = _e; } + { auto _e = stride_h(); _o->stride_h = _e; } + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = dilation_w_factor(); _o->dilation_w_factor = _e; } + { auto _e = dilation_h_factor(); _o->dilation_h_factor = _e; } +} + +inline flatbuffers::Offset Conv2DOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const Conv2DOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateConv2DOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateConv2DOptions(flatbuffers::FlatBufferBuilder &_fbb, const Conv2DOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const Conv2DOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _padding = _o->padding; + auto _stride_w = _o->stride_w; + auto _stride_h = _o->stride_h; + auto _fused_activation_function = _o->fused_activation_function; + auto _dilation_w_factor = _o->dilation_w_factor; + auto _dilation_h_factor = _o->dilation_h_factor; + return tflite::CreateConv2DOptions( + _fbb, + _padding, + _stride_w, + _stride_h, + _fused_activation_function, + _dilation_w_factor, + _dilation_h_factor); +} + +inline Pool2DOptionsT *Pool2DOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new Pool2DOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Pool2DOptions::UnPackTo(Pool2DOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = padding(); _o->padding = _e; } + { auto _e = stride_w(); _o->stride_w = _e; } + { auto _e = stride_h(); _o->stride_h = _e; } + { auto _e = filter_width(); _o->filter_width = _e; } + { auto _e = filter_height(); _o->filter_height = _e; } + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } +} + +inline flatbuffers::Offset Pool2DOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const Pool2DOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreatePool2DOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreatePool2DOptions(flatbuffers::FlatBufferBuilder &_fbb, const Pool2DOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const Pool2DOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _padding = _o->padding; + auto _stride_w = _o->stride_w; + auto _stride_h = _o->stride_h; + auto _filter_width = _o->filter_width; + auto _filter_height = _o->filter_height; + auto _fused_activation_function = _o->fused_activation_function; + return tflite::CreatePool2DOptions( + _fbb, + _padding, + _stride_w, + _stride_h, + _filter_width, + _filter_height, + _fused_activation_function); +} + +inline DepthwiseConv2DOptionsT *DepthwiseConv2DOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new DepthwiseConv2DOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void DepthwiseConv2DOptions::UnPackTo(DepthwiseConv2DOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = padding(); _o->padding = _e; } + { auto _e = stride_w(); _o->stride_w = _e; } + { auto _e = stride_h(); _o->stride_h = _e; } + { auto _e = depth_multiplier(); _o->depth_multiplier = _e; } + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = dilation_w_factor(); _o->dilation_w_factor = _e; } + { auto _e = dilation_h_factor(); _o->dilation_h_factor = _e; } +} + +inline flatbuffers::Offset DepthwiseConv2DOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DepthwiseConv2DOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateDepthwiseConv2DOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateDepthwiseConv2DOptions(flatbuffers::FlatBufferBuilder &_fbb, const DepthwiseConv2DOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DepthwiseConv2DOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _padding = _o->padding; + auto _stride_w = _o->stride_w; + auto _stride_h = _o->stride_h; + auto _depth_multiplier = _o->depth_multiplier; + auto _fused_activation_function = _o->fused_activation_function; + auto _dilation_w_factor = _o->dilation_w_factor; + auto _dilation_h_factor = _o->dilation_h_factor; + return tflite::CreateDepthwiseConv2DOptions( + _fbb, + _padding, + _stride_w, + _stride_h, + _depth_multiplier, + _fused_activation_function, + _dilation_w_factor, + _dilation_h_factor); +} + +inline ConcatEmbeddingsOptionsT *ConcatEmbeddingsOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ConcatEmbeddingsOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ConcatEmbeddingsOptions::UnPackTo(ConcatEmbeddingsOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = num_channels(); _o->num_channels = _e; } + { auto _e = num_columns_per_channel(); if (_e) { _o->num_columns_per_channel.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->num_columns_per_channel[_i] = _e->Get(_i); } } } + { auto _e = embedding_dim_per_channel(); if (_e) { _o->embedding_dim_per_channel.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->embedding_dim_per_channel[_i] = _e->Get(_i); } } } +} + +inline flatbuffers::Offset ConcatEmbeddingsOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ConcatEmbeddingsOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateConcatEmbeddingsOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateConcatEmbeddingsOptions(flatbuffers::FlatBufferBuilder &_fbb, const ConcatEmbeddingsOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ConcatEmbeddingsOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _num_channels = _o->num_channels; + auto _num_columns_per_channel = _o->num_columns_per_channel.size() ? _fbb.CreateVector(_o->num_columns_per_channel) : 0; + auto _embedding_dim_per_channel = _o->embedding_dim_per_channel.size() ? _fbb.CreateVector(_o->embedding_dim_per_channel) : 0; + return tflite::CreateConcatEmbeddingsOptions( + _fbb, + _num_channels, + _num_columns_per_channel, + _embedding_dim_per_channel); +} + +inline LSHProjectionOptionsT *LSHProjectionOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LSHProjectionOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LSHProjectionOptions::UnPackTo(LSHProjectionOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = type(); _o->type = _e; } +} + +inline flatbuffers::Offset LSHProjectionOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LSHProjectionOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLSHProjectionOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLSHProjectionOptions(flatbuffers::FlatBufferBuilder &_fbb, const LSHProjectionOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LSHProjectionOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _type = _o->type; + return tflite::CreateLSHProjectionOptions( + _fbb, + _type); +} + +inline SVDFOptionsT *SVDFOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SVDFOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SVDFOptions::UnPackTo(SVDFOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = rank(); _o->rank = _e; } + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset SVDFOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SVDFOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSVDFOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSVDFOptions(flatbuffers::FlatBufferBuilder &_fbb, const SVDFOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SVDFOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _rank = _o->rank; + auto _fused_activation_function = _o->fused_activation_function; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateSVDFOptions( + _fbb, + _rank, + _fused_activation_function, + _asymmetric_quantize_inputs); +} + +inline RNNOptionsT *RNNOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new RNNOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void RNNOptions::UnPackTo(RNNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset RNNOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const RNNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateRNNOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateRNNOptions(flatbuffers::FlatBufferBuilder &_fbb, const RNNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const RNNOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateRNNOptions( + _fbb, + _fused_activation_function, + _asymmetric_quantize_inputs); +} + +inline SequenceRNNOptionsT *SequenceRNNOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SequenceRNNOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SequenceRNNOptions::UnPackTo(SequenceRNNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = time_major(); _o->time_major = _e; } + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset SequenceRNNOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SequenceRNNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSequenceRNNOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSequenceRNNOptions(flatbuffers::FlatBufferBuilder &_fbb, const SequenceRNNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SequenceRNNOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _time_major = _o->time_major; + auto _fused_activation_function = _o->fused_activation_function; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateSequenceRNNOptions( + _fbb, + _time_major, + _fused_activation_function, + _asymmetric_quantize_inputs); +} + +inline BidirectionalSequenceRNNOptionsT *BidirectionalSequenceRNNOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new BidirectionalSequenceRNNOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void BidirectionalSequenceRNNOptions::UnPackTo(BidirectionalSequenceRNNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = time_major(); _o->time_major = _e; } + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = merge_outputs(); _o->merge_outputs = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset BidirectionalSequenceRNNOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceRNNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateBidirectionalSequenceRNNOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateBidirectionalSequenceRNNOptions(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceRNNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const BidirectionalSequenceRNNOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _time_major = _o->time_major; + auto _fused_activation_function = _o->fused_activation_function; + auto _merge_outputs = _o->merge_outputs; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateBidirectionalSequenceRNNOptions( + _fbb, + _time_major, + _fused_activation_function, + _merge_outputs, + _asymmetric_quantize_inputs); +} + +inline FullyConnectedOptionsT *FullyConnectedOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new FullyConnectedOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void FullyConnectedOptions::UnPackTo(FullyConnectedOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = weights_format(); _o->weights_format = _e; } + { auto _e = keep_num_dims(); _o->keep_num_dims = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset FullyConnectedOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const FullyConnectedOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateFullyConnectedOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateFullyConnectedOptions(flatbuffers::FlatBufferBuilder &_fbb, const FullyConnectedOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const FullyConnectedOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + auto _weights_format = _o->weights_format; + auto _keep_num_dims = _o->keep_num_dims; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateFullyConnectedOptions( + _fbb, + _fused_activation_function, + _weights_format, + _keep_num_dims, + _asymmetric_quantize_inputs); +} + +inline SoftmaxOptionsT *SoftmaxOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SoftmaxOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SoftmaxOptions::UnPackTo(SoftmaxOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = beta(); _o->beta = _e; } +} + +inline flatbuffers::Offset SoftmaxOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SoftmaxOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSoftmaxOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSoftmaxOptions(flatbuffers::FlatBufferBuilder &_fbb, const SoftmaxOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SoftmaxOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _beta = _o->beta; + return tflite::CreateSoftmaxOptions( + _fbb, + _beta); +} + +inline ConcatenationOptionsT *ConcatenationOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ConcatenationOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ConcatenationOptions::UnPackTo(ConcatenationOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = axis(); _o->axis = _e; } + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } +} + +inline flatbuffers::Offset ConcatenationOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ConcatenationOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateConcatenationOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateConcatenationOptions(flatbuffers::FlatBufferBuilder &_fbb, const ConcatenationOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ConcatenationOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _axis = _o->axis; + auto _fused_activation_function = _o->fused_activation_function; + return tflite::CreateConcatenationOptions( + _fbb, + _axis, + _fused_activation_function); +} + +inline AddOptionsT *AddOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new AddOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void AddOptions::UnPackTo(AddOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = pot_scale_int16(); _o->pot_scale_int16 = _e; } +} + +inline flatbuffers::Offset AddOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const AddOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateAddOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateAddOptions(flatbuffers::FlatBufferBuilder &_fbb, const AddOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const AddOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + auto _pot_scale_int16 = _o->pot_scale_int16; + return tflite::CreateAddOptions( + _fbb, + _fused_activation_function, + _pot_scale_int16); +} + +inline MulOptionsT *MulOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new MulOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void MulOptions::UnPackTo(MulOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } +} + +inline flatbuffers::Offset MulOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const MulOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateMulOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateMulOptions(flatbuffers::FlatBufferBuilder &_fbb, const MulOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const MulOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + return tflite::CreateMulOptions( + _fbb, + _fused_activation_function); +} + +inline L2NormOptionsT *L2NormOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new L2NormOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void L2NormOptions::UnPackTo(L2NormOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } +} + +inline flatbuffers::Offset L2NormOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const L2NormOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateL2NormOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateL2NormOptions(flatbuffers::FlatBufferBuilder &_fbb, const L2NormOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const L2NormOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + return tflite::CreateL2NormOptions( + _fbb, + _fused_activation_function); +} + +inline LocalResponseNormalizationOptionsT *LocalResponseNormalizationOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LocalResponseNormalizationOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LocalResponseNormalizationOptions::UnPackTo(LocalResponseNormalizationOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = radius(); _o->radius = _e; } + { auto _e = bias(); _o->bias = _e; } + { auto _e = alpha(); _o->alpha = _e; } + { auto _e = beta(); _o->beta = _e; } +} + +inline flatbuffers::Offset LocalResponseNormalizationOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LocalResponseNormalizationOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLocalResponseNormalizationOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLocalResponseNormalizationOptions(flatbuffers::FlatBufferBuilder &_fbb, const LocalResponseNormalizationOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LocalResponseNormalizationOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _radius = _o->radius; + auto _bias = _o->bias; + auto _alpha = _o->alpha; + auto _beta = _o->beta; + return tflite::CreateLocalResponseNormalizationOptions( + _fbb, + _radius, + _bias, + _alpha, + _beta); +} + +inline LSTMOptionsT *LSTMOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LSTMOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LSTMOptions::UnPackTo(LSTMOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = cell_clip(); _o->cell_clip = _e; } + { auto _e = proj_clip(); _o->proj_clip = _e; } + { auto _e = kernel_type(); _o->kernel_type = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset LSTMOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LSTMOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLSTMOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLSTMOptions(flatbuffers::FlatBufferBuilder &_fbb, const LSTMOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LSTMOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + auto _cell_clip = _o->cell_clip; + auto _proj_clip = _o->proj_clip; + auto _kernel_type = _o->kernel_type; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateLSTMOptions( + _fbb, + _fused_activation_function, + _cell_clip, + _proj_clip, + _kernel_type, + _asymmetric_quantize_inputs); +} + +inline UnidirectionalSequenceLSTMOptionsT *UnidirectionalSequenceLSTMOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new UnidirectionalSequenceLSTMOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void UnidirectionalSequenceLSTMOptions::UnPackTo(UnidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = cell_clip(); _o->cell_clip = _e; } + { auto _e = proj_clip(); _o->proj_clip = _e; } + { auto _e = time_major(); _o->time_major = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset UnidirectionalSequenceLSTMOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const UnidirectionalSequenceLSTMOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateUnidirectionalSequenceLSTMOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateUnidirectionalSequenceLSTMOptions(flatbuffers::FlatBufferBuilder &_fbb, const UnidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const UnidirectionalSequenceLSTMOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + auto _cell_clip = _o->cell_clip; + auto _proj_clip = _o->proj_clip; + auto _time_major = _o->time_major; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateUnidirectionalSequenceLSTMOptions( + _fbb, + _fused_activation_function, + _cell_clip, + _proj_clip, + _time_major, + _asymmetric_quantize_inputs); +} + +inline BidirectionalSequenceLSTMOptionsT *BidirectionalSequenceLSTMOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new BidirectionalSequenceLSTMOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void BidirectionalSequenceLSTMOptions::UnPackTo(BidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = cell_clip(); _o->cell_clip = _e; } + { auto _e = proj_clip(); _o->proj_clip = _e; } + { auto _e = merge_outputs(); _o->merge_outputs = _e; } + { auto _e = time_major(); _o->time_major = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset BidirectionalSequenceLSTMOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceLSTMOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateBidirectionalSequenceLSTMOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateBidirectionalSequenceLSTMOptions(flatbuffers::FlatBufferBuilder &_fbb, const BidirectionalSequenceLSTMOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const BidirectionalSequenceLSTMOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + auto _cell_clip = _o->cell_clip; + auto _proj_clip = _o->proj_clip; + auto _merge_outputs = _o->merge_outputs; + auto _time_major = _o->time_major; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateBidirectionalSequenceLSTMOptions( + _fbb, + _fused_activation_function, + _cell_clip, + _proj_clip, + _merge_outputs, + _time_major, + _asymmetric_quantize_inputs); +} + +inline ResizeBilinearOptionsT *ResizeBilinearOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ResizeBilinearOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ResizeBilinearOptions::UnPackTo(ResizeBilinearOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = align_corners(); _o->align_corners = _e; } + { auto _e = half_pixel_centers(); _o->half_pixel_centers = _e; } +} + +inline flatbuffers::Offset ResizeBilinearOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ResizeBilinearOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateResizeBilinearOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateResizeBilinearOptions(flatbuffers::FlatBufferBuilder &_fbb, const ResizeBilinearOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ResizeBilinearOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _align_corners = _o->align_corners; + auto _half_pixel_centers = _o->half_pixel_centers; + return tflite::CreateResizeBilinearOptions( + _fbb, + _align_corners, + _half_pixel_centers); +} + +inline ResizeNearestNeighborOptionsT *ResizeNearestNeighborOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ResizeNearestNeighborOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ResizeNearestNeighborOptions::UnPackTo(ResizeNearestNeighborOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = align_corners(); _o->align_corners = _e; } + { auto _e = half_pixel_centers(); _o->half_pixel_centers = _e; } +} + +inline flatbuffers::Offset ResizeNearestNeighborOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ResizeNearestNeighborOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateResizeNearestNeighborOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateResizeNearestNeighborOptions(flatbuffers::FlatBufferBuilder &_fbb, const ResizeNearestNeighborOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ResizeNearestNeighborOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _align_corners = _o->align_corners; + auto _half_pixel_centers = _o->half_pixel_centers; + return tflite::CreateResizeNearestNeighborOptions( + _fbb, + _align_corners, + _half_pixel_centers); +} + +inline CallOptionsT *CallOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new CallOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void CallOptions::UnPackTo(CallOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = subgraph(); _o->subgraph = _e; } +} + +inline flatbuffers::Offset CallOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const CallOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateCallOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateCallOptions(flatbuffers::FlatBufferBuilder &_fbb, const CallOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const CallOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _subgraph = _o->subgraph; + return tflite::CreateCallOptions( + _fbb, + _subgraph); +} + +inline PadOptionsT *PadOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new PadOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void PadOptions::UnPackTo(PadOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset PadOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const PadOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreatePadOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreatePadOptions(flatbuffers::FlatBufferBuilder &_fbb, const PadOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const PadOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreatePadOptions( + _fbb); +} + +inline PadV2OptionsT *PadV2Options::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new PadV2OptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void PadV2Options::UnPackTo(PadV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset PadV2Options::Pack(flatbuffers::FlatBufferBuilder &_fbb, const PadV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreatePadV2Options(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreatePadV2Options(flatbuffers::FlatBufferBuilder &_fbb, const PadV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const PadV2OptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreatePadV2Options( + _fbb); +} + +inline ReshapeOptionsT *ReshapeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ReshapeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ReshapeOptions::UnPackTo(ReshapeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = new_shape(); if (_e) { _o->new_shape.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->new_shape[_i] = _e->Get(_i); } } } +} + +inline flatbuffers::Offset ReshapeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReshapeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateReshapeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateReshapeOptions(flatbuffers::FlatBufferBuilder &_fbb, const ReshapeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ReshapeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _new_shape = _o->new_shape.size() ? _fbb.CreateVector(_o->new_shape) : 0; + return tflite::CreateReshapeOptions( + _fbb, + _new_shape); +} + +inline SpaceToBatchNDOptionsT *SpaceToBatchNDOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SpaceToBatchNDOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SpaceToBatchNDOptions::UnPackTo(SpaceToBatchNDOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset SpaceToBatchNDOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToBatchNDOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSpaceToBatchNDOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSpaceToBatchNDOptions(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToBatchNDOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SpaceToBatchNDOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateSpaceToBatchNDOptions( + _fbb); +} + +inline BatchToSpaceNDOptionsT *BatchToSpaceNDOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new BatchToSpaceNDOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void BatchToSpaceNDOptions::UnPackTo(BatchToSpaceNDOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset BatchToSpaceNDOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const BatchToSpaceNDOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateBatchToSpaceNDOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateBatchToSpaceNDOptions(flatbuffers::FlatBufferBuilder &_fbb, const BatchToSpaceNDOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const BatchToSpaceNDOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateBatchToSpaceNDOptions( + _fbb); +} + +inline SkipGramOptionsT *SkipGramOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SkipGramOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SkipGramOptions::UnPackTo(SkipGramOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = ngram_size(); _o->ngram_size = _e; } + { auto _e = max_skip_size(); _o->max_skip_size = _e; } + { auto _e = include_all_ngrams(); _o->include_all_ngrams = _e; } +} + +inline flatbuffers::Offset SkipGramOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SkipGramOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSkipGramOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSkipGramOptions(flatbuffers::FlatBufferBuilder &_fbb, const SkipGramOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SkipGramOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _ngram_size = _o->ngram_size; + auto _max_skip_size = _o->max_skip_size; + auto _include_all_ngrams = _o->include_all_ngrams; + return tflite::CreateSkipGramOptions( + _fbb, + _ngram_size, + _max_skip_size, + _include_all_ngrams); +} + +inline SpaceToDepthOptionsT *SpaceToDepthOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SpaceToDepthOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SpaceToDepthOptions::UnPackTo(SpaceToDepthOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = block_size(); _o->block_size = _e; } +} + +inline flatbuffers::Offset SpaceToDepthOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToDepthOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSpaceToDepthOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSpaceToDepthOptions(flatbuffers::FlatBufferBuilder &_fbb, const SpaceToDepthOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SpaceToDepthOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _block_size = _o->block_size; + return tflite::CreateSpaceToDepthOptions( + _fbb, + _block_size); +} + +inline DepthToSpaceOptionsT *DepthToSpaceOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new DepthToSpaceOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void DepthToSpaceOptions::UnPackTo(DepthToSpaceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = block_size(); _o->block_size = _e; } +} + +inline flatbuffers::Offset DepthToSpaceOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DepthToSpaceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateDepthToSpaceOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateDepthToSpaceOptions(flatbuffers::FlatBufferBuilder &_fbb, const DepthToSpaceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DepthToSpaceOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _block_size = _o->block_size; + return tflite::CreateDepthToSpaceOptions( + _fbb, + _block_size); +} + +inline SubOptionsT *SubOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SubOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SubOptions::UnPackTo(SubOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } + { auto _e = pot_scale_int16(); _o->pot_scale_int16 = _e; } +} + +inline flatbuffers::Offset SubOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SubOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSubOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSubOptions(flatbuffers::FlatBufferBuilder &_fbb, const SubOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SubOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + auto _pot_scale_int16 = _o->pot_scale_int16; + return tflite::CreateSubOptions( + _fbb, + _fused_activation_function, + _pot_scale_int16); +} + +inline DivOptionsT *DivOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new DivOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void DivOptions::UnPackTo(DivOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = fused_activation_function(); _o->fused_activation_function = _e; } +} + +inline flatbuffers::Offset DivOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DivOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateDivOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateDivOptions(flatbuffers::FlatBufferBuilder &_fbb, const DivOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DivOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _fused_activation_function = _o->fused_activation_function; + return tflite::CreateDivOptions( + _fbb, + _fused_activation_function); +} + +inline TopKV2OptionsT *TopKV2Options::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new TopKV2OptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void TopKV2Options::UnPackTo(TopKV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset TopKV2Options::Pack(flatbuffers::FlatBufferBuilder &_fbb, const TopKV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateTopKV2Options(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateTopKV2Options(flatbuffers::FlatBufferBuilder &_fbb, const TopKV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const TopKV2OptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateTopKV2Options( + _fbb); +} + +inline EmbeddingLookupSparseOptionsT *EmbeddingLookupSparseOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new EmbeddingLookupSparseOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void EmbeddingLookupSparseOptions::UnPackTo(EmbeddingLookupSparseOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = combiner(); _o->combiner = _e; } +} + +inline flatbuffers::Offset EmbeddingLookupSparseOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const EmbeddingLookupSparseOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateEmbeddingLookupSparseOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateEmbeddingLookupSparseOptions(flatbuffers::FlatBufferBuilder &_fbb, const EmbeddingLookupSparseOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const EmbeddingLookupSparseOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _combiner = _o->combiner; + return tflite::CreateEmbeddingLookupSparseOptions( + _fbb, + _combiner); +} + +inline GatherOptionsT *GatherOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new GatherOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void GatherOptions::UnPackTo(GatherOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = axis(); _o->axis = _e; } +} + +inline flatbuffers::Offset GatherOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const GatherOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateGatherOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateGatherOptions(flatbuffers::FlatBufferBuilder &_fbb, const GatherOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const GatherOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _axis = _o->axis; + return tflite::CreateGatherOptions( + _fbb, + _axis); +} + +inline TransposeOptionsT *TransposeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new TransposeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void TransposeOptions::UnPackTo(TransposeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset TransposeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const TransposeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateTransposeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateTransposeOptions(flatbuffers::FlatBufferBuilder &_fbb, const TransposeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const TransposeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateTransposeOptions( + _fbb); +} + +inline ExpOptionsT *ExpOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ExpOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ExpOptions::UnPackTo(ExpOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset ExpOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ExpOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateExpOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateExpOptions(flatbuffers::FlatBufferBuilder &_fbb, const ExpOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ExpOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateExpOptions( + _fbb); +} + +inline CosOptionsT *CosOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new CosOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void CosOptions::UnPackTo(CosOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset CosOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const CosOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateCosOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateCosOptions(flatbuffers::FlatBufferBuilder &_fbb, const CosOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const CosOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateCosOptions( + _fbb); +} + +inline ReducerOptionsT *ReducerOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ReducerOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ReducerOptions::UnPackTo(ReducerOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = keep_dims(); _o->keep_dims = _e; } +} + +inline flatbuffers::Offset ReducerOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReducerOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateReducerOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateReducerOptions(flatbuffers::FlatBufferBuilder &_fbb, const ReducerOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ReducerOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _keep_dims = _o->keep_dims; + return tflite::CreateReducerOptions( + _fbb, + _keep_dims); +} + +inline SqueezeOptionsT *SqueezeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SqueezeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SqueezeOptions::UnPackTo(SqueezeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = squeeze_dims(); if (_e) { _o->squeeze_dims.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->squeeze_dims[_i] = _e->Get(_i); } } } +} + +inline flatbuffers::Offset SqueezeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SqueezeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSqueezeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSqueezeOptions(flatbuffers::FlatBufferBuilder &_fbb, const SqueezeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SqueezeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _squeeze_dims = _o->squeeze_dims.size() ? _fbb.CreateVector(_o->squeeze_dims) : 0; + return tflite::CreateSqueezeOptions( + _fbb, + _squeeze_dims); +} + +inline SplitOptionsT *SplitOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SplitOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SplitOptions::UnPackTo(SplitOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = num_splits(); _o->num_splits = _e; } +} + +inline flatbuffers::Offset SplitOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SplitOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSplitOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSplitOptions(flatbuffers::FlatBufferBuilder &_fbb, const SplitOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SplitOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _num_splits = _o->num_splits; + return tflite::CreateSplitOptions( + _fbb, + _num_splits); +} + +inline SplitVOptionsT *SplitVOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SplitVOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SplitVOptions::UnPackTo(SplitVOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = num_splits(); _o->num_splits = _e; } +} + +inline flatbuffers::Offset SplitVOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SplitVOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSplitVOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSplitVOptions(flatbuffers::FlatBufferBuilder &_fbb, const SplitVOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SplitVOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _num_splits = _o->num_splits; + return tflite::CreateSplitVOptions( + _fbb, + _num_splits); +} + +inline StridedSliceOptionsT *StridedSliceOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new StridedSliceOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void StridedSliceOptions::UnPackTo(StridedSliceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = begin_mask(); _o->begin_mask = _e; } + { auto _e = end_mask(); _o->end_mask = _e; } + { auto _e = ellipsis_mask(); _o->ellipsis_mask = _e; } + { auto _e = new_axis_mask(); _o->new_axis_mask = _e; } + { auto _e = shrink_axis_mask(); _o->shrink_axis_mask = _e; } +} + +inline flatbuffers::Offset StridedSliceOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const StridedSliceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateStridedSliceOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateStridedSliceOptions(flatbuffers::FlatBufferBuilder &_fbb, const StridedSliceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const StridedSliceOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _begin_mask = _o->begin_mask; + auto _end_mask = _o->end_mask; + auto _ellipsis_mask = _o->ellipsis_mask; + auto _new_axis_mask = _o->new_axis_mask; + auto _shrink_axis_mask = _o->shrink_axis_mask; + return tflite::CreateStridedSliceOptions( + _fbb, + _begin_mask, + _end_mask, + _ellipsis_mask, + _new_axis_mask, + _shrink_axis_mask); +} + +inline LogSoftmaxOptionsT *LogSoftmaxOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LogSoftmaxOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LogSoftmaxOptions::UnPackTo(LogSoftmaxOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset LogSoftmaxOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogSoftmaxOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLogSoftmaxOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLogSoftmaxOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogSoftmaxOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LogSoftmaxOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateLogSoftmaxOptions( + _fbb); +} + +inline CastOptionsT *CastOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new CastOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void CastOptions::UnPackTo(CastOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = in_data_type(); _o->in_data_type = _e; } + { auto _e = out_data_type(); _o->out_data_type = _e; } +} + +inline flatbuffers::Offset CastOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const CastOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateCastOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateCastOptions(flatbuffers::FlatBufferBuilder &_fbb, const CastOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const CastOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _in_data_type = _o->in_data_type; + auto _out_data_type = _o->out_data_type; + return tflite::CreateCastOptions( + _fbb, + _in_data_type, + _out_data_type); +} + +inline DequantizeOptionsT *DequantizeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new DequantizeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void DequantizeOptions::UnPackTo(DequantizeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset DequantizeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DequantizeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateDequantizeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateDequantizeOptions(flatbuffers::FlatBufferBuilder &_fbb, const DequantizeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DequantizeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateDequantizeOptions( + _fbb); +} + +inline MaximumMinimumOptionsT *MaximumMinimumOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new MaximumMinimumOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void MaximumMinimumOptions::UnPackTo(MaximumMinimumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset MaximumMinimumOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const MaximumMinimumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateMaximumMinimumOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateMaximumMinimumOptions(flatbuffers::FlatBufferBuilder &_fbb, const MaximumMinimumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const MaximumMinimumOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateMaximumMinimumOptions( + _fbb); +} + +inline TileOptionsT *TileOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new TileOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void TileOptions::UnPackTo(TileOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset TileOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const TileOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateTileOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateTileOptions(flatbuffers::FlatBufferBuilder &_fbb, const TileOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const TileOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateTileOptions( + _fbb); +} + +inline ArgMaxOptionsT *ArgMaxOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ArgMaxOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ArgMaxOptions::UnPackTo(ArgMaxOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = output_type(); _o->output_type = _e; } +} + +inline flatbuffers::Offset ArgMaxOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ArgMaxOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateArgMaxOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateArgMaxOptions(flatbuffers::FlatBufferBuilder &_fbb, const ArgMaxOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ArgMaxOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _output_type = _o->output_type; + return tflite::CreateArgMaxOptions( + _fbb, + _output_type); +} + +inline ArgMinOptionsT *ArgMinOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ArgMinOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ArgMinOptions::UnPackTo(ArgMinOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = output_type(); _o->output_type = _e; } +} + +inline flatbuffers::Offset ArgMinOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateArgMinOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateArgMinOptions(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ArgMinOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _output_type = _o->output_type; + return tflite::CreateArgMinOptions( + _fbb, + _output_type); +} + +inline GreaterOptionsT *GreaterOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new GreaterOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void GreaterOptions::UnPackTo(GreaterOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset GreaterOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const GreaterOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateGreaterOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateGreaterOptions(flatbuffers::FlatBufferBuilder &_fbb, const GreaterOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const GreaterOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateGreaterOptions( + _fbb); +} + +inline GreaterEqualOptionsT *GreaterEqualOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new GreaterEqualOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void GreaterEqualOptions::UnPackTo(GreaterEqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset GreaterEqualOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const GreaterEqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateGreaterEqualOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateGreaterEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const GreaterEqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const GreaterEqualOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateGreaterEqualOptions( + _fbb); +} + +inline LessOptionsT *LessOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LessOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LessOptions::UnPackTo(LessOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset LessOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LessOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLessOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLessOptions(flatbuffers::FlatBufferBuilder &_fbb, const LessOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LessOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateLessOptions( + _fbb); +} + +inline LessEqualOptionsT *LessEqualOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LessEqualOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LessEqualOptions::UnPackTo(LessEqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset LessEqualOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LessEqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLessEqualOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLessEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const LessEqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LessEqualOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateLessEqualOptions( + _fbb); +} + +inline NegOptionsT *NegOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new NegOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void NegOptions::UnPackTo(NegOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset NegOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const NegOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateNegOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateNegOptions(flatbuffers::FlatBufferBuilder &_fbb, const NegOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const NegOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateNegOptions( + _fbb); +} + +inline SelectOptionsT *SelectOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SelectOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SelectOptions::UnPackTo(SelectOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset SelectOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SelectOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSelectOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSelectOptions(flatbuffers::FlatBufferBuilder &_fbb, const SelectOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SelectOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateSelectOptions( + _fbb); +} + +inline SliceOptionsT *SliceOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SliceOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SliceOptions::UnPackTo(SliceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset SliceOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SliceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSliceOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSliceOptions(flatbuffers::FlatBufferBuilder &_fbb, const SliceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SliceOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateSliceOptions( + _fbb); +} + +inline TransposeConvOptionsT *TransposeConvOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new TransposeConvOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void TransposeConvOptions::UnPackTo(TransposeConvOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = padding(); _o->padding = _e; } + { auto _e = stride_w(); _o->stride_w = _e; } + { auto _e = stride_h(); _o->stride_h = _e; } +} + +inline flatbuffers::Offset TransposeConvOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const TransposeConvOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateTransposeConvOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateTransposeConvOptions(flatbuffers::FlatBufferBuilder &_fbb, const TransposeConvOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const TransposeConvOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _padding = _o->padding; + auto _stride_w = _o->stride_w; + auto _stride_h = _o->stride_h; + return tflite::CreateTransposeConvOptions( + _fbb, + _padding, + _stride_w, + _stride_h); +} + +inline ExpandDimsOptionsT *ExpandDimsOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ExpandDimsOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ExpandDimsOptions::UnPackTo(ExpandDimsOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset ExpandDimsOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ExpandDimsOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateExpandDimsOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateExpandDimsOptions(flatbuffers::FlatBufferBuilder &_fbb, const ExpandDimsOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ExpandDimsOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateExpandDimsOptions( + _fbb); +} + +inline SparseToDenseOptionsT *SparseToDenseOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SparseToDenseOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SparseToDenseOptions::UnPackTo(SparseToDenseOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = validate_indices(); _o->validate_indices = _e; } +} + +inline flatbuffers::Offset SparseToDenseOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SparseToDenseOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSparseToDenseOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSparseToDenseOptions(flatbuffers::FlatBufferBuilder &_fbb, const SparseToDenseOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SparseToDenseOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _validate_indices = _o->validate_indices; + return tflite::CreateSparseToDenseOptions( + _fbb, + _validate_indices); +} + +inline EqualOptionsT *EqualOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new EqualOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void EqualOptions::UnPackTo(EqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset EqualOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const EqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateEqualOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const EqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const EqualOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateEqualOptions( + _fbb); +} + +inline NotEqualOptionsT *NotEqualOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new NotEqualOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void NotEqualOptions::UnPackTo(NotEqualOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset NotEqualOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const NotEqualOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateNotEqualOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateNotEqualOptions(flatbuffers::FlatBufferBuilder &_fbb, const NotEqualOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const NotEqualOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateNotEqualOptions( + _fbb); +} + +inline ShapeOptionsT *ShapeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ShapeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ShapeOptions::UnPackTo(ShapeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = out_type(); _o->out_type = _e; } +} + +inline flatbuffers::Offset ShapeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ShapeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateShapeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateShapeOptions(flatbuffers::FlatBufferBuilder &_fbb, const ShapeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ShapeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _out_type = _o->out_type; + return tflite::CreateShapeOptions( + _fbb, + _out_type); +} + +inline RankOptionsT *RankOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new RankOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void RankOptions::UnPackTo(RankOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset RankOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const RankOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateRankOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateRankOptions(flatbuffers::FlatBufferBuilder &_fbb, const RankOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const RankOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateRankOptions( + _fbb); +} + +inline PowOptionsT *PowOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new PowOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void PowOptions::UnPackTo(PowOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset PowOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreatePowOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreatePowOptions(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const PowOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreatePowOptions( + _fbb); +} + +inline FakeQuantOptionsT *FakeQuantOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new FakeQuantOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void FakeQuantOptions::UnPackTo(FakeQuantOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = min(); _o->min = _e; } + { auto _e = max(); _o->max = _e; } + { auto _e = num_bits(); _o->num_bits = _e; } + { auto _e = narrow_range(); _o->narrow_range = _e; } +} + +inline flatbuffers::Offset FakeQuantOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateFakeQuantOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateFakeQuantOptions(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const FakeQuantOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _min = _o->min; + auto _max = _o->max; + auto _num_bits = _o->num_bits; + auto _narrow_range = _o->narrow_range; + return tflite::CreateFakeQuantOptions( + _fbb, + _min, + _max, + _num_bits, + _narrow_range); +} + +inline PackOptionsT *PackOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new PackOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void PackOptions::UnPackTo(PackOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = values_count(); _o->values_count = _e; } + { auto _e = axis(); _o->axis = _e; } +} + +inline flatbuffers::Offset PackOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const PackOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreatePackOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreatePackOptions(flatbuffers::FlatBufferBuilder &_fbb, const PackOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const PackOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _values_count = _o->values_count; + auto _axis = _o->axis; + return tflite::CreatePackOptions( + _fbb, + _values_count, + _axis); +} + +inline LogicalOrOptionsT *LogicalOrOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LogicalOrOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LogicalOrOptions::UnPackTo(LogicalOrOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset LogicalOrOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogicalOrOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLogicalOrOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLogicalOrOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogicalOrOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LogicalOrOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateLogicalOrOptions( + _fbb); +} + +inline OneHotOptionsT *OneHotOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new OneHotOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void OneHotOptions::UnPackTo(OneHotOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = axis(); _o->axis = _e; } +} + +inline flatbuffers::Offset OneHotOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const OneHotOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateOneHotOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateOneHotOptions(flatbuffers::FlatBufferBuilder &_fbb, const OneHotOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const OneHotOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _axis = _o->axis; + return tflite::CreateOneHotOptions( + _fbb, + _axis); +} + +inline AbsOptionsT *AbsOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new AbsOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void AbsOptions::UnPackTo(AbsOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset AbsOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const AbsOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateAbsOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateAbsOptions(flatbuffers::FlatBufferBuilder &_fbb, const AbsOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const AbsOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateAbsOptions( + _fbb); +} + +inline HardSwishOptionsT *HardSwishOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new HardSwishOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void HardSwishOptions::UnPackTo(HardSwishOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset HardSwishOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const HardSwishOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateHardSwishOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateHardSwishOptions(flatbuffers::FlatBufferBuilder &_fbb, const HardSwishOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const HardSwishOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateHardSwishOptions( + _fbb); +} + +inline LogicalAndOptionsT *LogicalAndOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LogicalAndOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LogicalAndOptions::UnPackTo(LogicalAndOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset LogicalAndOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogicalAndOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLogicalAndOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLogicalAndOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogicalAndOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LogicalAndOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateLogicalAndOptions( + _fbb); +} + +inline LogicalNotOptionsT *LogicalNotOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LogicalNotOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LogicalNotOptions::UnPackTo(LogicalNotOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset LogicalNotOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LogicalNotOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLogicalNotOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLogicalNotOptions(flatbuffers::FlatBufferBuilder &_fbb, const LogicalNotOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LogicalNotOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateLogicalNotOptions( + _fbb); +} + +inline UnpackOptionsT *UnpackOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new UnpackOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void UnpackOptions::UnPackTo(UnpackOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = num(); _o->num = _e; } + { auto _e = axis(); _o->axis = _e; } +} + +inline flatbuffers::Offset UnpackOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const UnpackOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateUnpackOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateUnpackOptions(flatbuffers::FlatBufferBuilder &_fbb, const UnpackOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const UnpackOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _num = _o->num; + auto _axis = _o->axis; + return tflite::CreateUnpackOptions( + _fbb, + _num, + _axis); +} + +inline FloorDivOptionsT *FloorDivOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new FloorDivOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void FloorDivOptions::UnPackTo(FloorDivOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset FloorDivOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const FloorDivOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateFloorDivOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateFloorDivOptions(flatbuffers::FlatBufferBuilder &_fbb, const FloorDivOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const FloorDivOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateFloorDivOptions( + _fbb); +} + +inline SquareOptionsT *SquareOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SquareOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SquareOptions::UnPackTo(SquareOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset SquareOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SquareOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSquareOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSquareOptions(flatbuffers::FlatBufferBuilder &_fbb, const SquareOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SquareOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateSquareOptions( + _fbb); +} + +inline ZerosLikeOptionsT *ZerosLikeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ZerosLikeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ZerosLikeOptions::UnPackTo(ZerosLikeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset ZerosLikeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ZerosLikeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateZerosLikeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateZerosLikeOptions(flatbuffers::FlatBufferBuilder &_fbb, const ZerosLikeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ZerosLikeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateZerosLikeOptions( + _fbb); +} + +inline FillOptionsT *FillOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new FillOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void FillOptions::UnPackTo(FillOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset FillOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const FillOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateFillOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateFillOptions(flatbuffers::FlatBufferBuilder &_fbb, const FillOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const FillOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateFillOptions( + _fbb); +} + +inline FloorModOptionsT *FloorModOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new FloorModOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void FloorModOptions::UnPackTo(FloorModOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset FloorModOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const FloorModOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateFloorModOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateFloorModOptions(flatbuffers::FlatBufferBuilder &_fbb, const FloorModOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const FloorModOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateFloorModOptions( + _fbb); +} + +inline RangeOptionsT *RangeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new RangeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void RangeOptions::UnPackTo(RangeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset RangeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const RangeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateRangeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateRangeOptions(flatbuffers::FlatBufferBuilder &_fbb, const RangeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const RangeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateRangeOptions( + _fbb); +} + +inline LeakyReluOptionsT *LeakyReluOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new LeakyReluOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void LeakyReluOptions::UnPackTo(LeakyReluOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = alpha(); _o->alpha = _e; } +} + +inline flatbuffers::Offset LeakyReluOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const LeakyReluOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateLeakyReluOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateLeakyReluOptions(flatbuffers::FlatBufferBuilder &_fbb, const LeakyReluOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const LeakyReluOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _alpha = _o->alpha; + return tflite::CreateLeakyReluOptions( + _fbb, + _alpha); +} + +inline SquaredDifferenceOptionsT *SquaredDifferenceOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SquaredDifferenceOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SquaredDifferenceOptions::UnPackTo(SquaredDifferenceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset SquaredDifferenceOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SquaredDifferenceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSquaredDifferenceOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSquaredDifferenceOptions(flatbuffers::FlatBufferBuilder &_fbb, const SquaredDifferenceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SquaredDifferenceOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateSquaredDifferenceOptions( + _fbb); +} + +inline MirrorPadOptionsT *MirrorPadOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new MirrorPadOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void MirrorPadOptions::UnPackTo(MirrorPadOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = mode(); _o->mode = _e; } +} + +inline flatbuffers::Offset MirrorPadOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const MirrorPadOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateMirrorPadOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateMirrorPadOptions(flatbuffers::FlatBufferBuilder &_fbb, const MirrorPadOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const MirrorPadOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _mode = _o->mode; + return tflite::CreateMirrorPadOptions( + _fbb, + _mode); +} + +inline UniqueOptionsT *UniqueOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new UniqueOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void UniqueOptions::UnPackTo(UniqueOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = idx_out_type(); _o->idx_out_type = _e; } +} + +inline flatbuffers::Offset UniqueOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const UniqueOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateUniqueOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateUniqueOptions(flatbuffers::FlatBufferBuilder &_fbb, const UniqueOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const UniqueOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _idx_out_type = _o->idx_out_type; + return tflite::CreateUniqueOptions( + _fbb, + _idx_out_type); +} + +inline ReverseV2OptionsT *ReverseV2Options::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ReverseV2OptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ReverseV2Options::UnPackTo(ReverseV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset ReverseV2Options::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReverseV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateReverseV2Options(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateReverseV2Options(flatbuffers::FlatBufferBuilder &_fbb, const ReverseV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ReverseV2OptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateReverseV2Options( + _fbb); +} + +inline AddNOptionsT *AddNOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new AddNOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void AddNOptions::UnPackTo(AddNOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset AddNOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const AddNOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateAddNOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateAddNOptions(flatbuffers::FlatBufferBuilder &_fbb, const AddNOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const AddNOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateAddNOptions( + _fbb); +} + +inline GatherNdOptionsT *GatherNdOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new GatherNdOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void GatherNdOptions::UnPackTo(GatherNdOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset GatherNdOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const GatherNdOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateGatherNdOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateGatherNdOptions(flatbuffers::FlatBufferBuilder &_fbb, const GatherNdOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const GatherNdOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateGatherNdOptions( + _fbb); +} + +inline WhereOptionsT *WhereOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new WhereOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void WhereOptions::UnPackTo(WhereOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset WhereOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const WhereOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateWhereOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateWhereOptions(flatbuffers::FlatBufferBuilder &_fbb, const WhereOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const WhereOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateWhereOptions( + _fbb); +} + +inline ReverseSequenceOptionsT *ReverseSequenceOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ReverseSequenceOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ReverseSequenceOptions::UnPackTo(ReverseSequenceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = seq_dim(); _o->seq_dim = _e; } + { auto _e = batch_dim(); _o->batch_dim = _e; } +} + +inline flatbuffers::Offset ReverseSequenceOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ReverseSequenceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateReverseSequenceOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateReverseSequenceOptions(flatbuffers::FlatBufferBuilder &_fbb, const ReverseSequenceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ReverseSequenceOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _seq_dim = _o->seq_dim; + auto _batch_dim = _o->batch_dim; + return tflite::CreateReverseSequenceOptions( + _fbb, + _seq_dim, + _batch_dim); +} + +inline MatrixDiagOptionsT *MatrixDiagOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new MatrixDiagOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void MatrixDiagOptions::UnPackTo(MatrixDiagOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset MatrixDiagOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const MatrixDiagOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateMatrixDiagOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateMatrixDiagOptions(flatbuffers::FlatBufferBuilder &_fbb, const MatrixDiagOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const MatrixDiagOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateMatrixDiagOptions( + _fbb); +} + +inline QuantizeOptionsT *QuantizeOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new QuantizeOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void QuantizeOptions::UnPackTo(QuantizeOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset QuantizeOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const QuantizeOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateQuantizeOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateQuantizeOptions(flatbuffers::FlatBufferBuilder &_fbb, const QuantizeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const QuantizeOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateQuantizeOptions( + _fbb); +} + +inline MatrixSetDiagOptionsT *MatrixSetDiagOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new MatrixSetDiagOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void MatrixSetDiagOptions::UnPackTo(MatrixSetDiagOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset MatrixSetDiagOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const MatrixSetDiagOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateMatrixSetDiagOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateMatrixSetDiagOptions(flatbuffers::FlatBufferBuilder &_fbb, const MatrixSetDiagOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const MatrixSetDiagOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateMatrixSetDiagOptions( + _fbb); +} + +inline IfOptionsT *IfOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new IfOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void IfOptions::UnPackTo(IfOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = then_subgraph_index(); _o->then_subgraph_index = _e; } + { auto _e = else_subgraph_index(); _o->else_subgraph_index = _e; } +} + +inline flatbuffers::Offset IfOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const IfOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateIfOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateIfOptions(flatbuffers::FlatBufferBuilder &_fbb, const IfOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const IfOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _then_subgraph_index = _o->then_subgraph_index; + auto _else_subgraph_index = _o->else_subgraph_index; + return tflite::CreateIfOptions( + _fbb, + _then_subgraph_index, + _else_subgraph_index); +} + +inline CallOnceOptionsT *CallOnceOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new CallOnceOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void CallOnceOptions::UnPackTo(CallOnceOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = init_subgraph_index(); _o->init_subgraph_index = _e; } +} + +inline flatbuffers::Offset CallOnceOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const CallOnceOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateCallOnceOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateCallOnceOptions(flatbuffers::FlatBufferBuilder &_fbb, const CallOnceOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const CallOnceOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _init_subgraph_index = _o->init_subgraph_index; + return tflite::CreateCallOnceOptions( + _fbb, + _init_subgraph_index); +} + +inline WhileOptionsT *WhileOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new WhileOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void WhileOptions::UnPackTo(WhileOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = cond_subgraph_index(); _o->cond_subgraph_index = _e; } + { auto _e = body_subgraph_index(); _o->body_subgraph_index = _e; } +} + +inline flatbuffers::Offset WhileOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const WhileOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateWhileOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateWhileOptions(flatbuffers::FlatBufferBuilder &_fbb, const WhileOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const WhileOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _cond_subgraph_index = _o->cond_subgraph_index; + auto _body_subgraph_index = _o->body_subgraph_index; + return tflite::CreateWhileOptions( + _fbb, + _cond_subgraph_index, + _body_subgraph_index); +} + +inline NonMaxSuppressionV4OptionsT *NonMaxSuppressionV4Options::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new NonMaxSuppressionV4OptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void NonMaxSuppressionV4Options::UnPackTo(NonMaxSuppressionV4OptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset NonMaxSuppressionV4Options::Pack(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV4OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateNonMaxSuppressionV4Options(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateNonMaxSuppressionV4Options(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV4OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const NonMaxSuppressionV4OptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateNonMaxSuppressionV4Options( + _fbb); +} + +inline NonMaxSuppressionV5OptionsT *NonMaxSuppressionV5Options::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new NonMaxSuppressionV5OptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void NonMaxSuppressionV5Options::UnPackTo(NonMaxSuppressionV5OptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset NonMaxSuppressionV5Options::Pack(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV5OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateNonMaxSuppressionV5Options(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateNonMaxSuppressionV5Options(flatbuffers::FlatBufferBuilder &_fbb, const NonMaxSuppressionV5OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const NonMaxSuppressionV5OptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateNonMaxSuppressionV5Options( + _fbb); +} + +inline ScatterNdOptionsT *ScatterNdOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ScatterNdOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ScatterNdOptions::UnPackTo(ScatterNdOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset ScatterNdOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ScatterNdOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateScatterNdOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateScatterNdOptions(flatbuffers::FlatBufferBuilder &_fbb, const ScatterNdOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ScatterNdOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateScatterNdOptions( + _fbb); +} + +inline SelectV2OptionsT *SelectV2Options::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SelectV2OptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SelectV2Options::UnPackTo(SelectV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset SelectV2Options::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSelectV2Options(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSelectV2Options(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SelectV2OptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateSelectV2Options( + _fbb); +} + +inline DensifyOptionsT *DensifyOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new DensifyOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void DensifyOptions::UnPackTo(DensifyOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset DensifyOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateDensifyOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateDensifyOptions(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DensifyOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateDensifyOptions( + _fbb); +} + +inline SegmentSumOptionsT *SegmentSumOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SegmentSumOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SegmentSumOptions::UnPackTo(SegmentSumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset SegmentSumOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSegmentSumOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSegmentSumOptions(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SegmentSumOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateSegmentSumOptions( + _fbb); +} + +inline BatchMatMulOptionsT *BatchMatMulOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new BatchMatMulOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void BatchMatMulOptions::UnPackTo(BatchMatMulOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = adj_x(); _o->adj_x = _e; } + { auto _e = adj_y(); _o->adj_y = _e; } + { auto _e = asymmetric_quantize_inputs(); _o->asymmetric_quantize_inputs = _e; } +} + +inline flatbuffers::Offset BatchMatMulOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const BatchMatMulOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateBatchMatMulOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateBatchMatMulOptions(flatbuffers::FlatBufferBuilder &_fbb, const BatchMatMulOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const BatchMatMulOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _adj_x = _o->adj_x; + auto _adj_y = _o->adj_y; + auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs; + return tflite::CreateBatchMatMulOptions( + _fbb, + _adj_x, + _adj_y, + _asymmetric_quantize_inputs); +} + +inline CumsumOptionsT *CumsumOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new CumsumOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void CumsumOptions::UnPackTo(CumsumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = exclusive(); _o->exclusive = _e; } + { auto _e = reverse(); _o->reverse = _e; } +} + +inline flatbuffers::Offset CumsumOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const CumsumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateCumsumOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateCumsumOptions(flatbuffers::FlatBufferBuilder &_fbb, const CumsumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const CumsumOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _exclusive = _o->exclusive; + auto _reverse = _o->reverse; + return tflite::CreateCumsumOptions( + _fbb, + _exclusive, + _reverse); +} + +inline BroadcastToOptionsT *BroadcastToOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new BroadcastToOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void BroadcastToOptions::UnPackTo(BroadcastToOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset BroadcastToOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const BroadcastToOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateBroadcastToOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateBroadcastToOptions(flatbuffers::FlatBufferBuilder &_fbb, const BroadcastToOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const BroadcastToOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreateBroadcastToOptions( + _fbb); +} + +inline Rfft2dOptionsT *Rfft2dOptions::UnPack( + const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new Rfft2dOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Rfft2dOptions::UnPackTo( + Rfft2dOptionsT *_o, + const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset Rfft2dOptions::Pack( + flatbuffers::FlatBufferBuilder &_fbb, const Rfft2dOptionsT *_o, + const flatbuffers::rehasher_function_t *_rehasher) { + return CreateRfft2dOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateRfft2dOptions( + flatbuffers::FlatBufferBuilder &_fbb, const Rfft2dOptionsT *_o, + const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { + flatbuffers::FlatBufferBuilder *__fbb; + const Rfft2dOptionsT *__o; + const flatbuffers::rehasher_function_t *__rehasher; + } _va = {&_fbb, _o, _rehasher}; + (void)_va; + return tflite::CreateRfft2dOptions(_fbb); +} + +inline OperatorCodeT *OperatorCode::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new OperatorCodeT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void OperatorCode::UnPackTo(OperatorCodeT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = deprecated_builtin_code(); _o->deprecated_builtin_code = _e; } + { auto _e = custom_code(); if (_e) _o->custom_code = _e->str(); } + { auto _e = version(); _o->version = _e; } + { auto _e = builtin_code(); _o->builtin_code = _e; } +} + +inline flatbuffers::Offset OperatorCode::Pack(flatbuffers::FlatBufferBuilder &_fbb, const OperatorCodeT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateOperatorCode(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateOperatorCode(flatbuffers::FlatBufferBuilder &_fbb, const OperatorCodeT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const OperatorCodeT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _deprecated_builtin_code = _o->deprecated_builtin_code; + auto _custom_code = _o->custom_code.empty() ? 0 : _fbb.CreateString(_o->custom_code); + auto _version = _o->version; + auto _builtin_code = _o->builtin_code; + return tflite::CreateOperatorCode( + _fbb, + _deprecated_builtin_code, + _custom_code, + _version, + _builtin_code); +} + +inline OperatorT *Operator::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new OperatorT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Operator::UnPackTo(OperatorT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = opcode_index(); _o->opcode_index = _e; } + { auto _e = inputs(); if (_e) { _o->inputs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->inputs[_i] = _e->Get(_i); } } } + { auto _e = outputs(); if (_e) { _o->outputs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->outputs[_i] = _e->Get(_i); } } } + { auto _e = builtin_options_type(); _o->builtin_options.type = _e; } + { auto _e = builtin_options(); if (_e) _o->builtin_options.value = tflite::BuiltinOptionsUnion::UnPack(_e, builtin_options_type(), _resolver); } + { auto _e = custom_options(); if (_e) { _o->custom_options.resize(_e->size()); std::copy(_e->begin(), _e->end(), _o->custom_options.begin()); } } + { auto _e = custom_options_format(); _o->custom_options_format = _e; } + { auto _e = mutating_variable_inputs(); if (_e) { _o->mutating_variable_inputs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->mutating_variable_inputs[_i] = _e->Get(_i) != 0; } } } + { auto _e = intermediates(); if (_e) { _o->intermediates.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->intermediates[_i] = _e->Get(_i); } } } +} + +inline flatbuffers::Offset Operator::Pack(flatbuffers::FlatBufferBuilder &_fbb, const OperatorT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateOperator(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateOperator(flatbuffers::FlatBufferBuilder &_fbb, const OperatorT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const OperatorT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _opcode_index = _o->opcode_index; + auto _inputs = _o->inputs.size() ? _fbb.CreateVector(_o->inputs) : 0; + auto _outputs = _o->outputs.size() ? _fbb.CreateVector(_o->outputs) : 0; + auto _builtin_options_type = _o->builtin_options.type; + auto _builtin_options = _o->builtin_options.Pack(_fbb); + auto _custom_options = _o->custom_options.size() ? _fbb.CreateVector(_o->custom_options) : 0; + auto _custom_options_format = _o->custom_options_format; + auto _mutating_variable_inputs = _o->mutating_variable_inputs.size() ? _fbb.CreateVector(_o->mutating_variable_inputs) : 0; + auto _intermediates = _o->intermediates.size() ? _fbb.CreateVector(_o->intermediates) : 0; + return tflite::CreateOperator( + _fbb, + _opcode_index, + _inputs, + _outputs, + _builtin_options_type, + _builtin_options, + _custom_options, + _custom_options_format, + _mutating_variable_inputs, + _intermediates); +} + +inline SubGraphT *SubGraph::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SubGraphT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SubGraph::UnPackTo(SubGraphT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = tensors(); if (_e) { _o->tensors.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->tensors[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = inputs(); if (_e) { _o->inputs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->inputs[_i] = _e->Get(_i); } } } + { auto _e = outputs(); if (_e) { _o->outputs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->outputs[_i] = _e->Get(_i); } } } + { auto _e = operators(); if (_e) { _o->operators.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->operators[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = name(); if (_e) _o->name = _e->str(); } +} + +inline flatbuffers::Offset SubGraph::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SubGraphT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSubGraph(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSubGraph(flatbuffers::FlatBufferBuilder &_fbb, const SubGraphT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SubGraphT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _tensors = _o->tensors.size() ? _fbb.CreateVector> (_o->tensors.size(), [](size_t i, _VectorArgs *__va) { return CreateTensor(*__va->__fbb, __va->__o->tensors[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _inputs = _o->inputs.size() ? _fbb.CreateVector(_o->inputs) : 0; + auto _outputs = _o->outputs.size() ? _fbb.CreateVector(_o->outputs) : 0; + auto _operators = _o->operators.size() ? _fbb.CreateVector> (_o->operators.size(), [](size_t i, _VectorArgs *__va) { return CreateOperator(*__va->__fbb, __va->__o->operators[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _name = _o->name.empty() ? 0 : _fbb.CreateString(_o->name); + return tflite::CreateSubGraph( + _fbb, + _tensors, + _inputs, + _outputs, + _operators, + _name); +} + +inline BufferT *Buffer::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new BufferT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Buffer::UnPackTo(BufferT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = data(); if (_e) { _o->data.resize(_e->size()); std::copy(_e->begin(), _e->end(), _o->data.begin()); } } +} + +inline flatbuffers::Offset Buffer::Pack(flatbuffers::FlatBufferBuilder &_fbb, const BufferT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateBuffer(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateBuffer(flatbuffers::FlatBufferBuilder &_fbb, const BufferT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const BufferT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + _fbb.ForceVectorAlignment(_o->data.size(), sizeof(uint8_t), 16); + auto _data = _o->data.size() ? _fbb.CreateVector(_o->data) : 0; + return tflite::CreateBuffer( + _fbb, + _data); +} + +inline MetadataT *Metadata::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new MetadataT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Metadata::UnPackTo(MetadataT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = name(); if (_e) _o->name = _e->str(); } + { auto _e = buffer(); _o->buffer = _e; } +} + +inline flatbuffers::Offset Metadata::Pack(flatbuffers::FlatBufferBuilder &_fbb, const MetadataT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateMetadata(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateMetadata(flatbuffers::FlatBufferBuilder &_fbb, const MetadataT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const MetadataT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _name = _o->name.empty() ? 0 : _fbb.CreateString(_o->name); + auto _buffer = _o->buffer; + return tflite::CreateMetadata( + _fbb, + _name, + _buffer); +} + +inline TensorMapT *TensorMap::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new TensorMapT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void TensorMap::UnPackTo(TensorMapT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = name(); if (_e) _o->name = _e->str(); } + { auto _e = tensor_index(); _o->tensor_index = _e; } +} + +inline flatbuffers::Offset TensorMap::Pack(flatbuffers::FlatBufferBuilder &_fbb, const TensorMapT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateTensorMap(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateTensorMap(flatbuffers::FlatBufferBuilder &_fbb, const TensorMapT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const TensorMapT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _name = _o->name.empty() ? 0 : _fbb.CreateString(_o->name); + auto _tensor_index = _o->tensor_index; + return tflite::CreateTensorMap( + _fbb, + _name, + _tensor_index); +} + +inline SignatureDefT *SignatureDef::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new SignatureDefT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void SignatureDef::UnPackTo(SignatureDefT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = inputs(); if (_e) { _o->inputs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->inputs[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = outputs(); if (_e) { _o->outputs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->outputs[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = method_name(); if (_e) _o->method_name = _e->str(); } + { auto _e = key(); if (_e) _o->key = _e->str(); } +} + +inline flatbuffers::Offset SignatureDef::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SignatureDefT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateSignatureDef(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateSignatureDef(flatbuffers::FlatBufferBuilder &_fbb, const SignatureDefT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SignatureDefT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _inputs = _o->inputs.size() ? _fbb.CreateVector> (_o->inputs.size(), [](size_t i, _VectorArgs *__va) { return CreateTensorMap(*__va->__fbb, __va->__o->inputs[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _outputs = _o->outputs.size() ? _fbb.CreateVector> (_o->outputs.size(), [](size_t i, _VectorArgs *__va) { return CreateTensorMap(*__va->__fbb, __va->__o->outputs[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _method_name = _o->method_name.empty() ? 0 : _fbb.CreateString(_o->method_name); + auto _key = _o->key.empty() ? 0 : _fbb.CreateString(_o->key); + return tflite::CreateSignatureDef( + _fbb, + _inputs, + _outputs, + _method_name, + _key); +} + +inline ModelT *Model::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ModelT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void Model::UnPackTo(ModelT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = version(); _o->version = _e; } + { auto _e = operator_codes(); if (_e) { _o->operator_codes.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->operator_codes[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = subgraphs(); if (_e) { _o->subgraphs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->subgraphs[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = description(); if (_e) _o->description = _e->str(); } + { auto _e = buffers(); if (_e) { _o->buffers.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->buffers[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = metadata_buffer(); if (_e) { _o->metadata_buffer.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->metadata_buffer[_i] = _e->Get(_i); } } } + { auto _e = metadata(); if (_e) { _o->metadata.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->metadata[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } + { auto _e = signature_defs(); if (_e) { _o->signature_defs.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->signature_defs[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } } +} + +inline flatbuffers::Offset Model::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ModelT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateModel(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateModel(flatbuffers::FlatBufferBuilder &_fbb, const ModelT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ModelT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _version = _o->version; + auto _operator_codes = _o->operator_codes.size() ? _fbb.CreateVector> (_o->operator_codes.size(), [](size_t i, _VectorArgs *__va) { return CreateOperatorCode(*__va->__fbb, __va->__o->operator_codes[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _subgraphs = _o->subgraphs.size() ? _fbb.CreateVector> (_o->subgraphs.size(), [](size_t i, _VectorArgs *__va) { return CreateSubGraph(*__va->__fbb, __va->__o->subgraphs[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _description = _o->description.empty() ? 0 : _fbb.CreateString(_o->description); + auto _buffers = _o->buffers.size() ? _fbb.CreateVector> (_o->buffers.size(), [](size_t i, _VectorArgs *__va) { return CreateBuffer(*__va->__fbb, __va->__o->buffers[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _metadata_buffer = _o->metadata_buffer.size() ? _fbb.CreateVector(_o->metadata_buffer) : 0; + auto _metadata = _o->metadata.size() ? _fbb.CreateVector> (_o->metadata.size(), [](size_t i, _VectorArgs *__va) { return CreateMetadata(*__va->__fbb, __va->__o->metadata[i].get(), __va->__rehasher); }, &_va ) : 0; + auto _signature_defs = _o->signature_defs.size() ? _fbb.CreateVector> (_o->signature_defs.size(), [](size_t i, _VectorArgs *__va) { return CreateSignatureDef(*__va->__fbb, __va->__o->signature_defs[i].get(), __va->__rehasher); }, &_va ) : 0; + return tflite::CreateModel( + _fbb, + _version, + _operator_codes, + _subgraphs, + _description, + _buffers, + _metadata_buffer, + _metadata, + _signature_defs); +} + +inline bool VerifyQuantizationDetails(flatbuffers::Verifier &verifier, const void *obj, QuantizationDetails type) { + switch (type) { + case QuantizationDetails_NONE: { + return true; + } + case QuantizationDetails_CustomQuantization: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + default: return true; + } +} + +inline bool VerifyQuantizationDetailsVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector> *values, const flatbuffers::Vector *types) { + if (!values || !types) return !values && !types; + if (values->size() != types->size()) return false; + for (flatbuffers::uoffset_t i = 0; i < values->size(); ++i) { + if (!VerifyQuantizationDetails( + verifier, values->Get(i), types->GetEnum(i))) { + return false; + } + } + return true; +} + +inline void *QuantizationDetailsUnion::UnPack(const void *obj, QuantizationDetails type, const flatbuffers::resolver_function_t *resolver) { + switch (type) { + case QuantizationDetails_CustomQuantization: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + default: return nullptr; + } +} + +inline flatbuffers::Offset QuantizationDetailsUnion::Pack(flatbuffers::FlatBufferBuilder &_fbb, const flatbuffers::rehasher_function_t *_rehasher) const { + switch (type) { + case QuantizationDetails_CustomQuantization: { + auto ptr = reinterpret_cast(value); + return CreateCustomQuantization(_fbb, ptr, _rehasher).Union(); + } + default: return 0; + } +} + +inline QuantizationDetailsUnion::QuantizationDetailsUnion(const QuantizationDetailsUnion &u) FLATBUFFERS_NOEXCEPT : type(u.type), value(nullptr) { + switch (type) { + case QuantizationDetails_CustomQuantization: { + value = new tflite::CustomQuantizationT(*reinterpret_cast(u.value)); + break; + } + default: + break; + } +} + +inline void QuantizationDetailsUnion::Reset() { + switch (type) { + case QuantizationDetails_CustomQuantization: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + default: break; + } + value = nullptr; + type = QuantizationDetails_NONE; +} + +inline bool VerifySparseIndexVector(flatbuffers::Verifier &verifier, const void *obj, SparseIndexVector type) { + switch (type) { + case SparseIndexVector_NONE: { + return true; + } + case SparseIndexVector_Int32Vector: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case SparseIndexVector_Uint16Vector: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case SparseIndexVector_Uint8Vector: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + default: return true; + } +} + +inline bool VerifySparseIndexVectorVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector> *values, const flatbuffers::Vector *types) { + if (!values || !types) return !values && !types; + if (values->size() != types->size()) return false; + for (flatbuffers::uoffset_t i = 0; i < values->size(); ++i) { + if (!VerifySparseIndexVector( + verifier, values->Get(i), types->GetEnum(i))) { + return false; + } + } + return true; +} + +inline void *SparseIndexVectorUnion::UnPack(const void *obj, SparseIndexVector type, const flatbuffers::resolver_function_t *resolver) { + switch (type) { + case SparseIndexVector_Int32Vector: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case SparseIndexVector_Uint16Vector: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case SparseIndexVector_Uint8Vector: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + default: return nullptr; + } +} + +inline flatbuffers::Offset SparseIndexVectorUnion::Pack(flatbuffers::FlatBufferBuilder &_fbb, const flatbuffers::rehasher_function_t *_rehasher) const { + switch (type) { + case SparseIndexVector_Int32Vector: { + auto ptr = reinterpret_cast(value); + return CreateInt32Vector(_fbb, ptr, _rehasher).Union(); + } + case SparseIndexVector_Uint16Vector: { + auto ptr = reinterpret_cast(value); + return CreateUint16Vector(_fbb, ptr, _rehasher).Union(); + } + case SparseIndexVector_Uint8Vector: { + auto ptr = reinterpret_cast(value); + return CreateUint8Vector(_fbb, ptr, _rehasher).Union(); + } + default: return 0; + } +} + +inline SparseIndexVectorUnion::SparseIndexVectorUnion(const SparseIndexVectorUnion &u) FLATBUFFERS_NOEXCEPT : type(u.type), value(nullptr) { + switch (type) { + case SparseIndexVector_Int32Vector: { + value = new tflite::Int32VectorT(*reinterpret_cast(u.value)); + break; + } + case SparseIndexVector_Uint16Vector: { + value = new tflite::Uint16VectorT(*reinterpret_cast(u.value)); + break; + } + case SparseIndexVector_Uint8Vector: { + value = new tflite::Uint8VectorT(*reinterpret_cast(u.value)); + break; + } + default: + break; + } +} + +inline void SparseIndexVectorUnion::Reset() { + switch (type) { + case SparseIndexVector_Int32Vector: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case SparseIndexVector_Uint16Vector: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case SparseIndexVector_Uint8Vector: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + default: break; + } + value = nullptr; + type = SparseIndexVector_NONE; +} + +inline bool VerifyBuiltinOptions(flatbuffers::Verifier &verifier, const void *obj, BuiltinOptions type) { + switch (type) { + case BuiltinOptions_NONE: { + return true; + } + case BuiltinOptions_Conv2DOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_DepthwiseConv2DOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ConcatEmbeddingsOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LSHProjectionOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_Pool2DOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SVDFOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_RNNOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_FullyConnectedOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SoftmaxOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ConcatenationOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_AddOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_L2NormOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LocalResponseNormalizationOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LSTMOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ResizeBilinearOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_CallOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ReshapeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SkipGramOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SpaceToDepthOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_EmbeddingLookupSparseOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_MulOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_PadOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_GatherOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_BatchToSpaceNDOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SpaceToBatchNDOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_TransposeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ReducerOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SubOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_DivOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SqueezeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SequenceRNNOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_StridedSliceOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ExpOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_TopKV2Options: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SplitOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LogSoftmaxOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_CastOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_DequantizeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_MaximumMinimumOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ArgMaxOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LessOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_NegOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_PadV2Options: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_GreaterOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_GreaterEqualOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LessEqualOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SelectOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SliceOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_TransposeConvOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SparseToDenseOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_TileOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ExpandDimsOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_EqualOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_NotEqualOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ShapeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_PackOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LogicalOrOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_OneHotOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LogicalAndOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LogicalNotOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_UnpackOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_FloorDivOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SquareOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ZerosLikeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_FillOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_BidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_BidirectionalSequenceRNNOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_UnidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_FloorModOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_RangeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ResizeNearestNeighborOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_LeakyReluOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SquaredDifferenceOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_MirrorPadOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_AbsOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SplitVOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_UniqueOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ReverseV2Options: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_AddNOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_GatherNdOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_CosOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_WhereOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_RankOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ReverseSequenceOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_MatrixDiagOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_QuantizeOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_MatrixSetDiagOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_HardSwishOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_IfOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_WhileOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_DepthToSpaceOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_NonMaxSuppressionV4Options: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_NonMaxSuppressionV5Options: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ScatterNdOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SelectV2Options: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_DensifyOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_SegmentSumOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_BatchMatMulOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_CumsumOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_CallOnceOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_BroadcastToOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_Rfft2dOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + default: return true; + } +} + +inline bool VerifyBuiltinOptionsVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector> *values, const flatbuffers::Vector *types) { + if (!values || !types) return !values && !types; + if (values->size() != types->size()) return false; + for (flatbuffers::uoffset_t i = 0; i < values->size(); ++i) { + if (!VerifyBuiltinOptions( + verifier, values->Get(i), types->GetEnum(i))) { + return false; + } + } + return true; +} + +inline void *BuiltinOptionsUnion::UnPack(const void *obj, BuiltinOptions type, const flatbuffers::resolver_function_t *resolver) { + switch (type) { + case BuiltinOptions_Conv2DOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_DepthwiseConv2DOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ConcatEmbeddingsOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LSHProjectionOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_Pool2DOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SVDFOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_RNNOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_FullyConnectedOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SoftmaxOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ConcatenationOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_AddOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_L2NormOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LocalResponseNormalizationOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LSTMOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ResizeBilinearOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_CallOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ReshapeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SkipGramOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SpaceToDepthOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_EmbeddingLookupSparseOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_MulOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_PadOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_GatherOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_BatchToSpaceNDOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SpaceToBatchNDOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_TransposeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ReducerOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SubOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_DivOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SqueezeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SequenceRNNOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_StridedSliceOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ExpOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_TopKV2Options: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SplitOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LogSoftmaxOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_CastOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_DequantizeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_MaximumMinimumOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ArgMaxOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LessOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_NegOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_PadV2Options: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_GreaterOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_GreaterEqualOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LessEqualOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SelectOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SliceOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_TransposeConvOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SparseToDenseOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_TileOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ExpandDimsOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_EqualOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_NotEqualOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ShapeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_PackOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LogicalOrOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_OneHotOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LogicalAndOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LogicalNotOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_UnpackOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_FloorDivOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SquareOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ZerosLikeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_FillOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_BidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_BidirectionalSequenceRNNOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_UnidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_FloorModOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_RangeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ResizeNearestNeighborOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_LeakyReluOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SquaredDifferenceOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_MirrorPadOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_AbsOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SplitVOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_UniqueOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ReverseV2Options: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_AddNOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_GatherNdOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_CosOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_WhereOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_RankOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ReverseSequenceOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_MatrixDiagOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_QuantizeOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_MatrixSetDiagOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_HardSwishOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_IfOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_WhileOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_DepthToSpaceOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_NonMaxSuppressionV4Options: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_NonMaxSuppressionV5Options: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ScatterNdOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SelectV2Options: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_DensifyOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_SegmentSumOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_BatchMatMulOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_CumsumOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_CallOnceOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_BroadcastToOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_Rfft2dOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + default: return nullptr; + } +} + +inline flatbuffers::Offset BuiltinOptionsUnion::Pack(flatbuffers::FlatBufferBuilder &_fbb, const flatbuffers::rehasher_function_t *_rehasher) const { + switch (type) { + case BuiltinOptions_Conv2DOptions: { + auto ptr = reinterpret_cast(value); + return CreateConv2DOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_DepthwiseConv2DOptions: { + auto ptr = reinterpret_cast(value); + return CreateDepthwiseConv2DOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ConcatEmbeddingsOptions: { + auto ptr = reinterpret_cast(value); + return CreateConcatEmbeddingsOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LSHProjectionOptions: { + auto ptr = reinterpret_cast(value); + return CreateLSHProjectionOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_Pool2DOptions: { + auto ptr = reinterpret_cast(value); + return CreatePool2DOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SVDFOptions: { + auto ptr = reinterpret_cast(value); + return CreateSVDFOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_RNNOptions: { + auto ptr = reinterpret_cast(value); + return CreateRNNOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_FullyConnectedOptions: { + auto ptr = reinterpret_cast(value); + return CreateFullyConnectedOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SoftmaxOptions: { + auto ptr = reinterpret_cast(value); + return CreateSoftmaxOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ConcatenationOptions: { + auto ptr = reinterpret_cast(value); + return CreateConcatenationOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_AddOptions: { + auto ptr = reinterpret_cast(value); + return CreateAddOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_L2NormOptions: { + auto ptr = reinterpret_cast(value); + return CreateL2NormOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LocalResponseNormalizationOptions: { + auto ptr = reinterpret_cast(value); + return CreateLocalResponseNormalizationOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LSTMOptions: { + auto ptr = reinterpret_cast(value); + return CreateLSTMOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ResizeBilinearOptions: { + auto ptr = reinterpret_cast(value); + return CreateResizeBilinearOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_CallOptions: { + auto ptr = reinterpret_cast(value); + return CreateCallOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ReshapeOptions: { + auto ptr = reinterpret_cast(value); + return CreateReshapeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SkipGramOptions: { + auto ptr = reinterpret_cast(value); + return CreateSkipGramOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SpaceToDepthOptions: { + auto ptr = reinterpret_cast(value); + return CreateSpaceToDepthOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_EmbeddingLookupSparseOptions: { + auto ptr = reinterpret_cast(value); + return CreateEmbeddingLookupSparseOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_MulOptions: { + auto ptr = reinterpret_cast(value); + return CreateMulOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_PadOptions: { + auto ptr = reinterpret_cast(value); + return CreatePadOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_GatherOptions: { + auto ptr = reinterpret_cast(value); + return CreateGatherOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_BatchToSpaceNDOptions: { + auto ptr = reinterpret_cast(value); + return CreateBatchToSpaceNDOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SpaceToBatchNDOptions: { + auto ptr = reinterpret_cast(value); + return CreateSpaceToBatchNDOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_TransposeOptions: { + auto ptr = reinterpret_cast(value); + return CreateTransposeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ReducerOptions: { + auto ptr = reinterpret_cast(value); + return CreateReducerOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SubOptions: { + auto ptr = reinterpret_cast(value); + return CreateSubOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_DivOptions: { + auto ptr = reinterpret_cast(value); + return CreateDivOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SqueezeOptions: { + auto ptr = reinterpret_cast(value); + return CreateSqueezeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SequenceRNNOptions: { + auto ptr = reinterpret_cast(value); + return CreateSequenceRNNOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_StridedSliceOptions: { + auto ptr = reinterpret_cast(value); + return CreateStridedSliceOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ExpOptions: { + auto ptr = reinterpret_cast(value); + return CreateExpOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_TopKV2Options: { + auto ptr = reinterpret_cast(value); + return CreateTopKV2Options(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SplitOptions: { + auto ptr = reinterpret_cast(value); + return CreateSplitOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LogSoftmaxOptions: { + auto ptr = reinterpret_cast(value); + return CreateLogSoftmaxOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_CastOptions: { + auto ptr = reinterpret_cast(value); + return CreateCastOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_DequantizeOptions: { + auto ptr = reinterpret_cast(value); + return CreateDequantizeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_MaximumMinimumOptions: { + auto ptr = reinterpret_cast(value); + return CreateMaximumMinimumOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ArgMaxOptions: { + auto ptr = reinterpret_cast(value); + return CreateArgMaxOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LessOptions: { + auto ptr = reinterpret_cast(value); + return CreateLessOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_NegOptions: { + auto ptr = reinterpret_cast(value); + return CreateNegOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_PadV2Options: { + auto ptr = reinterpret_cast(value); + return CreatePadV2Options(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_GreaterOptions: { + auto ptr = reinterpret_cast(value); + return CreateGreaterOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_GreaterEqualOptions: { + auto ptr = reinterpret_cast(value); + return CreateGreaterEqualOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LessEqualOptions: { + auto ptr = reinterpret_cast(value); + return CreateLessEqualOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SelectOptions: { + auto ptr = reinterpret_cast(value); + return CreateSelectOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SliceOptions: { + auto ptr = reinterpret_cast(value); + return CreateSliceOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_TransposeConvOptions: { + auto ptr = reinterpret_cast(value); + return CreateTransposeConvOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SparseToDenseOptions: { + auto ptr = reinterpret_cast(value); + return CreateSparseToDenseOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_TileOptions: { + auto ptr = reinterpret_cast(value); + return CreateTileOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ExpandDimsOptions: { + auto ptr = reinterpret_cast(value); + return CreateExpandDimsOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_EqualOptions: { + auto ptr = reinterpret_cast(value); + return CreateEqualOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_NotEqualOptions: { + auto ptr = reinterpret_cast(value); + return CreateNotEqualOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ShapeOptions: { + auto ptr = reinterpret_cast(value); + return CreateShapeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(value); + return CreatePowOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(value); + return CreateArgMinOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(value); + return CreateFakeQuantOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_PackOptions: { + auto ptr = reinterpret_cast(value); + return CreatePackOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LogicalOrOptions: { + auto ptr = reinterpret_cast(value); + return CreateLogicalOrOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_OneHotOptions: { + auto ptr = reinterpret_cast(value); + return CreateOneHotOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LogicalAndOptions: { + auto ptr = reinterpret_cast(value); + return CreateLogicalAndOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LogicalNotOptions: { + auto ptr = reinterpret_cast(value); + return CreateLogicalNotOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_UnpackOptions: { + auto ptr = reinterpret_cast(value); + return CreateUnpackOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_FloorDivOptions: { + auto ptr = reinterpret_cast(value); + return CreateFloorDivOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SquareOptions: { + auto ptr = reinterpret_cast(value); + return CreateSquareOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ZerosLikeOptions: { + auto ptr = reinterpret_cast(value); + return CreateZerosLikeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_FillOptions: { + auto ptr = reinterpret_cast(value); + return CreateFillOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_BidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(value); + return CreateBidirectionalSequenceLSTMOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_BidirectionalSequenceRNNOptions: { + auto ptr = reinterpret_cast(value); + return CreateBidirectionalSequenceRNNOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_UnidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(value); + return CreateUnidirectionalSequenceLSTMOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_FloorModOptions: { + auto ptr = reinterpret_cast(value); + return CreateFloorModOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_RangeOptions: { + auto ptr = reinterpret_cast(value); + return CreateRangeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ResizeNearestNeighborOptions: { + auto ptr = reinterpret_cast(value); + return CreateResizeNearestNeighborOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_LeakyReluOptions: { + auto ptr = reinterpret_cast(value); + return CreateLeakyReluOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SquaredDifferenceOptions: { + auto ptr = reinterpret_cast(value); + return CreateSquaredDifferenceOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_MirrorPadOptions: { + auto ptr = reinterpret_cast(value); + return CreateMirrorPadOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_AbsOptions: { + auto ptr = reinterpret_cast(value); + return CreateAbsOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SplitVOptions: { + auto ptr = reinterpret_cast(value); + return CreateSplitVOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_UniqueOptions: { + auto ptr = reinterpret_cast(value); + return CreateUniqueOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ReverseV2Options: { + auto ptr = reinterpret_cast(value); + return CreateReverseV2Options(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_AddNOptions: { + auto ptr = reinterpret_cast(value); + return CreateAddNOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_GatherNdOptions: { + auto ptr = reinterpret_cast(value); + return CreateGatherNdOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_CosOptions: { + auto ptr = reinterpret_cast(value); + return CreateCosOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_WhereOptions: { + auto ptr = reinterpret_cast(value); + return CreateWhereOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_RankOptions: { + auto ptr = reinterpret_cast(value); + return CreateRankOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ReverseSequenceOptions: { + auto ptr = reinterpret_cast(value); + return CreateReverseSequenceOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_MatrixDiagOptions: { + auto ptr = reinterpret_cast(value); + return CreateMatrixDiagOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_QuantizeOptions: { + auto ptr = reinterpret_cast(value); + return CreateQuantizeOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_MatrixSetDiagOptions: { + auto ptr = reinterpret_cast(value); + return CreateMatrixSetDiagOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_HardSwishOptions: { + auto ptr = reinterpret_cast(value); + return CreateHardSwishOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_IfOptions: { + auto ptr = reinterpret_cast(value); + return CreateIfOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_WhileOptions: { + auto ptr = reinterpret_cast(value); + return CreateWhileOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_DepthToSpaceOptions: { + auto ptr = reinterpret_cast(value); + return CreateDepthToSpaceOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_NonMaxSuppressionV4Options: { + auto ptr = reinterpret_cast(value); + return CreateNonMaxSuppressionV4Options(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_NonMaxSuppressionV5Options: { + auto ptr = reinterpret_cast(value); + return CreateNonMaxSuppressionV5Options(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ScatterNdOptions: { + auto ptr = reinterpret_cast(value); + return CreateScatterNdOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SelectV2Options: { + auto ptr = reinterpret_cast(value); + return CreateSelectV2Options(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_DensifyOptions: { + auto ptr = reinterpret_cast(value); + return CreateDensifyOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_SegmentSumOptions: { + auto ptr = reinterpret_cast(value); + return CreateSegmentSumOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_BatchMatMulOptions: { + auto ptr = reinterpret_cast(value); + return CreateBatchMatMulOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_CumsumOptions: { + auto ptr = reinterpret_cast(value); + return CreateCumsumOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_CallOnceOptions: { + auto ptr = reinterpret_cast(value); + return CreateCallOnceOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_BroadcastToOptions: { + auto ptr = reinterpret_cast(value); + return CreateBroadcastToOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_Rfft2dOptions: { + auto ptr = reinterpret_cast(value); + return CreateRfft2dOptions(_fbb, ptr, _rehasher).Union(); + } + default: return 0; + } +} + +inline BuiltinOptionsUnion::BuiltinOptionsUnion(const BuiltinOptionsUnion &u) FLATBUFFERS_NOEXCEPT : type(u.type), value(nullptr) { + switch (type) { + case BuiltinOptions_Conv2DOptions: { + value = new tflite::Conv2DOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_DepthwiseConv2DOptions: { + value = new tflite::DepthwiseConv2DOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ConcatEmbeddingsOptions: { + value = new tflite::ConcatEmbeddingsOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LSHProjectionOptions: { + value = new tflite::LSHProjectionOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_Pool2DOptions: { + value = new tflite::Pool2DOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SVDFOptions: { + value = new tflite::SVDFOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_RNNOptions: { + value = new tflite::RNNOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_FullyConnectedOptions: { + value = new tflite::FullyConnectedOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SoftmaxOptions: { + value = new tflite::SoftmaxOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ConcatenationOptions: { + value = new tflite::ConcatenationOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_AddOptions: { + value = new tflite::AddOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_L2NormOptions: { + value = new tflite::L2NormOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LocalResponseNormalizationOptions: { + value = new tflite::LocalResponseNormalizationOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LSTMOptions: { + value = new tflite::LSTMOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ResizeBilinearOptions: { + value = new tflite::ResizeBilinearOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_CallOptions: { + value = new tflite::CallOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ReshapeOptions: { + value = new tflite::ReshapeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SkipGramOptions: { + value = new tflite::SkipGramOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SpaceToDepthOptions: { + value = new tflite::SpaceToDepthOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_EmbeddingLookupSparseOptions: { + value = new tflite::EmbeddingLookupSparseOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_MulOptions: { + value = new tflite::MulOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_PadOptions: { + value = new tflite::PadOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_GatherOptions: { + value = new tflite::GatherOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_BatchToSpaceNDOptions: { + value = new tflite::BatchToSpaceNDOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SpaceToBatchNDOptions: { + value = new tflite::SpaceToBatchNDOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_TransposeOptions: { + value = new tflite::TransposeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ReducerOptions: { + value = new tflite::ReducerOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SubOptions: { + value = new tflite::SubOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_DivOptions: { + value = new tflite::DivOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SqueezeOptions: { + value = new tflite::SqueezeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SequenceRNNOptions: { + value = new tflite::SequenceRNNOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_StridedSliceOptions: { + value = new tflite::StridedSliceOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ExpOptions: { + value = new tflite::ExpOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_TopKV2Options: { + value = new tflite::TopKV2OptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SplitOptions: { + value = new tflite::SplitOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LogSoftmaxOptions: { + value = new tflite::LogSoftmaxOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_CastOptions: { + value = new tflite::CastOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_DequantizeOptions: { + value = new tflite::DequantizeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_MaximumMinimumOptions: { + value = new tflite::MaximumMinimumOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ArgMaxOptions: { + value = new tflite::ArgMaxOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LessOptions: { + value = new tflite::LessOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_NegOptions: { + value = new tflite::NegOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_PadV2Options: { + value = new tflite::PadV2OptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_GreaterOptions: { + value = new tflite::GreaterOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_GreaterEqualOptions: { + value = new tflite::GreaterEqualOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LessEqualOptions: { + value = new tflite::LessEqualOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SelectOptions: { + value = new tflite::SelectOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SliceOptions: { + value = new tflite::SliceOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_TransposeConvOptions: { + value = new tflite::TransposeConvOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SparseToDenseOptions: { + value = new tflite::SparseToDenseOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_TileOptions: { + value = new tflite::TileOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ExpandDimsOptions: { + value = new tflite::ExpandDimsOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_EqualOptions: { + value = new tflite::EqualOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_NotEqualOptions: { + value = new tflite::NotEqualOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ShapeOptions: { + value = new tflite::ShapeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_PowOptions: { + value = new tflite::PowOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ArgMinOptions: { + value = new tflite::ArgMinOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_FakeQuantOptions: { + value = new tflite::FakeQuantOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_PackOptions: { + value = new tflite::PackOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LogicalOrOptions: { + value = new tflite::LogicalOrOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_OneHotOptions: { + value = new tflite::OneHotOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LogicalAndOptions: { + value = new tflite::LogicalAndOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LogicalNotOptions: { + value = new tflite::LogicalNotOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_UnpackOptions: { + value = new tflite::UnpackOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_FloorDivOptions: { + value = new tflite::FloorDivOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SquareOptions: { + value = new tflite::SquareOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ZerosLikeOptions: { + value = new tflite::ZerosLikeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_FillOptions: { + value = new tflite::FillOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_BidirectionalSequenceLSTMOptions: { + value = new tflite::BidirectionalSequenceLSTMOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_BidirectionalSequenceRNNOptions: { + value = new tflite::BidirectionalSequenceRNNOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_UnidirectionalSequenceLSTMOptions: { + value = new tflite::UnidirectionalSequenceLSTMOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_FloorModOptions: { + value = new tflite::FloorModOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_RangeOptions: { + value = new tflite::RangeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ResizeNearestNeighborOptions: { + value = new tflite::ResizeNearestNeighborOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_LeakyReluOptions: { + value = new tflite::LeakyReluOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SquaredDifferenceOptions: { + value = new tflite::SquaredDifferenceOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_MirrorPadOptions: { + value = new tflite::MirrorPadOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_AbsOptions: { + value = new tflite::AbsOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SplitVOptions: { + value = new tflite::SplitVOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_UniqueOptions: { + value = new tflite::UniqueOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ReverseV2Options: { + value = new tflite::ReverseV2OptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_AddNOptions: { + value = new tflite::AddNOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_GatherNdOptions: { + value = new tflite::GatherNdOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_CosOptions: { + value = new tflite::CosOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_WhereOptions: { + value = new tflite::WhereOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_RankOptions: { + value = new tflite::RankOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ReverseSequenceOptions: { + value = new tflite::ReverseSequenceOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_MatrixDiagOptions: { + value = new tflite::MatrixDiagOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_QuantizeOptions: { + value = new tflite::QuantizeOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_MatrixSetDiagOptions: { + value = new tflite::MatrixSetDiagOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_HardSwishOptions: { + value = new tflite::HardSwishOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_IfOptions: { + value = new tflite::IfOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_WhileOptions: { + value = new tflite::WhileOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_DepthToSpaceOptions: { + value = new tflite::DepthToSpaceOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_NonMaxSuppressionV4Options: { + value = new tflite::NonMaxSuppressionV4OptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_NonMaxSuppressionV5Options: { + value = new tflite::NonMaxSuppressionV5OptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ScatterNdOptions: { + value = new tflite::ScatterNdOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SelectV2Options: { + value = new tflite::SelectV2OptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_DensifyOptions: { + value = new tflite::DensifyOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_SegmentSumOptions: { + value = new tflite::SegmentSumOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_BatchMatMulOptions: { + value = new tflite::BatchMatMulOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_CumsumOptions: { + value = new tflite::CumsumOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_CallOnceOptions: { + value = new tflite::CallOnceOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_BroadcastToOptions: { + value = new tflite::BroadcastToOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_Rfft2dOptions: { + value = new tflite::Rfft2dOptionsT( + *reinterpret_cast(u.value)); + break; + } + default: + break; + } +} + +inline void BuiltinOptionsUnion::Reset() { + switch (type) { + case BuiltinOptions_Conv2DOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_DepthwiseConv2DOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ConcatEmbeddingsOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LSHProjectionOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_Pool2DOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SVDFOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_RNNOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_FullyConnectedOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SoftmaxOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ConcatenationOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_AddOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_L2NormOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LocalResponseNormalizationOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LSTMOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ResizeBilinearOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_CallOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ReshapeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SkipGramOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SpaceToDepthOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_EmbeddingLookupSparseOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_MulOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_PadOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_GatherOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_BatchToSpaceNDOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SpaceToBatchNDOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_TransposeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ReducerOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SubOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_DivOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SqueezeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SequenceRNNOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_StridedSliceOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ExpOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_TopKV2Options: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SplitOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LogSoftmaxOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_CastOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_DequantizeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_MaximumMinimumOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ArgMaxOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LessOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_NegOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_PadV2Options: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_GreaterOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_GreaterEqualOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LessEqualOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SelectOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SliceOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_TransposeConvOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SparseToDenseOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_TileOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ExpandDimsOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_EqualOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_NotEqualOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ShapeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_PackOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LogicalOrOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_OneHotOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LogicalAndOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LogicalNotOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_UnpackOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_FloorDivOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SquareOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ZerosLikeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_FillOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_BidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_BidirectionalSequenceRNNOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_UnidirectionalSequenceLSTMOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_FloorModOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_RangeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ResizeNearestNeighborOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_LeakyReluOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SquaredDifferenceOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_MirrorPadOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_AbsOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SplitVOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_UniqueOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ReverseV2Options: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_AddNOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_GatherNdOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_CosOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_WhereOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_RankOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ReverseSequenceOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_MatrixDiagOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_QuantizeOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_MatrixSetDiagOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_HardSwishOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_IfOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_WhileOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_DepthToSpaceOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_NonMaxSuppressionV4Options: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_NonMaxSuppressionV5Options: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ScatterNdOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SelectV2Options: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_DensifyOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_SegmentSumOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_BatchMatMulOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_CumsumOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_CallOnceOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_BroadcastToOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_Rfft2dOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + default: break; + } + value = nullptr; + type = BuiltinOptions_NONE; +} + +inline const tflite::Model *GetModel(const void *buf) { + return flatbuffers::GetRoot(buf); +} + +inline const tflite::Model *GetSizePrefixedModel(const void *buf) { + return flatbuffers::GetSizePrefixedRoot(buf); +} + +inline const char *ModelIdentifier() { + return "TFL3"; +} + +inline bool ModelBufferHasIdentifier(const void *buf) { + return flatbuffers::BufferHasIdentifier( + buf, ModelIdentifier()); +} + +inline bool VerifyModelBuffer( + flatbuffers::Verifier &verifier) { + return verifier.VerifyBuffer(ModelIdentifier()); +} + +inline bool VerifySizePrefixedModelBuffer( + flatbuffers::Verifier &verifier) { + return verifier.VerifySizePrefixedBuffer(ModelIdentifier()); +} + +inline const char *ModelExtension() { + return "tflite"; +} + +inline void FinishModelBuffer( + flatbuffers::FlatBufferBuilder &fbb, + flatbuffers::Offset root) { + fbb.Finish(root, ModelIdentifier()); +} + +inline void FinishSizePrefixedModelBuffer( + flatbuffers::FlatBufferBuilder &fbb, + flatbuffers::Offset root) { + fbb.FinishSizePrefixed(root, ModelIdentifier()); +} + +inline std::unique_ptr UnPackModel( + const void *buf, + const flatbuffers::resolver_function_t *res = nullptr) { + return std::unique_ptr(GetModel(buf)->UnPack(res)); +} + +inline std::unique_ptr UnPackSizePrefixedModel( + const void *buf, + const flatbuffers::resolver_function_t *res = nullptr) { + return std::unique_ptr(GetSizePrefixedModel(buf)->UnPack(res)); +} + +} // namespace tflite + +#endif // FLATBUFFERS_GENERATED_SCHEMA_TFLITE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/schema/schema_utils.cpp b/src/tensorflow_arm/tensorflow/lite/schema/schema_utils.cpp new file mode 100644 index 0000000..3fe213f --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/schema/schema_utils.cpp @@ -0,0 +1,65 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_arm/tensorflow/lite/schema/schema_utils.h" + +#include + +#include "tensorflow_arm/tensorflow/lite/kernels/internal/compatibility.h" + +namespace tflite { + +// The following GetBuiltinCode methods are the utility methods for reading +// builtin operatore code, ensuring compatibility issues between v3 and v3a +// schema. Always the maximum value of the two fields always will be the correct +// value as follows: +// +// - Supporting schema version v3 models +// +// The `builtin_code` field is not available in the v3 models. Flatbuffer +// library will feed zero value, which is the default value in the v3a schema. +// The actual builtin operatore code value will exist in the +// `deprecated_builtin_code` field. At the same time, it implies that +// `deprecated_builtin_code` >= `builtin_code` and the maximum value of the two +// fields will be same with `deprecated_builtin_code'. +// +// - Supporting builtin operator codes beyonds 127 +// +// New builtin operators, whose operator code is larger than 127, can not be +// assigned to the `deprecated_builtin_code` field. In such cases, the +// value of the `builtin_code` field should be used for the builtin operator +// code. In the case, the maximum value of the two fields will be the value of +// the `builtin_code` as the right value. + +BuiltinOperator GetBuiltinCode(const OperatorCode* op_code) { + // Caller should guarantee that the given argument value is not a nullptr. + TFLITE_DCHECK(op_code != nullptr); + + return std::max( + op_code->builtin_code(), + static_cast(op_code->deprecated_builtin_code())); +} + +BuiltinOperator GetBuiltinCode(const OperatorCodeT* op_code) { + // Caller should guarantee that the given argument value is not a nullptr. + TFLITE_DCHECK(op_code != nullptr); + + return std::max(op_code->builtin_code, static_cast( + op_code->deprecated_builtin_code)); +} + +} // namespace tflite + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/tensorflow/lite/schema/schema_utils.h b/src/tensorflow_arm/tensorflow/lite/schema/schema_utils.h new file mode 100644 index 0000000..cfe9ac2 --- /dev/null +++ b/src/tensorflow_arm/tensorflow/lite/schema/schema_utils.h @@ -0,0 +1,36 @@ +#if !defined(ESP32) +/* Copyright 2020 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_SCHEMA_SCHEMA_UTILS_H_ +#define TENSORFLOW_LITE_SCHEMA_SCHEMA_UTILS_H_ + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_arm/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +// The following methods are introduced to resolve op builtin code shortage +// problem. The new builtin operator will be assigned to the extended builtin +// code field in the flatbuffer schema. Those methods helps to hide builtin code +// details. +BuiltinOperator GetBuiltinCode(const OperatorCode *op_code); + +BuiltinOperator GetBuiltinCode(const OperatorCodeT *op_code); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_SCHEMA_SCHEMA_UTILS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/version.h b/src/tensorflow_arm/tensorflow/lite/version.h similarity index 91% rename from src/tensorflow/lite/version.h rename to src/tensorflow_arm/tensorflow/lite/version.h index f667447..4b585b4 100644 --- a/src/tensorflow/lite/version.h +++ b/src/tensorflow_arm/tensorflow/lite/version.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_VERSION_H_ #define TENSORFLOW_LITE_VERSION_H_ -#include "tensorflow/core/public/version.h" +#include "tensorflow_arm/tensorflow/core/public/version.h" // The version number of the Schema. Ideally all changes will be backward // compatible. If that ever changes, we must ensure that version is the first @@ -27,3 +28,5 @@ limitations under the License. #define TFLITE_VERSION_STRING TF_VERSION_STRING #endif // TENSORFLOW_LITE_VERSION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/flatbuffers/LICENSE.txt b/src/tensorflow_arm/third_party/flatbuffers/LICENSE.txt new file mode 100644 index 0000000..d645695 --- /dev/null +++ b/src/tensorflow_arm/third_party/flatbuffers/LICENSE.txt @@ -0,0 +1,202 @@ + + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. diff --git a/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h new file mode 100644 index 0000000..2a96074 --- /dev/null +++ b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h @@ -0,0 +1,430 @@ +#if !defined(ESP32) +#ifndef FLATBUFFERS_BASE_H_ +#define FLATBUFFERS_BASE_H_ + +// clang-format off + +// If activate should be declared and included first. +#if defined(FLATBUFFERS_MEMORY_LEAK_TRACKING) && \ + defined(_MSC_VER) && defined(_DEBUG) + // The _CRTDBG_MAP_ALLOC inside will replace + // calloc/free (etc) to its debug version using #define directives. + #define _CRTDBG_MAP_ALLOC + #include + #include + // Replace operator new by trace-enabled version. + #define DEBUG_NEW new(_NORMAL_BLOCK, __FILE__, __LINE__) + #define new DEBUG_NEW +#endif + +#if !defined(FLATBUFFERS_ASSERT) +#include +#define FLATBUFFERS_ASSERT assert +#elif defined(FLATBUFFERS_ASSERT_INCLUDE) +// Include file with forward declaration +#include FLATBUFFERS_ASSERT_INCLUDE +#endif + +#ifndef ARDUINO +#include +#endif + +#include +#include +#include + +#if defined(ARDUINO) && !defined(ARDUINOSTL_M_H) + #include +#else + #include +#endif + +#include +#include +#include +#include +#include +#include +#include + +#if defined(__unix__) && !defined(FLATBUFFERS_LOCALE_INDEPENDENT) + #include +#endif + +#ifdef _STLPORT_VERSION + #define FLATBUFFERS_CPP98_STL +#endif + +#ifdef __ANDROID__ + #include +#endif + +#if defined(__ICCARM__) +#include +#endif + +// Note the __clang__ check is needed, because clang presents itself +// as an older GNUC compiler (4.2). +// Clang 3.3 and later implement all of the ISO C++ 2011 standard. +// Clang 3.4 and later implement all of the ISO C++ 2014 standard. +// http://clang.llvm.org/cxx_status.html + +// Note the MSVC value '__cplusplus' may be incorrect: +// The '__cplusplus' predefined macro in the MSVC stuck at the value 199711L, +// indicating (erroneously!) that the compiler conformed to the C++98 Standard. +// This value should be correct starting from MSVC2017-15.7-Preview-3. +// The '__cplusplus' will be valid only if MSVC2017-15.7-P3 and the `/Zc:__cplusplus` switch is set. +// Workaround (for details see MSDN): +// Use the _MSC_VER and _MSVC_LANG definition instead of the __cplusplus for compatibility. +// The _MSVC_LANG macro reports the Standard version regardless of the '/Zc:__cplusplus' switch. + +#if defined(__GNUC__) && !defined(__clang__) + #define FLATBUFFERS_GCC (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) +#else + #define FLATBUFFERS_GCC 0 +#endif + +#if defined(__clang__) + #define FLATBUFFERS_CLANG (__clang_major__ * 10000 + __clang_minor__ * 100 + __clang_patchlevel__) +#else + #define FLATBUFFERS_CLANG 0 +#endif + +/// @cond FLATBUFFERS_INTERNAL +#if __cplusplus <= 199711L && \ + (!defined(_MSC_VER) || _MSC_VER < 1600) && \ + (!defined(__GNUC__) || \ + (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__ < 40400)) + #error A C++11 compatible compiler with support for the auto typing is \ + required for FlatBuffers. + #error __cplusplus _MSC_VER __GNUC__ __GNUC_MINOR__ __GNUC_PATCHLEVEL__ +#endif + +#if !defined(__clang__) && \ + defined(__GNUC__) && \ + (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__ < 40600) + // Backwards compatibility for g++ 4.4, and 4.5 which don't have the nullptr + // and constexpr keywords. Note the __clang__ check is needed, because clang + // presents itself as an older GNUC compiler. + #ifndef nullptr_t + const class nullptr_t { + public: + template inline operator T*() const { return 0; } + private: + void operator&() const; + } nullptr = {}; + #endif + #ifndef constexpr + #define constexpr const + #endif +#endif + +// The wire format uses a little endian encoding (since that's efficient for +// the common platforms). +#if defined(__s390x__) + #define FLATBUFFERS_LITTLEENDIAN 0 +#endif // __s390x__ +#if !defined(FLATBUFFERS_LITTLEENDIAN) + #if defined(__GNUC__) || defined(__clang__) || defined(__ICCARM__) + #if (defined(__BIG_ENDIAN__) || \ + (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)) + #define FLATBUFFERS_LITTLEENDIAN 0 + #else + #define FLATBUFFERS_LITTLEENDIAN 1 + #endif // __BIG_ENDIAN__ + #elif defined(_MSC_VER) + #if defined(_M_PPC) + #define FLATBUFFERS_LITTLEENDIAN 0 + #else + #define FLATBUFFERS_LITTLEENDIAN 1 + #endif + #else + #error Unable to determine endianness, define FLATBUFFERS_LITTLEENDIAN. + #endif +#endif // !defined(FLATBUFFERS_LITTLEENDIAN) + +#define FLATBUFFERS_VERSION_MAJOR 1 +#define FLATBUFFERS_VERSION_MINOR 12 +#define FLATBUFFERS_VERSION_REVISION 0 +#define FLATBUFFERS_STRING_EXPAND(X) #X +#define FLATBUFFERS_STRING(X) FLATBUFFERS_STRING_EXPAND(X) +namespace flatbuffers { + // Returns version as string "MAJOR.MINOR.REVISION". + const char* FLATBUFFERS_VERSION(); +} + +#if (!defined(_MSC_VER) || _MSC_VER > 1600) && \ + (!defined(__GNUC__) || (__GNUC__ * 100 + __GNUC_MINOR__ >= 407)) || \ + defined(__clang__) + #define FLATBUFFERS_FINAL_CLASS final + #define FLATBUFFERS_OVERRIDE override + #define FLATBUFFERS_EXPLICIT_CPP11 explicit + #define FLATBUFFERS_VTABLE_UNDERLYING_TYPE : flatbuffers::voffset_t +#else + #define FLATBUFFERS_FINAL_CLASS + #define FLATBUFFERS_OVERRIDE + #define FLATBUFFERS_EXPLICIT_CPP11 + #define FLATBUFFERS_VTABLE_UNDERLYING_TYPE +#endif + +#if (!defined(_MSC_VER) || _MSC_VER >= 1900) && \ + (!defined(__GNUC__) || (__GNUC__ * 100 + __GNUC_MINOR__ >= 406)) || \ + (defined(__cpp_constexpr) && __cpp_constexpr >= 200704) + #define FLATBUFFERS_CONSTEXPR constexpr + #define FLATBUFFERS_CONSTEXPR_CPP11 constexpr + #define FLATBUFFERS_CONSTEXPR_DEFINED +#else + #define FLATBUFFERS_CONSTEXPR const + #define FLATBUFFERS_CONSTEXPR_CPP11 +#endif + +// This macro is never used in code! +#if (defined(__cplusplus) && __cplusplus >= 201402L) || \ + (defined(__cpp_constexpr) && __cpp_constexpr >= 201304) + #define FLATBUFFERS_CONSTEXPR_CPP14 FLATBUFFERS_CONSTEXPR +#else + #define FLATBUFFERS_CONSTEXPR_CPP14 +#endif + +#if (defined(__GXX_EXPERIMENTAL_CXX0X__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 406)) || \ + (defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 190023026)) || \ + defined(__clang__) + #define FLATBUFFERS_NOEXCEPT noexcept +#else + #define FLATBUFFERS_NOEXCEPT +#endif + +// NOTE: the FLATBUFFERS_DELETE_FUNC macro may change the access mode to +// private, so be sure to put it at the end or reset access mode explicitly. +#if (!defined(_MSC_VER) || _MSC_FULL_VER >= 180020827) && \ + (!defined(__GNUC__) || (__GNUC__ * 100 + __GNUC_MINOR__ >= 404)) || \ + defined(__clang__) + #define FLATBUFFERS_DELETE_FUNC(func) func = delete; +#else + #define FLATBUFFERS_DELETE_FUNC(func) private: func; +#endif + +// Check if we can use template aliases +// Not possible if Microsoft Compiler before 2012 +// Possible is the language feature __cpp_alias_templates is defined well +// Or possible if the C++ std is C+11 or newer +#if (defined(_MSC_VER) && _MSC_VER > 1700 /* MSVC2012 */) \ + || (defined(__cpp_alias_templates) && __cpp_alias_templates >= 200704) \ + || (defined(__cplusplus) && __cplusplus >= 201103L) + #define FLATBUFFERS_TEMPLATES_ALIASES +#endif + +#ifndef FLATBUFFERS_HAS_STRING_VIEW + // Only provide flatbuffers::string_view if __has_include can be used + // to detect a header that provides an implementation + #if defined(__has_include) + // Check for std::string_view (in c++17) + #if __has_include() && (__cplusplus >= 201606 || (defined(_HAS_CXX17) && _HAS_CXX17)) + #include + namespace flatbuffers { + typedef std::string_view string_view; + } + #define FLATBUFFERS_HAS_STRING_VIEW 1 + // Check for std::experimental::string_view (in c++14, compiler-dependent) + #elif __has_include() && (__cplusplus >= 201411) + #include + namespace flatbuffers { + typedef std::experimental::string_view string_view; + } + #define FLATBUFFERS_HAS_STRING_VIEW 1 + // Check for absl::string_view + #elif __has_include("absl/strings/string_view.h") + #include "absl/strings/string_view.h" + namespace flatbuffers { + typedef absl::string_view string_view; + } + #define FLATBUFFERS_HAS_STRING_VIEW 1 + #endif + #endif // __has_include +#endif // !FLATBUFFERS_HAS_STRING_VIEW + +#ifndef FLATBUFFERS_HAS_NEW_STRTOD + // Modern (C++11) strtod and strtof functions are available for use. + // 1) nan/inf strings as argument of strtod; + // 2) hex-float as argument of strtod/strtof. + #if (defined(_MSC_VER) && _MSC_VER >= 1900) || \ + (defined(__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 409)) || \ + (defined(__clang__)) + #define FLATBUFFERS_HAS_NEW_STRTOD 1 + #endif +#endif // !FLATBUFFERS_HAS_NEW_STRTOD + +#ifndef FLATBUFFERS_LOCALE_INDEPENDENT + // Enable locale independent functions {strtof_l, strtod_l,strtoll_l, strtoull_l}. + #if ((defined(_MSC_VER) && _MSC_VER >= 1800) || \ + (defined(_XOPEN_VERSION) && (_XOPEN_VERSION>=700)) && (!defined(__ANDROID_API__) || (defined(__ANDROID_API__) && (__ANDROID_API__>=21)))) + #define FLATBUFFERS_LOCALE_INDEPENDENT 1 + #else + #define FLATBUFFERS_LOCALE_INDEPENDENT 0 + #endif +#endif // !FLATBUFFERS_LOCALE_INDEPENDENT + +// Suppress Undefined Behavior Sanitizer (recoverable only). Usage: +// - __supress_ubsan__("undefined") +// - __supress_ubsan__("signed-integer-overflow") +#if defined(__clang__) && (__clang_major__ > 3 || (__clang_major__ == 3 && __clang_minor__ >=7)) + #define __supress_ubsan__(type) __attribute__((no_sanitize(type))) +#elif defined(__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 409) + #define __supress_ubsan__(type) __attribute__((no_sanitize_undefined)) +#else + #define __supress_ubsan__(type) +#endif + +// This is constexpr function used for checking compile-time constants. +// Avoid `#pragma warning(disable: 4127) // C4127: expression is constant`. +template FLATBUFFERS_CONSTEXPR inline bool IsConstTrue(T t) { + return !!t; +} + +// Enable C++ attribute [[]] if std:c++17 or higher. +#if ((__cplusplus >= 201703L) \ + || (defined(_MSVC_LANG) && (_MSVC_LANG >= 201703L))) + // All attributes unknown to an implementation are ignored without causing an error. + #define FLATBUFFERS_ATTRIBUTE(attr) [[attr]] + + #define FLATBUFFERS_FALLTHROUGH() [[fallthrough]] +#else + #define FLATBUFFERS_ATTRIBUTE(attr) + + #if FLATBUFFERS_CLANG >= 30800 + #define FLATBUFFERS_FALLTHROUGH() [[clang::fallthrough]] + #elif FLATBUFFERS_GCC >= 70300 + #define FLATBUFFERS_FALLTHROUGH() [[gnu::fallthrough]] + #else + #define FLATBUFFERS_FALLTHROUGH() + #endif +#endif + +/// @endcond + +/// @file +namespace flatbuffers { + +/// @cond FLATBUFFERS_INTERNAL +// Our default offset / size type, 32bit on purpose on 64bit systems. +// Also, using a consistent offset type maintains compatibility of serialized +// offset values between 32bit and 64bit systems. +typedef uint32_t uoffset_t; + +// Signed offsets for references that can go in both directions. +typedef int32_t soffset_t; + +// Offset/index used in v-tables, can be changed to uint8_t in +// format forks to save a bit of space if desired. +typedef uint16_t voffset_t; + +typedef uintmax_t largest_scalar_t; + +// In 32bits, this evaluates to 2GB - 1 +#define FLATBUFFERS_MAX_BUFFER_SIZE ((1ULL << (sizeof(::flatbuffers::soffset_t) * 8 - 1)) - 1) + +// We support aligning the contents of buffers up to this size. +#define FLATBUFFERS_MAX_ALIGNMENT 16 + +#if defined(_MSC_VER) + #pragma warning(disable: 4351) // C4351: new behavior: elements of array ... will be default initialized + #pragma warning(push) + #pragma warning(disable: 4127) // C4127: conditional expression is constant +#endif + +template T EndianSwap(T t) { + #if defined(_MSC_VER) + #define FLATBUFFERS_BYTESWAP16 _byteswap_ushort + #define FLATBUFFERS_BYTESWAP32 _byteswap_ulong + #define FLATBUFFERS_BYTESWAP64 _byteswap_uint64 + #elif defined(__ICCARM__) + #define FLATBUFFERS_BYTESWAP16 __REV16 + #define FLATBUFFERS_BYTESWAP32 __REV + #define FLATBUFFERS_BYTESWAP64(x) \ + ((__REV(static_cast(x >> 32U))) | (static_cast(__REV(static_cast(x)))) << 32U) + #else + #if defined(__GNUC__) && __GNUC__ * 100 + __GNUC_MINOR__ < 408 && !defined(__clang__) + // __builtin_bswap16 was missing prior to GCC 4.8. + #define FLATBUFFERS_BYTESWAP16(x) \ + static_cast(__builtin_bswap32(static_cast(x) << 16)) + #else + #define FLATBUFFERS_BYTESWAP16 __builtin_bswap16 + #endif + #define FLATBUFFERS_BYTESWAP32 __builtin_bswap32 + #define FLATBUFFERS_BYTESWAP64 __builtin_bswap64 + #endif + if (sizeof(T) == 1) { // Compile-time if-then's. + return t; + } else if (sizeof(T) == 2) { + union { T t; uint16_t i; } u = { t }; + u.i = FLATBUFFERS_BYTESWAP16(u.i); + return u.t; + } else if (sizeof(T) == 4) { + union { T t; uint32_t i; } u = { t }; + u.i = FLATBUFFERS_BYTESWAP32(u.i); + return u.t; + } else if (sizeof(T) == 8) { + union { T t; uint64_t i; } u = { t }; + u.i = FLATBUFFERS_BYTESWAP64(u.i); + return u.t; + } else { + FLATBUFFERS_ASSERT(0); + return t; + } +} + +#if defined(_MSC_VER) + #pragma warning(pop) +#endif + + +template T EndianScalar(T t) { + #if FLATBUFFERS_LITTLEENDIAN + return t; + #else + return EndianSwap(t); + #endif +} + +template +// UBSAN: C++ aliasing type rules, see std::bit_cast<> for details. +__supress_ubsan__("alignment") +T ReadScalar(const void *p) { + return EndianScalar(*reinterpret_cast(p)); +} + +// See https://github.com/google/flatbuffers/issues/5950 + +#if (FLATBUFFERS_GCC >= 100000) && (FLATBUFFERS_GCC < 110000) + #pragma GCC diagnostic push + #pragma GCC diagnostic ignored "-Wstringop-overflow" +#endif + +template +// UBSAN: C++ aliasing type rules, see std::bit_cast<> for details. +__supress_ubsan__("alignment") +void WriteScalar(void *p, T t) { + *reinterpret_cast(p) = EndianScalar(t); +} + +template struct Offset; +template __supress_ubsan__("alignment") void WriteScalar(void *p, Offset t) { + *reinterpret_cast(p) = EndianScalar(t.o); +} + +#if (FLATBUFFERS_GCC >= 100000) && (FLATBUFFERS_GCC < 110000) + #pragma GCC diagnostic pop +#endif + +// Computes how many bytes you'd have to pad to be able to write an +// "scalar_size" scalar if the buffer had grown to "buf_size" (downwards in +// memory). +__supress_ubsan__("unsigned-integer-overflow") +inline size_t PaddingBytes(size_t buf_size, size_t scalar_size) { + return ((~buf_size) + 1) & (scalar_size - 1); +} + +} // namespace flatbuffers +#endif // FLATBUFFERS_BASE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h new file mode 100644 index 0000000..62c8bf3 --- /dev/null +++ b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flatbuffers.h @@ -0,0 +1,2830 @@ +#if !defined(ESP32) +/* + * Copyright 2014 Google Inc. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef FLATBUFFERS_H_ +#define FLATBUFFERS_H_ + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h" +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/stl_emulation.h" + +#ifndef FLATBUFFERS_CPP98_STL + #include +#endif + +#if defined(FLATBUFFERS_NAN_DEFAULTS) +# include +#endif + +namespace flatbuffers { +// Generic 'operator==' with conditional specialisations. +// T e - new value of a scalar field. +// T def - default of scalar (is known at compile-time). +template inline bool IsTheSameAs(T e, T def) { return e == def; } + +#if defined(FLATBUFFERS_NAN_DEFAULTS) && \ + defined(FLATBUFFERS_HAS_NEW_STRTOD) && (FLATBUFFERS_HAS_NEW_STRTOD > 0) +// Like `operator==(e, def)` with weak NaN if T=(float|double). +template inline bool IsFloatTheSameAs(T e, T def) { + return (e == def) || ((def != def) && (e != e)); +} +template<> inline bool IsTheSameAs(float e, float def) { + return IsFloatTheSameAs(e, def); +} +template<> inline bool IsTheSameAs(double e, double def) { + return IsFloatTheSameAs(e, def); +} +#endif + +// Check 'v' is out of closed range [low; high]. +// Workaround for GCC warning [-Werror=type-limits]: +// comparison is always true due to limited range of data type. +template +inline bool IsOutRange(const T &v, const T &low, const T &high) { + return (v < low) || (high < v); +} + +// Check 'v' is in closed range [low; high]. +template +inline bool IsInRange(const T &v, const T &low, const T &high) { + return !IsOutRange(v, low, high); +} + +// Wrapper for uoffset_t to allow safe template specialization. +// Value is allowed to be 0 to indicate a null object (see e.g. AddOffset). +template struct Offset { + uoffset_t o; + Offset() : o(0) {} + Offset(uoffset_t _o) : o(_o) {} + Offset Union() const { return Offset(o); } + bool IsNull() const { return !o; } +}; + +inline void EndianCheck() { + int endiantest = 1; + // If this fails, see FLATBUFFERS_LITTLEENDIAN above. + FLATBUFFERS_ASSERT(*reinterpret_cast(&endiantest) == + FLATBUFFERS_LITTLEENDIAN); + (void)endiantest; +} + +template FLATBUFFERS_CONSTEXPR size_t AlignOf() { + // clang-format off + #ifdef _MSC_VER + return __alignof(T); + #else + #ifndef alignof + return __alignof__(T); + #else + return alignof(T); + #endif + #endif + // clang-format on +} + +// When we read serialized data from memory, in the case of most scalars, +// we want to just read T, but in the case of Offset, we want to actually +// perform the indirection and return a pointer. +// The template specialization below does just that. +// It is wrapped in a struct since function templates can't overload on the +// return type like this. +// The typedef is for the convenience of callers of this function +// (avoiding the need for a trailing return decltype) +template struct IndirectHelper { + typedef T return_type; + typedef T mutable_return_type; + static const size_t element_stride = sizeof(T); + static return_type Read(const uint8_t *p, uoffset_t i) { + return EndianScalar((reinterpret_cast(p))[i]); + } +}; +template struct IndirectHelper> { + typedef const T *return_type; + typedef T *mutable_return_type; + static const size_t element_stride = sizeof(uoffset_t); + static return_type Read(const uint8_t *p, uoffset_t i) { + p += i * sizeof(uoffset_t); + return reinterpret_cast(p + ReadScalar(p)); + } +}; +template struct IndirectHelper { + typedef const T *return_type; + typedef T *mutable_return_type; + static const size_t element_stride = sizeof(T); + static return_type Read(const uint8_t *p, uoffset_t i) { + return reinterpret_cast(p + i * sizeof(T)); + } +}; + +// An STL compatible iterator implementation for Vector below, effectively +// calling Get() for every element. +template struct VectorIterator { + typedef std::random_access_iterator_tag iterator_category; + typedef IT value_type; + typedef ptrdiff_t difference_type; + typedef IT *pointer; + typedef IT &reference; + + VectorIterator(const uint8_t *data, uoffset_t i) + : data_(data + IndirectHelper::element_stride * i) {} + VectorIterator(const VectorIterator &other) : data_(other.data_) {} + VectorIterator() : data_(nullptr) {} + + VectorIterator &operator=(const VectorIterator &other) { + data_ = other.data_; + return *this; + } + + // clang-format off + #if !defined(FLATBUFFERS_CPP98_STL) + VectorIterator &operator=(VectorIterator &&other) { + data_ = other.data_; + return *this; + } + #endif // !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + + bool operator==(const VectorIterator &other) const { + return data_ == other.data_; + } + + bool operator<(const VectorIterator &other) const { + return data_ < other.data_; + } + + bool operator!=(const VectorIterator &other) const { + return data_ != other.data_; + } + + difference_type operator-(const VectorIterator &other) const { + return (data_ - other.data_) / IndirectHelper::element_stride; + } + + IT operator*() const { return IndirectHelper::Read(data_, 0); } + + IT operator->() const { return IndirectHelper::Read(data_, 0); } + + VectorIterator &operator++() { + data_ += IndirectHelper::element_stride; + return *this; + } + + VectorIterator operator++(int) { + VectorIterator temp(data_, 0); + data_ += IndirectHelper::element_stride; + return temp; + } + + VectorIterator operator+(const uoffset_t &offset) const { + return VectorIterator(data_ + offset * IndirectHelper::element_stride, + 0); + } + + VectorIterator &operator+=(const uoffset_t &offset) { + data_ += offset * IndirectHelper::element_stride; + return *this; + } + + VectorIterator &operator--() { + data_ -= IndirectHelper::element_stride; + return *this; + } + + VectorIterator operator--(int) { + VectorIterator temp(data_, 0); + data_ -= IndirectHelper::element_stride; + return temp; + } + + VectorIterator operator-(const uoffset_t &offset) const { + return VectorIterator(data_ - offset * IndirectHelper::element_stride, + 0); + } + + VectorIterator &operator-=(const uoffset_t &offset) { + data_ -= offset * IndirectHelper::element_stride; + return *this; + } + + private: + const uint8_t *data_; +}; + +template +struct VectorReverseIterator : public std::reverse_iterator { + explicit VectorReverseIterator(Iterator iter) + : std::reverse_iterator(iter) {} + + typename Iterator::value_type operator*() const { + return *(std::reverse_iterator::current); + } + + typename Iterator::value_type operator->() const { + return *(std::reverse_iterator::current); + } +}; + +struct String; + +// This is used as a helper type for accessing vectors. +// Vector::data() assumes the vector elements start after the length field. +template class Vector { + public: + typedef VectorIterator::mutable_return_type> + iterator; + typedef VectorIterator::return_type> + const_iterator; + typedef VectorReverseIterator reverse_iterator; + typedef VectorReverseIterator const_reverse_iterator; + + uoffset_t size() const { return EndianScalar(length_); } + + // Deprecated: use size(). Here for backwards compatibility. + FLATBUFFERS_ATTRIBUTE(deprecated("use size() instead")) + uoffset_t Length() const { return size(); } + + typedef typename IndirectHelper::return_type return_type; + typedef typename IndirectHelper::mutable_return_type mutable_return_type; + + return_type Get(uoffset_t i) const { + FLATBUFFERS_ASSERT(i < size()); + return IndirectHelper::Read(Data(), i); + } + + return_type operator[](uoffset_t i) const { return Get(i); } + + // If this is a Vector of enums, T will be its storage type, not the enum + // type. This function makes it convenient to retrieve value with enum + // type E. + template E GetEnum(uoffset_t i) const { + return static_cast(Get(i)); + } + + // If this a vector of unions, this does the cast for you. There's no check + // to make sure this is the right type! + template const U *GetAs(uoffset_t i) const { + return reinterpret_cast(Get(i)); + } + + // If this a vector of unions, this does the cast for you. There's no check + // to make sure this is actually a string! + const String *GetAsString(uoffset_t i) const { + return reinterpret_cast(Get(i)); + } + + const void *GetStructFromOffset(size_t o) const { + return reinterpret_cast(Data() + o); + } + + iterator begin() { return iterator(Data(), 0); } + const_iterator begin() const { return const_iterator(Data(), 0); } + + iterator end() { return iterator(Data(), size()); } + const_iterator end() const { return const_iterator(Data(), size()); } + + reverse_iterator rbegin() { return reverse_iterator(end() - 1); } + const_reverse_iterator rbegin() const { + return const_reverse_iterator(end() - 1); + } + + reverse_iterator rend() { return reverse_iterator(begin() - 1); } + const_reverse_iterator rend() const { + return const_reverse_iterator(begin() - 1); + } + + const_iterator cbegin() const { return begin(); } + + const_iterator cend() const { return end(); } + + const_reverse_iterator crbegin() const { return rbegin(); } + + const_reverse_iterator crend() const { return rend(); } + + // Change elements if you have a non-const pointer to this object. + // Scalars only. See reflection.h, and the documentation. + void Mutate(uoffset_t i, const T &val) { + FLATBUFFERS_ASSERT(i < size()); + WriteScalar(data() + i, val); + } + + // Change an element of a vector of tables (or strings). + // "val" points to the new table/string, as you can obtain from + // e.g. reflection::AddFlatBuffer(). + void MutateOffset(uoffset_t i, const uint8_t *val) { + FLATBUFFERS_ASSERT(i < size()); + static_assert(sizeof(T) == sizeof(uoffset_t), "Unrelated types"); + WriteScalar(data() + i, + static_cast(val - (Data() + i * sizeof(uoffset_t)))); + } + + // Get a mutable pointer to tables/strings inside this vector. + mutable_return_type GetMutableObject(uoffset_t i) const { + FLATBUFFERS_ASSERT(i < size()); + return const_cast(IndirectHelper::Read(Data(), i)); + } + + // The raw data in little endian format. Use with care. + const uint8_t *Data() const { + return reinterpret_cast(&length_ + 1); + } + + uint8_t *Data() { return reinterpret_cast(&length_ + 1); } + + // Similarly, but typed, much like std::vector::data + const T *data() const { return reinterpret_cast(Data()); } + T *data() { return reinterpret_cast(Data()); } + + template return_type LookupByKey(K key) const { + void *search_result = std::bsearch( + &key, Data(), size(), IndirectHelper::element_stride, KeyCompare); + + if (!search_result) { + return nullptr; // Key not found. + } + + const uint8_t *element = reinterpret_cast(search_result); + + return IndirectHelper::Read(element, 0); + } + + protected: + // This class is only used to access pre-existing data. Don't ever + // try to construct these manually. + Vector(); + + uoffset_t length_; + + private: + // This class is a pointer. Copying will therefore create an invalid object. + // Private and unimplemented copy constructor. + Vector(const Vector &); + Vector &operator=(const Vector &); + + template static int KeyCompare(const void *ap, const void *bp) { + const K *key = reinterpret_cast(ap); + const uint8_t *data = reinterpret_cast(bp); + auto table = IndirectHelper::Read(data, 0); + + // std::bsearch compares with the operands transposed, so we negate the + // result here. + return -table->KeyCompareWithValue(*key); + } +}; + +// Represent a vector much like the template above, but in this case we +// don't know what the element types are (used with reflection.h). +class VectorOfAny { + public: + uoffset_t size() const { return EndianScalar(length_); } + + const uint8_t *Data() const { + return reinterpret_cast(&length_ + 1); + } + uint8_t *Data() { return reinterpret_cast(&length_ + 1); } + + protected: + VectorOfAny(); + + uoffset_t length_; + + private: + VectorOfAny(const VectorOfAny &); + VectorOfAny &operator=(const VectorOfAny &); +}; + +#ifndef FLATBUFFERS_CPP98_STL +template +Vector> *VectorCast(Vector> *ptr) { + static_assert(std::is_base_of::value, "Unrelated types"); + return reinterpret_cast> *>(ptr); +} + +template +const Vector> *VectorCast(const Vector> *ptr) { + static_assert(std::is_base_of::value, "Unrelated types"); + return reinterpret_cast> *>(ptr); +} +#endif + +// Convenient helper function to get the length of any vector, regardless +// of whether it is null or not (the field is not set). +template static inline size_t VectorLength(const Vector *v) { + return v ? v->size() : 0; +} + +// This is used as a helper type for accessing arrays. +template class Array { + typedef + typename flatbuffers::integral_constant::value> + scalar_tag; + typedef + typename flatbuffers::conditional::type + IndirectHelperType; + + public: + typedef typename IndirectHelper::return_type return_type; + typedef VectorIterator const_iterator; + typedef VectorReverseIterator const_reverse_iterator; + + FLATBUFFERS_CONSTEXPR uint16_t size() const { return length; } + + return_type Get(uoffset_t i) const { + FLATBUFFERS_ASSERT(i < size()); + return IndirectHelper::Read(Data(), i); + } + + return_type operator[](uoffset_t i) const { return Get(i); } + + // If this is a Vector of enums, T will be its storage type, not the enum + // type. This function makes it convenient to retrieve value with enum + // type E. + template E GetEnum(uoffset_t i) const { + return static_cast(Get(i)); + } + + const_iterator begin() const { return const_iterator(Data(), 0); } + const_iterator end() const { return const_iterator(Data(), size()); } + + const_reverse_iterator rbegin() const { + return const_reverse_iterator(end()); + } + const_reverse_iterator rend() const { return const_reverse_iterator(end()); } + + const_iterator cbegin() const { return begin(); } + const_iterator cend() const { return end(); } + + const_reverse_iterator crbegin() const { return rbegin(); } + const_reverse_iterator crend() const { return rend(); } + + // Get a mutable pointer to elements inside this array. + // This method used to mutate arrays of structs followed by a @p Mutate + // operation. For primitive types use @p Mutate directly. + // @warning Assignments and reads to/from the dereferenced pointer are not + // automatically converted to the correct endianness. + typename flatbuffers::conditional::type + GetMutablePointer(uoffset_t i) const { + FLATBUFFERS_ASSERT(i < size()); + return const_cast(&data()[i]); + } + + // Change elements if you have a non-const pointer to this object. + void Mutate(uoffset_t i, const T &val) { MutateImpl(scalar_tag(), i, val); } + + // The raw data in little endian format. Use with care. + const uint8_t *Data() const { return data_; } + + uint8_t *Data() { return data_; } + + // Similarly, but typed, much like std::vector::data + const T *data() const { return reinterpret_cast(Data()); } + T *data() { return reinterpret_cast(Data()); } + + protected: + void MutateImpl(flatbuffers::integral_constant, uoffset_t i, + const T &val) { + FLATBUFFERS_ASSERT(i < size()); + WriteScalar(data() + i, val); + } + + void MutateImpl(flatbuffers::integral_constant, uoffset_t i, + const T &val) { + *(GetMutablePointer(i)) = val; + } + + // This class is only used to access pre-existing data. Don't ever + // try to construct these manually. + // 'constexpr' allows us to use 'size()' at compile time. + // @note Must not use 'FLATBUFFERS_CONSTEXPR' here, as const is not allowed on + // a constructor. +#if defined(__cpp_constexpr) + constexpr Array(); +#else + Array(); +#endif + + uint8_t data_[length * sizeof(T)]; + + private: + // This class is a pointer. Copying will therefore create an invalid object. + // Private and unimplemented copy constructor. + Array(const Array &); + Array &operator=(const Array &); +}; + +// Specialization for Array[struct] with access using Offset pointer. +// This specialization used by idl_gen_text.cpp. +template class Array, length> { + static_assert(flatbuffers::is_same::value, "unexpected type T"); + + public: + typedef const void *return_type; + + const uint8_t *Data() const { return data_; } + + // Make idl_gen_text.cpp::PrintContainer happy. + return_type operator[](uoffset_t) const { + FLATBUFFERS_ASSERT(false); + return nullptr; + } + + private: + // This class is only used to access pre-existing data. + Array(); + Array(const Array &); + Array &operator=(const Array &); + + uint8_t data_[1]; +}; + +// Lexicographically compare two strings (possibly containing nulls), and +// return true if the first is less than the second. +static inline bool StringLessThan(const char *a_data, uoffset_t a_size, + const char *b_data, uoffset_t b_size) { + const auto cmp = memcmp(a_data, b_data, (std::min)(a_size, b_size)); + return cmp == 0 ? a_size < b_size : cmp < 0; +} + +struct String : public Vector { + const char *c_str() const { return reinterpret_cast(Data()); } + std::string str() const { return std::string(c_str(), size()); } + + // clang-format off + #ifdef FLATBUFFERS_HAS_STRING_VIEW + flatbuffers::string_view string_view() const { + return flatbuffers::string_view(c_str(), size()); + } + #endif // FLATBUFFERS_HAS_STRING_VIEW + // clang-format on + + bool operator<(const String &o) const { + return StringLessThan(this->data(), this->size(), o.data(), o.size()); + } +}; + +// Convenience function to get std::string from a String returning an empty +// string on null pointer. +static inline std::string GetString(const String *str) { + return str ? str->str() : ""; +} + +// Convenience function to get char* from a String returning an empty string on +// null pointer. +static inline const char *GetCstring(const String *str) { + return str ? str->c_str() : ""; +} + +#ifdef FLATBUFFERS_HAS_STRING_VIEW +// Convenience function to get string_view from a String returning an empty +// string_view on null pointer. +static inline flatbuffers::string_view GetStringView(const String *str) { + return str ? str->string_view() : flatbuffers::string_view(); +} +#endif // FLATBUFFERS_HAS_STRING_VIEW + +// Allocator interface. This is flatbuffers-specific and meant only for +// `vector_downward` usage. +class Allocator { + public: + virtual ~Allocator() {} + + // Allocate `size` bytes of memory. + virtual uint8_t *allocate(size_t size) = 0; + + // Deallocate `size` bytes of memory at `p` allocated by this allocator. + virtual void deallocate(uint8_t *p, size_t size) = 0; + + // Reallocate `new_size` bytes of memory, replacing the old region of size + // `old_size` at `p`. In contrast to a normal realloc, this grows downwards, + // and is intended specifcally for `vector_downward` use. + // `in_use_back` and `in_use_front` indicate how much of `old_size` is + // actually in use at each end, and needs to be copied. + virtual uint8_t *reallocate_downward(uint8_t *old_p, size_t old_size, + size_t new_size, size_t in_use_back, + size_t in_use_front) { + FLATBUFFERS_ASSERT(new_size > old_size); // vector_downward only grows + uint8_t *new_p = allocate(new_size); + memcpy_downward(old_p, old_size, new_p, new_size, in_use_back, + in_use_front); + deallocate(old_p, old_size); + return new_p; + } + + protected: + // Called by `reallocate_downward` to copy memory from `old_p` of `old_size` + // to `new_p` of `new_size`. Only memory of size `in_use_front` and + // `in_use_back` will be copied from the front and back of the old memory + // allocation. + void memcpy_downward(uint8_t *old_p, size_t old_size, uint8_t *new_p, + size_t new_size, size_t in_use_back, + size_t in_use_front) { + memcpy(new_p + new_size - in_use_back, old_p + old_size - in_use_back, + in_use_back); + memcpy(new_p, old_p, in_use_front); + } +}; + +// DefaultAllocator uses new/delete to allocate memory regions +class DefaultAllocator : public Allocator { + public: + uint8_t *allocate(size_t size) FLATBUFFERS_OVERRIDE { + return new uint8_t[size]; + } + + void deallocate(uint8_t *p, size_t) FLATBUFFERS_OVERRIDE { delete[] p; } + + static void dealloc(void *p, size_t) { delete[] static_cast(p); } +}; + +// These functions allow for a null allocator to mean use the default allocator, +// as used by DetachedBuffer and vector_downward below. +// This is to avoid having a statically or dynamically allocated default +// allocator, or having to move it between the classes that may own it. +inline uint8_t *Allocate(Allocator *allocator, size_t size) { + return allocator ? allocator->allocate(size) + : DefaultAllocator().allocate(size); +} + +inline void Deallocate(Allocator *allocator, uint8_t *p, size_t size) { + if (allocator) + allocator->deallocate(p, size); + else + DefaultAllocator().deallocate(p, size); +} + +inline uint8_t *ReallocateDownward(Allocator *allocator, uint8_t *old_p, + size_t old_size, size_t new_size, + size_t in_use_back, size_t in_use_front) { + return allocator ? allocator->reallocate_downward(old_p, old_size, new_size, + in_use_back, in_use_front) + : DefaultAllocator().reallocate_downward( + old_p, old_size, new_size, in_use_back, in_use_front); +} + +// DetachedBuffer is a finished flatbuffer memory region, detached from its +// builder. The original memory region and allocator are also stored so that +// the DetachedBuffer can manage the memory lifetime. +class DetachedBuffer { + public: + DetachedBuffer() + : allocator_(nullptr), + own_allocator_(false), + buf_(nullptr), + reserved_(0), + cur_(nullptr), + size_(0) {} + + DetachedBuffer(Allocator *allocator, bool own_allocator, uint8_t *buf, + size_t reserved, uint8_t *cur, size_t sz) + : allocator_(allocator), + own_allocator_(own_allocator), + buf_(buf), + reserved_(reserved), + cur_(cur), + size_(sz) {} + + // clang-format off + #if !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + DetachedBuffer(DetachedBuffer &&other) + : allocator_(other.allocator_), + own_allocator_(other.own_allocator_), + buf_(other.buf_), + reserved_(other.reserved_), + cur_(other.cur_), + size_(other.size_) { + other.reset(); + } + // clang-format off + #endif // !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + + // clang-format off + #if !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + DetachedBuffer &operator=(DetachedBuffer &&other) { + if (this == &other) return *this; + + destroy(); + + allocator_ = other.allocator_; + own_allocator_ = other.own_allocator_; + buf_ = other.buf_; + reserved_ = other.reserved_; + cur_ = other.cur_; + size_ = other.size_; + + other.reset(); + + return *this; + } + // clang-format off + #endif // !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + + ~DetachedBuffer() { destroy(); } + + const uint8_t *data() const { return cur_; } + + uint8_t *data() { return cur_; } + + size_t size() const { return size_; } + + // clang-format off + #if 0 // disabled for now due to the ordering of classes in this header + template + bool Verify() const { + Verifier verifier(data(), size()); + return verifier.Verify(nullptr); + } + + template + const T* GetRoot() const { + return flatbuffers::GetRoot(data()); + } + + template + T* GetRoot() { + return flatbuffers::GetRoot(data()); + } + #endif + // clang-format on + + // clang-format off + #if !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + // These may change access mode, leave these at end of public section + FLATBUFFERS_DELETE_FUNC(DetachedBuffer(const DetachedBuffer &other)) + FLATBUFFERS_DELETE_FUNC( + DetachedBuffer &operator=(const DetachedBuffer &other)) + // clang-format off + #endif // !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + + protected: + Allocator *allocator_; + bool own_allocator_; + uint8_t *buf_; + size_t reserved_; + uint8_t *cur_; + size_t size_; + + inline void destroy() { + if (buf_) Deallocate(allocator_, buf_, reserved_); + if (own_allocator_ && allocator_) { delete allocator_; } + reset(); + } + + inline void reset() { + allocator_ = nullptr; + own_allocator_ = false; + buf_ = nullptr; + reserved_ = 0; + cur_ = nullptr; + size_ = 0; + } +}; + +// This is a minimal replication of std::vector functionality, +// except growing from higher to lower addresses. i.e push_back() inserts data +// in the lowest address in the vector. +// Since this vector leaves the lower part unused, we support a "scratch-pad" +// that can be stored there for temporary data, to share the allocated space. +// Essentially, this supports 2 std::vectors in a single buffer. +class vector_downward { + public: + explicit vector_downward(size_t initial_size, Allocator *allocator, + bool own_allocator, size_t buffer_minalign) + : allocator_(allocator), + own_allocator_(own_allocator), + initial_size_(initial_size), + buffer_minalign_(buffer_minalign), + reserved_(0), + buf_(nullptr), + cur_(nullptr), + scratch_(nullptr) {} + + // clang-format off + #if !defined(FLATBUFFERS_CPP98_STL) + vector_downward(vector_downward &&other) + #else + vector_downward(vector_downward &other) + #endif // defined(FLATBUFFERS_CPP98_STL) + // clang-format on + : allocator_(other.allocator_), + own_allocator_(other.own_allocator_), + initial_size_(other.initial_size_), + buffer_minalign_(other.buffer_minalign_), + reserved_(other.reserved_), + buf_(other.buf_), + cur_(other.cur_), + scratch_(other.scratch_) { + // No change in other.allocator_ + // No change in other.initial_size_ + // No change in other.buffer_minalign_ + other.own_allocator_ = false; + other.reserved_ = 0; + other.buf_ = nullptr; + other.cur_ = nullptr; + other.scratch_ = nullptr; + } + + // clang-format off + #if !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + vector_downward &operator=(vector_downward &&other) { + // Move construct a temporary and swap idiom + vector_downward temp(std::move(other)); + swap(temp); + return *this; + } + // clang-format off + #endif // defined(FLATBUFFERS_CPP98_STL) + // clang-format on + + ~vector_downward() { + clear_buffer(); + clear_allocator(); + } + + void reset() { + clear_buffer(); + clear(); + } + + void clear() { + if (buf_) { + cur_ = buf_ + reserved_; + } else { + reserved_ = 0; + cur_ = nullptr; + } + clear_scratch(); + } + + void clear_scratch() { scratch_ = buf_; } + + void clear_allocator() { + if (own_allocator_ && allocator_) { delete allocator_; } + allocator_ = nullptr; + own_allocator_ = false; + } + + void clear_buffer() { + if (buf_) Deallocate(allocator_, buf_, reserved_); + buf_ = nullptr; + } + + // Relinquish the pointer to the caller. + uint8_t *release_raw(size_t &allocated_bytes, size_t &offset) { + auto *buf = buf_; + allocated_bytes = reserved_; + offset = static_cast(cur_ - buf_); + + // release_raw only relinquishes the buffer ownership. + // Does not deallocate or reset the allocator. Destructor will do that. + buf_ = nullptr; + clear(); + return buf; + } + + // Relinquish the pointer to the caller. + DetachedBuffer release() { + // allocator ownership (if any) is transferred to DetachedBuffer. + DetachedBuffer fb(allocator_, own_allocator_, buf_, reserved_, cur_, + size()); + if (own_allocator_) { + allocator_ = nullptr; + own_allocator_ = false; + } + buf_ = nullptr; + clear(); + return fb; + } + + size_t ensure_space(size_t len) { + FLATBUFFERS_ASSERT(cur_ >= scratch_ && scratch_ >= buf_); + if (len > static_cast(cur_ - scratch_)) { reallocate(len); } + // Beyond this, signed offsets may not have enough range: + // (FlatBuffers > 2GB not supported). + FLATBUFFERS_ASSERT(size() < FLATBUFFERS_MAX_BUFFER_SIZE); + return len; + } + + inline uint8_t *make_space(size_t len) { + size_t space = ensure_space(len); + cur_ -= space; + return cur_; + } + + // Returns nullptr if using the DefaultAllocator. + Allocator *get_custom_allocator() { return allocator_; } + + uoffset_t size() const { + return static_cast(reserved_ - (cur_ - buf_)); + } + + uoffset_t scratch_size() const { + return static_cast(scratch_ - buf_); + } + + size_t capacity() const { return reserved_; } + + uint8_t *data() const { + FLATBUFFERS_ASSERT(cur_); + return cur_; + } + + uint8_t *scratch_data() const { + FLATBUFFERS_ASSERT(buf_); + return buf_; + } + + uint8_t *scratch_end() const { + FLATBUFFERS_ASSERT(scratch_); + return scratch_; + } + + uint8_t *data_at(size_t offset) const { return buf_ + reserved_ - offset; } + + void push(const uint8_t *bytes, size_t num) { + if (num > 0) { memcpy(make_space(num), bytes, num); } + } + + // Specialized version of push() that avoids memcpy call for small data. + template void push_small(const T &little_endian_t) { + make_space(sizeof(T)); + *reinterpret_cast(cur_) = little_endian_t; + } + + template void scratch_push_small(const T &t) { + ensure_space(sizeof(T)); + *reinterpret_cast(scratch_) = t; + scratch_ += sizeof(T); + } + + // fill() is most frequently called with small byte counts (<= 4), + // which is why we're using loops rather than calling memset. + void fill(size_t zero_pad_bytes) { + make_space(zero_pad_bytes); + for (size_t i = 0; i < zero_pad_bytes; i++) cur_[i] = 0; + } + + // Version for when we know the size is larger. + // Precondition: zero_pad_bytes > 0 + void fill_big(size_t zero_pad_bytes) { + memset(make_space(zero_pad_bytes), 0, zero_pad_bytes); + } + + void pop(size_t bytes_to_remove) { cur_ += bytes_to_remove; } + void scratch_pop(size_t bytes_to_remove) { scratch_ -= bytes_to_remove; } + + void swap(vector_downward &other) { + using std::swap; + swap(allocator_, other.allocator_); + swap(own_allocator_, other.own_allocator_); + swap(initial_size_, other.initial_size_); + swap(buffer_minalign_, other.buffer_minalign_); + swap(reserved_, other.reserved_); + swap(buf_, other.buf_); + swap(cur_, other.cur_); + swap(scratch_, other.scratch_); + } + + void swap_allocator(vector_downward &other) { + using std::swap; + swap(allocator_, other.allocator_); + swap(own_allocator_, other.own_allocator_); + } + + private: + // You shouldn't really be copying instances of this class. + FLATBUFFERS_DELETE_FUNC(vector_downward(const vector_downward &)) + FLATBUFFERS_DELETE_FUNC(vector_downward &operator=(const vector_downward &)) + + Allocator *allocator_; + bool own_allocator_; + size_t initial_size_; + size_t buffer_minalign_; + size_t reserved_; + uint8_t *buf_; + uint8_t *cur_; // Points at location between empty (below) and used (above). + uint8_t *scratch_; // Points to the end of the scratchpad in use. + + void reallocate(size_t len) { + auto old_reserved = reserved_; + auto old_size = size(); + auto old_scratch_size = scratch_size(); + reserved_ += + (std::max)(len, old_reserved ? old_reserved / 2 : initial_size_); + reserved_ = (reserved_ + buffer_minalign_ - 1) & ~(buffer_minalign_ - 1); + if (buf_) { + buf_ = ReallocateDownward(allocator_, buf_, old_reserved, reserved_, + old_size, old_scratch_size); + } else { + buf_ = Allocate(allocator_, reserved_); + } + cur_ = buf_ + reserved_ - old_size; + scratch_ = buf_ + old_scratch_size; + } +}; + +// Converts a Field ID to a virtual table offset. +inline voffset_t FieldIndexToOffset(voffset_t field_id) { + // Should correspond to what EndTable() below builds up. + const int fixed_fields = 2; // Vtable size and Object Size. + return static_cast((field_id + fixed_fields) * sizeof(voffset_t)); +} + +template +const T *data(const std::vector &v) { + // Eventually the returned pointer gets passed down to memcpy, so + // we need it to be non-null to avoid undefined behavior. + static uint8_t t; + return v.empty() ? reinterpret_cast(&t) : &v.front(); +} +template T *data(std::vector &v) { + // Eventually the returned pointer gets passed down to memcpy, so + // we need it to be non-null to avoid undefined behavior. + static uint8_t t; + return v.empty() ? reinterpret_cast(&t) : &v.front(); +} + +/// @endcond + +/// @addtogroup flatbuffers_cpp_api +/// @{ +/// @class FlatBufferBuilder +/// @brief Helper class to hold data needed in creation of a FlatBuffer. +/// To serialize data, you typically call one of the `Create*()` functions in +/// the generated code, which in turn call a sequence of `StartTable`/ +/// `PushElement`/`AddElement`/`EndTable`, or the builtin `CreateString`/ +/// `CreateVector` functions. Do this is depth-first order to build up a tree to +/// the root. `Finish()` wraps up the buffer ready for transport. +class FlatBufferBuilder { + public: + /// @brief Default constructor for FlatBufferBuilder. + /// @param[in] initial_size The initial size of the buffer, in bytes. Defaults + /// to `1024`. + /// @param[in] allocator An `Allocator` to use. If null will use + /// `DefaultAllocator`. + /// @param[in] own_allocator Whether the builder/vector should own the + /// allocator. Defaults to / `false`. + /// @param[in] buffer_minalign Force the buffer to be aligned to the given + /// minimum alignment upon reallocation. Only needed if you intend to store + /// types with custom alignment AND you wish to read the buffer in-place + /// directly after creation. + explicit FlatBufferBuilder( + size_t initial_size = 1024, Allocator *allocator = nullptr, + bool own_allocator = false, + size_t buffer_minalign = AlignOf()) + : buf_(initial_size, allocator, own_allocator, buffer_minalign), + num_field_loc(0), + max_voffset_(0), + nested(false), + finished(false), + minalign_(1), + force_defaults_(false), + dedup_vtables_(true), + string_pool(nullptr) { + EndianCheck(); + } + + // clang-format off + /// @brief Move constructor for FlatBufferBuilder. + #if !defined(FLATBUFFERS_CPP98_STL) + FlatBufferBuilder(FlatBufferBuilder &&other) + #else + FlatBufferBuilder(FlatBufferBuilder &other) + #endif // #if !defined(FLATBUFFERS_CPP98_STL) + : buf_(1024, nullptr, false, AlignOf()), + num_field_loc(0), + max_voffset_(0), + nested(false), + finished(false), + minalign_(1), + force_defaults_(false), + dedup_vtables_(true), + string_pool(nullptr) { + EndianCheck(); + // Default construct and swap idiom. + // Lack of delegating constructors in vs2010 makes it more verbose than needed. + Swap(other); + } + // clang-format on + + // clang-format off + #if !defined(FLATBUFFERS_CPP98_STL) + // clang-format on + /// @brief Move assignment operator for FlatBufferBuilder. + FlatBufferBuilder &operator=(FlatBufferBuilder &&other) { + // Move construct a temporary and swap idiom + FlatBufferBuilder temp(std::move(other)); + Swap(temp); + return *this; + } + // clang-format off + #endif // defined(FLATBUFFERS_CPP98_STL) + // clang-format on + + void Swap(FlatBufferBuilder &other) { + using std::swap; + buf_.swap(other.buf_); + swap(num_field_loc, other.num_field_loc); + swap(max_voffset_, other.max_voffset_); + swap(nested, other.nested); + swap(finished, other.finished); + swap(minalign_, other.minalign_); + swap(force_defaults_, other.force_defaults_); + swap(dedup_vtables_, other.dedup_vtables_); + swap(string_pool, other.string_pool); + } + + ~FlatBufferBuilder() { + if (string_pool) delete string_pool; + } + + void Reset() { + Clear(); // clear builder state + buf_.reset(); // deallocate buffer + } + + /// @brief Reset all the state in this FlatBufferBuilder so it can be reused + /// to construct another buffer. + void Clear() { + ClearOffsets(); + buf_.clear(); + nested = false; + finished = false; + minalign_ = 1; + if (string_pool) string_pool->clear(); + } + + /// @brief The current size of the serialized buffer, counting from the end. + /// @return Returns an `uoffset_t` with the current size of the buffer. + uoffset_t GetSize() const { return buf_.size(); } + + /// @brief Get the serialized buffer (after you call `Finish()`). + /// @return Returns an `uint8_t` pointer to the FlatBuffer data inside the + /// buffer. + uint8_t *GetBufferPointer() const { + Finished(); + return buf_.data(); + } + + /// @brief Get a pointer to an unfinished buffer. + /// @return Returns a `uint8_t` pointer to the unfinished buffer. + uint8_t *GetCurrentBufferPointer() const { return buf_.data(); } + + /// @brief Get the released pointer to the serialized buffer. + /// @warning Do NOT attempt to use this FlatBufferBuilder afterwards! + /// @return A `FlatBuffer` that owns the buffer and its allocator and + /// behaves similar to a `unique_ptr` with a deleter. + FLATBUFFERS_ATTRIBUTE(deprecated("use Release() instead")) + DetachedBuffer ReleaseBufferPointer() { + Finished(); + return buf_.release(); + } + + /// @brief Get the released DetachedBuffer. + /// @return A `DetachedBuffer` that owns the buffer and its allocator. + DetachedBuffer Release() { + Finished(); + return buf_.release(); + } + + /// @brief Get the released pointer to the serialized buffer. + /// @param size The size of the memory block containing + /// the serialized `FlatBuffer`. + /// @param offset The offset from the released pointer where the finished + /// `FlatBuffer` starts. + /// @return A raw pointer to the start of the memory block containing + /// the serialized `FlatBuffer`. + /// @remark If the allocator is owned, it gets deleted when the destructor is + /// called.. + uint8_t *ReleaseRaw(size_t &size, size_t &offset) { + Finished(); + return buf_.release_raw(size, offset); + } + + /// @brief get the minimum alignment this buffer needs to be accessed + /// properly. This is only known once all elements have been written (after + /// you call Finish()). You can use this information if you need to embed + /// a FlatBuffer in some other buffer, such that you can later read it + /// without first having to copy it into its own buffer. + size_t GetBufferMinAlignment() const { + Finished(); + return minalign_; + } + + /// @cond FLATBUFFERS_INTERNAL + void Finished() const { + // If you get this assert, you're attempting to get access a buffer + // which hasn't been finished yet. Be sure to call + // FlatBufferBuilder::Finish with your root table. + // If you really need to access an unfinished buffer, call + // GetCurrentBufferPointer instead. + FLATBUFFERS_ASSERT(finished); + } + /// @endcond + + /// @brief In order to save space, fields that are set to their default value + /// don't get serialized into the buffer. + /// @param[in] fd When set to `true`, always serializes default values that + /// are set. Optional fields which are not set explicitly, will still not be + /// serialized. + void ForceDefaults(bool fd) { force_defaults_ = fd; } + + /// @brief By default vtables are deduped in order to save space. + /// @param[in] dedup When set to `true`, dedup vtables. + void DedupVtables(bool dedup) { dedup_vtables_ = dedup; } + + /// @cond FLATBUFFERS_INTERNAL + void Pad(size_t num_bytes) { buf_.fill(num_bytes); } + + void TrackMinAlign(size_t elem_size) { + if (elem_size > minalign_) minalign_ = elem_size; + } + + void Align(size_t elem_size) { + TrackMinAlign(elem_size); + buf_.fill(PaddingBytes(buf_.size(), elem_size)); + } + + void PushFlatBuffer(const uint8_t *bytes, size_t size) { + PushBytes(bytes, size); + finished = true; + } + + void PushBytes(const uint8_t *bytes, size_t size) { buf_.push(bytes, size); } + + void PopBytes(size_t amount) { buf_.pop(amount); } + + template void AssertScalarT() { + // The code assumes power of 2 sizes and endian-swap-ability. + static_assert(flatbuffers::is_scalar::value, "T must be a scalar type"); + } + + // Write a single aligned scalar to the buffer + template uoffset_t PushElement(T element) { + AssertScalarT(); + T litle_endian_element = EndianScalar(element); + Align(sizeof(T)); + buf_.push_small(litle_endian_element); + return GetSize(); + } + + template uoffset_t PushElement(Offset off) { + // Special case for offsets: see ReferTo below. + return PushElement(ReferTo(off.o)); + } + + // When writing fields, we track where they are, so we can create correct + // vtables later. + void TrackField(voffset_t field, uoffset_t off) { + FieldLoc fl = { off, field }; + buf_.scratch_push_small(fl); + num_field_loc++; + max_voffset_ = (std::max)(max_voffset_, field); + } + + // Like PushElement, but additionally tracks the field this represents. + template void AddElement(voffset_t field, T e, T def) { + // We don't serialize values equal to the default. + if (IsTheSameAs(e, def) && !force_defaults_) return; + auto off = PushElement(e); + TrackField(field, off); + } + + template void AddElement(voffset_t field, T e) { + auto off = PushElement(e); + TrackField(field, off); + } + + template void AddOffset(voffset_t field, Offset off) { + if (off.IsNull()) return; // Don't store. + AddElement(field, ReferTo(off.o), static_cast(0)); + } + + template void AddStruct(voffset_t field, const T *structptr) { + if (!structptr) return; // Default, don't store. + Align(AlignOf()); + buf_.push_small(*structptr); + TrackField(field, GetSize()); + } + + void AddStructOffset(voffset_t field, uoffset_t off) { + TrackField(field, off); + } + + // Offsets initially are relative to the end of the buffer (downwards). + // This function converts them to be relative to the current location + // in the buffer (when stored here), pointing upwards. + uoffset_t ReferTo(uoffset_t off) { + // Align to ensure GetSize() below is correct. + Align(sizeof(uoffset_t)); + // Offset must refer to something already in buffer. + FLATBUFFERS_ASSERT(off && off <= GetSize()); + return GetSize() - off + static_cast(sizeof(uoffset_t)); + } + + void NotNested() { + // If you hit this, you're trying to construct a Table/Vector/String + // during the construction of its parent table (between the MyTableBuilder + // and table.Finish(). + // Move the creation of these sub-objects to above the MyTableBuilder to + // not get this assert. + // Ignoring this assert may appear to work in simple cases, but the reason + // it is here is that storing objects in-line may cause vtable offsets + // to not fit anymore. It also leads to vtable duplication. + FLATBUFFERS_ASSERT(!nested); + // If you hit this, fields were added outside the scope of a table. + FLATBUFFERS_ASSERT(!num_field_loc); + } + + // From generated code (or from the parser), we call StartTable/EndTable + // with a sequence of AddElement calls in between. + uoffset_t StartTable() { + NotNested(); + nested = true; + return GetSize(); + } + + // This finishes one serialized object by generating the vtable if it's a + // table, comparing it against existing vtables, and writing the + // resulting vtable offset. + uoffset_t EndTable(uoffset_t start) { + // If you get this assert, a corresponding StartTable wasn't called. + FLATBUFFERS_ASSERT(nested); + // Write the vtable offset, which is the start of any Table. + // We fill it's value later. + auto vtableoffsetloc = PushElement(0); + // Write a vtable, which consists entirely of voffset_t elements. + // It starts with the number of offsets, followed by a type id, followed + // by the offsets themselves. In reverse: + // Include space for the last offset and ensure empty tables have a + // minimum size. + max_voffset_ = + (std::max)(static_cast(max_voffset_ + sizeof(voffset_t)), + FieldIndexToOffset(0)); + buf_.fill_big(max_voffset_); + auto table_object_size = vtableoffsetloc - start; + // Vtable use 16bit offsets. + FLATBUFFERS_ASSERT(table_object_size < 0x10000); + WriteScalar(buf_.data() + sizeof(voffset_t), + static_cast(table_object_size)); + WriteScalar(buf_.data(), max_voffset_); + // Write the offsets into the table + for (auto it = buf_.scratch_end() - num_field_loc * sizeof(FieldLoc); + it < buf_.scratch_end(); it += sizeof(FieldLoc)) { + auto field_location = reinterpret_cast(it); + auto pos = static_cast(vtableoffsetloc - field_location->off); + // If this asserts, it means you've set a field twice. + FLATBUFFERS_ASSERT( + !ReadScalar(buf_.data() + field_location->id)); + WriteScalar(buf_.data() + field_location->id, pos); + } + ClearOffsets(); + auto vt1 = reinterpret_cast(buf_.data()); + auto vt1_size = ReadScalar(vt1); + auto vt_use = GetSize(); + // See if we already have generated a vtable with this exact same + // layout before. If so, make it point to the old one, remove this one. + if (dedup_vtables_) { + for (auto it = buf_.scratch_data(); it < buf_.scratch_end(); + it += sizeof(uoffset_t)) { + auto vt_offset_ptr = reinterpret_cast(it); + auto vt2 = reinterpret_cast(buf_.data_at(*vt_offset_ptr)); + auto vt2_size = ReadScalar(vt2); + if (vt1_size != vt2_size || 0 != memcmp(vt2, vt1, vt1_size)) continue; + vt_use = *vt_offset_ptr; + buf_.pop(GetSize() - vtableoffsetloc); + break; + } + } + // If this is a new vtable, remember it. + if (vt_use == GetSize()) { buf_.scratch_push_small(vt_use); } + // Fill the vtable offset we created above. + // The offset points from the beginning of the object to where the + // vtable is stored. + // Offsets default direction is downward in memory for future format + // flexibility (storing all vtables at the start of the file). + WriteScalar(buf_.data_at(vtableoffsetloc), + static_cast(vt_use) - + static_cast(vtableoffsetloc)); + + nested = false; + return vtableoffsetloc; + } + + FLATBUFFERS_ATTRIBUTE(deprecated("call the version above instead")) + uoffset_t EndTable(uoffset_t start, voffset_t /*numfields*/) { + return EndTable(start); + } + + // This checks a required field has been set in a given table that has + // just been constructed. + template void Required(Offset table, voffset_t field); + + uoffset_t StartStruct(size_t alignment) { + Align(alignment); + return GetSize(); + } + + uoffset_t EndStruct() { return GetSize(); } + + void ClearOffsets() { + buf_.scratch_pop(num_field_loc * sizeof(FieldLoc)); + num_field_loc = 0; + max_voffset_ = 0; + } + + // Aligns such that when "len" bytes are written, an object can be written + // after it with "alignment" without padding. + void PreAlign(size_t len, size_t alignment) { + TrackMinAlign(alignment); + buf_.fill(PaddingBytes(GetSize() + len, alignment)); + } + template void PreAlign(size_t len) { + AssertScalarT(); + PreAlign(len, sizeof(T)); + } + /// @endcond + + /// @brief Store a string in the buffer, which can contain any binary data. + /// @param[in] str A const char pointer to the data to be stored as a string. + /// @param[in] len The number of bytes that should be stored from `str`. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateString(const char *str, size_t len) { + NotNested(); + PreAlign(len + 1); // Always 0-terminated. + buf_.fill(1); + PushBytes(reinterpret_cast(str), len); + PushElement(static_cast(len)); + return Offset(GetSize()); + } + + /// @brief Store a string in the buffer, which is null-terminated. + /// @param[in] str A const char pointer to a C-string to add to the buffer. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateString(const char *str) { + return CreateString(str, strlen(str)); + } + + /// @brief Store a string in the buffer, which is null-terminated. + /// @param[in] str A char pointer to a C-string to add to the buffer. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateString(char *str) { + return CreateString(str, strlen(str)); + } + + /// @brief Store a string in the buffer, which can contain any binary data. + /// @param[in] str A const reference to a std::string to store in the buffer. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateString(const std::string &str) { + return CreateString(str.c_str(), str.length()); + } + + // clang-format off + #ifdef FLATBUFFERS_HAS_STRING_VIEW + /// @brief Store a string in the buffer, which can contain any binary data. + /// @param[in] str A const string_view to copy in to the buffer. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateString(flatbuffers::string_view str) { + return CreateString(str.data(), str.size()); + } + #endif // FLATBUFFERS_HAS_STRING_VIEW + // clang-format on + + /// @brief Store a string in the buffer, which can contain any binary data. + /// @param[in] str A const pointer to a `String` struct to add to the buffer. + /// @return Returns the offset in the buffer where the string starts + Offset CreateString(const String *str) { + return str ? CreateString(str->c_str(), str->size()) : 0; + } + + /// @brief Store a string in the buffer, which can contain any binary data. + /// @param[in] str A const reference to a std::string like type with support + /// of T::c_str() and T::length() to store in the buffer. + /// @return Returns the offset in the buffer where the string starts. + template Offset CreateString(const T &str) { + return CreateString(str.c_str(), str.length()); + } + + /// @brief Store a string in the buffer, which can contain any binary data. + /// If a string with this exact contents has already been serialized before, + /// instead simply returns the offset of the existing string. + /// @param[in] str A const char pointer to the data to be stored as a string. + /// @param[in] len The number of bytes that should be stored from `str`. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateSharedString(const char *str, size_t len) { + if (!string_pool) + string_pool = new StringOffsetMap(StringOffsetCompare(buf_)); + auto size_before_string = buf_.size(); + // Must first serialize the string, since the set is all offsets into + // buffer. + auto off = CreateString(str, len); + auto it = string_pool->find(off); + // If it exists we reuse existing serialized data! + if (it != string_pool->end()) { + // We can remove the string we serialized. + buf_.pop(buf_.size() - size_before_string); + return *it; + } + // Record this string for future use. + string_pool->insert(off); + return off; + } + + /// @brief Store a string in the buffer, which null-terminated. + /// If a string with this exact contents has already been serialized before, + /// instead simply returns the offset of the existing string. + /// @param[in] str A const char pointer to a C-string to add to the buffer. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateSharedString(const char *str) { + return CreateSharedString(str, strlen(str)); + } + + /// @brief Store a string in the buffer, which can contain any binary data. + /// If a string with this exact contents has already been serialized before, + /// instead simply returns the offset of the existing string. + /// @param[in] str A const reference to a std::string to store in the buffer. + /// @return Returns the offset in the buffer where the string starts. + Offset CreateSharedString(const std::string &str) { + return CreateSharedString(str.c_str(), str.length()); + } + + /// @brief Store a string in the buffer, which can contain any binary data. + /// If a string with this exact contents has already been serialized before, + /// instead simply returns the offset of the existing string. + /// @param[in] str A const pointer to a `String` struct to add to the buffer. + /// @return Returns the offset in the buffer where the string starts + Offset CreateSharedString(const String *str) { + return CreateSharedString(str->c_str(), str->size()); + } + + /// @cond FLATBUFFERS_INTERNAL + uoffset_t EndVector(size_t len) { + FLATBUFFERS_ASSERT(nested); // Hit if no corresponding StartVector. + nested = false; + return PushElement(static_cast(len)); + } + + void StartVector(size_t len, size_t elemsize) { + NotNested(); + nested = true; + PreAlign(len * elemsize); + PreAlign(len * elemsize, elemsize); // Just in case elemsize > uoffset_t. + } + + // Call this right before StartVector/CreateVector if you want to force the + // alignment to be something different than what the element size would + // normally dictate. + // This is useful when storing a nested_flatbuffer in a vector of bytes, + // or when storing SIMD floats, etc. + void ForceVectorAlignment(size_t len, size_t elemsize, size_t alignment) { + PreAlign(len * elemsize, alignment); + } + + // Similar to ForceVectorAlignment but for String fields. + void ForceStringAlignment(size_t len, size_t alignment) { + PreAlign((len + 1) * sizeof(char), alignment); + } + + /// @endcond + + /// @brief Serialize an array into a FlatBuffer `vector`. + /// @tparam T The data type of the array elements. + /// @param[in] v A pointer to the array of type `T` to serialize into the + /// buffer as a `vector`. + /// @param[in] len The number of elements to serialize. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template Offset> CreateVector(const T *v, size_t len) { + // If this assert hits, you're specifying a template argument that is + // causing the wrong overload to be selected, remove it. + AssertScalarT(); + StartVector(len, sizeof(T)); + if (len == 0) { + return Offset>(EndVector(len)); + } + // clang-format off + #if FLATBUFFERS_LITTLEENDIAN + PushBytes(reinterpret_cast(v), len * sizeof(T)); + #else + if (sizeof(T) == 1) { + PushBytes(reinterpret_cast(v), len); + } else { + for (auto i = len; i > 0; ) { + PushElement(v[--i]); + } + } + #endif + // clang-format on + return Offset>(EndVector(len)); + } + + template + Offset>> CreateVector(const Offset *v, size_t len) { + StartVector(len, sizeof(Offset)); + for (auto i = len; i > 0;) { PushElement(v[--i]); } + return Offset>>(EndVector(len)); + } + + /// @brief Serialize a `std::vector` into a FlatBuffer `vector`. + /// @tparam T The data type of the `std::vector` elements. + /// @param v A const reference to the `std::vector` to serialize into the + /// buffer as a `vector`. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template Offset> CreateVector(const std::vector &v) { + return CreateVector(data(v), v.size()); + } + + // vector may be implemented using a bit-set, so we can't access it as + // an array. Instead, read elements manually. + // Background: https://isocpp.org/blog/2012/11/on-vectorbool + Offset> CreateVector(const std::vector &v) { + StartVector(v.size(), sizeof(uint8_t)); + for (auto i = v.size(); i > 0;) { + PushElement(static_cast(v[--i])); + } + return Offset>(EndVector(v.size())); + } + + // clang-format off + #ifndef FLATBUFFERS_CPP98_STL + /// @brief Serialize values returned by a function into a FlatBuffer `vector`. + /// This is a convenience function that takes care of iteration for you. + /// @tparam T The data type of the `std::vector` elements. + /// @param f A function that takes the current iteration 0..vector_size-1 and + /// returns any type that you can construct a FlatBuffers vector out of. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template Offset> CreateVector(size_t vector_size, + const std::function &f) { + std::vector elems(vector_size); + for (size_t i = 0; i < vector_size; i++) elems[i] = f(i); + return CreateVector(elems); + } + #endif + // clang-format on + + /// @brief Serialize values returned by a function into a FlatBuffer `vector`. + /// This is a convenience function that takes care of iteration for you. + /// @tparam T The data type of the `std::vector` elements. + /// @param f A function that takes the current iteration 0..vector_size-1, + /// and the state parameter returning any type that you can construct a + /// FlatBuffers vector out of. + /// @param state State passed to f. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVector(size_t vector_size, F f, S *state) { + std::vector elems(vector_size); + for (size_t i = 0; i < vector_size; i++) elems[i] = f(i, state); + return CreateVector(elems); + } + + /// @brief Serialize a `std::vector` into a FlatBuffer `vector`. + /// This is a convenience function for a common case. + /// @param v A const reference to the `std::vector` to serialize into the + /// buffer as a `vector`. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + Offset>> CreateVectorOfStrings( + const std::vector &v) { + std::vector> offsets(v.size()); + for (size_t i = 0; i < v.size(); i++) offsets[i] = CreateString(v[i]); + return CreateVector(offsets); + } + + /// @brief Serialize an array of structs into a FlatBuffer `vector`. + /// @tparam T The data type of the struct array elements. + /// @param[in] v A pointer to the array of type `T` to serialize into the + /// buffer as a `vector`. + /// @param[in] len The number of elements to serialize. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfStructs(const T *v, size_t len) { + StartVector(len * sizeof(T) / AlignOf(), AlignOf()); + PushBytes(reinterpret_cast(v), sizeof(T) * len); + return Offset>(EndVector(len)); + } + + /// @brief Serialize an array of native structs into a FlatBuffer `vector`. + /// @tparam T The data type of the struct array elements. + /// @tparam S The data type of the native struct array elements. + /// @param[in] v A pointer to the array of type `S` to serialize into the + /// buffer as a `vector`. + /// @param[in] len The number of elements to serialize. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfNativeStructs(const S *v, + size_t len) { + extern T Pack(const S &); + std::vector vv(len); + std::transform(v, v + len, vv.begin(), Pack); + return CreateVectorOfStructs(data(vv), vv.size()); + } + + // clang-format off + #ifndef FLATBUFFERS_CPP98_STL + /// @brief Serialize an array of structs into a FlatBuffer `vector`. + /// @tparam T The data type of the struct array elements. + /// @param[in] filler A function that takes the current iteration 0..vector_size-1 + /// and a pointer to the struct that must be filled. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + /// This is mostly useful when flatbuffers are generated with mutation + /// accessors. + template Offset> CreateVectorOfStructs( + size_t vector_size, const std::function &filler) { + T* structs = StartVectorOfStructs(vector_size); + for (size_t i = 0; i < vector_size; i++) { + filler(i, structs); + structs++; + } + return EndVectorOfStructs(vector_size); + } + #endif + // clang-format on + + /// @brief Serialize an array of structs into a FlatBuffer `vector`. + /// @tparam T The data type of the struct array elements. + /// @param[in] f A function that takes the current iteration 0..vector_size-1, + /// a pointer to the struct that must be filled and the state argument. + /// @param[in] state Arbitrary state to pass to f. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + /// This is mostly useful when flatbuffers are generated with mutation + /// accessors. + template + Offset> CreateVectorOfStructs(size_t vector_size, F f, + S *state) { + T *structs = StartVectorOfStructs(vector_size); + for (size_t i = 0; i < vector_size; i++) { + f(i, structs, state); + structs++; + } + return EndVectorOfStructs(vector_size); + } + + /// @brief Serialize a `std::vector` of structs into a FlatBuffer `vector`. + /// @tparam T The data type of the `std::vector` struct elements. + /// @param[in] v A const reference to the `std::vector` of structs to + /// serialize into the buffer as a `vector`. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfStructs( + const std::vector &v) { + return CreateVectorOfStructs(data(v), v.size()); + } + + /// @brief Serialize a `std::vector` of native structs into a FlatBuffer + /// `vector`. + /// @tparam T The data type of the `std::vector` struct elements. + /// @tparam S The data type of the `std::vector` native struct elements. + /// @param[in] v A const reference to the `std::vector` of structs to + /// serialize into the buffer as a `vector`. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfNativeStructs( + const std::vector &v) { + return CreateVectorOfNativeStructs(data(v), v.size()); + } + + /// @cond FLATBUFFERS_INTERNAL + template struct StructKeyComparator { + bool operator()(const T &a, const T &b) const { + return a.KeyCompareLessThan(&b); + } + + FLATBUFFERS_DELETE_FUNC( + StructKeyComparator &operator=(const StructKeyComparator &)) + }; + /// @endcond + + /// @brief Serialize a `std::vector` of structs into a FlatBuffer `vector` + /// in sorted order. + /// @tparam T The data type of the `std::vector` struct elements. + /// @param[in] v A const reference to the `std::vector` of structs to + /// serialize into the buffer as a `vector`. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfSortedStructs(std::vector *v) { + return CreateVectorOfSortedStructs(data(*v), v->size()); + } + + /// @brief Serialize a `std::vector` of native structs into a FlatBuffer + /// `vector` in sorted order. + /// @tparam T The data type of the `std::vector` struct elements. + /// @tparam S The data type of the `std::vector` native struct elements. + /// @param[in] v A const reference to the `std::vector` of structs to + /// serialize into the buffer as a `vector`. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfSortedNativeStructs( + std::vector *v) { + return CreateVectorOfSortedNativeStructs(data(*v), v->size()); + } + + /// @brief Serialize an array of structs into a FlatBuffer `vector` in sorted + /// order. + /// @tparam T The data type of the struct array elements. + /// @param[in] v A pointer to the array of type `T` to serialize into the + /// buffer as a `vector`. + /// @param[in] len The number of elements to serialize. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfSortedStructs(T *v, size_t len) { + std::sort(v, v + len, StructKeyComparator()); + return CreateVectorOfStructs(v, len); + } + + /// @brief Serialize an array of native structs into a FlatBuffer `vector` in + /// sorted order. + /// @tparam T The data type of the struct array elements. + /// @tparam S The data type of the native struct array elements. + /// @param[in] v A pointer to the array of type `S` to serialize into the + /// buffer as a `vector`. + /// @param[in] len The number of elements to serialize. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset> CreateVectorOfSortedNativeStructs(S *v, + size_t len) { + extern T Pack(const S &); + typedef T (*Pack_t)(const S &); + std::vector vv(len); + std::transform(v, v + len, vv.begin(), static_cast(Pack)); + return CreateVectorOfSortedStructs(vv, len); + } + + /// @cond FLATBUFFERS_INTERNAL + template struct TableKeyComparator { + TableKeyComparator(vector_downward &buf) : buf_(buf) {} + TableKeyComparator(const TableKeyComparator &other) : buf_(other.buf_) {} + bool operator()(const Offset &a, const Offset &b) const { + auto table_a = reinterpret_cast(buf_.data_at(a.o)); + auto table_b = reinterpret_cast(buf_.data_at(b.o)); + return table_a->KeyCompareLessThan(table_b); + } + vector_downward &buf_; + + private: + FLATBUFFERS_DELETE_FUNC(TableKeyComparator &operator=(const TableKeyComparator &other)) + }; + /// @endcond + + /// @brief Serialize an array of `table` offsets as a `vector` in the buffer + /// in sorted order. + /// @tparam T The data type that the offset refers to. + /// @param[in] v An array of type `Offset` that contains the `table` + /// offsets to store in the buffer in sorted order. + /// @param[in] len The number of elements to store in the `vector`. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset>> CreateVectorOfSortedTables(Offset *v, + size_t len) { + std::sort(v, v + len, TableKeyComparator(buf_)); + return CreateVector(v, len); + } + + /// @brief Serialize an array of `table` offsets as a `vector` in the buffer + /// in sorted order. + /// @tparam T The data type that the offset refers to. + /// @param[in] v An array of type `Offset` that contains the `table` + /// offsets to store in the buffer in sorted order. + /// @return Returns a typed `Offset` into the serialized data indicating + /// where the vector is stored. + template + Offset>> CreateVectorOfSortedTables( + std::vector> *v) { + return CreateVectorOfSortedTables(data(*v), v->size()); + } + + /// @brief Specialized version of `CreateVector` for non-copying use cases. + /// Write the data any time later to the returned buffer pointer `buf`. + /// @param[in] len The number of elements to store in the `vector`. + /// @param[in] elemsize The size of each element in the `vector`. + /// @param[out] buf A pointer to a `uint8_t` pointer that can be + /// written to at a later time to serialize the data into a `vector` + /// in the buffer. + uoffset_t CreateUninitializedVector(size_t len, size_t elemsize, + uint8_t **buf) { + NotNested(); + StartVector(len, elemsize); + buf_.make_space(len * elemsize); + auto vec_start = GetSize(); + auto vec_end = EndVector(len); + *buf = buf_.data_at(vec_start); + return vec_end; + } + + /// @brief Specialized version of `CreateVector` for non-copying use cases. + /// Write the data any time later to the returned buffer pointer `buf`. + /// @tparam T The data type of the data that will be stored in the buffer + /// as a `vector`. + /// @param[in] len The number of elements to store in the `vector`. + /// @param[out] buf A pointer to a pointer of type `T` that can be + /// written to at a later time to serialize the data into a `vector` + /// in the buffer. + template + Offset> CreateUninitializedVector(size_t len, T **buf) { + AssertScalarT(); + return CreateUninitializedVector(len, sizeof(T), + reinterpret_cast(buf)); + } + + template + Offset> CreateUninitializedVectorOfStructs(size_t len, + T **buf) { + return CreateUninitializedVector(len, sizeof(T), + reinterpret_cast(buf)); + } + + // @brief Create a vector of scalar type T given as input a vector of scalar + // type U, useful with e.g. pre "enum class" enums, or any existing scalar + // data of the wrong type. + template + Offset> CreateVectorScalarCast(const U *v, size_t len) { + AssertScalarT(); + AssertScalarT(); + StartVector(len, sizeof(T)); + for (auto i = len; i > 0;) { PushElement(static_cast(v[--i])); } + return Offset>(EndVector(len)); + } + + /// @brief Write a struct by itself, typically to be part of a union. + template Offset CreateStruct(const T &structobj) { + NotNested(); + Align(AlignOf()); + buf_.push_small(structobj); + return Offset(GetSize()); + } + + /// @brief The length of a FlatBuffer file header. + static const size_t kFileIdentifierLength = 4; + + /// @brief Finish serializing a buffer by writing the root offset. + /// @param[in] file_identifier If a `file_identifier` is given, the buffer + /// will be prefixed with a standard FlatBuffers file header. + template + void Finish(Offset root, const char *file_identifier = nullptr) { + Finish(root.o, file_identifier, false); + } + + /// @brief Finish a buffer with a 32 bit size field pre-fixed (size of the + /// buffer following the size field). These buffers are NOT compatible + /// with standard buffers created by Finish, i.e. you can't call GetRoot + /// on them, you have to use GetSizePrefixedRoot instead. + /// All >32 bit quantities in this buffer will be aligned when the whole + /// size pre-fixed buffer is aligned. + /// These kinds of buffers are useful for creating a stream of FlatBuffers. + template + void FinishSizePrefixed(Offset root, + const char *file_identifier = nullptr) { + Finish(root.o, file_identifier, true); + } + + void SwapBufAllocator(FlatBufferBuilder &other) { + buf_.swap_allocator(other.buf_); + } + + protected: + // You shouldn't really be copying instances of this class. + FlatBufferBuilder(const FlatBufferBuilder &); + FlatBufferBuilder &operator=(const FlatBufferBuilder &); + + void Finish(uoffset_t root, const char *file_identifier, bool size_prefix) { + NotNested(); + buf_.clear_scratch(); + // This will cause the whole buffer to be aligned. + PreAlign((size_prefix ? sizeof(uoffset_t) : 0) + sizeof(uoffset_t) + + (file_identifier ? kFileIdentifierLength : 0), + minalign_); + if (file_identifier) { + FLATBUFFERS_ASSERT(strlen(file_identifier) == kFileIdentifierLength); + PushBytes(reinterpret_cast(file_identifier), + kFileIdentifierLength); + } + PushElement(ReferTo(root)); // Location of root. + if (size_prefix) { PushElement(GetSize()); } + finished = true; + } + + struct FieldLoc { + uoffset_t off; + voffset_t id; + }; + + vector_downward buf_; + + // Accumulating offsets of table members while it is being built. + // We store these in the scratch pad of buf_, after the vtable offsets. + uoffset_t num_field_loc; + // Track how much of the vtable is in use, so we can output the most compact + // possible vtable. + voffset_t max_voffset_; + + // Ensure objects are not nested. + bool nested; + + // Ensure the buffer is finished before it is being accessed. + bool finished; + + size_t minalign_; + + bool force_defaults_; // Serialize values equal to their defaults anyway. + + bool dedup_vtables_; + + struct StringOffsetCompare { + StringOffsetCompare(const vector_downward &buf) : buf_(&buf) {} + bool operator()(const Offset &a, const Offset &b) const { + auto stra = reinterpret_cast(buf_->data_at(a.o)); + auto strb = reinterpret_cast(buf_->data_at(b.o)); + return StringLessThan(stra->data(), stra->size(), strb->data(), + strb->size()); + } + const vector_downward *buf_; + }; + + // For use with CreateSharedString. Instantiated on first use only. + typedef std::set, StringOffsetCompare> StringOffsetMap; + StringOffsetMap *string_pool; + + private: + // Allocates space for a vector of structures. + // Must be completed with EndVectorOfStructs(). + template T *StartVectorOfStructs(size_t vector_size) { + StartVector(vector_size * sizeof(T) / AlignOf(), AlignOf()); + return reinterpret_cast(buf_.make_space(vector_size * sizeof(T))); + } + + // End the vector of structues in the flatbuffers. + // Vector should have previously be started with StartVectorOfStructs(). + template + Offset> EndVectorOfStructs(size_t vector_size) { + return Offset>(EndVector(vector_size)); + } +}; +/// @} + +/// @cond FLATBUFFERS_INTERNAL +// Helpers to get a typed pointer to the root object contained in the buffer. +template T *GetMutableRoot(void *buf) { + EndianCheck(); + return reinterpret_cast( + reinterpret_cast(buf) + + EndianScalar(*reinterpret_cast(buf))); +} + +template const T *GetRoot(const void *buf) { + return GetMutableRoot(const_cast(buf)); +} + +template const T *GetSizePrefixedRoot(const void *buf) { + return GetRoot(reinterpret_cast(buf) + sizeof(uoffset_t)); +} + +/// Helpers to get a typed pointer to objects that are currently being built. +/// @warning Creating new objects will lead to reallocations and invalidates +/// the pointer! +template +T *GetMutableTemporaryPointer(FlatBufferBuilder &fbb, Offset offset) { + return reinterpret_cast(fbb.GetCurrentBufferPointer() + fbb.GetSize() - + offset.o); +} + +template +const T *GetTemporaryPointer(FlatBufferBuilder &fbb, Offset offset) { + return GetMutableTemporaryPointer(fbb, offset); +} + +/// @brief Get a pointer to the the file_identifier section of the buffer. +/// @return Returns a const char pointer to the start of the file_identifier +/// characters in the buffer. The returned char * has length +/// 'flatbuffers::FlatBufferBuilder::kFileIdentifierLength'. +/// This function is UNDEFINED for FlatBuffers whose schema does not include +/// a file_identifier (likely points at padding or the start of a the root +/// vtable). +inline const char *GetBufferIdentifier(const void *buf, + bool size_prefixed = false) { + return reinterpret_cast(buf) + + ((size_prefixed) ? 2 * sizeof(uoffset_t) : sizeof(uoffset_t)); +} + +// Helper to see if the identifier in a buffer has the expected value. +inline bool BufferHasIdentifier(const void *buf, const char *identifier, + bool size_prefixed = false) { + return strncmp(GetBufferIdentifier(buf, size_prefixed), identifier, + FlatBufferBuilder::kFileIdentifierLength) == 0; +} + +// Helper class to verify the integrity of a FlatBuffer +class Verifier FLATBUFFERS_FINAL_CLASS { + public: + Verifier(const uint8_t *buf, size_t buf_len, uoffset_t _max_depth = 64, + uoffset_t _max_tables = 1000000, bool _check_alignment = true) + : buf_(buf), + size_(buf_len), + depth_(0), + max_depth_(_max_depth), + num_tables_(0), + max_tables_(_max_tables), + upper_bound_(0), + check_alignment_(_check_alignment) { + FLATBUFFERS_ASSERT(size_ < FLATBUFFERS_MAX_BUFFER_SIZE); + } + + // Central location where any verification failures register. + bool Check(bool ok) const { + // clang-format off + #ifdef FLATBUFFERS_DEBUG_VERIFICATION_FAILURE + FLATBUFFERS_ASSERT(ok); + #endif + #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE + if (!ok) + upper_bound_ = 0; + #endif + // clang-format on + return ok; + } + + // Verify any range within the buffer. + bool Verify(size_t elem, size_t elem_len) const { + // clang-format off + #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE + auto upper_bound = elem + elem_len; + if (upper_bound_ < upper_bound) + upper_bound_ = upper_bound; + #endif + // clang-format on + return Check(elem_len < size_ && elem <= size_ - elem_len); + } + + template bool VerifyAlignment(size_t elem) const { + return Check((elem & (sizeof(T) - 1)) == 0 || !check_alignment_); + } + + // Verify a range indicated by sizeof(T). + template bool Verify(size_t elem) const { + return VerifyAlignment(elem) && Verify(elem, sizeof(T)); + } + + bool VerifyFromPointer(const uint8_t *p, size_t len) { + auto o = static_cast(p - buf_); + return Verify(o, len); + } + + // Verify relative to a known-good base pointer. + bool Verify(const uint8_t *base, voffset_t elem_off, size_t elem_len) const { + return Verify(static_cast(base - buf_) + elem_off, elem_len); + } + + template + bool Verify(const uint8_t *base, voffset_t elem_off) const { + return Verify(static_cast(base - buf_) + elem_off, sizeof(T)); + } + + // Verify a pointer (may be NULL) of a table type. + template bool VerifyTable(const T *table) { + return !table || table->Verify(*this); + } + + // Verify a pointer (may be NULL) of any vector type. + template bool VerifyVector(const Vector *vec) const { + return !vec || VerifyVectorOrString(reinterpret_cast(vec), + sizeof(T)); + } + + // Verify a pointer (may be NULL) of a vector to struct. + template bool VerifyVector(const Vector *vec) const { + return VerifyVector(reinterpret_cast *>(vec)); + } + + // Verify a pointer (may be NULL) to string. + bool VerifyString(const String *str) const { + size_t end; + return !str || (VerifyVectorOrString(reinterpret_cast(str), + 1, &end) && + Verify(end, 1) && // Must have terminator + Check(buf_[end] == '\0')); // Terminating byte must be 0. + } + + // Common code between vectors and strings. + bool VerifyVectorOrString(const uint8_t *vec, size_t elem_size, + size_t *end = nullptr) const { + auto veco = static_cast(vec - buf_); + // Check we can read the size field. + if (!Verify(veco)) return false; + // Check the whole array. If this is a string, the byte past the array + // must be 0. + auto size = ReadScalar(vec); + auto max_elems = FLATBUFFERS_MAX_BUFFER_SIZE / elem_size; + if (!Check(size < max_elems)) + return false; // Protect against byte_size overflowing. + auto byte_size = sizeof(size) + elem_size * size; + if (end) *end = veco + byte_size; + return Verify(veco, byte_size); + } + + // Special case for string contents, after the above has been called. + bool VerifyVectorOfStrings(const Vector> *vec) const { + if (vec) { + for (uoffset_t i = 0; i < vec->size(); i++) { + if (!VerifyString(vec->Get(i))) return false; + } + } + return true; + } + + // Special case for table contents, after the above has been called. + template bool VerifyVectorOfTables(const Vector> *vec) { + if (vec) { + for (uoffset_t i = 0; i < vec->size(); i++) { + if (!vec->Get(i)->Verify(*this)) return false; + } + } + return true; + } + + __supress_ubsan__("unsigned-integer-overflow") bool VerifyTableStart( + const uint8_t *table) { + // Check the vtable offset. + auto tableo = static_cast(table - buf_); + if (!Verify(tableo)) return false; + // This offset may be signed, but doing the subtraction unsigned always + // gives the result we want. + auto vtableo = tableo - static_cast(ReadScalar(table)); + // Check the vtable size field, then check vtable fits in its entirety. + return VerifyComplexity() && Verify(vtableo) && + VerifyAlignment(ReadScalar(buf_ + vtableo)) && + Verify(vtableo, ReadScalar(buf_ + vtableo)); + } + + template + bool VerifyBufferFromStart(const char *identifier, size_t start) { + if (identifier && !Check((size_ >= 2 * sizeof(flatbuffers::uoffset_t) && + BufferHasIdentifier(buf_ + start, identifier)))) { + return false; + } + + // Call T::Verify, which must be in the generated code for this type. + auto o = VerifyOffset(start); + return o && reinterpret_cast(buf_ + start + o)->Verify(*this) + // clang-format off + #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE + && GetComputedSize() + #endif + ; + // clang-format on + } + + // Verify this whole buffer, starting with root type T. + template bool VerifyBuffer() { return VerifyBuffer(nullptr); } + + template bool VerifyBuffer(const char *identifier) { + return VerifyBufferFromStart(identifier, 0); + } + + template bool VerifySizePrefixedBuffer(const char *identifier) { + return Verify(0U) && + ReadScalar(buf_) == size_ - sizeof(uoffset_t) && + VerifyBufferFromStart(identifier, sizeof(uoffset_t)); + } + + uoffset_t VerifyOffset(size_t start) const { + if (!Verify(start)) return 0; + auto o = ReadScalar(buf_ + start); + // May not point to itself. + if (!Check(o != 0)) return 0; + // Can't wrap around / buffers are max 2GB. + if (!Check(static_cast(o) >= 0)) return 0; + // Must be inside the buffer to create a pointer from it (pointer outside + // buffer is UB). + if (!Verify(start + o, 1)) return 0; + return o; + } + + uoffset_t VerifyOffset(const uint8_t *base, voffset_t start) const { + return VerifyOffset(static_cast(base - buf_) + start); + } + + // Called at the start of a table to increase counters measuring data + // structure depth and amount, and possibly bails out with false if + // limits set by the constructor have been hit. Needs to be balanced + // with EndTable(). + bool VerifyComplexity() { + depth_++; + num_tables_++; + return Check(depth_ <= max_depth_ && num_tables_ <= max_tables_); + } + + // Called at the end of a table to pop the depth count. + bool EndTable() { + depth_--; + return true; + } + + // Returns the message size in bytes + size_t GetComputedSize() const { + // clang-format off + #ifdef FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE + uintptr_t size = upper_bound_; + // Align the size to uoffset_t + size = (size - 1 + sizeof(uoffset_t)) & ~(sizeof(uoffset_t) - 1); + return (size > size_) ? 0 : size; + #else + // Must turn on FLATBUFFERS_TRACK_VERIFIER_BUFFER_SIZE for this to work. + (void)upper_bound_; + FLATBUFFERS_ASSERT(false); + return 0; + #endif + // clang-format on + } + + private: + const uint8_t *buf_; + size_t size_; + uoffset_t depth_; + uoffset_t max_depth_; + uoffset_t num_tables_; + uoffset_t max_tables_; + mutable size_t upper_bound_; + bool check_alignment_; +}; + +// Convenient way to bundle a buffer and its length, to pass it around +// typed by its root. +// A BufferRef does not own its buffer. +struct BufferRefBase {}; // for std::is_base_of +template struct BufferRef : BufferRefBase { + BufferRef() : buf(nullptr), len(0), must_free(false) {} + BufferRef(uint8_t *_buf, uoffset_t _len) + : buf(_buf), len(_len), must_free(false) {} + + ~BufferRef() { + if (must_free) free(buf); + } + + const T *GetRoot() const { return flatbuffers::GetRoot(buf); } + + bool Verify() { + Verifier verifier(buf, len); + return verifier.VerifyBuffer(nullptr); + } + + uint8_t *buf; + uoffset_t len; + bool must_free; +}; + +// "structs" are flat structures that do not have an offset table, thus +// always have all members present and do not support forwards/backwards +// compatible extensions. + +class Struct FLATBUFFERS_FINAL_CLASS { + public: + template T GetField(uoffset_t o) const { + return ReadScalar(&data_[o]); + } + + template T GetStruct(uoffset_t o) const { + return reinterpret_cast(&data_[o]); + } + + const uint8_t *GetAddressOf(uoffset_t o) const { return &data_[o]; } + uint8_t *GetAddressOf(uoffset_t o) { return &data_[o]; } + + private: + // private constructor & copy constructor: you obtain instances of this + // class by pointing to existing data only + Struct(); + Struct(const Struct &); + Struct &operator=(const Struct &); + + uint8_t data_[1]; +}; + +// "tables" use an offset table (possibly shared) that allows fields to be +// omitted and added at will, but uses an extra indirection to read. +class Table { + public: + const uint8_t *GetVTable() const { + return data_ - ReadScalar(data_); + } + + // This gets the field offset for any of the functions below it, or 0 + // if the field was not present. + voffset_t GetOptionalFieldOffset(voffset_t field) const { + // The vtable offset is always at the start. + auto vtable = GetVTable(); + // The first element is the size of the vtable (fields + type id + itself). + auto vtsize = ReadScalar(vtable); + // If the field we're accessing is outside the vtable, we're reading older + // data, so it's the same as if the offset was 0 (not present). + return field < vtsize ? ReadScalar(vtable + field) : 0; + } + + template T GetField(voffset_t field, T defaultval) const { + auto field_offset = GetOptionalFieldOffset(field); + return field_offset ? ReadScalar(data_ + field_offset) : defaultval; + } + + template P GetPointer(voffset_t field) { + auto field_offset = GetOptionalFieldOffset(field); + auto p = data_ + field_offset; + return field_offset ? reinterpret_cast

(p + ReadScalar(p)) + : nullptr; + } + template P GetPointer(voffset_t field) const { + return const_cast(this)->GetPointer

(field); + } + + template P GetStruct(voffset_t field) const { + auto field_offset = GetOptionalFieldOffset(field); + auto p = const_cast(data_ + field_offset); + return field_offset ? reinterpret_cast

(p) : nullptr; + } + + template + flatbuffers::Optional GetOptional(voffset_t field) const { + auto field_offset = GetOptionalFieldOffset(field); + auto p = data_ + field_offset; + return field_offset ? Optional(static_cast(ReadScalar(p))) + : Optional(); + } + + template bool SetField(voffset_t field, T val, T def) { + auto field_offset = GetOptionalFieldOffset(field); + if (!field_offset) return IsTheSameAs(val, def); + WriteScalar(data_ + field_offset, val); + return true; + } + template bool SetField(voffset_t field, T val) { + auto field_offset = GetOptionalFieldOffset(field); + if (!field_offset) return false; + WriteScalar(data_ + field_offset, val); + return true; + } + + bool SetPointer(voffset_t field, const uint8_t *val) { + auto field_offset = GetOptionalFieldOffset(field); + if (!field_offset) return false; + WriteScalar(data_ + field_offset, + static_cast(val - (data_ + field_offset))); + return true; + } + + uint8_t *GetAddressOf(voffset_t field) { + auto field_offset = GetOptionalFieldOffset(field); + return field_offset ? data_ + field_offset : nullptr; + } + const uint8_t *GetAddressOf(voffset_t field) const { + return const_cast

(this)->GetAddressOf(field); + } + + bool CheckField(voffset_t field) const { + return GetOptionalFieldOffset(field) != 0; + } + + // Verify the vtable of this table. + // Call this once per table, followed by VerifyField once per field. + bool VerifyTableStart(Verifier &verifier) const { + return verifier.VerifyTableStart(data_); + } + + // Verify a particular field. + template + bool VerifyField(const Verifier &verifier, voffset_t field) const { + // Calling GetOptionalFieldOffset should be safe now thanks to + // VerifyTable(). + auto field_offset = GetOptionalFieldOffset(field); + // Check the actual field. + return !field_offset || verifier.Verify(data_, field_offset); + } + + // VerifyField for required fields. + template + bool VerifyFieldRequired(const Verifier &verifier, voffset_t field) const { + auto field_offset = GetOptionalFieldOffset(field); + return verifier.Check(field_offset != 0) && + verifier.Verify(data_, field_offset); + } + + // Versions for offsets. + bool VerifyOffset(const Verifier &verifier, voffset_t field) const { + auto field_offset = GetOptionalFieldOffset(field); + return !field_offset || verifier.VerifyOffset(data_, field_offset); + } + + bool VerifyOffsetRequired(const Verifier &verifier, voffset_t field) const { + auto field_offset = GetOptionalFieldOffset(field); + return verifier.Check(field_offset != 0) && + verifier.VerifyOffset(data_, field_offset); + } + + private: + // private constructor & copy constructor: you obtain instances of this + // class by pointing to existing data only + Table(); + Table(const Table &other); + Table &operator=(const Table &); + + uint8_t data_[1]; +}; + +// This specialization allows avoiding warnings like: +// MSVC C4800: type: forcing value to bool 'true' or 'false'. +template<> +inline flatbuffers::Optional Table::GetOptional( + voffset_t field) const { + auto field_offset = GetOptionalFieldOffset(field); + auto p = data_ + field_offset; + return field_offset ? Optional(ReadScalar(p) != 0) + : Optional(); +} + +template +void FlatBufferBuilder::Required(Offset table, voffset_t field) { + auto table_ptr = reinterpret_cast(buf_.data_at(table.o)); + bool ok = table_ptr->GetOptionalFieldOffset(field) != 0; + // If this fails, the caller will show what field needs to be set. + FLATBUFFERS_ASSERT(ok); + (void)ok; +} + +/// @brief This can compute the start of a FlatBuffer from a root pointer, i.e. +/// it is the opposite transformation of GetRoot(). +/// This may be useful if you want to pass on a root and have the recipient +/// delete the buffer afterwards. +inline const uint8_t *GetBufferStartFromRootPointer(const void *root) { + auto table = reinterpret_cast(root); + auto vtable = table->GetVTable(); + // Either the vtable is before the root or after the root. + auto start = (std::min)(vtable, reinterpret_cast(root)); + // Align to at least sizeof(uoffset_t). + start = reinterpret_cast(reinterpret_cast(start) & + ~(sizeof(uoffset_t) - 1)); + // Additionally, there may be a file_identifier in the buffer, and the root + // offset. The buffer may have been aligned to any size between + // sizeof(uoffset_t) and FLATBUFFERS_MAX_ALIGNMENT (see "force_align"). + // Sadly, the exact alignment is only known when constructing the buffer, + // since it depends on the presence of values with said alignment properties. + // So instead, we simply look at the next uoffset_t values (root, + // file_identifier, and alignment padding) to see which points to the root. + // None of the other values can "impersonate" the root since they will either + // be 0 or four ASCII characters. + static_assert(FlatBufferBuilder::kFileIdentifierLength == sizeof(uoffset_t), + "file_identifier is assumed to be the same size as uoffset_t"); + for (auto possible_roots = FLATBUFFERS_MAX_ALIGNMENT / sizeof(uoffset_t) + 1; + possible_roots; possible_roots--) { + start -= sizeof(uoffset_t); + if (ReadScalar(start) + start == + reinterpret_cast(root)) + return start; + } + // We didn't find the root, either the "root" passed isn't really a root, + // or the buffer is corrupt. + // Assert, because calling this function with bad data may cause reads + // outside of buffer boundaries. + FLATBUFFERS_ASSERT(false); + return nullptr; +} + +/// @brief This return the prefixed size of a FlatBuffer. +inline uoffset_t GetPrefixedSize(const uint8_t *buf) { + return ReadScalar(buf); +} + +// Base class for native objects (FlatBuffer data de-serialized into native +// C++ data structures). +// Contains no functionality, purely documentative. +struct NativeTable {}; + +/// @brief Function types to be used with resolving hashes into objects and +/// back again. The resolver gets a pointer to a field inside an object API +/// object that is of the type specified in the schema using the attribute +/// `cpp_type` (it is thus important whatever you write to this address +/// matches that type). The value of this field is initially null, so you +/// may choose to implement a delayed binding lookup using this function +/// if you wish. The resolver does the opposite lookup, for when the object +/// is being serialized again. +typedef uint64_t hash_value_t; +// clang-format off +#ifdef FLATBUFFERS_CPP98_STL + typedef void (*resolver_function_t)(void **pointer_adr, hash_value_t hash); + typedef hash_value_t (*rehasher_function_t)(void *pointer); +#else + typedef std::function + resolver_function_t; + typedef std::function rehasher_function_t; +#endif +// clang-format on + +// Helper function to test if a field is present, using any of the field +// enums in the generated code. +// `table` must be a generated table type. Since this is a template parameter, +// this is not typechecked to be a subclass of Table, so beware! +// Note: this function will return false for fields equal to the default +// value, since they're not stored in the buffer (unless force_defaults was +// used). +template +bool IsFieldPresent(const T *table, typename T::FlatBuffersVTableOffset field) { + // Cast, since Table is a private baseclass of any table types. + return reinterpret_cast(table)->CheckField( + static_cast(field)); +} + +// Utility function for reverse lookups on the EnumNames*() functions +// (in the generated C++ code) +// names must be NULL terminated. +inline int LookupEnum(const char **names, const char *name) { + for (const char **p = names; *p; p++) + if (!strcmp(*p, name)) return static_cast(p - names); + return -1; +} + +// These macros allow us to layout a struct with a guarantee that they'll end +// up looking the same on different compilers and platforms. +// It does this by disallowing the compiler to do any padding, and then +// does padding itself by inserting extra padding fields that make every +// element aligned to its own size. +// Additionally, it manually sets the alignment of the struct as a whole, +// which is typically its largest element, or a custom size set in the schema +// by the force_align attribute. +// These are used in the generated code only. + +// clang-format off +#if defined(_MSC_VER) + #define FLATBUFFERS_MANUALLY_ALIGNED_STRUCT(alignment) \ + __pragma(pack(1)) \ + struct __declspec(align(alignment)) + #define FLATBUFFERS_STRUCT_END(name, size) \ + __pragma(pack()) \ + static_assert(sizeof(name) == size, "compiler breaks packing rules") +#elif defined(__GNUC__) || defined(__clang__) || defined(__ICCARM__) + #define FLATBUFFERS_MANUALLY_ALIGNED_STRUCT(alignment) \ + _Pragma("pack(1)") \ + struct __attribute__((aligned(alignment))) + #define FLATBUFFERS_STRUCT_END(name, size) \ + _Pragma("pack()") \ + static_assert(sizeof(name) == size, "compiler breaks packing rules") +#else + #error Unknown compiler, please define structure alignment macros +#endif +// clang-format on + +// Minimal reflection via code generation. +// Besides full-fat reflection (see reflection.h) and parsing/printing by +// loading schemas (see idl.h), we can also have code generation for mimimal +// reflection data which allows pretty-printing and other uses without needing +// a schema or a parser. +// Generate code with --reflect-types (types only) or --reflect-names (names +// also) to enable. +// See minireflect.h for utilities using this functionality. + +// These types are organized slightly differently as the ones in idl.h. +enum SequenceType { ST_TABLE, ST_STRUCT, ST_UNION, ST_ENUM }; + +// Scalars have the same order as in idl.h +// clang-format off +#define FLATBUFFERS_GEN_ELEMENTARY_TYPES(ET) \ + ET(ET_UTYPE) \ + ET(ET_BOOL) \ + ET(ET_CHAR) \ + ET(ET_UCHAR) \ + ET(ET_SHORT) \ + ET(ET_USHORT) \ + ET(ET_INT) \ + ET(ET_UINT) \ + ET(ET_LONG) \ + ET(ET_ULONG) \ + ET(ET_FLOAT) \ + ET(ET_DOUBLE) \ + ET(ET_STRING) \ + ET(ET_SEQUENCE) // See SequenceType. + +enum ElementaryType { + #define FLATBUFFERS_ET(E) E, + FLATBUFFERS_GEN_ELEMENTARY_TYPES(FLATBUFFERS_ET) + #undef FLATBUFFERS_ET +}; + +inline const char * const *ElementaryTypeNames() { + static const char * const names[] = { + #define FLATBUFFERS_ET(E) #E, + FLATBUFFERS_GEN_ELEMENTARY_TYPES(FLATBUFFERS_ET) + #undef FLATBUFFERS_ET + }; + return names; +} +// clang-format on + +// Basic type info cost just 16bits per field! +struct TypeCode { + uint16_t base_type : 4; // ElementaryType + uint16_t is_repeating : 1; // Either vector (in table) or array (in struct) + int16_t sequence_ref : 11; // Index into type_refs below, or -1 for none. +}; + +static_assert(sizeof(TypeCode) == 2, "TypeCode"); + +struct TypeTable; + +// Signature of the static method present in each type. +typedef const TypeTable *(*TypeFunction)(); + +struct TypeTable { + SequenceType st; + size_t num_elems; // of type_codes, values, names (but not type_refs). + const TypeCode *type_codes; // num_elems count + const TypeFunction *type_refs; // less than num_elems entries (see TypeCode). + const int16_t *array_sizes; // less than num_elems entries (see TypeCode). + const int64_t *values; // Only set for non-consecutive enum/union or structs. + const char *const *names; // Only set if compiled with --reflect-names. +}; + +// String which identifies the current version of FlatBuffers. +// flatbuffer_version_string is used by Google developers to identify which +// applications uploaded to Google Play are using this library. This allows +// the development team at Google to determine the popularity of the library. +// How it works: Applications that are uploaded to the Google Play Store are +// scanned for this version string. We track which applications are using it +// to measure popularity. You are free to remove it (of course) but we would +// appreciate if you left it in. + +// Weak linkage is culled by VS & doesn't work on cygwin. +// clang-format off +#if !defined(_WIN32) && !defined(__CYGWIN__) + +extern volatile __attribute__((weak)) const char *flatbuffer_version_string; +volatile __attribute__((weak)) const char *flatbuffer_version_string = + "FlatBuffers " + FLATBUFFERS_STRING(FLATBUFFERS_VERSION_MAJOR) "." + FLATBUFFERS_STRING(FLATBUFFERS_VERSION_MINOR) "." + FLATBUFFERS_STRING(FLATBUFFERS_VERSION_REVISION); + +#endif // !defined(_WIN32) && !defined(__CYGWIN__) + +#define FLATBUFFERS_DEFINE_BITMASK_OPERATORS(E, T)\ + inline E operator | (E lhs, E rhs){\ + return E(T(lhs) | T(rhs));\ + }\ + inline E operator & (E lhs, E rhs){\ + return E(T(lhs) & T(rhs));\ + }\ + inline E operator ^ (E lhs, E rhs){\ + return E(T(lhs) ^ T(rhs));\ + }\ + inline E operator ~ (E lhs){\ + return E(~T(lhs));\ + }\ + inline E operator |= (E &lhs, E rhs){\ + lhs = lhs | rhs;\ + return lhs;\ + }\ + inline E operator &= (E &lhs, E rhs){\ + lhs = lhs & rhs;\ + return lhs;\ + }\ + inline E operator ^= (E &lhs, E rhs){\ + lhs = lhs ^ rhs;\ + return lhs;\ + }\ + inline bool operator !(E rhs) \ + {\ + return !bool(T(rhs)); \ + } +/// @endcond +} // namespace flatbuffers + +// clang-format on + +#endif // FLATBUFFERS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flexbuffers.h b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flexbuffers.h new file mode 100644 index 0000000..f4d9aba --- /dev/null +++ b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/flexbuffers.h @@ -0,0 +1,1634 @@ +#if !defined(ESP32) +/* + * Copyright 2017 Google Inc. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef FLATBUFFERS_FLEXBUFFERS_H_ +#define FLATBUFFERS_FLEXBUFFERS_H_ + +#include +// Used to select STL variant. +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h" +// We use the basic binary writing functions from the regular FlatBuffers. +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/util.h" + +#ifdef _MSC_VER +# include +#endif + +#if defined(_MSC_VER) +# pragma warning(push) +# pragma warning(disable : 4127) // C4127: conditional expression is constant +#endif + +namespace flexbuffers { + +class Reference; +class Map; + +// These are used in the lower 2 bits of a type field to determine the size of +// the elements (and or size field) of the item pointed to (e.g. vector). +enum BitWidth { + BIT_WIDTH_8 = 0, + BIT_WIDTH_16 = 1, + BIT_WIDTH_32 = 2, + BIT_WIDTH_64 = 3, +}; + +// These are used as the upper 6 bits of a type field to indicate the actual +// type. +enum Type { + FBT_NULL = 0, + FBT_INT = 1, + FBT_UINT = 2, + FBT_FLOAT = 3, + // Types above stored inline, types below store an offset. + FBT_KEY = 4, + FBT_STRING = 5, + FBT_INDIRECT_INT = 6, + FBT_INDIRECT_UINT = 7, + FBT_INDIRECT_FLOAT = 8, + FBT_MAP = 9, + FBT_VECTOR = 10, // Untyped. + FBT_VECTOR_INT = 11, // Typed any size (stores no type table). + FBT_VECTOR_UINT = 12, + FBT_VECTOR_FLOAT = 13, + FBT_VECTOR_KEY = 14, + // DEPRECATED, use FBT_VECTOR or FBT_VECTOR_KEY instead. + // Read test.cpp/FlexBuffersDeprecatedTest() for details on why. + FBT_VECTOR_STRING_DEPRECATED = 15, + FBT_VECTOR_INT2 = 16, // Typed tuple (no type table, no size field). + FBT_VECTOR_UINT2 = 17, + FBT_VECTOR_FLOAT2 = 18, + FBT_VECTOR_INT3 = 19, // Typed triple (no type table, no size field). + FBT_VECTOR_UINT3 = 20, + FBT_VECTOR_FLOAT3 = 21, + FBT_VECTOR_INT4 = 22, // Typed quad (no type table, no size field). + FBT_VECTOR_UINT4 = 23, + FBT_VECTOR_FLOAT4 = 24, + FBT_BLOB = 25, + FBT_BOOL = 26, + FBT_VECTOR_BOOL = + 36, // To Allow the same type of conversion of type to vector type +}; + +inline bool IsInline(Type t) { return t <= FBT_FLOAT || t == FBT_BOOL; } + +inline bool IsTypedVectorElementType(Type t) { + return (t >= FBT_INT && t <= FBT_STRING) || t == FBT_BOOL; +} + +inline bool IsTypedVector(Type t) { + return (t >= FBT_VECTOR_INT && t <= FBT_VECTOR_STRING_DEPRECATED) || + t == FBT_VECTOR_BOOL; +} + +inline bool IsFixedTypedVector(Type t) { + return t >= FBT_VECTOR_INT2 && t <= FBT_VECTOR_FLOAT4; +} + +inline Type ToTypedVector(Type t, size_t fixed_len = 0) { + FLATBUFFERS_ASSERT(IsTypedVectorElementType(t)); + switch (fixed_len) { + case 0: return static_cast(t - FBT_INT + FBT_VECTOR_INT); + case 2: return static_cast(t - FBT_INT + FBT_VECTOR_INT2); + case 3: return static_cast(t - FBT_INT + FBT_VECTOR_INT3); + case 4: return static_cast(t - FBT_INT + FBT_VECTOR_INT4); + default: FLATBUFFERS_ASSERT(0); return FBT_NULL; + } +} + +inline Type ToTypedVectorElementType(Type t) { + FLATBUFFERS_ASSERT(IsTypedVector(t)); + return static_cast(t - FBT_VECTOR_INT + FBT_INT); +} + +inline Type ToFixedTypedVectorElementType(Type t, uint8_t *len) { + FLATBUFFERS_ASSERT(IsFixedTypedVector(t)); + auto fixed_type = t - FBT_VECTOR_INT2; + *len = static_cast(fixed_type / 3 + + 2); // 3 types each, starting from length 2. + return static_cast(fixed_type % 3 + FBT_INT); +} + +// TODO: implement proper support for 8/16bit floats, or decide not to +// support them. +typedef int16_t half; +typedef int8_t quarter; + +// TODO: can we do this without conditionals using intrinsics or inline asm +// on some platforms? Given branch prediction the method below should be +// decently quick, but it is the most frequently executed function. +// We could do an (unaligned) 64-bit read if we ifdef out the platforms for +// which that doesn't work (or where we'd read into un-owned memory). +template +R ReadSizedScalar(const uint8_t *data, uint8_t byte_width) { + return byte_width < 4 + ? (byte_width < 2 + ? static_cast(flatbuffers::ReadScalar(data)) + : static_cast(flatbuffers::ReadScalar(data))) + : (byte_width < 8 + ? static_cast(flatbuffers::ReadScalar(data)) + : static_cast(flatbuffers::ReadScalar(data))); +} + +inline int64_t ReadInt64(const uint8_t *data, uint8_t byte_width) { + return ReadSizedScalar( + data, byte_width); +} + +inline uint64_t ReadUInt64(const uint8_t *data, uint8_t byte_width) { + // This is the "hottest" function (all offset lookups use this), so worth + // optimizing if possible. + // TODO: GCC apparently replaces memcpy by a rep movsb, but only if count is a + // constant, which here it isn't. Test if memcpy is still faster than + // the conditionals in ReadSizedScalar. Can also use inline asm. + // clang-format off + #if defined(_MSC_VER) && (defined(_M_X64) || defined _M_IX86) + uint64_t u = 0; + __movsb(reinterpret_cast(&u), + reinterpret_cast(data), byte_width); + return flatbuffers::EndianScalar(u); + #else + return ReadSizedScalar( + data, byte_width); + #endif + // clang-format on +} + +inline double ReadDouble(const uint8_t *data, uint8_t byte_width) { + return ReadSizedScalar(data, + byte_width); +} + +inline const uint8_t *Indirect(const uint8_t *offset, uint8_t byte_width) { + return offset - ReadUInt64(offset, byte_width); +} + +template const uint8_t *Indirect(const uint8_t *offset) { + return offset - flatbuffers::ReadScalar(offset); +} + +inline BitWidth WidthU(uint64_t u) { +#define FLATBUFFERS_GET_FIELD_BIT_WIDTH(value, width) \ + { \ + if (!((u) & ~((1ULL << (width)) - 1ULL))) return BIT_WIDTH_##width; \ + } + FLATBUFFERS_GET_FIELD_BIT_WIDTH(u, 8); + FLATBUFFERS_GET_FIELD_BIT_WIDTH(u, 16); + FLATBUFFERS_GET_FIELD_BIT_WIDTH(u, 32); +#undef FLATBUFFERS_GET_FIELD_BIT_WIDTH + return BIT_WIDTH_64; +} + +inline BitWidth WidthI(int64_t i) { + auto u = static_cast(i) << 1; + return WidthU(i >= 0 ? u : ~u); +} + +inline BitWidth WidthF(double f) { + return static_cast(static_cast(f)) == f ? BIT_WIDTH_32 + : BIT_WIDTH_64; +} + +// Base class of all types below. +// Points into the data buffer and allows access to one type. +class Object { + public: + Object(const uint8_t *data, uint8_t byte_width) + : data_(data), byte_width_(byte_width) {} + + protected: + const uint8_t *data_; + uint8_t byte_width_; +}; + +// Object that has a size, obtained either from size prefix, or elsewhere. +class Sized : public Object { + public: + // Size prefix. + Sized(const uint8_t *data, uint8_t byte_width) + : Object(data, byte_width), size_(read_size()) {} + // Manual size. + Sized(const uint8_t *data, uint8_t byte_width, size_t sz) + : Object(data, byte_width), size_(sz) {} + size_t size() const { return size_; } + // Access size stored in `byte_width_` bytes before data_ pointer. + size_t read_size() const { + return static_cast(ReadUInt64(data_ - byte_width_, byte_width_)); + } + + protected: + size_t size_; +}; + +class String : public Sized { + public: + // Size prefix. + String(const uint8_t *data, uint8_t byte_width) : Sized(data, byte_width) {} + // Manual size. + String(const uint8_t *data, uint8_t byte_width, size_t sz) + : Sized(data, byte_width, sz) {} + + size_t length() const { return size(); } + const char *c_str() const { return reinterpret_cast(data_); } + std::string str() const { return std::string(c_str(), size()); } + + static String EmptyString() { + static const char *empty_string = ""; + return String(reinterpret_cast(empty_string), 1, 0); + } + bool IsTheEmptyString() const { return data_ == EmptyString().data_; } +}; + +class Blob : public Sized { + public: + Blob(const uint8_t *data_buf, uint8_t byte_width) + : Sized(data_buf, byte_width) {} + + static Blob EmptyBlob() { + static const uint8_t empty_blob[] = { 0 /*len*/ }; + return Blob(empty_blob + 1, 1); + } + bool IsTheEmptyBlob() const { return data_ == EmptyBlob().data_; } + const uint8_t *data() const { return data_; } +}; + +class Vector : public Sized { + public: + Vector(const uint8_t *data, uint8_t byte_width) : Sized(data, byte_width) {} + + Reference operator[](size_t i) const; + + static Vector EmptyVector() { + static const uint8_t empty_vector[] = { 0 /*len*/ }; + return Vector(empty_vector + 1, 1); + } + bool IsTheEmptyVector() const { return data_ == EmptyVector().data_; } +}; + +class TypedVector : public Sized { + public: + TypedVector(const uint8_t *data, uint8_t byte_width, Type element_type) + : Sized(data, byte_width), type_(element_type) {} + + Reference operator[](size_t i) const; + + static TypedVector EmptyTypedVector() { + static const uint8_t empty_typed_vector[] = { 0 /*len*/ }; + return TypedVector(empty_typed_vector + 1, 1, FBT_INT); + } + bool IsTheEmptyVector() const { + return data_ == TypedVector::EmptyTypedVector().data_; + } + + Type ElementType() { return type_; } + + friend Reference; + + private: + Type type_; + + friend Map; +}; + +class FixedTypedVector : public Object { + public: + FixedTypedVector(const uint8_t *data, uint8_t byte_width, Type element_type, + uint8_t len) + : Object(data, byte_width), type_(element_type), len_(len) {} + + Reference operator[](size_t i) const; + + static FixedTypedVector EmptyFixedTypedVector() { + static const uint8_t fixed_empty_vector[] = { 0 /* unused */ }; + return FixedTypedVector(fixed_empty_vector, 1, FBT_INT, 0); + } + bool IsTheEmptyFixedTypedVector() const { + return data_ == FixedTypedVector::EmptyFixedTypedVector().data_; + } + + Type ElementType() { return type_; } + uint8_t size() { return len_; } + + private: + Type type_; + uint8_t len_; +}; + +class Map : public Vector { + public: + Map(const uint8_t *data, uint8_t byte_width) : Vector(data, byte_width) {} + + Reference operator[](const char *key) const; + Reference operator[](const std::string &key) const; + + Vector Values() const { return Vector(data_, byte_width_); } + + TypedVector Keys() const { + const size_t num_prefixed_fields = 3; + auto keys_offset = data_ - byte_width_ * num_prefixed_fields; + return TypedVector(Indirect(keys_offset, byte_width_), + static_cast( + ReadUInt64(keys_offset + byte_width_, byte_width_)), + FBT_KEY); + } + + static Map EmptyMap() { + static const uint8_t empty_map[] = { + 0 /*keys_len*/, 0 /*keys_offset*/, 1 /*keys_width*/, 0 /*len*/ + }; + return Map(empty_map + 4, 1); + } + + bool IsTheEmptyMap() const { return data_ == EmptyMap().data_; } +}; + +template +void AppendToString(std::string &s, T &&v, bool keys_quoted) { + s += "[ "; + for (size_t i = 0; i < v.size(); i++) { + if (i) s += ", "; + v[i].ToString(true, keys_quoted, s); + } + s += " ]"; +} + +class Reference { + public: + Reference() + : data_(nullptr), + parent_width_(0), + byte_width_(BIT_WIDTH_8), + type_(FBT_NULL) {} + + Reference(const uint8_t *data, uint8_t parent_width, uint8_t byte_width, + Type type) + : data_(data), + parent_width_(parent_width), + byte_width_(byte_width), + type_(type) {} + + Reference(const uint8_t *data, uint8_t parent_width, uint8_t packed_type) + : data_(data), parent_width_(parent_width) { + byte_width_ = 1U << static_cast(packed_type & 3); + type_ = static_cast(packed_type >> 2); + } + + Type GetType() const { return type_; } + + bool IsNull() const { return type_ == FBT_NULL; } + bool IsBool() const { return type_ == FBT_BOOL; } + bool IsInt() const { return type_ == FBT_INT || type_ == FBT_INDIRECT_INT; } + bool IsUInt() const { + return type_ == FBT_UINT || type_ == FBT_INDIRECT_UINT; + } + bool IsIntOrUint() const { return IsInt() || IsUInt(); } + bool IsFloat() const { + return type_ == FBT_FLOAT || type_ == FBT_INDIRECT_FLOAT; + } + bool IsNumeric() const { return IsIntOrUint() || IsFloat(); } + bool IsString() const { return type_ == FBT_STRING; } + bool IsKey() const { return type_ == FBT_KEY; } + bool IsVector() const { return type_ == FBT_VECTOR || type_ == FBT_MAP; } + bool IsUntypedVector() const { return type_ == FBT_VECTOR; } + bool IsTypedVector() const { return flexbuffers::IsTypedVector(type_); } + bool IsFixedTypedVector() const { + return flexbuffers::IsFixedTypedVector(type_); + } + bool IsAnyVector() const { + return (IsTypedVector() || IsFixedTypedVector() || IsVector()); + } + bool IsMap() const { return type_ == FBT_MAP; } + bool IsBlob() const { return type_ == FBT_BLOB; } + bool AsBool() const { + return (type_ == FBT_BOOL ? ReadUInt64(data_, parent_width_) + : AsUInt64()) != 0; + } + + // Reads any type as a int64_t. Never fails, does most sensible conversion. + // Truncates floats, strings are attempted to be parsed for a number, + // vectors/maps return their size. Returns 0 if all else fails. + int64_t AsInt64() const { + if (type_ == FBT_INT) { + // A fast path for the common case. + return ReadInt64(data_, parent_width_); + } else + switch (type_) { + case FBT_INDIRECT_INT: return ReadInt64(Indirect(), byte_width_); + case FBT_UINT: return ReadUInt64(data_, parent_width_); + case FBT_INDIRECT_UINT: return ReadUInt64(Indirect(), byte_width_); + case FBT_FLOAT: + return static_cast(ReadDouble(data_, parent_width_)); + case FBT_INDIRECT_FLOAT: + return static_cast(ReadDouble(Indirect(), byte_width_)); + case FBT_NULL: return 0; + case FBT_STRING: return flatbuffers::StringToInt(AsString().c_str()); + case FBT_VECTOR: return static_cast(AsVector().size()); + case FBT_BOOL: return ReadInt64(data_, parent_width_); + default: + // Convert other things to int. + return 0; + } + } + + // TODO: could specialize these to not use AsInt64() if that saves + // extension ops in generated code, and use a faster op than ReadInt64. + int32_t AsInt32() const { return static_cast(AsInt64()); } + int16_t AsInt16() const { return static_cast(AsInt64()); } + int8_t AsInt8() const { return static_cast(AsInt64()); } + + uint64_t AsUInt64() const { + if (type_ == FBT_UINT) { + // A fast path for the common case. + return ReadUInt64(data_, parent_width_); + } else + switch (type_) { + case FBT_INDIRECT_UINT: return ReadUInt64(Indirect(), byte_width_); + case FBT_INT: return ReadInt64(data_, parent_width_); + case FBT_INDIRECT_INT: return ReadInt64(Indirect(), byte_width_); + case FBT_FLOAT: + return static_cast(ReadDouble(data_, parent_width_)); + case FBT_INDIRECT_FLOAT: + return static_cast(ReadDouble(Indirect(), byte_width_)); + case FBT_NULL: return 0; + case FBT_STRING: return flatbuffers::StringToUInt(AsString().c_str()); + case FBT_VECTOR: return static_cast(AsVector().size()); + case FBT_BOOL: return ReadUInt64(data_, parent_width_); + default: + // Convert other things to uint. + return 0; + } + } + + uint32_t AsUInt32() const { return static_cast(AsUInt64()); } + uint16_t AsUInt16() const { return static_cast(AsUInt64()); } + uint8_t AsUInt8() const { return static_cast(AsUInt64()); } + + double AsDouble() const { + if (type_ == FBT_FLOAT) { + // A fast path for the common case. + return ReadDouble(data_, parent_width_); + } else + switch (type_) { + case FBT_INDIRECT_FLOAT: return ReadDouble(Indirect(), byte_width_); + case FBT_INT: + return static_cast(ReadInt64(data_, parent_width_)); + case FBT_UINT: + return static_cast(ReadUInt64(data_, parent_width_)); + case FBT_INDIRECT_INT: + return static_cast(ReadInt64(Indirect(), byte_width_)); + case FBT_INDIRECT_UINT: + return static_cast(ReadUInt64(Indirect(), byte_width_)); + case FBT_NULL: return 0.0; + case FBT_STRING: { +#if 1 +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wnull-dereference" + // TODO(b/173239141): Patched via micro/tools/make/flexbuffers_download.sh + // Introduce a segfault for an unsupported code path for TFLM. + return *(static_cast(nullptr)); +#pragma GCC diagnostic pop +#else + // This is the original code + double d; + flatbuffers::StringToNumber(AsString().c_str(), &d); + return d; +#endif + } + case FBT_VECTOR: return static_cast(AsVector().size()); + case FBT_BOOL: + return static_cast(ReadUInt64(data_, parent_width_)); + default: + // Convert strings and other things to float. + return 0; + } + } + + float AsFloat() const { return static_cast(AsDouble()); } + + const char *AsKey() const { + if (type_ == FBT_KEY || type_ == FBT_STRING) { + return reinterpret_cast(Indirect()); + } else { + return ""; + } + } + + // This function returns the empty string if you try to read something that + // is not a string or key. + String AsString() const { + if (type_ == FBT_STRING) { + return String(Indirect(), byte_width_); + } else if (type_ == FBT_KEY) { + auto key = Indirect(); + return String(key, byte_width_, + strlen(reinterpret_cast(key))); + } else { + return String::EmptyString(); + } + } + + // Unlike AsString(), this will convert any type to a std::string. + std::string ToString() const { + std::string s; + ToString(false, false, s); + return s; + } + + // Convert any type to a JSON-like string. strings_quoted determines if + // string values at the top level receive "" quotes (inside other values + // they always do). keys_quoted determines if keys are quoted, at any level. + // TODO(wvo): add further options to have indentation/newlines. + void ToString(bool strings_quoted, bool keys_quoted, std::string &s) const { + if (type_ == FBT_STRING) { + String str(Indirect(), byte_width_); + if (strings_quoted) { + flatbuffers::EscapeString(str.c_str(), str.length(), &s, true, false); + } else { + s.append(str.c_str(), str.length()); + } + } else if (IsKey()) { + auto str = AsKey(); + if (keys_quoted) { + flatbuffers::EscapeString(str, strlen(str), &s, true, false); + } else { + s += str; + } + } else if (IsInt()) { + s += flatbuffers::NumToString(AsInt64()); + } else if (IsUInt()) { + s += flatbuffers::NumToString(AsUInt64()); + } else if (IsFloat()) { + s += flatbuffers::NumToString(AsDouble()); + } else if (IsNull()) { + s += "null"; + } else if (IsBool()) { + s += AsBool() ? "true" : "false"; + } else if (IsMap()) { + s += "{ "; + auto m = AsMap(); + auto keys = m.Keys(); + auto vals = m.Values(); + for (size_t i = 0; i < keys.size(); i++) { + keys[i].ToString(true, keys_quoted, s); + s += ": "; + vals[i].ToString(true, keys_quoted, s); + if (i < keys.size() - 1) s += ", "; + } + s += " }"; + } else if (IsVector()) { + AppendToString(s, AsVector(), keys_quoted); + } else if (IsTypedVector()) { + AppendToString(s, AsTypedVector(), keys_quoted); + } else if (IsFixedTypedVector()) { + AppendToString(s, AsFixedTypedVector(), keys_quoted); + } else if (IsBlob()) { + auto blob = AsBlob(); + flatbuffers::EscapeString(reinterpret_cast(blob.data()), + blob.size(), &s, true, false); + } else { + s += "(?)"; + } + } + + // This function returns the empty blob if you try to read a not-blob. + // Strings can be viewed as blobs too. + Blob AsBlob() const { + if (type_ == FBT_BLOB || type_ == FBT_STRING) { + return Blob(Indirect(), byte_width_); + } else { + return Blob::EmptyBlob(); + } + } + + // This function returns the empty vector if you try to read a not-vector. + // Maps can be viewed as vectors too. + Vector AsVector() const { + if (type_ == FBT_VECTOR || type_ == FBT_MAP) { + return Vector(Indirect(), byte_width_); + } else { + return Vector::EmptyVector(); + } + } + + TypedVector AsTypedVector() const { + if (IsTypedVector()) { + auto tv = + TypedVector(Indirect(), byte_width_, ToTypedVectorElementType(type_)); + if (tv.type_ == FBT_STRING) { + // These can't be accessed as strings, since we don't know the bit-width + // of the size field, see the declaration of + // FBT_VECTOR_STRING_DEPRECATED above for details. + // We change the type here to be keys, which are a subtype of strings, + // and will ignore the size field. This will truncate strings with + // embedded nulls. + tv.type_ = FBT_KEY; + } + return tv; + } else { + return TypedVector::EmptyTypedVector(); + } + } + + FixedTypedVector AsFixedTypedVector() const { + if (IsFixedTypedVector()) { + uint8_t len = 0; + auto vtype = ToFixedTypedVectorElementType(type_, &len); + return FixedTypedVector(Indirect(), byte_width_, vtype, len); + } else { + return FixedTypedVector::EmptyFixedTypedVector(); + } + } + + Map AsMap() const { + if (type_ == FBT_MAP) { + return Map(Indirect(), byte_width_); + } else { + return Map::EmptyMap(); + } + } + + template T As() const; + + // Experimental: Mutation functions. + // These allow scalars in an already created buffer to be updated in-place. + // Since by default scalars are stored in the smallest possible space, + // the new value may not fit, in which case these functions return false. + // To avoid this, you can construct the values you intend to mutate using + // Builder::ForceMinimumBitWidth. + bool MutateInt(int64_t i) { + if (type_ == FBT_INT) { + return Mutate(data_, i, parent_width_, WidthI(i)); + } else if (type_ == FBT_INDIRECT_INT) { + return Mutate(Indirect(), i, byte_width_, WidthI(i)); + } else if (type_ == FBT_UINT) { + auto u = static_cast(i); + return Mutate(data_, u, parent_width_, WidthU(u)); + } else if (type_ == FBT_INDIRECT_UINT) { + auto u = static_cast(i); + return Mutate(Indirect(), u, byte_width_, WidthU(u)); + } else { + return false; + } + } + + bool MutateBool(bool b) { + return type_ == FBT_BOOL && Mutate(data_, b, parent_width_, BIT_WIDTH_8); + } + + bool MutateUInt(uint64_t u) { + if (type_ == FBT_UINT) { + return Mutate(data_, u, parent_width_, WidthU(u)); + } else if (type_ == FBT_INDIRECT_UINT) { + return Mutate(Indirect(), u, byte_width_, WidthU(u)); + } else if (type_ == FBT_INT) { + auto i = static_cast(u); + return Mutate(data_, i, parent_width_, WidthI(i)); + } else if (type_ == FBT_INDIRECT_INT) { + auto i = static_cast(u); + return Mutate(Indirect(), i, byte_width_, WidthI(i)); + } else { + return false; + } + } + + bool MutateFloat(float f) { + if (type_ == FBT_FLOAT) { + return MutateF(data_, f, parent_width_, BIT_WIDTH_32); + } else if (type_ == FBT_INDIRECT_FLOAT) { + return MutateF(Indirect(), f, byte_width_, BIT_WIDTH_32); + } else { + return false; + } + } + + bool MutateFloat(double d) { + if (type_ == FBT_FLOAT) { + return MutateF(data_, d, parent_width_, WidthF(d)); + } else if (type_ == FBT_INDIRECT_FLOAT) { + return MutateF(Indirect(), d, byte_width_, WidthF(d)); + } else { + return false; + } + } + + bool MutateString(const char *str, size_t len) { + auto s = AsString(); + if (s.IsTheEmptyString()) return false; + // This is very strict, could allow shorter strings, but that creates + // garbage. + if (s.length() != len) return false; + memcpy(const_cast(s.c_str()), str, len); + return true; + } + bool MutateString(const char *str) { return MutateString(str, strlen(str)); } + bool MutateString(const std::string &str) { + return MutateString(str.data(), str.length()); + } + + private: + const uint8_t *Indirect() const { + return flexbuffers::Indirect(data_, parent_width_); + } + + template + bool Mutate(const uint8_t *dest, T t, size_t byte_width, + BitWidth value_width) { + auto fits = static_cast(static_cast(1U) << value_width) <= + byte_width; + if (fits) { + t = flatbuffers::EndianScalar(t); + memcpy(const_cast(dest), &t, byte_width); + } + return fits; + } + + template + bool MutateF(const uint8_t *dest, T t, size_t byte_width, + BitWidth value_width) { + if (byte_width == sizeof(double)) + return Mutate(dest, static_cast(t), byte_width, value_width); + if (byte_width == sizeof(float)) + return Mutate(dest, static_cast(t), byte_width, value_width); + FLATBUFFERS_ASSERT(false); + return false; + } + + const uint8_t *data_; + uint8_t parent_width_; + uint8_t byte_width_; + Type type_; +}; + +// Template specialization for As(). +template<> inline bool Reference::As() const { return AsBool(); } + +template<> inline int8_t Reference::As() const { return AsInt8(); } +template<> inline int16_t Reference::As() const { return AsInt16(); } +template<> inline int32_t Reference::As() const { return AsInt32(); } +template<> inline int64_t Reference::As() const { return AsInt64(); } + +template<> inline uint8_t Reference::As() const { return AsUInt8(); } +template<> inline uint16_t Reference::As() const { + return AsUInt16(); +} +template<> inline uint32_t Reference::As() const { + return AsUInt32(); +} +template<> inline uint64_t Reference::As() const { + return AsUInt64(); +} + +template<> inline double Reference::As() const { return AsDouble(); } +template<> inline float Reference::As() const { return AsFloat(); } + +template<> inline String Reference::As() const { return AsString(); } +template<> inline std::string Reference::As() const { + return AsString().str(); +} + +template<> inline Blob Reference::As() const { return AsBlob(); } +template<> inline Vector Reference::As() const { return AsVector(); } +template<> inline TypedVector Reference::As() const { + return AsTypedVector(); +} +template<> inline FixedTypedVector Reference::As() const { + return AsFixedTypedVector(); +} +template<> inline Map Reference::As() const { return AsMap(); } + +inline uint8_t PackedType(BitWidth bit_width, Type type) { + return static_cast(bit_width | (type << 2)); +} + +inline uint8_t NullPackedType() { return PackedType(BIT_WIDTH_8, FBT_NULL); } + +// Vector accessors. +// Note: if you try to access outside of bounds, you get a Null value back +// instead. Normally this would be an assert, but since this is "dynamically +// typed" data, you may not want that (someone sends you a 2d vector and you +// wanted 3d). +// The Null converts seamlessly into a default value for any other type. +// TODO(wvo): Could introduce an #ifdef that makes this into an assert? +inline Reference Vector::operator[](size_t i) const { + auto len = size(); + if (i >= len) return Reference(nullptr, 1, NullPackedType()); + auto packed_type = (data_ + len * byte_width_)[i]; + auto elem = data_ + i * byte_width_; + return Reference(elem, byte_width_, packed_type); +} + +inline Reference TypedVector::operator[](size_t i) const { + auto len = size(); + if (i >= len) return Reference(nullptr, 1, NullPackedType()); + auto elem = data_ + i * byte_width_; + return Reference(elem, byte_width_, 1, type_); +} + +inline Reference FixedTypedVector::operator[](size_t i) const { + if (i >= len_) return Reference(nullptr, 1, NullPackedType()); + auto elem = data_ + i * byte_width_; + return Reference(elem, byte_width_, 1, type_); +} + +template int KeyCompare(const void *key, const void *elem) { + auto str_elem = reinterpret_cast( + Indirect(reinterpret_cast(elem))); + auto skey = reinterpret_cast(key); + return strcmp(skey, str_elem); +} + +inline Reference Map::operator[](const char *key) const { + auto keys = Keys(); + // We can't pass keys.byte_width_ to the comparison function, so we have + // to pick the right one ahead of time. + int (*comp)(const void *, const void *) = nullptr; + switch (keys.byte_width_) { + case 1: comp = KeyCompare; break; + case 2: comp = KeyCompare; break; + case 4: comp = KeyCompare; break; + case 8: comp = KeyCompare; break; + } + auto res = std::bsearch(key, keys.data_, keys.size(), keys.byte_width_, comp); + if (!res) return Reference(nullptr, 1, NullPackedType()); + auto i = (reinterpret_cast(res) - keys.data_) / keys.byte_width_; + return (*static_cast(this))[i]; +} + +inline Reference Map::operator[](const std::string &key) const { + return (*this)[key.c_str()]; +} + +inline Reference GetRoot(const uint8_t *buffer, size_t size) { + // See Finish() below for the serialization counterpart of this. + // The root starts at the end of the buffer, so we parse backwards from there. + auto end = buffer + size; + auto byte_width = *--end; + auto packed_type = *--end; + end -= byte_width; // The root data item. + return Reference(end, byte_width, packed_type); +} + +inline Reference GetRoot(const std::vector &buffer) { + return GetRoot(flatbuffers::vector_data(buffer), buffer.size()); +} + +// Flags that configure how the Builder behaves. +// The "Share" flags determine if the Builder automatically tries to pool +// this type. Pooling can reduce the size of serialized data if there are +// multiple maps of the same kind, at the expense of slightly slower +// serialization (the cost of lookups) and more memory use (std::set). +// By default this is on for keys, but off for strings. +// Turn keys off if you have e.g. only one map. +// Turn strings on if you expect many non-unique string values. +// Additionally, sharing key vectors can save space if you have maps with +// identical field populations. +enum BuilderFlag { + BUILDER_FLAG_NONE = 0, + BUILDER_FLAG_SHARE_KEYS = 1, + BUILDER_FLAG_SHARE_STRINGS = 2, + BUILDER_FLAG_SHARE_KEYS_AND_STRINGS = 3, + BUILDER_FLAG_SHARE_KEY_VECTORS = 4, + BUILDER_FLAG_SHARE_ALL = 7, +}; + +class Builder FLATBUFFERS_FINAL_CLASS { + public: + Builder(size_t initial_size = 256, + BuilderFlag flags = BUILDER_FLAG_SHARE_KEYS) + : buf_(initial_size), + finished_(false), + flags_(flags), + force_min_bit_width_(BIT_WIDTH_8), + key_pool(KeyOffsetCompare(buf_)), + string_pool(StringOffsetCompare(buf_)) { + buf_.clear(); + } + + /// @brief Get the serialized buffer (after you call `Finish()`). + /// @return Returns a vector owned by this class. + const std::vector &GetBuffer() const { + Finished(); + return buf_; + } + + // Size of the buffer. Does not include unfinished values. + size_t GetSize() const { return buf_.size(); } + + // Reset all state so we can re-use the buffer. + void Clear() { + buf_.clear(); + stack_.clear(); + finished_ = false; + // flags_ remains as-is; + force_min_bit_width_ = BIT_WIDTH_8; + key_pool.clear(); + string_pool.clear(); + } + + // All value constructing functions below have two versions: one that + // takes a key (for placement inside a map) and one that doesn't (for inside + // vectors and elsewhere). + + void Null() { stack_.push_back(Value()); } + void Null(const char *key) { + Key(key); + Null(); + } + + void Int(int64_t i) { stack_.push_back(Value(i, FBT_INT, WidthI(i))); } + void Int(const char *key, int64_t i) { + Key(key); + Int(i); + } + + void UInt(uint64_t u) { stack_.push_back(Value(u, FBT_UINT, WidthU(u))); } + void UInt(const char *key, uint64_t u) { + Key(key); + UInt(u); + } + + void Float(float f) { stack_.push_back(Value(f)); } + void Float(const char *key, float f) { + Key(key); + Float(f); + } + + void Double(double f) { stack_.push_back(Value(f)); } + void Double(const char *key, double d) { + Key(key); + Double(d); + } + + void Bool(bool b) { stack_.push_back(Value(b)); } + void Bool(const char *key, bool b) { + Key(key); + Bool(b); + } + + void IndirectInt(int64_t i) { PushIndirect(i, FBT_INDIRECT_INT, WidthI(i)); } + void IndirectInt(const char *key, int64_t i) { + Key(key); + IndirectInt(i); + } + + void IndirectUInt(uint64_t u) { + PushIndirect(u, FBT_INDIRECT_UINT, WidthU(u)); + } + void IndirectUInt(const char *key, uint64_t u) { + Key(key); + IndirectUInt(u); + } + + void IndirectFloat(float f) { + PushIndirect(f, FBT_INDIRECT_FLOAT, BIT_WIDTH_32); + } + void IndirectFloat(const char *key, float f) { + Key(key); + IndirectFloat(f); + } + + void IndirectDouble(double f) { + PushIndirect(f, FBT_INDIRECT_FLOAT, WidthF(f)); + } + void IndirectDouble(const char *key, double d) { + Key(key); + IndirectDouble(d); + } + + size_t Key(const char *str, size_t len) { + auto sloc = buf_.size(); + WriteBytes(str, len + 1); + if (flags_ & BUILDER_FLAG_SHARE_KEYS) { + auto it = key_pool.find(sloc); + if (it != key_pool.end()) { + // Already in the buffer. Remove key we just serialized, and use + // existing offset instead. + buf_.resize(sloc); + sloc = *it; + } else { + key_pool.insert(sloc); + } + } + stack_.push_back(Value(static_cast(sloc), FBT_KEY, BIT_WIDTH_8)); + return sloc; + } + + size_t Key(const char *str) { return Key(str, strlen(str)); } + size_t Key(const std::string &str) { return Key(str.c_str(), str.size()); } + + size_t String(const char *str, size_t len) { + auto reset_to = buf_.size(); + auto sloc = CreateBlob(str, len, 1, FBT_STRING); + if (flags_ & BUILDER_FLAG_SHARE_STRINGS) { + StringOffset so(sloc, len); + auto it = string_pool.find(so); + if (it != string_pool.end()) { + // Already in the buffer. Remove string we just serialized, and use + // existing offset instead. + buf_.resize(reset_to); + sloc = it->first; + stack_.back().u_ = sloc; + } else { + string_pool.insert(so); + } + } + return sloc; + } + size_t String(const char *str) { return String(str, strlen(str)); } + size_t String(const std::string &str) { + return String(str.c_str(), str.size()); + } + void String(const flexbuffers::String &str) { + String(str.c_str(), str.length()); + } + + void String(const char *key, const char *str) { + Key(key); + String(str); + } + void String(const char *key, const std::string &str) { + Key(key); + String(str); + } + void String(const char *key, const flexbuffers::String &str) { + Key(key); + String(str); + } + + size_t Blob(const void *data, size_t len) { + return CreateBlob(data, len, 0, FBT_BLOB); + } + size_t Blob(const std::vector &v) { + return CreateBlob(flatbuffers::vector_data(v), v.size(), 0, FBT_BLOB); + } + + // TODO(wvo): support all the FlexBuffer types (like flexbuffers::String), + // e.g. Vector etc. Also in overloaded versions. + // Also some FlatBuffers types? + + size_t StartVector() { return stack_.size(); } + size_t StartVector(const char *key) { + Key(key); + return stack_.size(); + } + size_t StartMap() { return stack_.size(); } + size_t StartMap(const char *key) { + Key(key); + return stack_.size(); + } + + // TODO(wvo): allow this to specify an aligment greater than the natural + // alignment. + size_t EndVector(size_t start, bool typed, bool fixed) { + auto vec = CreateVector(start, stack_.size() - start, 1, typed, fixed); + // Remove temp elements and return vector. + stack_.resize(start); + stack_.push_back(vec); + return static_cast(vec.u_); + } + + size_t EndMap(size_t start) { + // We should have interleaved keys and values on the stack. + // Make sure it is an even number: + auto len = stack_.size() - start; + FLATBUFFERS_ASSERT(!(len & 1)); + len /= 2; + // Make sure keys are all strings: + for (auto key = start; key < stack_.size(); key += 2) { + FLATBUFFERS_ASSERT(stack_[key].type_ == FBT_KEY); + } + // Now sort values, so later we can do a binary search lookup. + // We want to sort 2 array elements at a time. + struct TwoValue { + Value key; + Value val; + }; + // TODO(wvo): strict aliasing? + // TODO(wvo): allow the caller to indicate the data is already sorted + // for maximum efficiency? With an assert to check sortedness to make sure + // we're not breaking binary search. + // Or, we can track if the map is sorted as keys are added which would be + // be quite cheap (cheaper than checking it here), so we can skip this + // step automatically when appliccable, and encourage people to write in + // sorted fashion. + // std::sort is typically already a lot faster on sorted data though. + auto dict = + reinterpret_cast(flatbuffers::vector_data(stack_) + start); + std::sort(dict, dict + len, + [&](const TwoValue &a, const TwoValue &b) -> bool { + auto as = reinterpret_cast( + flatbuffers::vector_data(buf_) + a.key.u_); + auto bs = reinterpret_cast( + flatbuffers::vector_data(buf_) + b.key.u_); + auto comp = strcmp(as, bs); + // If this assertion hits, you've added two keys with the same + // value to this map. + // TODO: Have to check for pointer equality, as some sort + // implementation apparently call this function with the same + // element?? Why? + FLATBUFFERS_ASSERT(comp || &a == &b); + return comp < 0; + }); + // First create a vector out of all keys. + // TODO(wvo): if kBuilderFlagShareKeyVectors is true, see if we can share + // the first vector. + auto keys = CreateVector(start, len, 2, true, false); + auto vec = CreateVector(start + 1, len, 2, false, false, &keys); + // Remove temp elements and return map. + stack_.resize(start); + stack_.push_back(vec); + return static_cast(vec.u_); + } + + template size_t Vector(F f) { + auto start = StartVector(); + f(); + return EndVector(start, false, false); + } + template size_t Vector(F f, T &state) { + auto start = StartVector(); + f(state); + return EndVector(start, false, false); + } + template size_t Vector(const char *key, F f) { + auto start = StartVector(key); + f(); + return EndVector(start, false, false); + } + template + size_t Vector(const char *key, F f, T &state) { + auto start = StartVector(key); + f(state); + return EndVector(start, false, false); + } + + template void Vector(const T *elems, size_t len) { + if (flatbuffers::is_scalar::value) { + // This path should be a lot quicker and use less space. + ScalarVector(elems, len, false); + } else { + auto start = StartVector(); + for (size_t i = 0; i < len; i++) Add(elems[i]); + EndVector(start, false, false); + } + } + template + void Vector(const char *key, const T *elems, size_t len) { + Key(key); + Vector(elems, len); + } + template void Vector(const std::vector &vec) { + Vector(flatbuffers::vector_data(vec), vec.size()); + } + + template size_t TypedVector(F f) { + auto start = StartVector(); + f(); + return EndVector(start, true, false); + } + template size_t TypedVector(F f, T &state) { + auto start = StartVector(); + f(state); + return EndVector(start, true, false); + } + template size_t TypedVector(const char *key, F f) { + auto start = StartVector(key); + f(); + return EndVector(start, true, false); + } + template + size_t TypedVector(const char *key, F f, T &state) { + auto start = StartVector(key); + f(state); + return EndVector(start, true, false); + } + + template size_t FixedTypedVector(const T *elems, size_t len) { + // We only support a few fixed vector lengths. Anything bigger use a + // regular typed vector. + FLATBUFFERS_ASSERT(len >= 2 && len <= 4); + // And only scalar values. + static_assert(flatbuffers::is_scalar::value, "Unrelated types"); + return ScalarVector(elems, len, true); + } + + template + size_t FixedTypedVector(const char *key, const T *elems, size_t len) { + Key(key); + return FixedTypedVector(elems, len); + } + + template size_t Map(F f) { + auto start = StartMap(); + f(); + return EndMap(start); + } + template size_t Map(F f, T &state) { + auto start = StartMap(); + f(state); + return EndMap(start); + } + template size_t Map(const char *key, F f) { + auto start = StartMap(key); + f(); + return EndMap(start); + } + template size_t Map(const char *key, F f, T &state) { + auto start = StartMap(key); + f(state); + return EndMap(start); + } + template void Map(const std::map &map) { + auto start = StartMap(); + for (auto it = map.begin(); it != map.end(); ++it) + Add(it->first.c_str(), it->second); + EndMap(start); + } + + // If you wish to share a value explicitly (a value not shared automatically + // through one of the BUILDER_FLAG_SHARE_* flags) you can do so with these + // functions. Or if you wish to turn those flags off for performance reasons + // and still do some explicit sharing. For example: + // builder.IndirectDouble(M_PI); + // auto id = builder.LastValue(); // Remember where we stored it. + // .. more code goes here .. + // builder.ReuseValue(id); // Refers to same double by offset. + // LastValue works regardless of whether the value has a key or not. + // Works on any data type. + struct Value; + Value LastValue() { return stack_.back(); } + void ReuseValue(Value v) { stack_.push_back(v); } + void ReuseValue(const char *key, Value v) { + Key(key); + ReuseValue(v); + } + + // Overloaded Add that tries to call the correct function above. + void Add(int8_t i) { Int(i); } + void Add(int16_t i) { Int(i); } + void Add(int32_t i) { Int(i); } + void Add(int64_t i) { Int(i); } + void Add(uint8_t u) { UInt(u); } + void Add(uint16_t u) { UInt(u); } + void Add(uint32_t u) { UInt(u); } + void Add(uint64_t u) { UInt(u); } + void Add(float f) { Float(f); } + void Add(double d) { Double(d); } + void Add(bool b) { Bool(b); } + void Add(const char *str) { String(str); } + void Add(const std::string &str) { String(str); } + void Add(const flexbuffers::String &str) { String(str); } + + template void Add(const std::vector &vec) { Vector(vec); } + + template void Add(const char *key, const T &t) { + Key(key); + Add(t); + } + + template void Add(const std::map &map) { + Map(map); + } + + template void operator+=(const T &t) { Add(t); } + + // This function is useful in combination with the Mutate* functions above. + // It forces elements of vectors and maps to have a minimum size, such that + // they can later be updated without failing. + // Call with no arguments to reset. + void ForceMinimumBitWidth(BitWidth bw = BIT_WIDTH_8) { + force_min_bit_width_ = bw; + } + + void Finish() { + // If you hit this assert, you likely have objects that were never included + // in a parent. You need to have exactly one root to finish a buffer. + // Check your Start/End calls are matched, and all objects are inside + // some other object. + FLATBUFFERS_ASSERT(stack_.size() == 1); + + // Write root value. + auto byte_width = Align(stack_[0].ElemWidth(buf_.size(), 0)); + WriteAny(stack_[0], byte_width); + // Write root type. + Write(stack_[0].StoredPackedType(), 1); + // Write root size. Normally determined by parent, but root has no parent :) + Write(byte_width, 1); + + finished_ = true; + } + + private: + void Finished() const { + // If you get this assert, you're attempting to get access a buffer + // which hasn't been finished yet. Be sure to call + // Builder::Finish with your root object. + FLATBUFFERS_ASSERT(finished_); + } + + // Align to prepare for writing a scalar with a certain size. + uint8_t Align(BitWidth alignment) { + auto byte_width = 1U << alignment; + buf_.insert(buf_.end(), flatbuffers::PaddingBytes(buf_.size(), byte_width), + 0); + return static_cast(byte_width); + } + + void WriteBytes(const void *val, size_t size) { + buf_.insert(buf_.end(), reinterpret_cast(val), + reinterpret_cast(val) + size); + } + + template void Write(T val, size_t byte_width) { + FLATBUFFERS_ASSERT(sizeof(T) >= byte_width); + val = flatbuffers::EndianScalar(val); + WriteBytes(&val, byte_width); + } + + void WriteDouble(double f, uint8_t byte_width) { + switch (byte_width) { + case 8: Write(f, byte_width); break; + case 4: Write(static_cast(f), byte_width); break; + // case 2: Write(static_cast(f), byte_width); break; + // case 1: Write(static_cast(f), byte_width); break; + default: FLATBUFFERS_ASSERT(0); + } + } + + void WriteOffset(uint64_t o, uint8_t byte_width) { + auto reloff = buf_.size() - o; + FLATBUFFERS_ASSERT(byte_width == 8 || reloff < 1ULL << (byte_width * 8)); + Write(reloff, byte_width); + } + + template void PushIndirect(T val, Type type, BitWidth bit_width) { + auto byte_width = Align(bit_width); + auto iloc = buf_.size(); + Write(val, byte_width); + stack_.push_back(Value(static_cast(iloc), type, bit_width)); + } + + static BitWidth WidthB(size_t byte_width) { + switch (byte_width) { + case 1: return BIT_WIDTH_8; + case 2: return BIT_WIDTH_16; + case 4: return BIT_WIDTH_32; + case 8: return BIT_WIDTH_64; + default: FLATBUFFERS_ASSERT(false); return BIT_WIDTH_64; + } + } + + template static Type GetScalarType() { + static_assert(flatbuffers::is_scalar::value, "Unrelated types"); + return flatbuffers::is_floating_point::value + ? FBT_FLOAT + : flatbuffers::is_same::value + ? FBT_BOOL + : (flatbuffers::is_unsigned::value ? FBT_UINT + : FBT_INT); + } + + public: + // This was really intended to be private, except for LastValue/ReuseValue. + struct Value { + union { + int64_t i_; + uint64_t u_; + double f_; + }; + + Type type_; + + // For scalars: of itself, for vector: of its elements, for string: length. + BitWidth min_bit_width_; + + Value() : i_(0), type_(FBT_NULL), min_bit_width_(BIT_WIDTH_8) {} + + Value(bool b) + : u_(static_cast(b)), + type_(FBT_BOOL), + min_bit_width_(BIT_WIDTH_8) {} + + Value(int64_t i, Type t, BitWidth bw) + : i_(i), type_(t), min_bit_width_(bw) {} + Value(uint64_t u, Type t, BitWidth bw) + : u_(u), type_(t), min_bit_width_(bw) {} + + Value(float f) + : f_(static_cast(f)), + type_(FBT_FLOAT), + min_bit_width_(BIT_WIDTH_32) {} + Value(double f) : f_(f), type_(FBT_FLOAT), min_bit_width_(WidthF(f)) {} + + uint8_t StoredPackedType(BitWidth parent_bit_width_ = BIT_WIDTH_8) const { + return PackedType(StoredWidth(parent_bit_width_), type_); + } + + BitWidth ElemWidth(size_t buf_size, size_t elem_index) const { + if (IsInline(type_)) { + return min_bit_width_; + } else { + // We have an absolute offset, but want to store a relative offset + // elem_index elements beyond the current buffer end. Since whether + // the relative offset fits in a certain byte_width depends on + // the size of the elements before it (and their alignment), we have + // to test for each size in turn. + for (size_t byte_width = 1; + byte_width <= sizeof(flatbuffers::largest_scalar_t); + byte_width *= 2) { + // Where are we going to write this offset? + auto offset_loc = buf_size + + flatbuffers::PaddingBytes(buf_size, byte_width) + + elem_index * byte_width; + // Compute relative offset. + auto offset = offset_loc - u_; + // Does it fit? + auto bit_width = WidthU(offset); + if (static_cast(static_cast(1U) << bit_width) == + byte_width) + return bit_width; + } + FLATBUFFERS_ASSERT(false); // Must match one of the sizes above. + return BIT_WIDTH_64; + } + } + + BitWidth StoredWidth(BitWidth parent_bit_width_ = BIT_WIDTH_8) const { + if (IsInline(type_)) { + return (std::max)(min_bit_width_, parent_bit_width_); + } else { + return min_bit_width_; + } + } + }; + + private: + void WriteAny(const Value &val, uint8_t byte_width) { + switch (val.type_) { + case FBT_NULL: + case FBT_INT: Write(val.i_, byte_width); break; + case FBT_BOOL: + case FBT_UINT: Write(val.u_, byte_width); break; + case FBT_FLOAT: WriteDouble(val.f_, byte_width); break; + default: WriteOffset(val.u_, byte_width); break; + } + } + + size_t CreateBlob(const void *data, size_t len, size_t trailing, Type type) { + auto bit_width = WidthU(len); + auto byte_width = Align(bit_width); + Write(len, byte_width); + auto sloc = buf_.size(); + WriteBytes(data, len + trailing); + stack_.push_back(Value(static_cast(sloc), type, bit_width)); + return sloc; + } + + template + size_t ScalarVector(const T *elems, size_t len, bool fixed) { + auto vector_type = GetScalarType(); + auto byte_width = sizeof(T); + auto bit_width = WidthB(byte_width); + // If you get this assert, you're trying to write a vector with a size + // field that is bigger than the scalars you're trying to write (e.g. a + // byte vector > 255 elements). For such types, write a "blob" instead. + // TODO: instead of asserting, could write vector with larger elements + // instead, though that would be wasteful. + FLATBUFFERS_ASSERT(WidthU(len) <= bit_width); + Align(bit_width); + if (!fixed) Write(len, byte_width); + auto vloc = buf_.size(); + for (size_t i = 0; i < len; i++) Write(elems[i], byte_width); + stack_.push_back(Value(static_cast(vloc), + ToTypedVector(vector_type, fixed ? len : 0), + bit_width)); + return vloc; + } + + Value CreateVector(size_t start, size_t vec_len, size_t step, bool typed, + bool fixed, const Value *keys = nullptr) { + FLATBUFFERS_ASSERT( + !fixed || + typed); // typed=false, fixed=true combination is not supported. + // Figure out smallest bit width we can store this vector with. + auto bit_width = (std::max)(force_min_bit_width_, WidthU(vec_len)); + auto prefix_elems = 1; + if (keys) { + // If this vector is part of a map, we will pre-fix an offset to the keys + // to this vector. + bit_width = (std::max)(bit_width, keys->ElemWidth(buf_.size(), 0)); + prefix_elems += 2; + } + Type vector_type = FBT_KEY; + // Check bit widths and types for all elements. + for (size_t i = start; i < stack_.size(); i += step) { + auto elem_width = + stack_[i].ElemWidth(buf_.size(), i - start + prefix_elems); + bit_width = (std::max)(bit_width, elem_width); + if (typed) { + if (i == start) { + vector_type = stack_[i].type_; + } else { + // If you get this assert, you are writing a typed vector with + // elements that are not all the same type. + FLATBUFFERS_ASSERT(vector_type == stack_[i].type_); + } + } + } + // If you get this assert, your fixed types are not one of: + // Int / UInt / Float / Key. + FLATBUFFERS_ASSERT(!fixed || IsTypedVectorElementType(vector_type)); + auto byte_width = Align(bit_width); + // Write vector. First the keys width/offset if available, and size. + if (keys) { + WriteOffset(keys->u_, byte_width); + Write(1ULL << keys->min_bit_width_, byte_width); + } + if (!fixed) Write(vec_len, byte_width); + // Then the actual data. + auto vloc = buf_.size(); + for (size_t i = start; i < stack_.size(); i += step) { + WriteAny(stack_[i], byte_width); + } + // Then the types. + if (!typed) { + for (size_t i = start; i < stack_.size(); i += step) { + buf_.push_back(stack_[i].StoredPackedType(bit_width)); + } + } + return Value(static_cast(vloc), + keys ? FBT_MAP + : (typed ? ToTypedVector(vector_type, fixed ? vec_len : 0) + : FBT_VECTOR), + bit_width); + } + + // You shouldn't really be copying instances of this class. + Builder(const Builder &); + Builder &operator=(const Builder &); + + std::vector buf_; + std::vector stack_; + + bool finished_; + + BuilderFlag flags_; + + BitWidth force_min_bit_width_; + + struct KeyOffsetCompare { + explicit KeyOffsetCompare(const std::vector &buf) : buf_(&buf) {} + bool operator()(size_t a, size_t b) const { + auto stra = + reinterpret_cast(flatbuffers::vector_data(*buf_) + a); + auto strb = + reinterpret_cast(flatbuffers::vector_data(*buf_) + b); + return strcmp(stra, strb) < 0; + } + const std::vector *buf_; + }; + + typedef std::pair StringOffset; + struct StringOffsetCompare { + explicit StringOffsetCompare(const std::vector &buf) + : buf_(&buf) {} + bool operator()(const StringOffset &a, const StringOffset &b) const { + auto stra = reinterpret_cast( + flatbuffers::vector_data(*buf_) + a.first); + auto strb = reinterpret_cast( + flatbuffers::vector_data(*buf_) + b.first); + return strncmp(stra, strb, (std::min)(a.second, b.second) + 1) < 0; + } + const std::vector *buf_; + }; + + typedef std::set KeyOffsetMap; + typedef std::set StringOffsetMap; + + KeyOffsetMap key_pool; + StringOffsetMap string_pool; +}; + +} // namespace flexbuffers + +#if defined(_MSC_VER) +# pragma warning(pop) +#endif + +#endif // FLATBUFFERS_FLEXBUFFERS_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/stl_emulation.h b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/stl_emulation.h new file mode 100644 index 0000000..4214a47 --- /dev/null +++ b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/stl_emulation.h @@ -0,0 +1,452 @@ +#if !defined(ESP32) +/* + * Copyright 2017 Google Inc. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef FLATBUFFERS_STL_EMULATION_H_ +#define FLATBUFFERS_STL_EMULATION_H_ + +// clang-format off +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h" + +#include +#include +#include +#include +#include + +// Detect C++17 compatible compiler. +// __cplusplus >= 201703L - a compiler has support of 'static inline' variables. +#if defined(FLATBUFFERS_USE_STD_OPTIONAL) \ + || (defined(__cplusplus) && __cplusplus >= 201703L) \ + || (defined(_MSVC_LANG) && (_MSVC_LANG >= 201703L)) + #include + #ifndef FLATBUFFERS_USE_STD_OPTIONAL + #define FLATBUFFERS_USE_STD_OPTIONAL + #endif +#endif + +#if defined(_STLPORT_VERSION) && !defined(FLATBUFFERS_CPP98_STL) + #define FLATBUFFERS_CPP98_STL +#endif // defined(_STLPORT_VERSION) && !defined(FLATBUFFERS_CPP98_STL) + +#if defined(FLATBUFFERS_CPP98_STL) + #include +#endif // defined(FLATBUFFERS_CPP98_STL) + +// This header provides backwards compatibility for C++98 STLs like stlport. +namespace flatbuffers { + +// Retrieve ::back() from a string in a way that is compatible with pre C++11 +// STLs (e.g stlport). +inline char& string_back(std::string &value) { + return value[value.length() - 1]; +} + +inline char string_back(const std::string &value) { + return value[value.length() - 1]; +} + +// Helper method that retrieves ::data() from a vector in a way that is +// compatible with pre C++11 STLs (e.g stlport). +template inline T *vector_data(std::vector &vector) { + // In some debug environments, operator[] does bounds checking, so &vector[0] + // can't be used. + return vector.empty() ? nullptr : &vector[0]; +} + +template inline const T *vector_data( + const std::vector &vector) { + return vector.empty() ? nullptr : &vector[0]; +} + +template +inline void vector_emplace_back(std::vector *vector, V &&data) { + #if defined(FLATBUFFERS_CPP98_STL) + vector->push_back(data); + #else + vector->emplace_back(std::forward(data)); + #endif // defined(FLATBUFFERS_CPP98_STL) +} + +#ifndef FLATBUFFERS_CPP98_STL + #if defined(FLATBUFFERS_TEMPLATES_ALIASES) + template + using numeric_limits = std::numeric_limits; + #else + template class numeric_limits : + public std::numeric_limits {}; + #endif // defined(FLATBUFFERS_TEMPLATES_ALIASES) +#else + template class numeric_limits : + public std::numeric_limits { + public: + // Android NDK fix. + static T lowest() { + return std::numeric_limits::min(); + } + }; + + template <> class numeric_limits : + public std::numeric_limits { + public: + static float lowest() { return -FLT_MAX; } + }; + + template <> class numeric_limits : + public std::numeric_limits { + public: + static double lowest() { return -DBL_MAX; } + }; + + template <> class numeric_limits { + public: + static unsigned long long min() { return 0ULL; } + static unsigned long long max() { return ~0ULL; } + static unsigned long long lowest() { + return numeric_limits::min(); + } + }; + + template <> class numeric_limits { + public: + static long long min() { + return static_cast(1ULL << ((sizeof(long long) << 3) - 1)); + } + static long long max() { + return static_cast( + (1ULL << ((sizeof(long long) << 3) - 1)) - 1); + } + static long long lowest() { + return numeric_limits::min(); + } + }; +#endif // FLATBUFFERS_CPP98_STL + +#if defined(FLATBUFFERS_TEMPLATES_ALIASES) + #ifndef FLATBUFFERS_CPP98_STL + template using is_scalar = std::is_scalar; + template using is_same = std::is_same; + template using is_floating_point = std::is_floating_point; + template using is_unsigned = std::is_unsigned; + template using is_enum = std::is_enum; + template using make_unsigned = std::make_unsigned; + template + using conditional = std::conditional; + template + using integral_constant = std::integral_constant; + #else + // Map C++ TR1 templates defined by stlport. + template using is_scalar = std::tr1::is_scalar; + template using is_same = std::tr1::is_same; + template using is_floating_point = + std::tr1::is_floating_point; + template using is_unsigned = std::tr1::is_unsigned; + template using is_enum = std::tr1::is_enum; + // Android NDK doesn't have std::make_unsigned or std::tr1::make_unsigned. + template struct make_unsigned { + static_assert(is_unsigned::value, "Specialization not implemented!"); + using type = T; + }; + template<> struct make_unsigned { using type = unsigned char; }; + template<> struct make_unsigned { using type = unsigned short; }; + template<> struct make_unsigned { using type = unsigned int; }; + template<> struct make_unsigned { using type = unsigned long; }; + template<> + struct make_unsigned { using type = unsigned long long; }; + template + using conditional = std::tr1::conditional; + template + using integral_constant = std::tr1::integral_constant; + #endif // !FLATBUFFERS_CPP98_STL +#else + // MSVC 2010 doesn't support C++11 aliases. + template struct is_scalar : public std::is_scalar {}; + template struct is_same : public std::is_same {}; + template struct is_floating_point : + public std::is_floating_point {}; + template struct is_unsigned : public std::is_unsigned {}; + template struct is_enum : public std::is_enum {}; + template struct make_unsigned : public std::make_unsigned {}; + template + struct conditional : public std::conditional {}; + template + struct integral_constant : public std::integral_constant {}; +#endif // defined(FLATBUFFERS_TEMPLATES_ALIASES) + +#ifndef FLATBUFFERS_CPP98_STL + #if defined(FLATBUFFERS_TEMPLATES_ALIASES) + template using unique_ptr = std::unique_ptr; + #else + // MSVC 2010 doesn't support C++11 aliases. + // We're manually "aliasing" the class here as we want to bring unique_ptr + // into the flatbuffers namespace. We have unique_ptr in the flatbuffers + // namespace we have a completely independent implementation (see below) + // for C++98 STL implementations. + template class unique_ptr : public std::unique_ptr { + public: + unique_ptr() {} + explicit unique_ptr(T* p) : std::unique_ptr(p) {} + unique_ptr(std::unique_ptr&& u) { *this = std::move(u); } + unique_ptr(unique_ptr&& u) { *this = std::move(u); } + unique_ptr& operator=(std::unique_ptr&& u) { + std::unique_ptr::reset(u.release()); + return *this; + } + unique_ptr& operator=(unique_ptr&& u) { + std::unique_ptr::reset(u.release()); + return *this; + } + unique_ptr& operator=(T* p) { + return std::unique_ptr::operator=(p); + } + }; + #endif // defined(FLATBUFFERS_TEMPLATES_ALIASES) +#else + // Very limited implementation of unique_ptr. + // This is provided simply to allow the C++ code generated from the default + // settings to function in C++98 environments with no modifications. + template class unique_ptr { + public: + typedef T element_type; + + unique_ptr() : ptr_(nullptr) {} + explicit unique_ptr(T* p) : ptr_(p) {} + unique_ptr(unique_ptr&& u) : ptr_(nullptr) { reset(u.release()); } + unique_ptr(const unique_ptr& u) : ptr_(nullptr) { + reset(const_cast(&u)->release()); + } + ~unique_ptr() { reset(); } + + unique_ptr& operator=(const unique_ptr& u) { + reset(const_cast(&u)->release()); + return *this; + } + + unique_ptr& operator=(unique_ptr&& u) { + reset(u.release()); + return *this; + } + + unique_ptr& operator=(T* p) { + reset(p); + return *this; + } + + const T& operator*() const { return *ptr_; } + T* operator->() const { return ptr_; } + T* get() const noexcept { return ptr_; } + explicit operator bool() const { return ptr_ != nullptr; } + + // modifiers + T* release() { + T* value = ptr_; + ptr_ = nullptr; + return value; + } + + void reset(T* p = nullptr) { + T* value = ptr_; + ptr_ = p; + if (value) delete value; + } + + void swap(unique_ptr& u) { + T* temp_ptr = ptr_; + ptr_ = u.ptr_; + u.ptr_ = temp_ptr; + } + + private: + T* ptr_; + }; + + template bool operator==(const unique_ptr& x, + const unique_ptr& y) { + return x.get() == y.get(); + } + + template bool operator==(const unique_ptr& x, + const D* y) { + return static_cast(x.get()) == y; + } + + template bool operator==(const unique_ptr& x, intptr_t y) { + return reinterpret_cast(x.get()) == y; + } + + template bool operator!=(const unique_ptr& x, decltype(nullptr)) { + return !!x; + } + + template bool operator!=(decltype(nullptr), const unique_ptr& x) { + return !!x; + } + + template bool operator==(const unique_ptr& x, decltype(nullptr)) { + return !x; + } + + template bool operator==(decltype(nullptr), const unique_ptr& x) { + return !x; + } + +#endif // !FLATBUFFERS_CPP98_STL + +#ifdef FLATBUFFERS_USE_STD_OPTIONAL +template +using Optional = std::optional; +using nullopt_t = std::nullopt_t; +inline constexpr nullopt_t nullopt = std::nullopt; + +#else +// Limited implementation of Optional type for a scalar T. +// This implementation limited by trivial types compatible with +// std::is_arithmetic or std::is_enum type traits. + +// A tag to indicate an empty flatbuffers::optional. +struct nullopt_t { + explicit FLATBUFFERS_CONSTEXPR_CPP11 nullopt_t(int) {} +}; + +#if defined(FLATBUFFERS_CONSTEXPR_DEFINED) + namespace internal { + template struct nullopt_holder { + static constexpr nullopt_t instance_ = nullopt_t(0); + }; + template + constexpr nullopt_t nullopt_holder::instance_; + } + static constexpr const nullopt_t &nullopt = internal::nullopt_holder::instance_; + +#else + namespace internal { + template struct nullopt_holder { + static const nullopt_t instance_; + }; + template + const nullopt_t nullopt_holder::instance_ = nullopt_t(0); + } + static const nullopt_t &nullopt = internal::nullopt_holder::instance_; + +#endif + +template +class Optional FLATBUFFERS_FINAL_CLASS { + // Non-scalar 'T' would extremely complicated Optional. + // Use is_scalar checking because flatbuffers flatbuffers::is_arithmetic + // isn't implemented. + static_assert(flatbuffers::is_scalar::value, "unexpected type T"); + + public: + ~Optional() {} + + FLATBUFFERS_CONSTEXPR_CPP11 Optional() FLATBUFFERS_NOEXCEPT + : value_(), has_value_(false) {} + + FLATBUFFERS_CONSTEXPR_CPP11 Optional(nullopt_t) FLATBUFFERS_NOEXCEPT + : value_(), has_value_(false) {} + + FLATBUFFERS_CONSTEXPR_CPP11 Optional(T val) FLATBUFFERS_NOEXCEPT + : value_(val), has_value_(true) {} + + FLATBUFFERS_CONSTEXPR_CPP11 Optional(const Optional &other) FLATBUFFERS_NOEXCEPT + : value_(other.value_), has_value_(other.has_value_) {} + + FLATBUFFERS_CONSTEXPR_CPP14 Optional &operator=(const Optional &other) FLATBUFFERS_NOEXCEPT { + value_ = other.value_; + has_value_ = other.has_value_; + return *this; + } + + FLATBUFFERS_CONSTEXPR_CPP14 Optional &operator=(nullopt_t) FLATBUFFERS_NOEXCEPT { + value_ = T(); + has_value_ = false; + return *this; + } + + FLATBUFFERS_CONSTEXPR_CPP14 Optional &operator=(T val) FLATBUFFERS_NOEXCEPT { + value_ = val; + has_value_ = true; + return *this; + } + + void reset() FLATBUFFERS_NOEXCEPT { + *this = nullopt; + } + + void swap(Optional &other) FLATBUFFERS_NOEXCEPT { + std::swap(value_, other.value_); + std::swap(has_value_, other.has_value_); + } + + FLATBUFFERS_CONSTEXPR_CPP11 FLATBUFFERS_EXPLICIT_CPP11 operator bool() const FLATBUFFERS_NOEXCEPT { + return has_value_; + } + + FLATBUFFERS_CONSTEXPR_CPP11 bool has_value() const FLATBUFFERS_NOEXCEPT { + return has_value_; + } + + FLATBUFFERS_CONSTEXPR_CPP11 const T& operator*() const FLATBUFFERS_NOEXCEPT { + return value_; + } + + const T& value() const { + FLATBUFFERS_ASSERT(has_value()); + return value_; + } + + T value_or(T default_value) const FLATBUFFERS_NOEXCEPT { + return has_value() ? value_ : default_value; + } + + private: + T value_; + bool has_value_; +}; + +template +FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const Optional& opt, nullopt_t) FLATBUFFERS_NOEXCEPT { + return !opt; +} +template +FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(nullopt_t, const Optional& opt) FLATBUFFERS_NOEXCEPT { + return !opt; +} + +template +FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const Optional& lhs, const U& rhs) FLATBUFFERS_NOEXCEPT { + return static_cast(lhs) && (*lhs == rhs); +} + +template +FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const T& lhs, const Optional& rhs) FLATBUFFERS_NOEXCEPT { + return static_cast(rhs) && (lhs == *rhs); +} + +template +FLATBUFFERS_CONSTEXPR_CPP11 bool operator==(const Optional& lhs, const Optional& rhs) FLATBUFFERS_NOEXCEPT { + return static_cast(lhs) != static_cast(rhs) + ? false + : !static_cast(lhs) ? false : (*lhs == *rhs); +} +#endif // FLATBUFFERS_USE_STD_OPTIONAL + +} // namespace flatbuffers + +#endif // FLATBUFFERS_STL_EMULATION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/util.h b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/util.h new file mode 100644 index 0000000..c00718b --- /dev/null +++ b/src/tensorflow_arm/third_party/flatbuffers/include/flatbuffers/util.h @@ -0,0 +1,695 @@ +#if !defined(ESP32) +/* + * Copyright 2014 Google Inc. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef FLATBUFFERS_UTIL_H_ +#define FLATBUFFERS_UTIL_H_ + +#include + +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/base.h" +#include "tensorflow_arm/third_party/flatbuffers/include/flatbuffers/stl_emulation.h" + +#ifndef FLATBUFFERS_PREFER_PRINTF +# include +#else // FLATBUFFERS_PREFER_PRINTF +# include +# include +#endif // FLATBUFFERS_PREFER_PRINTF + +#include +#include + +namespace flatbuffers { + +// @locale-independent functions for ASCII characters set. + +// Fast checking that character lies in closed range: [a <= x <= b] +// using one compare (conditional branch) operator. +inline bool check_ascii_range(char x, char a, char b) { + FLATBUFFERS_ASSERT(a <= b); + // (Hacker's Delight): `a <= x <= b` <=> `(x-a) <={u} (b-a)`. + // The x, a, b will be promoted to int and subtracted without overflow. + return static_cast(x - a) <= static_cast(b - a); +} + +// Case-insensitive isalpha +inline bool is_alpha(char c) { + // ASCII only: alpha to upper case => reset bit 0x20 (~0x20 = 0xDF). + return check_ascii_range(c & 0xDF, 'a' & 0xDF, 'z' & 0xDF); +} + +// Check (case-insensitive) that `c` is equal to alpha. +inline bool is_alpha_char(char c, char alpha) { + FLATBUFFERS_ASSERT(is_alpha(alpha)); + // ASCII only: alpha to upper case => reset bit 0x20 (~0x20 = 0xDF). + return ((c & 0xDF) == (alpha & 0xDF)); +} + +// https://en.cppreference.com/w/cpp/string/byte/isxdigit +// isdigit and isxdigit are the only standard narrow character classification +// functions that are not affected by the currently installed C locale. although +// some implementations (e.g. Microsoft in 1252 codepage) may classify +// additional single-byte characters as digits. +inline bool is_digit(char c) { return check_ascii_range(c, '0', '9'); } + +inline bool is_xdigit(char c) { + // Replace by look-up table. + return is_digit(c) || check_ascii_range(c & 0xDF, 'a' & 0xDF, 'f' & 0xDF); +} + +// Case-insensitive isalnum +inline bool is_alnum(char c) { return is_alpha(c) || is_digit(c); } + +inline char CharToUpper(char c) { + return static_cast(::toupper(static_cast(c))); +} + +inline char CharToLower(char c) { + return static_cast(::tolower(static_cast(c))); +} + +// @end-locale-independent functions for ASCII character set + +#ifdef FLATBUFFERS_PREFER_PRINTF +template size_t IntToDigitCount(T t) { + size_t digit_count = 0; + // Count the sign for negative numbers + if (t < 0) digit_count++; + // Count a single 0 left of the dot for fractional numbers + if (-1 < t && t < 1) digit_count++; + // Count digits until fractional part + T eps = std::numeric_limits::epsilon(); + while (t <= (-1 + eps) || (1 - eps) <= t) { + t /= 10; + digit_count++; + } + return digit_count; +} + +template size_t NumToStringWidth(T t, int precision = 0) { + size_t string_width = IntToDigitCount(t); + // Count the dot for floating point numbers + if (precision) string_width += (precision + 1); + return string_width; +} + +template +std::string NumToStringImplWrapper(T t, const char *fmt, int precision = 0) { + size_t string_width = NumToStringWidth(t, precision); + std::string s(string_width, 0x00); + // Allow snprintf to use std::string trailing null to detect buffer overflow + snprintf(const_cast(s.data()), (s.size() + 1), fmt, string_width, t); + return s; +} +#endif // FLATBUFFERS_PREFER_PRINTF + +// Convert an integer or floating point value to a string. +// In contrast to std::stringstream, "char" values are +// converted to a string of digits, and we don't use scientific notation. +template std::string NumToString(T t) { + // clang-format off + + #ifndef FLATBUFFERS_PREFER_PRINTF + std::stringstream ss; + ss << t; + return ss.str(); + #else // FLATBUFFERS_PREFER_PRINTF + auto v = static_cast(t); + return NumToStringImplWrapper(v, "%.*lld"); + #endif // FLATBUFFERS_PREFER_PRINTF + // clang-format on +} +// Avoid char types used as character data. +template<> inline std::string NumToString(signed char t) { + return NumToString(static_cast(t)); +} +template<> inline std::string NumToString(unsigned char t) { + return NumToString(static_cast(t)); +} +template<> inline std::string NumToString(char t) { + return NumToString(static_cast(t)); +} +#if defined(FLATBUFFERS_CPP98_STL) +template<> inline std::string NumToString(long long t) { + char buf[21]; // (log((1 << 63) - 1) / log(10)) + 2 + snprintf(buf, sizeof(buf), "%lld", t); + return std::string(buf); +} + +template<> +inline std::string NumToString(unsigned long long t) { + char buf[22]; // (log((1 << 63) - 1) / log(10)) + 1 + snprintf(buf, sizeof(buf), "%llu", t); + return std::string(buf); +} +#endif // defined(FLATBUFFERS_CPP98_STL) + +// Special versions for floats/doubles. +template std::string FloatToString(T t, int precision) { + // clang-format off + + #ifndef FLATBUFFERS_PREFER_PRINTF + // to_string() prints different numbers of digits for floats depending on + // platform and isn't available on Android, so we use stringstream + std::stringstream ss; + // Use std::fixed to suppress scientific notation. + ss << std::fixed; + // Default precision is 6, we want that to be higher for doubles. + ss << std::setprecision(precision); + ss << t; + auto s = ss.str(); + #else // FLATBUFFERS_PREFER_PRINTF + auto v = static_cast(t); + auto s = NumToStringImplWrapper(v, "%0.*f", precision); + #endif // FLATBUFFERS_PREFER_PRINTF + // clang-format on + // Sadly, std::fixed turns "1" into "1.00000", so here we undo that. + auto p = s.find_last_not_of('0'); + if (p != std::string::npos) { + // Strip trailing zeroes. If it is a whole number, keep one zero. + s.resize(p + (s[p] == '.' ? 2 : 1)); + } + return s; +} + +template<> inline std::string NumToString(double t) { + return FloatToString(t, 12); +} +template<> inline std::string NumToString(float t) { + return FloatToString(t, 6); +} + +// Convert an integer value to a hexadecimal string. +// The returned string length is always xdigits long, prefixed by 0 digits. +// For example, IntToStringHex(0x23, 8) returns the string "00000023". +inline std::string IntToStringHex(int i, int xdigits) { + FLATBUFFERS_ASSERT(i >= 0); + // clang-format off + + #ifndef FLATBUFFERS_PREFER_PRINTF + std::stringstream ss; + ss << std::setw(xdigits) << std::setfill('0') << std::hex << std::uppercase + << i; + return ss.str(); + #else // FLATBUFFERS_PREFER_PRINTF + return NumToStringImplWrapper(i, "%.*X", xdigits); + #endif // FLATBUFFERS_PREFER_PRINTF + // clang-format on +} + +// clang-format off +// Use locale independent functions {strtod_l, strtof_l, strtoll_l, strtoull_l}. +#if defined(FLATBUFFERS_LOCALE_INDEPENDENT) && (FLATBUFFERS_LOCALE_INDEPENDENT > 0) + class ClassicLocale { + #ifdef _MSC_VER + typedef _locale_t locale_type; + #else + typedef locale_t locale_type; // POSIX.1-2008 locale_t type + #endif + ClassicLocale(); + ~ClassicLocale(); + locale_type locale_; + static ClassicLocale instance_; + public: + static locale_type Get() { return instance_.locale_; } + }; + + #ifdef _MSC_VER + #define __strtoull_impl(s, pe, b) _strtoui64_l(s, pe, b, ClassicLocale::Get()) + #define __strtoll_impl(s, pe, b) _strtoi64_l(s, pe, b, ClassicLocale::Get()) + #define __strtod_impl(s, pe) _strtod_l(s, pe, ClassicLocale::Get()) + #define __strtof_impl(s, pe) _strtof_l(s, pe, ClassicLocale::Get()) + #else + #define __strtoull_impl(s, pe, b) strtoull_l(s, pe, b, ClassicLocale::Get()) + #define __strtoll_impl(s, pe, b) strtoll_l(s, pe, b, ClassicLocale::Get()) + #define __strtod_impl(s, pe) strtod_l(s, pe, ClassicLocale::Get()) + #define __strtof_impl(s, pe) strtof_l(s, pe, ClassicLocale::Get()) + #endif +#else + #define __strtod_impl(s, pe) strtod(s, pe) + #define __strtof_impl(s, pe) static_cast(strtod(s, pe)) + #ifdef _MSC_VER + #define __strtoull_impl(s, pe, b) _strtoui64(s, pe, b) + #define __strtoll_impl(s, pe, b) _strtoi64(s, pe, b) + #else + #define __strtoull_impl(s, pe, b) strtoull(s, pe, b) + #define __strtoll_impl(s, pe, b) strtoll(s, pe, b) + #endif +#endif + +inline void strtoval_impl(int64_t *val, const char *str, char **endptr, + int base) { + *val = __strtoll_impl(str, endptr, base); +} + +inline void strtoval_impl(uint64_t *val, const char *str, char **endptr, + int base) { + *val = __strtoull_impl(str, endptr, base); +} + +inline void strtoval_impl(double *val, const char *str, char **endptr) { + *val = __strtod_impl(str, endptr); +} + +// UBSAN: double to float is safe if numeric_limits::is_iec559 is true. +__supress_ubsan__("float-cast-overflow") +inline void strtoval_impl(float *val, const char *str, char **endptr) { + *val = __strtof_impl(str, endptr); +} +#undef __strtoull_impl +#undef __strtoll_impl +#undef __strtod_impl +#undef __strtof_impl +// clang-format on + +// Adaptor for strtoull()/strtoll(). +// Flatbuffers accepts numbers with any count of leading zeros (-009 is -9), +// while strtoll with base=0 interprets first leading zero as octal prefix. +// In future, it is possible to add prefixed 0b0101. +// 1) Checks errno code for overflow condition (out of range). +// 2) If base <= 0, function try to detect base of number by prefix. +// +// Return value (like strtoull and strtoll, but reject partial result): +// - If successful, an integer value corresponding to the str is returned. +// - If full string conversion can't be performed, 0 is returned. +// - If the converted value falls out of range of corresponding return type, a +// range error occurs. In this case value MAX(T)/MIN(T) is returned. +template +inline bool StringToIntegerImpl(T *val, const char *const str, + const int base = 0, + const bool check_errno = true) { + // T is int64_t or uint64_T + FLATBUFFERS_ASSERT(str); + if (base <= 0) { + auto s = str; + while (*s && !is_digit(*s)) s++; + if (s[0] == '0' && is_alpha_char(s[1], 'X')) + return StringToIntegerImpl(val, str, 16, check_errno); + // if a prefix not match, try base=10 + return StringToIntegerImpl(val, str, 10, check_errno); + } else { + if (check_errno) errno = 0; // clear thread-local errno + auto endptr = str; + strtoval_impl(val, str, const_cast(&endptr), base); + if ((*endptr != '\0') || (endptr == str)) { + *val = 0; // erase partial result + return false; // invalid string + } + // errno is out-of-range, return MAX/MIN + if (check_errno && errno) return false; + return true; + } +} + +template +inline bool StringToFloatImpl(T *val, const char *const str) { + // Type T must be either float or double. + FLATBUFFERS_ASSERT(str && val); + auto end = str; + strtoval_impl(val, str, const_cast(&end)); + auto done = (end != str) && (*end == '\0'); + if (!done) *val = 0; // erase partial result + return done; +} + +// Convert a string to an instance of T. +// Return value (matched with StringToInteger64Impl and strtod): +// - If successful, a numeric value corresponding to the str is returned. +// - If full string conversion can't be performed, 0 is returned. +// - If the converted value falls out of range of corresponding return type, a +// range error occurs. In this case value MAX(T)/MIN(T) is returned. +template inline bool StringToNumber(const char *s, T *val) { + FLATBUFFERS_ASSERT(s && val); + int64_t i64; + // The errno check isn't needed, will return MAX/MIN on overflow. + if (StringToIntegerImpl(&i64, s, 0, false)) { + const int64_t max = (flatbuffers::numeric_limits::max)(); + const int64_t min = flatbuffers::numeric_limits::lowest(); + if (i64 > max) { + *val = static_cast(max); + return false; + } + if (i64 < min) { + // For unsigned types return max to distinguish from + // "no conversion can be performed" when 0 is returned. + *val = static_cast(flatbuffers::is_unsigned::value ? max : min); + return false; + } + *val = static_cast(i64); + return true; + } + *val = 0; + return false; +} + +template<> inline bool StringToNumber(const char *str, int64_t *val) { + return StringToIntegerImpl(val, str); +} + +template<> +inline bool StringToNumber(const char *str, uint64_t *val) { + if (!StringToIntegerImpl(val, str)) return false; + // The strtoull accepts negative numbers: + // If the minus sign was part of the input sequence, the numeric value + // calculated from the sequence of digits is negated as if by unary minus + // in the result type, which applies unsigned integer wraparound rules. + // Fix this behaviour (except -0). + if (*val) { + auto s = str; + while (*s && !is_digit(*s)) s++; + s = (s > str) ? (s - 1) : s; // step back to one symbol + if (*s == '-') { + // For unsigned types return the max to distinguish from + // "no conversion can be performed". + *val = (flatbuffers::numeric_limits::max)(); + return false; + } + } + return true; +} + +template<> inline bool StringToNumber(const char *s, float *val) { + return StringToFloatImpl(val, s); +} + +template<> inline bool StringToNumber(const char *s, double *val) { + return StringToFloatImpl(val, s); +} + +inline int64_t StringToInt(const char *s, int base = 10) { + int64_t val; + return StringToIntegerImpl(&val, s, base) ? val : 0; +} + +inline uint64_t StringToUInt(const char *s, int base = 10) { + uint64_t val; + return StringToIntegerImpl(&val, s, base) ? val : 0; +} + +typedef bool (*LoadFileFunction)(const char *filename, bool binary, + std::string *dest); +typedef bool (*FileExistsFunction)(const char *filename); + +LoadFileFunction SetLoadFileFunction(LoadFileFunction load_file_function); + +FileExistsFunction SetFileExistsFunction( + FileExistsFunction file_exists_function); + +// Check if file "name" exists. +bool FileExists(const char *name); + +// Check if "name" exists and it is also a directory. +bool DirExists(const char *name); + +// Load file "name" into "buf" returning true if successful +// false otherwise. If "binary" is false data is read +// using ifstream's text mode, otherwise data is read with +// no transcoding. +bool LoadFile(const char *name, bool binary, std::string *buf); + +// Save data "buf" of length "len" bytes into a file +// "name" returning true if successful, false otherwise. +// If "binary" is false data is written using ifstream's +// text mode, otherwise data is written with no +// transcoding. +bool SaveFile(const char *name, const char *buf, size_t len, bool binary); + +// Save data "buf" into file "name" returning true if +// successful, false otherwise. If "binary" is false +// data is written using ifstream's text mode, otherwise +// data is written with no transcoding. +inline bool SaveFile(const char *name, const std::string &buf, bool binary) { + return SaveFile(name, buf.c_str(), buf.size(), binary); +} + +// Functionality for minimalistic portable path handling. + +// The functions below behave correctly regardless of whether posix ('/') or +// Windows ('/' or '\\') separators are used. + +// Any new separators inserted are always posix. +FLATBUFFERS_CONSTEXPR char kPathSeparator = '/'; + +// Returns the path with the extension, if any, removed. +std::string StripExtension(const std::string &filepath); + +// Returns the extension, if any. +std::string GetExtension(const std::string &filepath); + +// Return the last component of the path, after the last separator. +std::string StripPath(const std::string &filepath); + +// Strip the last component of the path + separator. +std::string StripFileName(const std::string &filepath); + +// Concatenates a path with a filename, regardless of whether the path +// ends in a separator or not. +std::string ConCatPathFileName(const std::string &path, + const std::string &filename); + +// Replaces any '\\' separators with '/' +std::string PosixPath(const char *path); + +// This function ensure a directory exists, by recursively +// creating dirs for any parts of the path that don't exist yet. +void EnsureDirExists(const std::string &filepath); + +// Obtains the absolute path from any other path. +// Returns the input path if the absolute path couldn't be resolved. +std::string AbsolutePath(const std::string &filepath); + +// To and from UTF-8 unicode conversion functions + +// Convert a unicode code point into a UTF-8 representation by appending it +// to a string. Returns the number of bytes generated. +inline int ToUTF8(uint32_t ucc, std::string *out) { + FLATBUFFERS_ASSERT(!(ucc & 0x80000000)); // Top bit can't be set. + // 6 possible encodings: http://en.wikipedia.org/wiki/UTF-8 + for (int i = 0; i < 6; i++) { + // Max bits this encoding can represent. + uint32_t max_bits = 6 + i * 5 + static_cast(!i); + if (ucc < (1u << max_bits)) { // does it fit? + // Remaining bits not encoded in the first byte, store 6 bits each + uint32_t remain_bits = i * 6; + // Store first byte: + (*out) += static_cast((0xFE << (max_bits - remain_bits)) | + (ucc >> remain_bits)); + // Store remaining bytes: + for (int j = i - 1; j >= 0; j--) { + (*out) += static_cast(((ucc >> (j * 6)) & 0x3F) | 0x80); + } + return i + 1; // Return the number of bytes added. + } + } + FLATBUFFERS_ASSERT(0); // Impossible to arrive here. + return -1; +} + +// Converts whatever prefix of the incoming string corresponds to a valid +// UTF-8 sequence into a unicode code. The incoming pointer will have been +// advanced past all bytes parsed. +// returns -1 upon corrupt UTF-8 encoding (ignore the incoming pointer in +// this case). +inline int FromUTF8(const char **in) { + int len = 0; + // Count leading 1 bits. + for (int mask = 0x80; mask >= 0x04; mask >>= 1) { + if (**in & mask) { + len++; + } else { + break; + } + } + if ((static_cast(**in) << len) & 0x80) + return -1; // Bit after leading 1's must be 0. + if (!len) return *(*in)++; + // UTF-8 encoded values with a length are between 2 and 4 bytes. + if (len < 2 || len > 4) { return -1; } + // Grab initial bits of the code. + int ucc = *(*in)++ & ((1 << (7 - len)) - 1); + for (int i = 0; i < len - 1; i++) { + if ((**in & 0xC0) != 0x80) return -1; // Upper bits must 1 0. + ucc <<= 6; + ucc |= *(*in)++ & 0x3F; // Grab 6 more bits of the code. + } + // UTF-8 cannot encode values between 0xD800 and 0xDFFF (reserved for + // UTF-16 surrogate pairs). + if (ucc >= 0xD800 && ucc <= 0xDFFF) { return -1; } + // UTF-8 must represent code points in their shortest possible encoding. + switch (len) { + case 2: + // Two bytes of UTF-8 can represent code points from U+0080 to U+07FF. + if (ucc < 0x0080 || ucc > 0x07FF) { return -1; } + break; + case 3: + // Three bytes of UTF-8 can represent code points from U+0800 to U+FFFF. + if (ucc < 0x0800 || ucc > 0xFFFF) { return -1; } + break; + case 4: + // Four bytes of UTF-8 can represent code points from U+10000 to U+10FFFF. + if (ucc < 0x10000 || ucc > 0x10FFFF) { return -1; } + break; + } + return ucc; +} + +#ifndef FLATBUFFERS_PREFER_PRINTF +// Wraps a string to a maximum length, inserting new lines where necessary. Any +// existing whitespace will be collapsed down to a single space. A prefix or +// suffix can be provided, which will be inserted before or after a wrapped +// line, respectively. +inline std::string WordWrap(const std::string in, size_t max_length, + const std::string wrapped_line_prefix, + const std::string wrapped_line_suffix) { + std::istringstream in_stream(in); + std::string wrapped, line, word; + + in_stream >> word; + line = word; + + while (in_stream >> word) { + if ((line.length() + 1 + word.length() + wrapped_line_suffix.length()) < + max_length) { + line += " " + word; + } else { + wrapped += line + wrapped_line_suffix + "\n"; + line = wrapped_line_prefix + word; + } + } + wrapped += line; + + return wrapped; +} +#endif // !FLATBUFFERS_PREFER_PRINTF + +inline bool EscapeString(const char *s, size_t length, std::string *_text, + bool allow_non_utf8, bool natural_utf8) { + std::string &text = *_text; + text += "\""; + for (uoffset_t i = 0; i < length; i++) { + char c = s[i]; + switch (c) { + case '\n': text += "\\n"; break; + case '\t': text += "\\t"; break; + case '\r': text += "\\r"; break; + case '\b': text += "\\b"; break; + case '\f': text += "\\f"; break; + case '\"': text += "\\\""; break; + case '\\': text += "\\\\"; break; + default: + if (c >= ' ' && c <= '~') { + text += c; + } else { + // Not printable ASCII data. Let's see if it's valid UTF-8 first: + const char *utf8 = s + i; + int ucc = FromUTF8(&utf8); + if (ucc < 0) { + if (allow_non_utf8) { + text += "\\x"; + text += IntToStringHex(static_cast(c), 2); + } else { + // There are two cases here: + // + // 1) We reached here by parsing an IDL file. In that case, + // we previously checked for non-UTF-8, so we shouldn't reach + // here. + // + // 2) We reached here by someone calling GenerateText() + // on a previously-serialized flatbuffer. The data might have + // non-UTF-8 Strings, or might be corrupt. + // + // In both cases, we have to give up and inform the caller + // they have no JSON. + return false; + } + } else { + if (natural_utf8) { + // utf8 points to past all utf-8 bytes parsed + text.append(s + i, static_cast(utf8 - s - i)); + } else if (ucc <= 0xFFFF) { + // Parses as Unicode within JSON's \uXXXX range, so use that. + text += "\\u"; + text += IntToStringHex(ucc, 4); + } else if (ucc <= 0x10FFFF) { + // Encode Unicode SMP values to a surrogate pair using two \u + // escapes. + uint32_t base = ucc - 0x10000; + auto high_surrogate = (base >> 10) + 0xD800; + auto low_surrogate = (base & 0x03FF) + 0xDC00; + text += "\\u"; + text += IntToStringHex(high_surrogate, 4); + text += "\\u"; + text += IntToStringHex(low_surrogate, 4); + } + // Skip past characters recognized. + i = static_cast(utf8 - s - 1); + } + } + break; + } + } + text += "\""; + return true; +} + +inline std::string BufferToHexText(const void *buffer, size_t buffer_size, + size_t max_length, + const std::string &wrapped_line_prefix, + const std::string &wrapped_line_suffix) { + std::string text = wrapped_line_prefix; + size_t start_offset = 0; + const char *s = reinterpret_cast(buffer); + for (size_t i = 0; s && i < buffer_size; i++) { + // Last iteration or do we have more? + bool have_more = i + 1 < buffer_size; + text += "0x"; + text += IntToStringHex(static_cast(s[i]), 2); + if (have_more) { text += ','; } + // If we have more to process and we reached max_length + if (have_more && + text.size() + wrapped_line_suffix.size() >= start_offset + max_length) { + text += wrapped_line_suffix; + text += '\n'; + start_offset = text.size(); + text += wrapped_line_prefix; + } + } + text += wrapped_line_suffix; + return text; +} + +// Remove paired quotes in a string: "text"|'text' -> text. +std::string RemoveStringQuotes(const std::string &s); + +// Change th global C-locale to locale with name . +// Returns an actual locale name in <_value>, useful if locale_name is "" or +// null. +bool SetGlobalTestLocale(const char *locale_name, + std::string *_value = nullptr); + +// Read (or test) a value of environment variable. +bool ReadEnvironmentVariable(const char *var_name, + std::string *_value = nullptr); + +// MSVC specific: Send all assert reports to STDOUT to prevent CI hangs. +void SetupDefaultCRTReportMode(); + +} // namespace flatbuffers + +#endif // FLATBUFFERS_UTIL_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/gemmlowp/LICENSE b/src/tensorflow_arm/third_party/gemmlowp/LICENSE new file mode 100644 index 0000000..d645695 --- /dev/null +++ b/src/tensorflow_arm/third_party/gemmlowp/LICENSE @@ -0,0 +1,202 @@ + + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. diff --git a/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h b/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h new file mode 100644 index 0000000..db83da4 --- /dev/null +++ b/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h @@ -0,0 +1,903 @@ +#if !defined(ESP32) +// Copyright 2015 The Gemmlowp Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// fixedpoint.h: fixed-point arithmetic, with basic operations and +// a few math functions such as tanh. + +#ifndef GEMMLOWP_INTERNAL_FIXEDPOINT_H_ +#define GEMMLOWP_INTERNAL_FIXEDPOINT_H_ + +#include +#include +#include +#include +#include + +#include "../internal/detect_platform.h" + +namespace gemmlowp { + +// Part 1: Low-level integer-arithmetic primitives. +// The implementations here are generic implementations valid for +// scalar types (e.g. std::int32_t). Architecture-specific SIMD types +// (e.g. NEON int32x4_t) may be supported by providing +// specializations for them in separate files. +// +// The purpose of these primitives is two-fold: +// - They will be used to implement higher-level fixed-point +// abstractions, namely the FixedPoint class and its arithmetic +// operators. +// - They will be directly used to implement some more involved +// fixed-point computations, e.g. the fixed-point implementation +// of math functions such as tanh. + +// Some compile-time traits around raw types to handle SIMD aspects: +// number of lanes, underlying scalar type. +template +struct FixedPointRawTypeTraits {}; + +template <> +struct FixedPointRawTypeTraits { + typedef std::int32_t ScalarRawType; + static constexpr int kLanes = 1; +}; + +template <> +struct FixedPointRawTypeTraits { + typedef std::int16_t ScalarRawType; + static constexpr int kLanes = 1; +}; + +// Returns a SIMD value duplicating a scalar value across all lanes. +template +tRawType Dup(typename FixedPointRawTypeTraits::ScalarRawType x) { + return x; +} + +// Plain bit-wise AND +template +tIntegerType BitAnd(tIntegerType a, tIntegerType b) { + return a & b; +} + +// Plain bit-wise OR +template +tIntegerType BitOr(tIntegerType a, tIntegerType b) { + return a | b; +} + +// Plain bit-wise XOR +template +tIntegerType BitXor(tIntegerType a, tIntegerType b) { + return a ^ b; +} + +// Plain bit-wise NOT +template +tIntegerType BitNot(tIntegerType a) { + return ~a; +} + +// Integer addition. Not saturating. Overflow is undefined behavior. +template +tIntegerType Add(tIntegerType a, tIntegerType b) { + return a + b; +} + +// Integer subtraction. Not saturating. Overflow is undefined behavior. +template +tIntegerType Mul(tIntegerType a, tIntegerType b) { + return a * b; +} + +template +tIntegerType Sub(tIntegerType a, tIntegerType b) { + return a - b; +} + +// Integer unary negative. Not saturating. Overflow is undefined behavior. +template +tIntegerType Neg(tIntegerType a) { + return -a; +} + +// Integer arithmetic left-shift, equivalent to multiplying with a power of two. +// Negative values are OK. In case of overflow, no Undefined +// Behavior, but the results are implementation-defined (in practice, +// they currently are saturated, but we make no commitment to that). The idea +// is that the caller will want to implement the overflowing cases with +// saturation with compare-and-mask, so we don't care about the results +// in the overflow case, we just want to avoid undefined behavior. +// +// tIntegerType may be int32 or any narrower signed type. +template +tIntegerType ShiftLeft(tIntegerType a, int offset) { + const std::int64_t wide_a = static_cast(a); + const std::int64_t wide_shifted = wide_a * (1 << offset); + const auto min = std::numeric_limits::min(); + const auto max = std::numeric_limits::max(); + return wide_shifted < min + ? min + : wide_shifted > max ? max + : static_cast(wide_shifted); +} + +// Integer arithmetic right-shift. Not rounding. +// Relying on implementation-defined, but in-practice-consistent, +// C++ compiler behavior. +template +tIntegerType ShiftRight(tIntegerType a, int offset) { + return a >> offset; +} + +// Each bit of the result is set to the corresponding bit of either then_val or +// else_val depending on whether the corresponding bit of if_mask is set. +// Equivalent to the VBSL instruction in ARM NEON. +template +tIntegerType SelectUsingMask(tIntegerType if_mask, tIntegerType then_val, + tIntegerType else_val) { + return BitXor(BitAnd(if_mask, then_val), BitAnd(BitNot(if_mask), else_val)); +} + +// For each input scalar, the corresponding bits of the result are set if the +// input scalar is non-zero. +template +tIntegerType MaskIfNonZero(tIntegerType a) { + static constexpr tIntegerType zero = 0; + return a ? BitNot(zero) : zero; +} + +// For each input scalar, the corresponding bits of the result are set if the +// input scalar is zero. +template +tIntegerType MaskIfZero(tIntegerType a) { + return MaskIfNonZero(!a); +} + +// For each pair of input scalars, the corresponding bits of the result are +// set if the input scalars are equal. +template +tIntegerType MaskIfEqual(tIntegerType a, tIntegerType b) { + return MaskIfNonZero(a == b); +} + +// For each pair of input scalars, the corresponding bits of the result are +// set if the input scalars are not equal. +template +tIntegerType MaskIfNotEqual(tIntegerType a, tIntegerType b) { + return MaskIfNonZero(a != b); +} + +// For each pair of input scalars, the corresponding bits of the result are +// set if the input scalars a, b satisfy a > b. +template +tIntegerType MaskIfGreaterThan(tIntegerType a, tIntegerType b) { + return MaskIfNonZero(a > b); +} + +// For each pair of input scalars, the corresponding bits of the result are +// set if the input scalars a, b satisfy a >= b. +template +tIntegerType MaskIfGreaterThanOrEqual(tIntegerType a, tIntegerType b) { + return MaskIfNonZero(a >= b); +} + +// For each pair of input scalars, the corresponding bits of the result are +// set if the input scalars a, b satisfy a < b. +template +tIntegerType MaskIfLessThan(tIntegerType a, tIntegerType b) { + return MaskIfNonZero(a < b); +} + +// For each pair of input scalars, the corresponding bits of the result are +// set if the input scalars a, b satisfy a <= b. +template +tIntegerType MaskIfLessThanOrEqual(tIntegerType a, tIntegerType b) { + return MaskIfNonZero(a <= b); +} + +// Returns true if all of the input scalars are nonzero. +// This function may currently assume that each of the input scalars has either +// all or none of its bits set. Otherwise, its behavior is currently undefined. +template +bool All(tIntegerType a) { + return a; +} + +// Returns true if any of the input scalars are nonzero. +// This function may currently assume that each of the input scalars has either +// all or none of its bits set. Otherwise, its behavior is currently undefined. +template +bool Any(tIntegerType a) { + return a; +} + +// Returns (a+b)/2, rounded to the nearest integer. +// Equivalent to VRHADD in the ARM NEON instruction set. +template +IntegerType RoundingHalfSum(IntegerType a, IntegerType b) { + static_assert(std::is_same::value, "unimplemented"); + (void)b; + return a; +} + +template <> +inline std::int32_t RoundingHalfSum(std::int32_t a, std::int32_t b) { + std::int64_t a64 = a; + std::int64_t b64 = b; + std::int64_t sum = a64 + b64; + std::int64_t sign = sum >= 0 ? 1 : -1; + return static_cast((sum + sign) / 2); +} + +template <> +inline std::int16_t RoundingHalfSum(std::int16_t a, std::int16_t b) { + std::int32_t a32 = a; + std::int32_t b32 = b; + std::int32_t sum = a32 + b32; + std::int32_t sign = sum >= 0 ? 1 : -1; + return static_cast((sum + sign) / 2); +} + +template +IntegerType SaturatingAdd(IntegerType a, IntegerType b) { + static_assert(std::is_same::value, "unimplemented"); + (void)b; + return a; +} + +// So far this is only needed for int16. +template <> +inline std::int16_t SaturatingAdd(std::int16_t a, std::int16_t b) { + std::int32_t a32 = a; + std::int32_t b32 = b; + std::int32_t sum = a32 + b32; + return static_cast( + std::min(static_cast(32767), + std::max(static_cast(-32768), sum))); +} + +// Returns a+b, saturating if the integers are 16bit or narrower, +// otherwise just a plain addition. +template +struct AddSaturatingIf16BitImpl { + static IntegerType Run(IntegerType a, IntegerType b) { return Add(a, b); } +}; +template +struct AddSaturatingIf16BitImpl { + static IntegerType Run(IntegerType a, IntegerType b) { + return SaturatingAdd(a, b); + } +}; +template +IntegerType AddSaturatingIf16Bit(IntegerType a, IntegerType b) { + using ScalarType = + typename FixedPointRawTypeTraits::ScalarRawType; + return AddSaturatingIf16BitImpl::Run(a, + b); +} + +// Returns the integer that represents the product of two fixed-point +// numbers, interpreting all integers as fixed-point values in the +// interval [-1, 1), rounding to the nearest value, and saturating +// -1 * -1 to the maximum value (since 1 is not in the half-open +// interval [-1, 1)). +// +// [The explanation below specializes to std::int32_t for example purpose.] +// +// The mapping between IntegerType and the interval [-1, 1) is unique and +// implied by IntegerType, which is assumed to be signed. For example, +// for IntegerType==std::int32_t, the mapping is +// real_value = integer_value / 2^31. +// So in this case, and leaving aside rounding and saturating, this +// function computes ((a / 2^31) * (b / 2^31)) * 2^31, which simplifies to +// (a * b) / 2^31. +// +// The 'doubling' part in the name of this function comes from the fact that +// this operation is very close to a "multiply-high" operation, keeping only +// the top half bits, except that that would be effectively computing +// (a * b) / 2^32, +// so here we are computing 2x that, since +// 1/2^31 = 2 * 1/2^32. +// The idea is to use all of the available 32 bits in the destination int32 +// value. +// +// [End of the explanation specializing to int32.] +// +// This is equivalent to the VQRDMULH instruction in ARM NEON. +template +IntegerType SaturatingRoundingDoublingHighMul(IntegerType a, IntegerType b) { + static_assert(std::is_same::value, "unimplemented"); + (void)b; + return a; +} + +// This function implements the same computation as the ARMv7 NEON VQRDMULH +// instruction. +template <> +inline std::int32_t SaturatingRoundingDoublingHighMul(std::int32_t a, + std::int32_t b) { + bool overflow = a == b && a == std::numeric_limits::min(); + std::int64_t a_64(a); + std::int64_t b_64(b); + std::int64_t ab_64 = a_64 * b_64; + std::int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30)); + std::int32_t ab_x2_high32 = + static_cast((ab_64 + nudge) / (1ll << 31)); + return overflow ? std::numeric_limits::max() : ab_x2_high32; +} + +template <> +inline std::int16_t SaturatingRoundingDoublingHighMul(std::int16_t a, + std::int16_t b) { + bool overflow = a == b && a == std::numeric_limits::min(); + std::int32_t a_32(a); + std::int32_t b_32(b); + std::int32_t ab_32 = a_32 * b_32; + std::int16_t nudge = ab_32 >= 0 ? (1 << 14) : (1 - (1 << 14)); + std::int16_t ab_x2_high16 = + static_cast((ab_32 + nudge) / (1 << 15)); + return overflow ? std::numeric_limits::max() : ab_x2_high16; +} + +// Correctly-rounded-to-nearest division by a power-of-two. +// Also known as a rounding arithmetic right shift. +template +inline IntegerType RoundingDivideByPOT(IntegerType x, int exponent) { + assert(exponent >= 0); + assert(exponent <= 31); + const IntegerType mask = Dup((1ll << exponent) - 1); + const IntegerType zero = Dup(0); + const IntegerType one = Dup(1); + const IntegerType remainder = BitAnd(x, mask); + const IntegerType threshold = + Add(ShiftRight(mask, 1), BitAnd(MaskIfLessThan(x, zero), one)); + return Add(ShiftRight(x, exponent), + BitAnd(MaskIfGreaterThan(remainder, threshold), one)); +} + +// Returns the product of a run-time integer value by a compile-time power +// of two, with either a positive exponent (equivalent to an arithmetic +// left shift, saturating) or a negative exponent (equivalent to an arithmetic +// right shift, rounding to nearest). +template 0 ? 1 : Exponent < 0 ? -1 : 0)> +struct ImplSaturatingRoundingMultiplyByPOT {}; + +template +struct ImplSaturatingRoundingMultiplyByPOT { + static IntegerType eval(IntegerType x) { return x; } +}; + +template +struct ImplSaturatingRoundingMultiplyByPOT { + static IntegerType eval(IntegerType x) { + using ScalarIntegerType = + typename FixedPointRawTypeTraits::ScalarRawType; + const IntegerType min = + Dup(std::numeric_limits::min()); + const IntegerType max = + Dup(std::numeric_limits::max()); + const int ScalarIntegerTypeBits = 8 * sizeof(ScalarIntegerType); + + const std::int32_t threshold = + ((1 << (ScalarIntegerTypeBits - 1 - Exponent)) - 1); + const IntegerType positive_mask = + MaskIfGreaterThan(x, Dup(threshold)); + const IntegerType negative_mask = + MaskIfLessThan(x, Dup(-threshold)); + + IntegerType result = ShiftLeft(x, Exponent); + result = SelectUsingMask(positive_mask, max, result); + result = SelectUsingMask(negative_mask, min, result); + return result; + } +}; + +template +struct ImplSaturatingRoundingMultiplyByPOT { + static IntegerType eval(IntegerType x) { + return RoundingDivideByPOT(x, -Exponent); + } +}; + +template +IntegerType SaturatingRoundingMultiplyByPOT(IntegerType x) { + return ImplSaturatingRoundingMultiplyByPOT::eval(x); +} + +// Part 2: the FixedPoint class. + +// A FixedPoint object represents a fixed-point value stored in the underlying +// integer type tRawType, if tRawType is a plain scalar integer type. +// Alternatively, tRawType may be a SIMD type (e.g. NEON int32x4_t) in which +// case a FixedPoint object represents a corresponding SIMD vector of fixed +// point values. +// +// tIntegerBits describes the range of the fixed-point format: if +// tIntegerBits == m then the range of representable values is the half-open +// interval [-2^m; 2^m) where the open boundary on the right side means that +// 2^m is not representable (how close the maximum representable value is to +// it, depends on bit-depth of tRawType). +// +// In "Q format notation", +// https://en.wikipedia.org/wiki/Q_(number_format) +// we are describing the format +// Qm.n +// where +// m = tIntegerBits +// and +// n = NumberOfBits(tRawType) - (m + 1) +// Note that the (m + 1) in the above line is because we adopt the convention +// that we count the integer bits exclusively of the sign bit; so (m + 1) is +// the total number of integer bits inclusive of the sign bit. +// +// Accordingly, the number of integral representable values in our range +// [-2^m ; 2^m) +// is equal to 2^(m+1). +template +class FixedPoint { + public: + typedef tRawType RawType; + + typedef FixedPointRawTypeTraits RawTypeTraits; + typedef typename RawTypeTraits::ScalarRawType ScalarRawType; + + static constexpr int kTotalBits = 8 * sizeof(ScalarRawType); + static constexpr int kIntegerBits = tIntegerBits; + static constexpr int kFractionalBits = kTotalBits - 1 - kIntegerBits; + static_assert(kIntegerBits >= 0 && kIntegerBits < kTotalBits, + "bad IntegerBits"); + + typedef FixedPoint ScalarFixedPointType; + + static const ScalarRawType ScalarRawMin() { + return std::numeric_limits::min(); + } + + static const ScalarRawType ScalarRawMax() { + return std::numeric_limits::max(); + } + + static const ScalarRawType RawMin() { + return VectorFromScalar(ScalarRawMin()); + } + + static const ScalarRawType RawMax() { + return VectorFromScalar(ScalarRawMax()); + } + + static FixedPoint FromRaw(RawType x) { + FixedPoint retval; + retval.raw() = x; + return retval; + } + + static FixedPoint FromScalarRaw(ScalarRawType x) { + FixedPoint retval; + retval.raw() = Dup(x); + return retval; + } + + static FixedPoint FromScalarFixedPoint(ScalarFixedPointType x) { + return FromScalarRaw(x.raw()); + } + + template + static FixedPoint ConstantPOT() { + static constexpr int kOffset = kFractionalBits + Exponent; + static_assert( + kOffset < 31, + "Constant not exactly representable in this fixed-point format"); + return FromScalarRaw(ScalarRawType(1) << kOffset); + } + + static FixedPoint Zero() { return FromScalarRaw(0); } + + static FixedPoint One() { + return FromScalarRaw( + kIntegerBits == 0 + ? ScalarRawMax() + : (ScalarRawType(1) << (kIntegerBits == 0 ? 0 : kFractionalBits))); + } + + static FixedPoint FromDouble(double x) { + const double min_bound = static_cast(ScalarRawMin()); + const double max_bound = static_cast(ScalarRawMax()); + return FromScalarRaw(static_cast(std::min( + std::max(round(x * static_cast(1ll << kFractionalBits)), + min_bound), + max_bound))); + } + + RawType raw() const { return i_; } + RawType& raw() { return i_; } + + private: + RawType i_; +}; + +// Part 3: implementation of arithmetic operators for the +// FixedPoint class, and a few related functions. + +// A FixedPoint multiplication is just a +// SaturatingRoundingDoublingHighMul operation on the underlying +// raw integer values. The IntegerBits simply add up, as is obvious +// from the fact that the range is [-2^IntegerBits, 2^IntegerBits). +template +FixedPoint operator*( + FixedPoint a, + FixedPoint b) { + FixedPoint c; + c.raw() = SaturatingRoundingDoublingHighMul(a.raw(), b.raw()); + return c; +} + +// Tweaking IntegerBits gives exact multiplication by a power of two. +template +FixedPoint ExactMulByPot( + FixedPoint a) { + FixedPoint c; + c.raw() = a.raw(); + return c; +} + +// If we want to leave IntegerBits fixed, then multiplication +// by a power of two has to be saturating/rounding, not exact anymore. +template +FixedPoint SaturatingRoundingMultiplyByPOT( + FixedPoint a) { + return FixedPoint::FromRaw( + SaturatingRoundingMultiplyByPOT(a.raw())); +} + +// Generic arithmetic operators. + +#define MAKE_FIXEDPOINT_UNARY_FUNC(FuncName, ImplFuncName) \ + template \ + FixedPoint FuncName( \ + FixedPoint a) { \ + return FixedPoint::FromRaw(ImplFuncName(a.raw())); \ + } + +#define MAKE_FIXEDPOINT_BINARY_FUNC(FuncName, ImplFuncName) \ + template \ + FixedPoint FuncName( \ + FixedPoint a, \ + FixedPoint b) { \ + return FixedPoint::FromRaw( \ + ImplFuncName(a.raw(), b.raw())); \ + } + +MAKE_FIXEDPOINT_UNARY_FUNC(operator-, Neg) +MAKE_FIXEDPOINT_UNARY_FUNC(operator~, BitNot) +MAKE_FIXEDPOINT_BINARY_FUNC(operator+, Add) +MAKE_FIXEDPOINT_BINARY_FUNC(operator-, Sub) +MAKE_FIXEDPOINT_BINARY_FUNC(operator&, BitAnd) +MAKE_FIXEDPOINT_BINARY_FUNC(operator^, BitXor) +MAKE_FIXEDPOINT_BINARY_FUNC(operator|, BitOr) +MAKE_FIXEDPOINT_BINARY_FUNC(RoundingHalfSum, RoundingHalfSum) + +#undef MAKE_FIXEDPOINT_UNARY_FUNC +#undef MAKE_FIXEDPOINT_BINARY_FUNC + +#define MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW(FuncName) \ + template \ + tRawType FuncName(FixedPoint a) { \ + return FuncName(a.raw()); \ + } + +#define MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(FuncName) \ + template \ + tRawType FuncName(FixedPoint a, \ + FixedPoint b) { \ + return FuncName(a.raw(), b.raw()); \ + } + +MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW(MaskIfZero) +MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW(MaskIfNonZero) +MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfEqual) +MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfNotEqual) +MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfGreaterThan) +MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfGreaterThanOrEqual) +MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfLessThan) +MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfLessThanOrEqual) + +#undef MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW +#undef MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW + +template +FixedPoint SelectUsingMask( + tRawType if_mask, FixedPoint then_val, + FixedPoint else_val) { + return FixedPoint::FromRaw( + SelectUsingMask(if_mask, then_val.raw(), else_val.raw())); +} + +template +bool operator==(FixedPoint a, + FixedPoint b) { + return All(MaskIfEqual(a.raw(), b.raw())); +} + +template +bool operator!=(FixedPoint a, + FixedPoint b) { + return !(a == b); +} + +template +FixedPoint SaturatingAdd( + FixedPoint a, + FixedPoint b) { + return FixedPoint::FromRaw( + SaturatingAdd(a.raw(), b.raw())); +} + +template +FixedPoint AddSaturatingIf16Bit( + FixedPoint a, + FixedPoint b) { + return FixedPoint::FromRaw( + AddSaturatingIf16Bit(a.raw(), b.raw())); +} + +// Conversion to floating-point. +template +double ToDouble(FixedPoint x) { + static_assert(FixedPointRawTypeTraits::kLanes == 1, + "not applicable to SIMD types"); + typedef FixedPoint F; + return x.raw() / static_cast(1ll << F::kFractionalBits); +} + +// Rescale changes the number of IntegerBits and updates the underlying +// raw integer value accordingly. +template +FixedPoint Rescale( + FixedPoint x) { + static constexpr int kExponent = tIntegerBitsSrc - tIntegerBitsDst; + FixedPoint result; + result.raw() = SaturatingRoundingMultiplyByPOT(x.raw()); + return result; +} + +// CheckedFixedPointConstant allows to specify fixed-point constants +// initialized as real numbers, in a way that does not compile floating-point +// arithmetic in production code, yet still checks agreement with the +// floating-point expressions when asserts are enabled. +// +// The raw integer value provided is always a int32, encoding a 32-bit +// fixed-point value, regardless of the actual Scalar type. This allows +// writing generic code that applies just as well to the 32-bit and 16-bit +// cases. In the 16-bit case, the raw integer value is internally +// rounding-shifted by 16 bits to the right. +template +inline typename FixedPointType::ScalarRawType RescaleConstantInitializer( + std::int32_t int32_value) { + typedef typename FixedPointType::ScalarRawType ScalarRawType; + static constexpr int ScalarTypeBits = 8 * sizeof(ScalarRawType); + return static_cast( + RoundingDivideByPOT(int32_value, 32 - ScalarTypeBits)); +} +#ifdef GEMMLOWP_ENABLE_FIXEDPOINT_CONSTANTS_CHECKS +template +FixedPointType CheckedFixedPointConstant(std::int32_t raw_value, + double double_value) { + const FixedPointType result = FixedPointType::FromScalarRaw(raw_value); + assert(result == FixedPointType::FromDouble(double_value)); + return result; +} +#define GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(FixedPointType, \ + ScalarRawInt32Value, DoubleValue) \ + (gemmlowp::CheckedFixedPointConstant( \ + gemmlowp::RescaleConstantInitializer( \ + ScalarRawInt32Value), \ + DoubleValue)) + +#else +#define GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(FixedPointType, \ + ScalarRawInt32Value, DoubleValue) \ + (FixedPointType::FromScalarRaw( \ + gemmlowp::RescaleConstantInitializer( \ + ScalarRawInt32Value))) +#endif + +// Implementation of exponential function. + +// Returns exp(x) for x in [-1/4, 0). +template +FixedPoint exp_on_interval_between_negative_one_quarter_and_0_excl( + FixedPoint a) { + typedef FixedPoint F; + const F constant_term = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F, 1895147668, std::exp(-1.0 / 8.0)); + const F constant_1_over_3 = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F, 715827883, 1.0 / 3.0); + // We're evaluating a Taylor expansion around -1/8, so we do the change of + // variable: x = a + 1/8. + // In fixed-point with 0 integer bits, 1/8 is represented by 1 << 28. + F x = a + F::template ConstantPOT<-3>(); + F x2 = x * x; + F x3 = x2 * x; + F x4 = x2 * x2; + F x4_over_4 = SaturatingRoundingMultiplyByPOT<-2>(x4); + F x4_over_24_plus_x3_over_6_plus_x2_over_2 = + SaturatingRoundingMultiplyByPOT<-1>( + ((x4_over_4 + x3) * constant_1_over_3) + x2); + return AddSaturatingIf16Bit( + constant_term, + constant_term * (x + x4_over_24_plus_x3_over_6_plus_x2_over_2)); +} + +// Returns exp(x) for x < 0. +template +FixedPoint exp_on_negative_values( + FixedPoint a) { + typedef FixedPoint InputF; + typedef FixedPoint ResultF; + static constexpr int kFractionalBits = InputF::kFractionalBits; + static constexpr int kIntegerBits = InputF::kIntegerBits; + const InputF kOneQuarter = InputF::template ConstantPOT<-2>(); + InputF mask = kOneQuarter - InputF::FromScalarRaw(1); + InputF a_mod_quarter_minus_one_quarter = (a & mask) - kOneQuarter; + ResultF result = exp_on_interval_between_negative_one_quarter_and_0_excl( + Rescale<0>(a_mod_quarter_minus_one_quarter)); + tRawType remainder = (a_mod_quarter_minus_one_quarter - a).raw(); + +#define GEMMLOWP_EXP_BARREL_SHIFTER(Exponent, FixedPointMultiplier) \ + if (kIntegerBits > Exponent) { \ + const ResultF kMultiplier = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT( \ + ResultF, FixedPointMultiplier, std::exp(-std::pow(2.0, Exponent))); \ + static constexpr int kShiftAmount = \ + kIntegerBits > Exponent ? kFractionalBits + Exponent : 0; \ + result = SelectUsingMask( \ + MaskIfNonZero(BitAnd(remainder, Dup(1 << kShiftAmount))), \ + result * kMultiplier, result); \ + } + + GEMMLOWP_EXP_BARREL_SHIFTER(-2, 1672461947); + GEMMLOWP_EXP_BARREL_SHIFTER(-1, 1302514674); + GEMMLOWP_EXP_BARREL_SHIFTER(+0, 790015084); + GEMMLOWP_EXP_BARREL_SHIFTER(+1, 290630308); + GEMMLOWP_EXP_BARREL_SHIFTER(+2, 39332535); + GEMMLOWP_EXP_BARREL_SHIFTER(+3, 720401); + GEMMLOWP_EXP_BARREL_SHIFTER(+4, 242); + +#undef GEMMLOWP_EXP_BARREL_SHIFTER + + static constexpr int clampB = kIntegerBits > 5 ? 36 - kIntegerBits : 0; + if (kIntegerBits > 5) { + const InputF clamp = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(InputF, -(1 << clampB), -32.0); + result = SelectUsingMask(MaskIfLessThan(a, clamp), ResultF::Zero(), result); + } + + result = SelectUsingMask(MaskIfZero(a), ResultF::One(), result); + return result; +} + +// Implementation of tanh: (1 - exp(-2x)) / (1 + exp(-2x)). + +// Returns (1 - x) / (1 + x) for x in (0, 1). +template +FixedPoint one_minus_x_over_one_plus_x_for_x_in_0_1( + FixedPoint a) { + typedef FixedPoint F0; + typedef FixedPoint F2; + F0 half_denominator = RoundingHalfSum(a, F0::One()); + // Newton-Raphson division + // https://en.wikipedia.org/wiki/Division_algorithm#Newton.E2.80.93Raphson_division + // Refer to that page for the logic behind the 48/17 and 32/17 constants. + const F2 constant_48_over_17 = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, 1515870810, 48.0 / 17.0); + const F2 constant_neg_32_over_17 = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, -1010580540, -32.0 / 17.0); + F2 x = constant_48_over_17 + half_denominator * constant_neg_32_over_17; + for (int i = 0; i < 3; i++) { + F2 half_denominator_times_x = half_denominator * x; + F2 one_minus_half_denominator_times_x = + F2::One() - half_denominator_times_x; + x = x + Rescale<2>(x * one_minus_half_denominator_times_x); + } + return Rescale<0>(x - F2::One()); +} + +// Returns -tanh(x) for x < 0. +template +FixedPoint neg_tanh_on_negative_values( + FixedPoint a) { + return one_minus_x_over_one_plus_x_for_x_in_0_1( + exp_on_negative_values(ExactMulByPot<1>(a))); +} + +// Returns tanh(x) for any x. +template +FixedPoint tanh(FixedPoint a) { + typedef FixedPoint InputF; + typedef FixedPoint ResultF; + tRawType mask_if_negative = MaskIfLessThan(a, InputF::Zero()); + tRawType mask_if_zero = MaskIfZero(a); + InputF n = SelectUsingMask(mask_if_negative, a, -a); + ResultF t = neg_tanh_on_negative_values(n); + return SelectUsingMask(mask_if_zero, ResultF::Zero(), + SelectUsingMask(mask_if_negative, -t, t)); +} + +// Implementation of logistic function. + +// Returns 1 / (1 + x) for x in (0, 1). +template +FixedPoint one_over_one_plus_x_for_x_in_0_1( + FixedPoint a) { + typedef FixedPoint F0; + typedef FixedPoint F2; + F0 half_denominator = RoundingHalfSum(a, F0::One()); + // Newton-Raphson division + // https://en.wikipedia.org/wiki/Division_algorithm#Newton.E2.80.93Raphson_division + // Refer to that page for the logic behind the 48/17 and 32/17 constants. + const F2 constant_48_over_17 = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, 1515870810, 48.0 / 17.0); + const F2 constant_neg_32_over_17 = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, -1010580540, -32.0 / 17.0); + F2 x = constant_48_over_17 + half_denominator * constant_neg_32_over_17; + for (int i = 0; i < 3; i++) { + F2 half_denominator_times_x = half_denominator * x; + F2 one_minus_half_denominator_times_x = + F2::One() - half_denominator_times_x; + x = x + Rescale<2>(x * one_minus_half_denominator_times_x); + } + return Rescale<0>(ExactMulByPot<-1>(x)); +} + +// Returns logistic(x) = 1 / (1 + exp(-x)) for x > 0. +template +FixedPoint logistic_on_positive_values( + FixedPoint a) { + return one_over_one_plus_x_for_x_in_0_1(exp_on_negative_values(-a)); +} + +// Returns logistic(x) = 1 / (1 + exp(-x)) for any x. +template +FixedPoint logistic(FixedPoint a) { + typedef FixedPoint InputF; + typedef FixedPoint ResultF; + tRawType mask_if_positive = MaskIfGreaterThan(a, InputF::Zero()); + tRawType mask_if_zero = MaskIfZero(a); + InputF abs_input = SelectUsingMask(mask_if_positive, a, -a); + ResultF result_if_positive = logistic_on_positive_values(abs_input); + ResultF result_if_negative = ResultF::One() - result_if_positive; + const ResultF one_half = + GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(ResultF, 1 << 30, 0.5); + return SelectUsingMask(mask_if_zero, one_half, + SelectUsingMask(mask_if_positive, result_if_positive, + result_if_negative)); +} + +} // end namespace gemmlowp + +#ifdef GEMMLOWP_NEON +#include "./fixedpoint_neon.h" +#elif defined(GEMMLOWP_AVX2) +#include "./fixedpoint_avx.h" +#elif defined(GEMMLOWP_SSE4) +#include "./fixedpoint_sse.h" +#elif defined(GEMMLOWP_MSA) +#include "./fixedpoint_msa.h" +#endif + +#endif // GEMMLOWP_INTERNAL_FIXEDPOINT_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_neon.h b/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_neon.h new file mode 100644 index 0000000..c010008 --- /dev/null +++ b/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_neon.h @@ -0,0 +1,334 @@ +#if !defined(ESP32) +// Copyright 2015 The Gemmlowp Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// fixedpoint_neon.h: optimized NEON specializations of the templates +// in fixedpoint.h. + +#ifndef GEMMLOWP_INTERNAL_FIXEDPOINT_NEON_H_ +#define GEMMLOWP_INTERNAL_FIXEDPOINT_NEON_H_ + +#include + +namespace gemmlowp { + +template <> +struct FixedPointRawTypeTraits { + typedef std::int32_t ScalarRawType; + static constexpr int kLanes = 4; +}; + +template <> +struct FixedPointRawTypeTraits { + typedef std::int16_t ScalarRawType; + static constexpr int kLanes = 8; +}; + +template <> +inline int32x4_t BitAnd(int32x4_t a, int32x4_t b) { + return vandq_s32(a, b); +} + +template <> +inline int16x8_t BitAnd(int16x8_t a, int16x8_t b) { + return vandq_s16(a, b); +} + +template <> +inline int32x4_t BitOr(int32x4_t a, int32x4_t b) { + return vorrq_s32(a, b); +} + +template <> +inline int16x8_t BitOr(int16x8_t a, int16x8_t b) { + return vorrq_s16(a, b); +} + +template <> +inline int32x4_t BitXor(int32x4_t a, int32x4_t b) { + return veorq_s32(a, b); +} + +template <> +inline int16x8_t BitXor(int16x8_t a, int16x8_t b) { + return veorq_s16(a, b); +} + +template <> +inline int32x4_t BitNot(int32x4_t a) { + return veorq_s32(a, vdupq_n_s32(-1)); +} + +template <> +inline int16x8_t BitNot(int16x8_t a) { + return veorq_s16(a, vdupq_n_s16(-1)); +} + +template <> +inline int32x4_t Add(int32x4_t a, int32x4_t b) { + return vaddq_s32(a, b); +} + +template <> +inline int16x8_t Add(int16x8_t a, int16x8_t b) { + return vaddq_s16(a, b); +} + +template <> +inline int32x4_t Sub(int32x4_t a, int32x4_t b) { + return vsubq_s32(a, b); +} + +template <> +inline int16x8_t Sub(int16x8_t a, int16x8_t b) { + return vsubq_s16(a, b); +} + +template <> +inline int32x4_t Neg(int32x4_t a) { + return vnegq_s32(a); +} + +template <> +inline int16x8_t Neg(int16x8_t a) { + return vnegq_s16(a); +} + +template <> +inline int32x4_t ShiftLeft(int32x4_t a, int offset) { + return vshlq_s32(a, vdupq_n_s32(offset)); +} + +template <> +inline int16x8_t ShiftLeft(int16x8_t a, int offset) { + return vshlq_s16(a, vdupq_n_s16(offset)); +} + +template <> +inline int32x4_t ShiftRight(int32x4_t a, int offset) { + return vshlq_s32(a, vdupq_n_s32(-offset)); +} + +template <> +inline int16x8_t ShiftRight(int16x8_t a, int offset) { + return vshlq_s16(a, vdupq_n_s16(-offset)); +} + +template <> +inline int32x4_t SelectUsingMask(int32x4_t if_mask, int32x4_t then_val, + int32x4_t else_val) { + return vbslq_s32(vreinterpretq_u32_s32(if_mask), then_val, else_val); +} + +template <> +inline int16x8_t SelectUsingMask(int16x8_t if_mask, int16x8_t then_val, + int16x8_t else_val) { + return vbslq_s16(vreinterpretq_u16_s16(if_mask), then_val, else_val); +} + +template <> +inline int32x4_t MaskIfEqual(int32x4_t a, int32x4_t b) { + return vreinterpretq_s32_u32(vceqq_s32(a, b)); +} + +template <> +inline int16x8_t MaskIfEqual(int16x8_t a, int16x8_t b) { + return vreinterpretq_s16_u16(vceqq_s16(a, b)); +} + +template <> +inline int32x4_t MaskIfNotEqual(int32x4_t a, int32x4_t b) { + return BitNot(MaskIfEqual(a, b)); +} + +template <> +inline int16x8_t MaskIfNotEqual(int16x8_t a, int16x8_t b) { + return BitNot(MaskIfEqual(a, b)); +} + +template <> +inline int32x4_t MaskIfZero(int32x4_t a) { + return MaskIfEqual(a, vdupq_n_s32(0)); +} + +template <> +inline int16x8_t MaskIfZero(int16x8_t a) { + return MaskIfEqual(a, vdupq_n_s16(0)); +} + +template <> +inline int32x4_t MaskIfNonZero(int32x4_t a) { + return vreinterpretq_s32_u32(vtstq_s32(a, a)); +} + +template <> +inline int16x8_t MaskIfNonZero(int16x8_t a) { + return vreinterpretq_s16_u16(vtstq_s16(a, a)); +} + +template <> +inline int32x4_t MaskIfGreaterThan(int32x4_t a, int32x4_t b) { + return vreinterpretq_s32_u32(vcgtq_s32(a, b)); +} + +template <> +inline int16x8_t MaskIfGreaterThan(int16x8_t a, int16x8_t b) { + return vreinterpretq_s16_u16(vcgtq_s16(a, b)); +} + +template <> +inline int32x4_t MaskIfGreaterThanOrEqual(int32x4_t a, int32x4_t b) { + return vreinterpretq_s32_u32(vcgeq_s32(a, b)); +} + +template <> +inline int16x8_t MaskIfGreaterThanOrEqual(int16x8_t a, int16x8_t b) { + return vreinterpretq_s16_u16(vcgeq_s16(a, b)); +} + +template <> +inline int32x4_t MaskIfLessThan(int32x4_t a, int32x4_t b) { + return vreinterpretq_s32_u32(vcltq_s32(a, b)); +} + +template <> +inline int16x8_t MaskIfLessThan(int16x8_t a, int16x8_t b) { + return vreinterpretq_s16_u16(vcltq_s16(a, b)); +} + +template <> +inline int32x4_t MaskIfLessThanOrEqual(int32x4_t a, int32x4_t b) { + return vreinterpretq_s32_u32(vcleq_s32(a, b)); +} + +template <> +inline int16x8_t MaskIfLessThanOrEqual(int16x8_t a, int16x8_t b) { + return vreinterpretq_s16_u16(vcleq_s16(a, b)); +} + +template <> +inline bool All(int32x4_t a) { + a = vandq_s32(a, vextq_s32(a, a, 1)); + a = vandq_s32(a, vextq_s32(a, a, 2)); + return vgetq_lane_s32(a, 0); +} + +template <> +inline bool All(int16x8_t a) { + a = vandq_s16(a, vextq_s16(a, a, 1)); + a = vandq_s16(a, vextq_s16(a, a, 2)); + a = vandq_s16(a, vextq_s16(a, a, 4)); + return vgetq_lane_s16(a, 0); +} + +template <> +inline bool Any(int32x4_t a) { + a = vorrq_s32(a, vextq_s32(a, a, 1)); + a = vorrq_s32(a, vextq_s32(a, a, 2)); + return vgetq_lane_s32(a, 0); +} + +template <> +inline bool Any(int16x8_t a) { + a = vorrq_s16(a, vextq_s16(a, a, 1)); + a = vorrq_s16(a, vextq_s16(a, a, 2)); + a = vorrq_s16(a, vextq_s16(a, a, 4)); + return vgetq_lane_s16(a, 0); +} + +template <> +inline int32x4_t RoundingHalfSum(int32x4_t a, int32x4_t b) { + return vrhaddq_s32(a, b); +} + +template <> +inline int16x8_t RoundingHalfSum(int16x8_t a, int16x8_t b) { + return vrhaddq_s16(a, b); +} + +template <> +inline int32x4_t SaturatingRoundingDoublingHighMul(int32x4_t a, int32x4_t b) { + return vqrdmulhq_s32(a, b); +} + +template <> +inline int16x8_t SaturatingRoundingDoublingHighMul(int16x8_t a, int16x8_t b) { + return vqrdmulhq_s16(a, b); +} + +template <> +inline int32x4_t RoundingDivideByPOT(int32x4_t x, int exponent) { + const int32x4_t shift_vec = vdupq_n_s32(-exponent); + const int32x4_t fixup = vshrq_n_s32(vandq_s32(x, shift_vec), 31); + const int32x4_t fixed_up_x = vqaddq_s32(x, fixup); + return vrshlq_s32(fixed_up_x, shift_vec); +} + +template <> +inline int16x8_t RoundingDivideByPOT(int16x8_t x, int exponent) { + const int16x8_t shift_vec = vdupq_n_s16(-exponent); + const int16x8_t fixup = vshrq_n_s16(vandq_s16(x, shift_vec), 15); + const int16x8_t fixed_up_x = vqaddq_s16(x, fixup); + return vrshlq_s16(fixed_up_x, shift_vec); +} + +template +struct ImplSaturatingRoundingMultiplyByPOT { + static int32x4_t eval(int32x4_t x) { return vqshlq_n_s32(x, Exponent); } +}; + +template +struct ImplSaturatingRoundingMultiplyByPOT { + static int32x4_t eval(int32x4_t x) { + const int32x4_t fixup = vshrq_n_s32(x, 31); + const int32x4_t fixed_up_x = vqaddq_s32(x, fixup); + return vrshrq_n_s32(fixed_up_x, -Exponent); + } +}; + +template +struct ImplSaturatingRoundingMultiplyByPOT { + static int16x8_t eval(int16x8_t x) { return vqshlq_n_s16(x, Exponent); } +}; + +template +struct ImplSaturatingRoundingMultiplyByPOT { + static int16x8_t eval(int16x8_t x) { + const int16x8_t fixup = vshrq_n_s16(x, 15); + const int16x8_t fixed_up_x = vqaddq_s16(x, fixup); + return vrshrq_n_s16(fixed_up_x, -Exponent); + } +}; + +template <> +inline int32x4_t Dup(std::int32_t x) { + return vdupq_n_s32(x); +} + +template <> +inline int16x8_t Dup(std::int16_t x) { + return vdupq_n_s16(x); +} + +// So far this is only needed for int16. +template <> +inline int16x8_t SaturatingAdd(int16x8_t a, int16x8_t b) { + return vqaddq_s16(a, b); +} + +} // end namespace gemmlowp + +#endif // GEMMLOWP_INTERNAL_FIXEDPOINT_NEON_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_sse.h b/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_sse.h new file mode 100644 index 0000000..26153eb --- /dev/null +++ b/src/tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint_sse.h @@ -0,0 +1,387 @@ +#if !defined(ESP32) +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// fixedpoint_SSE.h: optimized SSE specializations of the templates +// in fixedpoint.h. + +#ifndef GEMMLOWP_INTERNAL_FIXEDPOINT_SSE_H_ +#define GEMMLOWP_INTERNAL_FIXEDPOINT_SSE_H_ + +#include +#include "tensorflow_arm/third_party/gemmlowp/fixedpoint/fixedpoint.h" + +namespace gemmlowp { + +// SSE intrinsics are not finely typed: there is a single __m128i vector +// type that does not distinguish between "int32x4" and "int16x8" use +// cases, unlike the NEON equivalents. Because we had initially focused +// on int32x4, we did not pay attention and specialized these fixedpoint +// templates directly for __m128i hardcoding the int32x4 semantics, +// not leaving room for int16x8 semantics. Amending that by adding a separate +// data type, int16x8_m128i, that wraps __m128i while being a separate +// type. +struct int16x8_m128i { + int16x8_m128i() {} + explicit int16x8_m128i(__m128i w) : v(w) {} + ~int16x8_m128i() {} + + __m128i v; +}; + +template <> +struct FixedPointRawTypeTraits<__m128i> { + typedef std::int32_t ScalarRawType; + static constexpr int kLanes = 4; +}; + +template <> +struct FixedPointRawTypeTraits { + typedef std::int16_t ScalarRawType; + static constexpr int kLanes = 8; +}; + +template <> +inline __m128i BitAnd(__m128i a, __m128i b) { + return _mm_and_si128(a, b); +} + +template <> +inline int16x8_m128i BitAnd(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_and_si128(a.v, b.v)); +} + +template <> +inline __m128i BitOr(__m128i a, __m128i b) { + return _mm_or_si128(a, b); +} + +template <> +inline int16x8_m128i BitOr(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_or_si128(a.v, b.v)); +} + +template <> +inline __m128i BitXor(__m128i a, __m128i b) { + return _mm_xor_si128(a, b); +} + +template <> +inline int16x8_m128i BitXor(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_xor_si128(a.v, b.v)); +} + +template <> +inline __m128i BitNot(__m128i a) { + return _mm_andnot_si128(a, _mm_set1_epi32(-1)); +} + +template <> +inline int16x8_m128i BitNot(int16x8_m128i a) { + return int16x8_m128i(_mm_andnot_si128(a.v, _mm_set1_epi16(-1))); +} + +template <> +inline __m128i Add(__m128i a, __m128i b) { + return _mm_add_epi32(a, b); +} + +template <> +inline int16x8_m128i Add(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_add_epi16(a.v, b.v)); +} + +template <> +inline __m128i Mul(__m128i a, __m128i b) { + return _mm_mullo_epi32(a, b); +} + +template <> +inline int16x8_m128i Mul(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_mullo_epi16(a.v, b.v)); +} + +template <> +inline __m128i Sub(__m128i a, __m128i b) { + return _mm_sub_epi32(a, b); +} + +template <> +inline int16x8_m128i Sub(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_sub_epi16(a.v, b.v)); +} + +template <> +inline __m128i Neg(__m128i a) { + return _mm_sign_epi32(a, _mm_set1_epi32(-1)); +} + +template <> +inline int16x8_m128i Neg(int16x8_m128i a) { + return int16x8_m128i(_mm_sign_epi16(a.v, _mm_set1_epi16(-1))); +} + +template <> +inline __m128i ShiftLeft(__m128i a, int offset) { + return _mm_slli_epi32(a, offset); +} + +template <> +inline int16x8_m128i ShiftLeft(int16x8_m128i a, int offset) { + return int16x8_m128i(_mm_slli_epi16(a.v, offset)); +} + +template <> +inline __m128i ShiftRight(__m128i a, int offset) { + return _mm_srai_epi32(a, offset); +} + +template <> +inline int16x8_m128i ShiftRight(int16x8_m128i a, int offset) { + return int16x8_m128i(_mm_srai_epi16(a.v, offset)); +} + +template <> +inline __m128i SelectUsingMask(__m128i if_mask, __m128i then_val, + __m128i else_val) { + // borrowed from Intel's arm_neon_sse.h header. + return _mm_or_si128(_mm_and_si128(if_mask, then_val), + _mm_andnot_si128(if_mask, else_val)); +} + +template <> +inline int16x8_m128i SelectUsingMask(int16x8_m128i if_mask, + int16x8_m128i then_val, + int16x8_m128i else_val) { + // borrowed from Intel's arm_neon_sse.h header. + return int16x8_m128i(SelectUsingMask(if_mask.v, then_val.v, else_val.v)); +} + +template <> +inline __m128i MaskIfEqual(__m128i a, __m128i b) { + return _mm_cmpeq_epi32(a, b); +} + +template <> +inline int16x8_m128i MaskIfEqual(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_cmpeq_epi16(a.v, b.v)); +} + +template <> +inline __m128i MaskIfNotEqual(__m128i a, __m128i b) { + return BitNot(MaskIfEqual(a, b)); +} + +template <> +inline int16x8_m128i MaskIfNotEqual(int16x8_m128i a, int16x8_m128i b) { + return BitNot(MaskIfEqual(a, b)); +} + +template <> +inline __m128i MaskIfZero(__m128i a) { + return MaskIfEqual(a, _mm_set1_epi32(0)); +} + +template <> +inline int16x8_m128i MaskIfZero(int16x8_m128i a) { + return MaskIfEqual(a, int16x8_m128i(_mm_set1_epi16(0))); +} + +template <> +inline __m128i MaskIfNonZero(__m128i a) { + return MaskIfNotEqual(a, _mm_set1_epi32(0)); +} + +template <> +inline int16x8_m128i MaskIfNonZero(int16x8_m128i a) { + return MaskIfNotEqual(a, int16x8_m128i(_mm_set1_epi16(0))); +} + +template <> +inline __m128i MaskIfGreaterThan(__m128i a, __m128i b) { + return _mm_cmpgt_epi32(a, b); +} + +template <> +inline int16x8_m128i MaskIfGreaterThan(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_cmpgt_epi16(a.v, b.v)); +} + +template <> +inline __m128i MaskIfLessThan(__m128i a, __m128i b) { + return _mm_cmplt_epi32(a, b); +} + +template <> +inline int16x8_m128i MaskIfLessThan(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_cmplt_epi16(a.v, b.v)); +} + +template <> +inline __m128i MaskIfGreaterThanOrEqual(__m128i a, __m128i b) { + return BitNot(MaskIfLessThan(a, b)); +} + +template <> +inline int16x8_m128i MaskIfGreaterThanOrEqual(int16x8_m128i a, + int16x8_m128i b) { + return BitNot(MaskIfLessThan(a, b)); +} + +template <> +inline __m128i MaskIfLessThanOrEqual(__m128i a, __m128i b) { + return BitNot(MaskIfGreaterThan(a, b)); +} + +template <> +inline int16x8_m128i MaskIfLessThanOrEqual(int16x8_m128i a, int16x8_m128i b) { + return BitNot(MaskIfGreaterThan(a, b)); +} + +/* Assumptions: + - All and Any are used on masks. + - masks are all_ones for true lanes, all_zeroes otherwise. +Hence, All means all 128bits set, and Any means any bit set. +*/ + +template <> +inline bool All(__m128i a) { + return _mm_testc_si128(a, a); +} + +template <> +inline bool All(int16x8_m128i a) { + return _mm_testc_si128(a.v, a.v); +} + +template <> +inline bool Any(__m128i a) { + return !_mm_testz_si128(a, a); +} + +template <> +inline bool Any(int16x8_m128i a) { + return !_mm_testz_si128(a.v, a.v); +} + +template <> +inline __m128i RoundingHalfSum(__m128i a, __m128i b) { + /* __m128i round_bit_mask, a_over_2, b_over_2, round_bit, sum; */ + /* We divide the inputs before the add to avoid the overflow and costly test + */ + /* of checking if an overflow occured on signed add */ + /* round_bit_mask = _mm_set1_epi32(1); */ + /* a_over_2 = _mm_srai_epi32(a, 1); */ + /* b_over_2 = _mm_srai_epi32(b, 1); */ + /* sum = Add(a_over_2, b_over_2); */ + /* round_bit = _mm_sign_epi32(BitAnd(BitOr(a,b), round_bit_mask), sum); */ + /* return Add(sum, round_bit); */ + + /* Other possibility detecting overflow and xor the sign if an overflow + * happened*/ + __m128i one, sign_bit_mask, sum, rounded_half_sum, overflow, result; + one = _mm_set1_epi32(1); + sign_bit_mask = _mm_set1_epi32(0x80000000); + sum = Add(a, b); + rounded_half_sum = _mm_srai_epi32(Add(sum, one), 1); + overflow = + BitAnd(BitAnd(BitXor(a, rounded_half_sum), BitXor(b, rounded_half_sum)), + sign_bit_mask); + result = BitXor(rounded_half_sum, overflow); + return result; +} + +template <> +inline int16x8_m128i RoundingHalfSum(int16x8_m128i a, int16x8_m128i b) { + // Idea: go to unsigned to use _mm_avg_epu16, + // borrowed from Intel's arm_neon_sse.h header. + __m128i constant_neg_32768 = _mm_set1_epi16(-32768); + __m128i a_unsigned = _mm_sub_epi16(a.v, constant_neg_32768); + __m128i b_unsigned = _mm_sub_epi16(b.v, constant_neg_32768); + __m128i avg_unsigned = _mm_avg_epu16(a_unsigned, b_unsigned); + __m128i avg = _mm_add_epi16(avg_unsigned, constant_neg_32768); + return int16x8_m128i(avg); +} + +template <> +inline __m128i SaturatingRoundingDoublingHighMul(__m128i a, __m128i b) { + __m128i min, saturation_mask, a0_a2, a1_a3, b0_b2, b1_b3; + __m128i a0b0_a2b2, a1b1_a3b3, a0b0_a2b2_rounded, a1b1_a3b3_rounded; + __m128i a0b0_a2b2_rounded_2x, a1b1_a3b3_rounded_2x, result; + __m128i nudge; + + // saturation only happen if a == b == INT_MIN + min = _mm_set1_epi32(std::numeric_limits::min()); + saturation_mask = BitAnd(MaskIfEqual(a, b), MaskIfEqual(a, min)); + + // a = a0 | a1 | a2 | a3 + // b = b0 | b1 | b2 | b3 + a0_a2 = a; + a1_a3 = _mm_srli_si128(a, 4); + b0_b2 = b; + b1_b3 = _mm_srli_si128(b, 4); + + a0b0_a2b2 = _mm_mul_epi32(a0_a2, b0_b2); + a1b1_a3b3 = _mm_mul_epi32(a1_a3, b1_b3); + + // do the rounding and take into account that it will be doubled + nudge = _mm_set1_epi64x(1 << 30); + a0b0_a2b2_rounded = _mm_add_epi64(a0b0_a2b2, nudge); + a1b1_a3b3_rounded = _mm_add_epi64(a1b1_a3b3, nudge); + + // do the doubling + a0b0_a2b2_rounded_2x = _mm_slli_epi64(a0b0_a2b2_rounded, 1); + a1b1_a3b3_rounded_2x = _mm_slli_epi64(a1b1_a3b3_rounded, 1); + + // get the high part of the products + result = _mm_blend_epi16(_mm_srli_si128(a0b0_a2b2_rounded_2x, 4), + a1b1_a3b3_rounded_2x, 0xcc); + + // saturate those which overflowed + return SelectUsingMask(saturation_mask, min, result); +} + +template <> +inline int16x8_m128i SaturatingRoundingDoublingHighMul(int16x8_m128i a, + int16x8_m128i b) { + // Idea: use _mm_mulhrs_epi16 then saturate with a bit-operation, + // borrowed from Intel's arm_neon_sse.h header. + __m128i result_unsaturated = _mm_mulhrs_epi16(a.v, b.v); + __m128i saturation_mask = + _mm_cmpeq_epi16(result_unsaturated, _mm_set1_epi16(0x8000)); + __m128i result = _mm_xor_si128(result_unsaturated, saturation_mask); + return int16x8_m128i(result); +} + +template <> +inline __m128i Dup<__m128i>(std::int32_t x) { + return _mm_set1_epi32(x); +} + +template <> +inline int16x8_m128i Dup(std::int16_t x) { + return int16x8_m128i(_mm_set1_epi16(x)); +} + +// So far this is only needed for int16. +template <> +inline int16x8_m128i SaturatingAdd(int16x8_m128i a, int16x8_m128i b) { + return int16x8_m128i(_mm_adds_epi16(a.v, b.v)); +} + +} // end namespace gemmlowp + +#endif // GEMMLOWP_INTERNAL_FIXEDPOINT_SSE_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/gemmlowp/internal/detect_platform.h b/src/tensorflow_arm/third_party/gemmlowp/internal/detect_platform.h new file mode 100644 index 0000000..70e1e96 --- /dev/null +++ b/src/tensorflow_arm/third_party/gemmlowp/internal/detect_platform.h @@ -0,0 +1,169 @@ +#if !defined(ESP32) +// Copyright 2018 The Gemmlowp Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// detect_platform.h: Sets up macros that control architecture-specific +// features of gemmlowp's implementation. + +#ifndef GEMMLOWP_INTERNAL_DETECT_PLATFORM_H_ +#define GEMMLOWP_INTERNAL_DETECT_PLATFORM_H_ + +// Our inline assembly path assume GCC/Clang syntax. +// Native Client doesn't seem to support inline assembly(?). +#if defined(__GNUC__) && !defined(__native_client__) +#define GEMMLOWP_ALLOW_INLINE_ASM +#endif + +// Define macro statement that avoids inlining for GCC. +// For non-GCC, define as empty macro. +#if defined(__GNUC__) +#define GEMMLOWP_NOINLINE __attribute__((noinline)) +#else +#define GEMMLOWP_NOINLINE +#endif + +// Detect ARM, 32-bit or 64-bit +#ifdef __arm__ +#define GEMMLOWP_ARM_32 +#endif + +#ifdef __aarch64__ +#define GEMMLOWP_ARM_64 +#endif + +#if defined(GEMMLOWP_ARM_32) || defined(GEMMLOWP_ARM_64) +#define GEMMLOWP_ARM +#endif + +// Detect MIPS, 32-bit or 64-bit +#if defined(__mips) && !defined(__LP64__) +#define GEMMLOWP_MIPS_32 +#endif + +#if defined(__mips) && defined(__LP64__) +#define GEMMLOWP_MIPS_64 +#endif + +#if defined(GEMMLOWP_MIPS_32) || defined(GEMMLOWP_MIPS_64) +#define GEMMLOWP_MIPS +#endif + +// Detect x86, 32-bit or 64-bit +#if defined(__i386__) || defined(_M_IX86) || defined(_X86_) || defined(__i386) +#define GEMMLOWP_X86_32 +#endif + +#if defined(__x86_64__) || defined(_M_X64) || defined(__amd64) +#define GEMMLOWP_X86_64 +#endif + +#if defined(GEMMLOWP_X86_32) || defined(GEMMLOWP_X86_64) +#define GEMMLOWP_X86 +#endif + +// Some of our optimized paths use inline assembly and for +// now we don't bother enabling some other optimized paths using intrinddics +// where we can't use inline assembly paths. +#ifdef GEMMLOWP_ALLOW_INLINE_ASM + +// Detect NEON. It's important to check for both tokens. +#if (defined __ARM_NEON) || (defined __ARM_NEON__) +#define GEMMLOWP_NEON +#endif + +// Convenience NEON tokens for 32-bit or 64-bit +#if defined(GEMMLOWP_NEON) && defined(GEMMLOWP_ARM_32) +#define GEMMLOWP_NEON_32 +#endif + +#if defined(GEMMLOWP_NEON) && defined(GEMMLOWP_ARM_64) +#define GEMMLOWP_NEON_64 +#endif + +// Detect MIPS MSA. +// Limit MSA optimizations to little-endian CPUs for now. +// TODO: Perhaps, eventually support MSA optimizations on big-endian CPUs? +#if defined(GEMMLOWP_MIPS) && (__mips_isa_rev >= 5) && defined(__mips_msa) && \ + defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) +#define GEMMLOWP_MSA +#endif + +// Convenience MIPS MSA tokens for 32-bit or 64-bit. +#if defined(GEMMLOWP_MSA) && defined(GEMMLOWP_MIPS_32) +#define GEMMLOWP_MSA_32 +#endif + +#if defined(GEMMLOWP_MSA) && defined(GEMMLOWP_MIPS_64) +#define GEMMLOWP_MSA_64 +#endif + +// compiler define for AVX2 -D GEMMLOWP_ENABLE_AVX2 +// Detect AVX2 +#if defined(__AVX2__) && defined(GEMMLOWP_ENABLE_AVX2) +#define GEMMLOWP_AVX2 +// Detect SSE4. +// MSVC does not have __SSE4_1__ macro, but will enable SSE4 +// when AVX is turned on. +#elif defined(__SSE4_1__) || (defined(_MSC_VER) && defined(__AVX__)) +#define GEMMLOWP_SSE4 +// Detect SSE3. +#elif defined(__SSE3__) +#define GEMMLOWP_SSE3 +#endif + +// Convenience SSE4 tokens for 32-bit or 64-bit +#if defined(GEMMLOWP_SSE4) && defined(GEMMLOWP_X86_32) && \ + !defined(GEMMLOWP_DISABLE_SSE4) +#define GEMMLOWP_SSE4_32 +#endif + +#if defined(GEMMLOWP_SSE3) && defined(GEMMLOWP_X86_32) +#define GEMMLOWP_SSE3_32 +#endif + +#if defined(GEMMLOWP_SSE4) && defined(GEMMLOWP_X86_64) && \ + !defined(GEMMLOWP_DISABLE_SSE4) +#define GEMMLOWP_SSE4_64 +#endif + +#if defined(GEMMLOWP_SSE3) && defined(GEMMLOWP_X86_64) +#define GEMMLOWP_SSE3_64 +#endif + +#if defined(GEMMLOWP_AVX2) && defined(GEMMLOWP_X86_64) +#define GEMMLOWP_AVX2_64 +#endif + +#if defined(__has_feature) +#if __has_feature(memory_sanitizer) +#include +#define GEMMLOWP_MARK_MEMORY_AS_INITIALIZED __msan_unpoison +#elif __has_feature(address_sanitizer) +#include +#define GEMMLOWP_MARK_MEMORY_AS_INITIALIZED __asan_unpoison_memory_region +#endif +#endif + +#endif // GEMMLOWP_ALLOW_INLINE_ASM + +// Detect Android. Don't conflate with ARM - we care about tuning +// for non-ARM Android devices too. This can be used in conjunction +// with x86 to tune differently for mobile x86 CPUs (Atom) vs. desktop x86 CPUs. +#if defined(__ANDROID__) || defined(ANDROID) +#define GEMMLOWP_ANDROID +#endif + +#endif // GEMMLOWP_INTERNAL_DETECT_PLATFORM_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/kissfft/COPYING b/src/tensorflow_arm/third_party/kissfft/COPYING new file mode 100644 index 0000000..2fc6685 --- /dev/null +++ b/src/tensorflow_arm/third_party/kissfft/COPYING @@ -0,0 +1,11 @@ +Copyright (c) 2003-2010 Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/src/tensorflow_arm/third_party/kissfft/_kiss_fft_guts.h b/src/tensorflow_arm/third_party/kissfft/_kiss_fft_guts.h new file mode 100644 index 0000000..f8ca2f8 --- /dev/null +++ b/src/tensorflow_arm/third_party/kissfft/_kiss_fft_guts.h @@ -0,0 +1,167 @@ +#if !defined(ESP32) +/* +Copyright (c) 2003-2010, Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +/* kiss_fft.h + defines kiss_fft_scalar as either short or a float type + and defines + typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */ +#include "tensorflow_arm/third_party/kissfft/kiss_fft.h" +#include + +#define MAXFACTORS 32 +/* e.g. an fft of length 128 has 4 factors + as far as kissfft is concerned + 4*4*4*2 + */ + +struct kiss_fft_state{ + int nfft; + int inverse; + int factors[2*MAXFACTORS]; + kiss_fft_cpx twiddles[1]; +}; + +/* + Explanation of macros dealing with complex math: + + C_MUL(m,a,b) : m = a*b + C_FIXDIV( c , div ) : if a fixed point impl., c /= div. noop otherwise + C_SUB( res, a,b) : res = a - b + C_SUBFROM( res , a) : res -= a + C_ADDTO( res , a) : res += a + * */ +#ifdef FIXED_POINT +#if (FIXED_POINT==32) +# define FRACBITS 31 +# define SAMPPROD int64_t +#define SAMP_MAX 2147483647 +#else +# define FRACBITS 15 +# define SAMPPROD int32_t +#define SAMP_MAX 32767 +#endif + +#define SAMP_MIN -SAMP_MAX + +#if defined(CHECK_OVERFLOW) +# define CHECK_OVERFLOW_OP(a,op,b) \ + if ( (SAMPPROD)(a) op (SAMPPROD)(b) > SAMP_MAX || (SAMPPROD)(a) op (SAMPPROD)(b) < SAMP_MIN ) { \ + fprintf(stderr,"WARNING:overflow @ " __FILE__ "(%d): (%d " #op" %d) = %ld\n",__LINE__,(a),(b),(SAMPPROD)(a) op (SAMPPROD)(b) ); } +#endif + + +# define smul(a,b) ( (SAMPPROD)(a)*(b) ) +# define sround( x ) (kiss_fft_scalar)( ( (x) + (1<<(FRACBITS-1)) ) >> FRACBITS ) + +# define S_MUL(a,b) sround( smul(a,b) ) + +# define C_MUL(m,a,b) \ + do{ (m).r = sround( smul((a).r,(b).r) - smul((a).i,(b).i) ); \ + (m).i = sround( smul((a).r,(b).i) + smul((a).i,(b).r) ); }while(0) + +# define DIVSCALAR(x,k) \ + (x) = sround( smul( x, SAMP_MAX/k ) ) + +# define C_FIXDIV(c,div) \ + do { DIVSCALAR( (c).r , div); \ + DIVSCALAR( (c).i , div); }while (0) + +# define C_MULBYSCALAR( c, s ) \ + do{ (c).r = sround( smul( (c).r , s ) ) ;\ + (c).i = sround( smul( (c).i , s ) ) ; }while(0) + +#else /* not FIXED_POINT*/ + +# define S_MUL(a,b) ( (a)*(b) ) +#define C_MUL(m,a,b) \ + do{ (m).r = (a).r*(b).r - (a).i*(b).i;\ + (m).i = (a).r*(b).i + (a).i*(b).r; }while(0) +# define C_FIXDIV(c,div) /* NOOP */ +# define C_MULBYSCALAR( c, s ) \ + do{ (c).r *= (s);\ + (c).i *= (s); }while(0) +#endif + +#ifndef CHECK_OVERFLOW_OP +# define CHECK_OVERFLOW_OP(a,op,b) /* noop */ +#endif + +#define C_ADD( res, a,b)\ + do { \ + CHECK_OVERFLOW_OP((a).r,+,(b).r)\ + CHECK_OVERFLOW_OP((a).i,+,(b).i)\ + (res).r=(a).r+(b).r; (res).i=(a).i+(b).i; \ + }while(0) +#define C_SUB( res, a,b)\ + do { \ + CHECK_OVERFLOW_OP((a).r,-,(b).r)\ + CHECK_OVERFLOW_OP((a).i,-,(b).i)\ + (res).r=(a).r-(b).r; (res).i=(a).i-(b).i; \ + }while(0) +#define C_ADDTO( res , a)\ + do { \ + CHECK_OVERFLOW_OP((res).r,+,(a).r)\ + CHECK_OVERFLOW_OP((res).i,+,(a).i)\ + (res).r += (a).r; (res).i += (a).i;\ + }while(0) + +#define C_SUBFROM( res , a)\ + do {\ + CHECK_OVERFLOW_OP((res).r,-,(a).r)\ + CHECK_OVERFLOW_OP((res).i,-,(a).i)\ + (res).r -= (a).r; (res).i -= (a).i; \ + }while(0) + + +#ifdef FIXED_POINT +# define KISS_FFT_COS(phase) floor(.5+SAMP_MAX * cos (phase)) +# define KISS_FFT_SIN(phase) floor(.5+SAMP_MAX * sin (phase)) +# define HALF_OF(x) ((x)>>1) +#elif defined(USE_SIMD) +# define KISS_FFT_COS(phase) _mm_set1_ps( cos(phase) ) +# define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) ) +# define HALF_OF(x) ((x)*_mm_set1_ps(.5)) +#else +# define KISS_FFT_COS(phase) (kiss_fft_scalar) cos(phase) +# define KISS_FFT_SIN(phase) (kiss_fft_scalar) sin(phase) +# define HALF_OF(x) ((x)*.5) +#endif + +#define kf_cexp(x,phase) \ + do{ \ + (x)->r = KISS_FFT_COS(phase);\ + (x)->i = KISS_FFT_SIN(phase);\ + }while(0) + + +/* a debugging function */ +#define pcpx(c)\ + fprintf(stderr,"%g + %gi\n",(double)((c)->r),(double)((c)->i) ) + + +#ifdef KISS_FFT_USE_ALLOCA +// define this to allow use of alloca instead of malloc for temporary buffers +// Temporary buffers are used in two case: +// 1. FFT sizes that have "bad" factors. i.e. not 2,3 and 5 +// 2. "in-place" FFTs. Notice the quotes, since kissfft does not really do an in-place transform. +#include +#define KISS_FFT_TMP_ALLOC(nbytes) alloca(nbytes) +#define KISS_FFT_TMP_FREE(ptr) +#else +#define KISS_FFT_TMP_ALLOC(nbytes) KISS_FFT_MALLOC(nbytes) +#define KISS_FFT_TMP_FREE(ptr) KISS_FFT_FREE(ptr) +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/kissfft/kiss_fft.h b/src/tensorflow_arm/third_party/kissfft/kiss_fft.h new file mode 100644 index 0000000..6f91a4b --- /dev/null +++ b/src/tensorflow_arm/third_party/kissfft/kiss_fft.h @@ -0,0 +1,133 @@ +#if !defined(ESP32) +#ifndef KISS_FFT_H +#define KISS_FFT_H + +#include +#include +#include +//#include /* Patched by helper_functions.inc for Arduino compatibility */ + +#ifdef __cplusplus +extern "C" { +#endif + +/* + ATTENTION! + If you would like a : + -- a utility that will handle the caching of fft objects + -- real-only (no imaginary time component ) FFT + -- a multi-dimensional FFT + -- a command-line utility to perform ffts + -- a command-line utility to perform fast-convolution filtering + + Then see kfc.h kiss_fftr.h kiss_fftnd.h fftutil.c kiss_fastfir.c + in the tools/ directory. +*/ + +#ifdef USE_SIMD +# include +# define kiss_fft_scalar __m128 +#define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes,16) +#define KISS_FFT_FREE _mm_free +#else +#define KISS_FFT_MALLOC(X) (void*)(0) /* Patched. */ +#define KISS_FFT_FREE(X) /* Patched. */ +#endif + + +// Patched automatically by download_dependencies.sh so default is 16 bit. +#ifndef FIXED_POINT +#define FIXED_POINT (16) +#endif +// End patch. + +#ifdef FIXED_POINT +#include /* Patched. */ +# if (FIXED_POINT == 32) +# define kiss_fft_scalar int32_t +# else +# define kiss_fft_scalar int16_t +# endif +#else +# ifndef kiss_fft_scalar +/* default is float */ +# define kiss_fft_scalar float +# endif +#endif + +typedef struct { + kiss_fft_scalar r; + kiss_fft_scalar i; +}kiss_fft_cpx; + +typedef struct kiss_fft_state* kiss_fft_cfg; + +/* + * kiss_fft_alloc + * + * Initialize a FFT (or IFFT) algorithm's cfg/state buffer. + * + * typical usage: kiss_fft_cfg mycfg=kiss_fft_alloc(1024,0,NULL,NULL); + * + * The return value from fft_alloc is a cfg buffer used internally + * by the fft routine or NULL. + * + * If lenmem is NULL, then kiss_fft_alloc will allocate a cfg buffer using malloc. + * The returned value should be free()d when done to avoid memory leaks. + * + * The state can be placed in a user supplied buffer 'mem': + * If lenmem is not NULL and mem is not NULL and *lenmem is large enough, + * then the function places the cfg in mem and the size used in *lenmem + * and returns mem. + * + * If lenmem is not NULL and ( mem is NULL or *lenmem is not large enough), + * then the function returns NULL and places the minimum cfg + * buffer size in *lenmem. + * */ + +kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem); + +/* + * kiss_fft(cfg,in_out_buf) + * + * Perform an FFT on a complex input buffer. + * for a forward FFT, + * fin should be f[0] , f[1] , ... ,f[nfft-1] + * fout will be F[0] , F[1] , ... ,F[nfft-1] + * Note that each element is complex and can be accessed like + f[k].r and f[k].i + * */ +void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout); + +/* + A more generic version of the above function. It reads its input from every Nth sample. + * */ +void kiss_fft_stride(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int fin_stride); + +/* If kiss_fft_alloc allocated a buffer, it is one contiguous + buffer and can be simply free()d when no longer needed*/ +#define kiss_fft_free free + +/* + Cleans up some memory that gets managed internally. Not necessary to call, but it might clean up + your compiler output to call this before you exit. +*/ +void kiss_fft_cleanup(void); + + +/* + * Returns the smallest integer k, such that k>=n and k has only "fast" factors (2,3,5) + */ +int kiss_fft_next_fast_size(int n); + +/* for real ffts, we need an even size */ +#define kiss_fftr_next_fast_size_real(n) \ + (kiss_fft_next_fast_size( ((n)+1)>>1)<<1) + +#ifdef __cplusplus +} +#endif + +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_arm/third_party/kissfft/tools/kiss_fftr.h b/src/tensorflow_arm/third_party/kissfft/tools/kiss_fftr.h new file mode 100644 index 0000000..159bbf8 --- /dev/null +++ b/src/tensorflow_arm/third_party/kissfft/tools/kiss_fftr.h @@ -0,0 +1,49 @@ +#if !defined(ESP32) +#ifndef KISS_FTR_H +#define KISS_FTR_H + +#include "tensorflow_arm/third_party/kissfft/kiss_fft.h" +#ifdef __cplusplus +extern "C" { +#endif + + +/* + + Real optimized version can save about 45% cpu time vs. complex fft of a real seq. + + + + */ + +typedef struct kiss_fftr_state *kiss_fftr_cfg; + + +kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem, size_t * lenmem); +/* + nfft must be even + + If you don't care to allocate space, use mem = lenmem = NULL +*/ + + +void kiss_fftr(kiss_fftr_cfg cfg,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata); +/* + input timedata has nfft scalar points + output freqdata has nfft/2+1 complex points +*/ + +void kiss_fftri(kiss_fftr_cfg cfg,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata); +/* + input freqdata has nfft/2+1 complex points + output timedata has nfft scalar points +*/ + +#define kiss_fftr_free free + +#ifdef __cplusplus +} +#endif +#endif + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/experimental/ruy/profiler/instrumentation.h b/src/tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h similarity index 96% rename from src/tensorflow/lite/experimental/ruy/profiler/instrumentation.h rename to src/tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h index cb0e702..7fff51e 100644 --- a/src/tensorflow/lite/experimental/ruy/profiler/instrumentation.h +++ b/src/tensorflow_arm/third_party/ruy/ruy/profiler/instrumentation.h @@ -1,3 +1,4 @@ +#if !defined(ESP32) /* Copyright 2020 Google LLC. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,8 +14,8 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_EXPERIMENTAL_RUY_PROFILER_INSTRUMENTATION_H_ -#define TENSORFLOW_LITE_EXPERIMENTAL_RUY_PROFILER_INSTRUMENTATION_H_ +#ifndef RUY_RUY_PROFILER_INSTRUMENTATION_H_ +#define RUY_RUY_PROFILER_INSTRUMENTATION_H_ #ifdef RUY_PROFILER #include @@ -200,4 +201,6 @@ class ScopeLabel { } // namespace profiler } // namespace ruy -#endif // TENSORFLOW_LITE_EXPERIMENTAL_RUY_PROFILER_INSTRUMENTATION_H_ +#endif // RUY_RUY_PROFILER_INSTRUMENTATION_H_ + +#endif // end of #if !defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/core/public/version.h b/src/tensorflow_esp32/tensorflow/core/public/version.h similarity index 98% rename from src/tensorflow/core/public/version.h rename to src/tensorflow_esp32/tensorflow/core/public/version.h index 10d6b54..566366e 100644 --- a/src/tensorflow/core/public/version.h +++ b/src/tensorflow_esp32/tensorflow/core/public/version.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2015 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -22,7 +23,7 @@ limitations under the License. // tensorflow/tools/pip_package/setup.py #define TF_MAJOR_VERSION 2 #define TF_MINOR_VERSION 1 -#define TF_PATCH_VERSION 0 +#define TF_PATCH_VERSION 1 // TF_VERSION_SUFFIX is non-empty for pre-releases (e.g. "-alpha", "-alpha.1", // "-beta", "-rc", "-rc.1") @@ -137,3 +138,5 @@ extern int tf_cxx11_abi_flag(); extern int tf_monolithic_build(); #endif // TENSORFLOW_CORE_PUBLIC_VERSION_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/c/builtin_op_data.h b/src/tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h similarity index 94% rename from src/tensorflow/lite/c/builtin_op_data.h rename to src/tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h index 65717c4..7346332 100644 --- a/src/tensorflow/lite/c/builtin_op_data.h +++ b/src/tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,16 +18,12 @@ limitations under the License. #include -#include "tensorflow/lite/c/common.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" #ifdef __cplusplus extern "C" { #endif // __cplusplus -// TfLiteReshapeParams can't have dynamic data so we fix the maximum possible -// number of dimensions. -#define TFLITE_RESHAPE_PARAMS_MAX_DIMENSION_COUNT 8 - // TODO(aselle): Consider using "if this then that" for testing. // Useful placeholder to put in otherwise empty structs to avoid size warnings. @@ -75,16 +72,12 @@ typedef enum { } TfLiteFusedActivation; typedef struct { - // Parameters for CONV_2D version 1. TfLitePadding padding; int stride_width; int stride_height; - TfLiteFusedActivation activation; - - // Parameters for CONV_2D version 2. - // Note: Version 2 supports dilation values not equal to 1. int dilation_width_factor; int dilation_height_factor; + TfLiteFusedActivation activation; } TfLiteConvParams; typedef struct { @@ -257,10 +250,6 @@ typedef struct { typedef struct { bool align_corners; - // half_pixel_centers assumes pixels are of half the actual dimensions, and - // yields more accurate resizes. Corresponds to the same argument for the - // original TensorFlow op in TF2.0. - bool half_pixel_centers; } TfLiteResizeBilinearParams; typedef struct { @@ -278,7 +267,7 @@ typedef struct { typedef struct { // TODO(ahentz): We can't have dynamic data in this struct, at least not yet. // For now we will fix the maximum possible number of dimensions. - int shape[TFLITE_RESHAPE_PARAMS_MAX_DIMENSION_COUNT]; + int shape[8]; int num_dimensions; } TfLiteReshapeParams; @@ -432,3 +421,5 @@ typedef struct { #endif // __cplusplus #endif // TENSORFLOW_LITE_C_BUILTIN_OP_DATA_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/c/c_api_internal.c b/src/tensorflow_esp32/tensorflow/lite/c/c_api_internal.c new file mode 100644 index 0000000..c7137e3 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/c/c_api_internal.c @@ -0,0 +1,188 @@ +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#ifndef TF_LITE_STATIC_MEMORY +#include +#include +#endif // TF_LITE_STATIC_MEMORY + +int TfLiteIntArrayGetSizeInBytes(int size) { + static TfLiteIntArray dummy; + return sizeof(dummy) + sizeof(dummy.data[0]) * size; +} + +int TfLiteIntArrayEqual(const TfLiteIntArray* a, const TfLiteIntArray* b) { + if (a == b) return 1; + if (a == NULL || b == NULL) return 0; + return TfLiteIntArrayEqualsArray(a, b->size, b->data); +} + +int TfLiteIntArrayEqualsArray(const TfLiteIntArray* a, int b_size, + const int b_data[]) { + if (a == NULL) return (b_size == 0); + if (a->size != b_size) return 0; + int i = 0; + for (; i < a->size; i++) + if (a->data[i] != b_data[i]) return 0; + return 1; +} + +#ifndef TF_LITE_STATIC_MEMORY + +TfLiteIntArray* TfLiteIntArrayCreate(int size) { + TfLiteIntArray* ret = + (TfLiteIntArray*)malloc(TfLiteIntArrayGetSizeInBytes(size)); + ret->size = size; + return ret; +} + +TfLiteIntArray* TfLiteIntArrayCopy(const TfLiteIntArray* src) { + if (!src) return NULL; + TfLiteIntArray* ret = TfLiteIntArrayCreate(src->size); + if (ret) { + memcpy(ret->data, src->data, src->size * sizeof(int)); + } + return ret; +} + +void TfLiteIntArrayFree(TfLiteIntArray* a) { free(a); } + +#endif // TF_LITE_STATIC_MEMORY + +int TfLiteFloatArrayGetSizeInBytes(int size) { + static TfLiteFloatArray dummy; + return sizeof(dummy) + sizeof(dummy.data[0]) * size; +} + +#ifndef TF_LITE_STATIC_MEMORY + +TfLiteFloatArray* TfLiteFloatArrayCreate(int size) { + TfLiteFloatArray* ret = + (TfLiteFloatArray*)malloc(TfLiteFloatArrayGetSizeInBytes(size)); + ret->size = size; + return ret; +} + +void TfLiteFloatArrayFree(TfLiteFloatArray* a) { free(a); } + +void TfLiteTensorDataFree(TfLiteTensor* t) { + if (t->allocation_type == kTfLiteDynamic) { + free(t->data.raw); + } + t->data.raw = NULL; +} + +void TfLiteQuantizationFree(TfLiteQuantization* quantization) { + if (quantization->type == kTfLiteAffineQuantization) { + TfLiteAffineQuantization* q_params = + (TfLiteAffineQuantization*)(quantization->params); + if (q_params->scale) { + TfLiteFloatArrayFree(q_params->scale); + q_params->scale = NULL; + } + if (q_params->zero_point) { + TfLiteIntArrayFree(q_params->zero_point); + q_params->zero_point = NULL; + } + free(q_params); + } + quantization->params = NULL; + quantization->type = kTfLiteNoQuantization; +} + +void TfLiteTensorFree(TfLiteTensor* t) { + TfLiteTensorDataFree(t); + if (t->dims) TfLiteIntArrayFree(t->dims); + t->dims = NULL; + + TfLiteQuantizationFree(&t->quantization); +} + +void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims, + TfLiteQuantizationParams quantization, char* buffer, + size_t size, TfLiteAllocationType allocation_type, + const void* allocation, bool is_variable, + TfLiteTensor* tensor) { + TfLiteTensorFree(tensor); + tensor->type = type; + tensor->name = name; + tensor->dims = dims; + tensor->params = quantization; + tensor->data.raw = buffer; + tensor->bytes = size; + tensor->allocation_type = allocation_type; + tensor->allocation = allocation; + tensor->is_variable = is_variable; + + tensor->quantization.type = kTfLiteNoQuantization; + tensor->quantization.params = NULL; +} + +void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor) { + if (tensor->allocation_type != kTfLiteDynamic) { + return; + } + if (!tensor->data.raw) { + tensor->data.raw = malloc(num_bytes); + } else if (num_bytes > tensor->bytes) { + tensor->data.raw = realloc(tensor->data.raw, num_bytes); + } + tensor->bytes = num_bytes; +} +#endif // TF_LITE_STATIC_MEMORY + +const char* TfLiteTypeGetName(TfLiteType type) { + switch (type) { + case kTfLiteNoType: + return "NOTYPE"; + case kTfLiteFloat32: + return "FLOAT32"; + case kTfLiteInt16: + return "INT16"; + case kTfLiteInt32: + return "INT32"; + case kTfLiteUInt8: + return "UINT8"; + case kTfLiteInt8: + return "INT8"; + case kTfLiteInt64: + return "INT64"; + case kTfLiteBool: + return "BOOL"; + case kTfLiteComplex64: + return "COMPLEX64"; + case kTfLiteString: + return "STRING"; + case kTfLiteFloat16: + return "FLOAT16"; + } + return "Unknown type"; +} + +TfLiteDelegate TfLiteDelegateCreate() { + TfLiteDelegate d = { + .data_ = NULL, + .Prepare = NULL, + .CopyFromBufferHandle = NULL, + .CopyToBufferHandle = NULL, + .FreeBufferHandle = NULL, + .flags = kTfLiteDelegateFlagsNone, + }; + return d; +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/c/common.h b/src/tensorflow_esp32/tensorflow/lite/c/c_api_internal.h similarity index 80% rename from src/tensorflow/lite/c/common.h rename to src/tensorflow_esp32/tensorflow/lite/c/c_api_internal.h index 4d7fe8c..5579cc6 100644 --- a/src/tensorflow/lite/c/common.h +++ b/src/tensorflow_esp32/tensorflow/lite/c/c_api_internal.h @@ -1,4 +1,5 @@ -/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -12,11 +13,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ - -// This file defines common C types and APIs for implementing operations, -// delegates and other constructs in TensorFlow Lite. The actual operations and -// delegtes can be defined using C++, but the interface between the interpreter -// and the operations are C. +// This file defines a C API for implementing operations in tflite. +// These operations can be defined using c++ but the interface between +// the interpreter and the operations are C. // // Summary of abstractions // TF_LITE_ENSURE - Self-sufficient error checking @@ -26,12 +25,10 @@ limitations under the License. // TfLiteTensor - tensor (a multidimensional array) // TfLiteNode - a single node or operation // TfLiteRegistration - the implementation of a conceptual operation. -// TfLiteDelegate - allows delegation of nodes to alternative backends. // // Some abstractions in this file are created and managed by Interpreter. - -#ifndef TENSORFLOW_LITE_C_COMMON_H_ -#define TENSORFLOW_LITE_C_COMMON_H_ +#ifndef TENSORFLOW_LITE_C_C_API_INTERNAL_H_ +#define TENSORFLOW_LITE_C_C_API_INTERNAL_H_ #include #include @@ -41,13 +38,13 @@ limitations under the License. extern "C" { #endif // __cplusplus -typedef enum TfLiteStatus { kTfLiteOk = 0, kTfLiteError = 1 } TfLiteStatus; +typedef enum { kTfLiteOk = 0, kTfLiteError = 1 } TfLiteStatus; // The list of external context types known to TF Lite. This list exists solely // to avoid conflicts and to ensure ops can share the external contexts they // need. Access to the external contexts is controled by one of the // corresponding support files. -typedef enum TfLiteExternalContextType { +typedef enum { kTfLiteEigenContext = 0, // include eigen_support.h to use. kTfLiteGemmLowpContext = 1, // include gemm_support.h to use. kTfLiteEdgeTpuContext = 2, // Placeholder for Edge TPU support. @@ -66,16 +63,16 @@ struct TfLiteRegistration; // about about the actual contexts, but it keeps a list of them, and is able to // refresh them if configurations like the number of recommended threads // change. -typedef struct TfLiteExternalContext { +typedef struct { TfLiteExternalContextType type; TfLiteStatus (*Refresh)(struct TfLiteContext* context); } TfLiteExternalContext; -#define kTfLiteOptionalTensor (-1) +#define kOptionalTensor (-1) // Fixed size list of integers. Used for dimensions and inputs/outputs tensor // indices -typedef struct TfLiteIntArray { +typedef struct { int size; // gcc 6.1+ have a bug where flexible members aren't properly handled // https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c @@ -91,11 +88,9 @@ typedef struct TfLiteIntArray { // in bytes. int TfLiteIntArrayGetSizeInBytes(int size); -#ifndef TF_LITE_STATIC_MEMORY // Create a array of a given `size` (uninitialized entries). // This returns a pointer, that you must free using TfLiteIntArrayFree(). TfLiteIntArray* TfLiteIntArrayCreate(int size); -#endif // Check if two intarrays are equal. Returns 1 if they are equal, 0 otherwise. int TfLiteIntArrayEqual(const TfLiteIntArray* a, const TfLiteIntArray* b); @@ -104,17 +99,15 @@ int TfLiteIntArrayEqual(const TfLiteIntArray* a, const TfLiteIntArray* b); int TfLiteIntArrayEqualsArray(const TfLiteIntArray* a, int b_size, const int b_data[]); -#ifndef TF_LITE_STATIC_MEMORY // Create a copy of an array passed as `src`. // You are expected to free memory with TfLiteIntArrayFree TfLiteIntArray* TfLiteIntArrayCopy(const TfLiteIntArray* src); // Free memory of array `a`. void TfLiteIntArrayFree(TfLiteIntArray* a); -#endif // TF_LITE_STATIC_MEMORY // Fixed size list of floats. Used for per-channel quantization. -typedef struct TfLiteFloatArray { +typedef struct { int size; // gcc 6.1+ have a bug where flexible members aren't properly handled // https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c @@ -130,14 +123,12 @@ typedef struct TfLiteFloatArray { // in bytes. int TfLiteFloatArrayGetSizeInBytes(int size); -#ifndef TF_LITE_STATIC_MEMORY // Create a array of a given `size` (uninitialized entries). // This returns a pointer, that you must free using TfLiteFloatArrayFree(). TfLiteFloatArray* TfLiteFloatArrayCreate(int size); // Free memory of array `a`. void TfLiteFloatArrayFree(TfLiteFloatArray* a); -#endif // TF_LITE_STATIC_MEMORY // Since we must not depend on any libraries, define a minimal subset of // error macros while avoiding names that have pre-conceived meanings like @@ -202,12 +193,12 @@ void TfLiteFloatArrayFree(TfLiteFloatArray* a); } while (0) // Single-precision complex data type compatible with the C99 definition. -typedef struct TfLiteComplex64 { +typedef struct { float re, im; // real and imaginary parts, respectively. } TfLiteComplex64; // Half precision data type compatible with the C99 definition. -typedef struct TfLiteFloat16 { +typedef struct { uint16_t data; } TfLiteFloat16; @@ -230,7 +221,7 @@ typedef enum { const char* TfLiteTypeGetName(TfLiteType type); // SupportedQuantizationTypes. -typedef enum TfLiteQuantizationType { +typedef enum { // No quantization. kTfLiteNoQuantization = 0, // Affine quantization (with support for per-channel quantization). @@ -239,7 +230,7 @@ typedef enum TfLiteQuantizationType { } TfLiteQuantizationType; // Structure specifying the quantization used by the tensor, if-any. -typedef struct TfLiteQuantization { +typedef struct { // The type of quantization held by params. TfLiteQuantizationType type; // Holds a reference to one of the quantization param structures specified @@ -253,7 +244,7 @@ typedef struct TfLiteQuantization { // Parameters for asymmetric quantization. Quantized values can be converted // back to float using: // real_value = scale * (quantized_value - zero_point) -typedef struct TfLiteQuantizationParams { +typedef struct { float scale; int32_t zero_point; } TfLiteQuantizationParams; @@ -265,20 +256,18 @@ typedef struct TfLiteQuantizationParams { // For a particular value in quantized_dimension, quantized values can be // converted back to float using: // real_value = scale * (quantized_value - zero_point) -typedef struct TfLiteAffineQuantization { +typedef struct { TfLiteFloatArray* scale; TfLiteIntArray* zero_point; int32_t quantized_dimension; } TfLiteAffineQuantization; -/* A union of pointers that points to memory for a given tensor. */ -typedef union TfLitePtrUnion { - /* Do not access these members directly, if possible, use - * GetTensorData(tensor) instead, otherwise only access .data, as other - * members are deprecated. */ +// A union of pointers that points to memory for a given tensor. +typedef union { int32_t* i32; int64_t* i64; float* f; + // Placeholder for 16b float type. Use uint16* in the pointer union for now. TfLiteFloat16* f16; char* raw; const char* raw_const; @@ -287,14 +276,12 @@ typedef union TfLitePtrUnion { int16_t* i16; TfLiteComplex64* c64; int8_t* int8; - /* Only use this member. */ - void* data; } TfLitePtrUnion; // Memory allocation strategies. kTfLiteMmapRo is for read-only memory-mapped // data (or data externally allocated). kTfLiteArenaRw is arena allocated // data. kTfLiteDynamic is for tensors that are allocated during evaluation. -typedef enum TfLiteAllocationType { +typedef enum { kTfLiteMemNone = 0, kTfLiteMmapRo, kTfLiteArenaRw, @@ -309,32 +296,9 @@ enum { kTfLiteNullBufferHandle = -1, }; -// Storage format of each dimension in a sparse tensor. -typedef enum TfLiteDimensionType { - kTfLiteDimDense = 0, - kTfLiteDimSparseCSR, -} TfLiteDimensionType; - -// Metadata to encode each dimension in a sparse tensor. -typedef struct TfLiteDimensionMetadata { - TfLiteDimensionType format; - int dense_size; - TfLiteIntArray* array_segments; - TfLiteIntArray* array_indices; -} TfLiteDimensionMetadata; - -// Parameters used to encode a sparse tensor. For detailed explanation of each -// field please refer to lite/schema/schema.fbs. -typedef struct TfLiteSparsity { - TfLiteIntArray* traversal_order; - TfLiteIntArray* block_map; - TfLiteDimensionMetadata* dim_metadata; - int dim_metadata_size; -} TfLiteSparsity; - // An tensor in the interpreter system which is a wrapper around a buffer of // data including a dimensionality (or NULL if not currently defined). -typedef struct TfLiteTensor { +typedef struct { // The data type specification for data stored in `data`. This affects // what member of `data` union should be used. TfLiteType type; @@ -386,23 +350,14 @@ typedef struct TfLiteTensor { // Quantization information. Replaces params field above. TfLiteQuantization quantization; - - // Parameters used to encode a sparse tensor. - // This is optional. The field is NULL if a tensor is dense. - // WARNING: This is an experimental interface that is subject to change. - TfLiteSparsity* sparsity; } TfLiteTensor; -#ifndef TF_LITE_STATIC_MEMORY // Free data memory of tensor `t`. void TfLiteTensorDataFree(TfLiteTensor* t); // Free quantization data. void TfLiteQuantizationFree(TfLiteQuantization* quantization); -// Free sparsity parameters. -void TfLiteSparsityFree(TfLiteSparsity* sparsity); - // Free memory of tensor `t`. void TfLiteTensorFree(TfLiteTensor* t); @@ -416,12 +371,11 @@ void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims, // Resize the allocated data of a (dynamic) tensor. Tensors with allocation // types other than kTfLiteDynamic will be ignored. void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor); -#endif // TF_LITE_STATIC_MEMORY // A structure representing an instance of a node. // This structure only exhibits the inputs, outputs and user defined data, not // other features like the type. -typedef struct TfLiteNode { +typedef struct { // Inputs to this node expressed as indices into the simulator's tensors. TfLiteIntArray* inputs; @@ -455,20 +409,6 @@ typedef struct TfLiteNode { struct TfLiteDelegate* delegate; } TfLiteNode; -// WARNING: This is an experimental interface that is subject to change. -// -// Currently, TfLiteDelegateParams has to be allocated in a way that it's -// trivially destructable. It will be stored as `builtin_data` field in -// `TfLiteNode` of the delegate node. -// -// See also the `CreateDelegateParams` function in `interpreter.cc` details. -typedef struct TfLiteDelegateParams { - struct TfLiteDelegate* delegate; - TfLiteIntArray* nodes_to_replace; - TfLiteIntArray* input_tensors; - TfLiteIntArray* output_tensors; -} TfLiteDelegateParams; - typedef struct TfLiteContext { // Number of tensors in the context. size_t tensors_size; @@ -543,68 +483,6 @@ typedef struct TfLiteContext { // Pointer to the op-level profiler, if set; nullptr otherwise. void* profiler; - - // Allocate persistent buffer which has the same life time as the interpreter. - // The memory is allocated from heap for TFL, and from tail in TFLM. - // If *ptr is not nullptr, the pointer will be reallocated. - // This method is only available in Prepare stage. - // WARNING: This is an experimental interface that is subject to change. - TfLiteStatus (*AllocatePersistentBuffer)(struct TfLiteContext* ctx, - size_t bytes, void** ptr); - - // Allocate a buffer which will be deallocated right after invoke phase. - // The memory is allocated from heap in TFL, and from volatile arena in TFLM. - // This method is only available in invoke stage. - // NOTE: If possible use RequestScratchBufferInArena method to avoid memory - // allocation during inference time. - // WARNING: This is an experimental interface that is subject to change. - TfLiteStatus (*AllocateBufferForEval)(struct TfLiteContext* ctx, size_t bytes, - void** ptr); - - // Request a scratch buffer in the arena through static memory planning. - // This method is only available in Prepare stage and the buffer is allocated - // by the interpreter between Prepare and Eval stage. In Eval stage, - // GetScratchBuffer API can be used to fetch the address. - // WARNING: This is an experimental interface that is subject to change. - TfLiteStatus (*RequestScratchBufferInArena)(struct TfLiteContext* ctx, - size_t bytes, int* buffer_idx); - - // Get the scratch buffer pointer. - // This method is only available in Eval stage. - // WARNING: This is an experimental interface that is subject to change. - void* (*GetScratchBuffer)(struct TfLiteContext* ctx, int buffer_idx); - - // Resize the memory pointer of the `tensor`. This method behaves the same as - // `ResizeTensor`, except that it makes a copy of the shape array internally - // so the shape array could be deallocated right afterwards. - // WARNING: This is an experimental interface that is subject to change. - TfLiteStatus (*ResizeTensorExplicit)(struct TfLiteContext* ctx, - TfLiteTensor* tensor, int dims, - const int* shape); - - // This method provides a preview of post-delegation partitioning. Each - // TfLiteDelegateParams in the referenced array corresponds to one instance of - // the delegate kernel. - // Example usage: - // - // TfLiteIntArray* nodes_to_replace = ...; - // TfLiteDelegateParams* params_array; - // int num_partitions = 0; - // TF_LITE_ENSURE_STATUS(context->PreviewDelegatePartitioning( - // context, delegate, nodes_to_replace, ¶ms_array, &num_partitions)); - // for (int idx = 0; idx < num_partitions; idx++) { - // const auto& partition_params = params_array[idx]; - // ... - // } - // - // NOTE: The context owns the memory referenced by partition_params_array. It - // will be cleared with another call to PreviewDelegateParitioning, or after - // TfLiteDelegateParams::Prepare returns. - // - // WARNING: This is an experimental interface that is subject to change. - TfLiteStatus (*PreviewDelegatePartitioning)( - struct TfLiteContext* context, const TfLiteIntArray* nodes_to_replace, - TfLiteDelegateParams** partition_params_array, int* num_partitions); } TfLiteContext; typedef struct TfLiteRegistration { @@ -668,7 +546,7 @@ typedef struct TfLiteRegistration { // The flags used in `TfLiteDelegate`. Note that this is a bitmask, so the // values should be 1, 2, 4, 8, ...etc. -typedef enum TfLiteDelegateFlags { +typedef enum { kTfLiteDelegateFlagsNone = 0, // The flag is set if the delegate can handle dynamic sized tensors. // For example, the output shape of a `Resize` op with non-constant shape @@ -728,7 +606,23 @@ typedef struct TfLiteDelegate { // values. TfLiteDelegate TfLiteDelegateCreate(); +// WARNING: This is an experimental interface that is subject to change. +// +// Currently, TfLiteDelegateParams has to be allocated in a way that it's +// trivially destructable. It will be stored as `builtin_data` field in +// `TfLiteNode` of the delegate node. +// +// See also the `CreateDelegateParams` function in `interpreter.cc` details. +typedef struct { + TfLiteDelegate* delegate; + TfLiteIntArray* nodes_to_replace; + TfLiteIntArray* input_tensors; + TfLiteIntArray* output_tensors; +} TfLiteDelegateParams; + #ifdef __cplusplus } // extern "C" #endif // __cplusplus -#endif // TENSORFLOW_LITE_C_COMMON_H_ +#endif // TENSORFLOW_LITE_C_C_API_INTERNAL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/core/api/error_reporter.cpp b/src/tensorflow_esp32/tensorflow/lite/core/api/error_reporter.cpp new file mode 100644 index 0000000..b8e1d4b --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/error_reporter.cpp @@ -0,0 +1,41 @@ +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include + +namespace tflite { + +int ErrorReporter::Report(const char* format, ...) { + va_list args; + va_start(args, format); + int code = Report(format, args); + va_end(args); + return code; +} + +// TODO(aselle): Make the name of ReportError on context the same, so +// we can use the ensure functions w/o a context and w/ a reporter. +int ErrorReporter::ReportError(void*, const char* format, ...) { + va_list args; + va_start(args, format); + int code = Report(format, args); + va_end(args); + return code; +} + +} // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/error_reporter.h b/src/tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h similarity index 96% rename from src/tensorflow/lite/core/api/error_reporter.h rename to src/tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h index fcde1c4..cabebe8 100644 --- a/src/tensorflow/lite/core/api/error_reporter.h +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -43,3 +44,5 @@ class ErrorReporter { } // namespace tflite #endif // TENSORFLOW_LITE_CORE_API_ERROR_REPORTER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/flatbuffer_conversions.cpp b/src/tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.cpp similarity index 97% rename from src/tensorflow/lite/core/api/flatbuffer_conversions.cpp rename to src/tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.cpp index 3ce8bab..5c61337 100644 --- a/src/tensorflow/lite/core/api/flatbuffer_conversions.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,13 +14,13 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h" #include -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { @@ -68,14 +69,14 @@ TfLiteStatus FlatBufferIntVectorToArray( op_name); return kTfLiteError; } else { - size_t num_dimensions = flat_vector->size(); + int num_dimensions = flat_vector->size(); if (num_dimensions > max_size_of_buffer / sizeof(int)) { error_reporter->Report( "Found too many dimensions in the input array of operation '%s'.\n", op_name); return kTfLiteError; } else { - for (size_t i = 0; i < num_dimensions; ++i) { + for (int i = 0; i < num_dimensions; ++i) { buffer[i] = flat_vector->Get(i); } } @@ -476,7 +477,6 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, if (const auto* schema_params = op->builtin_options_as_ResizeBilinearOptions()) { params->align_corners = schema_params->align_corners(); - params->half_pixel_centers = schema_params->half_pixel_centers(); } *builtin_data = reinterpret_cast(params.release()); break; @@ -500,15 +500,10 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, auto params = safe_allocator.Allocate(); if (const auto* schema_params = op->builtin_options_as_ReshapeOptions()) { auto* new_shape = schema_params->new_shape(); - // TODO(b/147203660): We need to figure out when dynamic reshape - // (new_shape is a tensor) happens, why the option is not a nullptr. - // But nonethless, we should only copy when new_shape is not a nullptr. - if (new_shape) { - TF_LITE_ENSURE_STATUS(FlatBufferIntVectorToArray( - sizeof(params->shape), new_shape, params->shape, error_reporter, - "reshape")); - params->num_dimensions = new_shape->size(); - } + TF_LITE_ENSURE_STATUS(FlatBufferIntVectorToArray( + sizeof(params->shape), new_shape, params->shape, error_reporter, + "reshape")); + params->num_dimensions = new_shape->size(); } *builtin_data = reinterpret_cast(params.release()); break; @@ -797,7 +792,6 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, case BuiltinOperator_ROUND: case BuiltinOperator_RSQRT: case BuiltinOperator_SELECT: - case BuiltinOperator_SELECT_V2: case BuiltinOperator_SIN: case BuiltinOperator_SLICE: case BuiltinOperator_SPACE_TO_BATCH_ND: @@ -826,11 +820,11 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, case BuiltinOperator_NON_MAX_SUPPRESSION_V4: case BuiltinOperator_NON_MAX_SUPPRESSION_V5: case BuiltinOperator_SCATTER_ND: - case BuiltinOperator_DENSIFY: - case BuiltinOperator_SEGMENT_SUM: break; } return kTfLiteOk; } // NOLINT[readability/fn_size] } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/flatbuffer_conversions.h b/src/tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h similarity index 89% rename from src/tensorflow/lite/core/api/flatbuffer_conversions.h rename to src/tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h index a6ed2e9..77609a0 100644 --- a/src/tensorflow/lite/core/api/flatbuffer_conversions.h +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -19,10 +20,10 @@ limitations under the License. // flatbuffer serialization format into in-memory values that are used by the // runtime API and interpreter. -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/core/api/op_resolver.h" -#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { @@ -66,3 +67,5 @@ TfLiteStatus ConvertTensorType(TensorType tensor_type, TfLiteType* type, } // namespace tflite #endif // TENSORFLOW_LITE_CORE_API_FLATBUFFER_CONVERSIONS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/op_resolver.cpp b/src/tensorflow_esp32/tensorflow/lite/core/api/op_resolver.cpp similarity index 94% rename from src/tensorflow/lite/core/api/op_resolver.cpp rename to src/tensorflow_esp32/tensorflow/lite/core/api/op_resolver.cpp index 284f003..c92015c 100644 --- a/src/tensorflow/lite/core/api/op_resolver.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/op_resolver.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,7 +14,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/core/api/op_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h" namespace tflite { @@ -57,3 +58,5 @@ TfLiteStatus GetRegistrationFromOpCode( } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/op_resolver.h b/src/tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h similarity index 88% rename from src/tensorflow/lite/core/api/op_resolver.h rename to src/tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h index 1294b7b..1945a3b 100644 --- a/src/tensorflow/lite/core/api/op_resolver.h +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,9 @@ limitations under the License. #ifndef TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_ #define TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { @@ -46,3 +47,5 @@ TfLiteStatus GetRegistrationFromOpCode(const OperatorCode* opcode, } // namespace tflite #endif // TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/core/api/tensor_utils.cpp b/src/tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.cpp similarity index 90% rename from src/tensorflow/lite/core/api/tensor_utils.cpp rename to src/tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.cpp index d8d6fc4..7470637 100644 --- a/src/tensorflow/lite/core/api/tensor_utils.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,7 +14,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/core/api/tensor_utils.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h" #include @@ -37,7 +38,7 @@ TfLiteStatus ResetVariableTensor(TfLiteTensor* tensor) { memset(tensor->data.raw, value, tensor->bytes); #else char* raw_ptr = tensor->data.raw; - for (size_t i = 0; i < tensor->bytes; ++i) { + for (int i = 0; i < tensor->bytes; ++i) { *raw_ptr = value; raw_ptr++; } @@ -46,3 +47,5 @@ TfLiteStatus ResetVariableTensor(TfLiteTensor* tensor) { } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h b/src/tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h new file mode 100644 index 0000000..92ceaa2 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h @@ -0,0 +1,31 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_ +#define TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_ + +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" + +namespace tflite { + +// Resets a variable tensor to the default value. +TfLiteStatus ResetVariableTensor(TfLiteTensor* tensor); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h new file mode 100644 index 0000000..3ea946e --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h @@ -0,0 +1,35 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_COMPATIBILITY_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_COMPATIBILITY_H_ + +// C++ will automatically create class-specific delete operators for virtual +// objects, which by default call the global delete function. For embedded +// applications we want to avoid this, and won't be calling new/delete on these +// objects, so we need to override the default implementation with one that does +// nothing to avoid linking in ::delete(). +// This macro needs to be included in all subclasses of a virtual base class in +// the private section. +#ifdef TF_LITE_STATIC_MEMORY +#define TF_LITE_REMOVE_VIRTUAL_DELETE \ + void operator delete(void* p) {} +#else +#define TF_LITE_REMOVE_VIRTUAL_DELETE +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_COMPATIBILITY_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/debug_log.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.cpp similarity index 88% rename from src/tensorflow/lite/micro/debug_log.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.cpp index 7ef582b..37251e6 100644 --- a/src/tensorflow/lite/micro/debug_log.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -31,11 +32,13 @@ limitations under the License. // To add an equivalent function for your own platform, create your own // implementation file, and place it in a subfolder with named after the OS // you're targeting. For example, see the Cortex M bare metal version in -// tensorflow/lite/micro/bluepill/debug_log.cc or the mbed one on -// tensorflow/lite/micro/mbed/debug_log.cc. +// tensorflow/lite/experimental/micro/bluepill/debug_log.cc or the mbed one on +// tensorflow/lite/experimental/micro/mbed/debug_log.cc. -#include "tensorflow/lite/micro/debug_log.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.h" #include extern "C" void DebugLog(const char* s) { fprintf(stderr, "%s", s); } + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.h new file mode 100644 index 0000000..4e52054 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.h @@ -0,0 +1,26 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_DEBUG_LOG_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_DEBUG_LOG_H_ + +// This function should be implemented by each target platform, and provide a +// way for strings to be output to some text stream. For more information, see +// tensorflow/lite/experimental/micro/debug_log.cc. +extern "C" void DebugLog(const char* s); + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_DEBUG_LOG_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/debug_log_numbers.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.cpp similarity index 96% rename from src/tensorflow/lite/micro/debug_log_numbers.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.cpp index bb6e1d4..e16b29a 100644 --- a/src/tensorflow/lite/micro/debug_log_numbers.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -19,9 +20,9 @@ limitations under the License. // of DebugLog() and then get the numerical variations without requiring any // more code. -#include "tensorflow/lite/micro/debug_log_numbers.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.h" -#include "tensorflow/lite/micro/debug_log.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.h" namespace { @@ -183,3 +184,5 @@ extern "C" void DebugLogFloat(float i) { FastFloatToBufferLeft(i, number_string); DebugLog(number_string); } + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/debug_log_numbers.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.h similarity index 77% rename from src/tensorflow/lite/micro/debug_log_numbers.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.h index 8cd16d4..1ddda19 100644 --- a/src/tensorflow/lite/micro/debug_log_numbers.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,8 +13,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_DEBUG_LOG_NUMBERS_H_ -#define TENSORFLOW_LITE_MICRO_DEBUG_LOG_NUMBERS_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_DEBUG_LOG_NUMBERS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_DEBUG_LOG_NUMBERS_H_ #include @@ -25,4 +26,6 @@ void DebugLogHex(uint32_t i); void DebugLogFloat(float i); } -#endif // TENSORFLOW_LITE_MICRO_DEBUG_LOG_NUMBERS_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_DEBUG_LOG_NUMBERS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/activation_utils.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/activation_utils.h similarity index 63% rename from src/tensorflow/lite/micro/kernels/activation_utils.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/activation_utils.h index 9887087..c256bf7 100644 --- a/src/tensorflow/lite/micro/kernels/activation_utils.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/activation_utils.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,19 +14,14 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_KERNELS_ACTIVATION_UTILS_H_ -#define TENSORFLOW_LITE_MICRO_KERNELS_ACTIVATION_UTILS_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_ACTIVATION_UTILS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_ACTIVATION_UTILS_H_ +#include #include +#include -#include "tensorflow/lite/c/builtin_op_data.h" - -#ifndef _max -#define _max(a, b) ((a) > (b) ? (a) : (b)) -#endif -#ifndef _min -#define _min(a, b) ((a) > (b) ? (b) : (a)) -#endif +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" namespace tflite { namespace ops { @@ -37,18 +33,19 @@ inline float ActivationValFloat(TfLiteFusedActivation act, float a) { case kTfLiteActNone: return a; case kTfLiteActRelu: - return _max(0.0f, a); + return a < 0.f ? 0.f : a; case kTfLiteActRelu1: - return _max(-1.0f, _min(a, 1.0f)); + return a < 0.f ? 0.f : ((a > 1.f) ? 1.f : a); case kTfLiteActRelu6: - return _max(0.0f, _min(a, 6.0f)); + return a < 0.f ? 0.f : ((a > 6.f) ? 6.f : a); case kTfLiteActTanh: - return tanh(a); + return (expf(a) - expf(-a)) / (expf(a) + expf(-a)); case kTfLiteActSignBit: - //return signbit(a); - return a > 0 ? 1 : 0; + return std::signbit(a); case kTfLiteActSigmoid: - return 1.0f / (1.0f + exp(-a)); + return 1.f / (1.f + expf(-a)); + default: + return a; } } @@ -56,4 +53,6 @@ inline float ActivationValFloat(TfLiteFusedActivation act, float a) { } // namespace ops } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_KERNELS_ACTIVATION_UTILS_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_ACTIVATION_UTILS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/activations.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/activations.cpp similarity index 85% rename from src/tensorflow/lite/micro/kernels/activations.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/activations.cpp index b111641..3ad39c1 100644 --- a/src/tensorflow/lite/micro/kernels/activations.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/activations.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,14 +14,14 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" -#include "tensorflow/lite/micro/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -160,19 +161,21 @@ TfLiteStatus Relu6Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace activations TfLiteRegistration* Register_RELU() { - static TfLiteRegistration r = {}; - r.prepare = activations::ReluPrepare; - r.invoke = activations::ReluEval; + static TfLiteRegistration r = {/*init=*/nullptr, + /*free=*/nullptr, activations::ReluPrepare, + activations::ReluEval}; return &r; } TfLiteRegistration* Register_RELU6() { - static TfLiteRegistration r = {}; - r.prepare = activations::Relu6Prepare; - r.invoke = activations::Relu6Eval; + static TfLiteRegistration r = {/*init=*/nullptr, + /*free=*/nullptr, activations::Relu6Prepare, + activations::Relu6Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/add.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/add.cpp similarity index 82% rename from src/tensorflow/lite/micro/kernels/add.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/add.cpp index 38c7a17..cb4fbd9 100644 --- a/src/tensorflow/lite/micro/kernels/add.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/add.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,16 +14,16 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/add.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/add.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/reference/integer_ops/add.h" -#include "tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/add.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -77,15 +78,14 @@ TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteAddParams* params, data->output_offset = output->params.zero_point; data->left_shift = 20; const double twice_max_input_scale = - 2 * static_cast( - std::max(input1->params.scale, input2->params.scale)); + 2 * std::max(input1->params.scale, input2->params.scale); const double real_input1_multiplier = - static_cast(input1->params.scale) / twice_max_input_scale; + input1->params.scale / twice_max_input_scale; const double real_input2_multiplier = - static_cast(input2->params.scale) / twice_max_input_scale; + input2->params.scale / twice_max_input_scale; const double real_output_multiplier = twice_max_input_scale / - ((1 << data->left_shift) * static_cast(output->params.scale)); + ((1 << data->left_shift) * output->params.scale); QuantizeMultiplierSmallerThanOneExp( real_input1_multiplier, &data->input1_multiplier, &data->input1_shift); @@ -96,9 +96,15 @@ TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteAddParams* params, QuantizeMultiplierSmallerThanOneExp( real_output_multiplier, &data->output_multiplier, &data->output_shift); - TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( - context, params->activation, output, &data->output_activation_min, - &data->output_activation_max)); + if (output->type == kTfLiteUInt8) { + CalculateActivationRangeUint8(params->activation, output, + &data->output_activation_min, + &data->output_activation_max); + } else { + CalculateActivationRangeInt8(params->activation, output, + &data->output_activation_min, + &data->output_activation_max); + } } return kTfLiteOk; @@ -198,14 +204,12 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace add TfLiteRegistration* Register_ADD() { - static TfLiteRegistration r = {}; - r.init = add::Init; - r.free = add::Free; - r.prepare = add::Prepare; - r.invoke = add::Eval; + static TfLiteRegistration r = {add::Init, add::Free, add::Prepare, add::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/all_ops_resolver.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.cpp similarity index 65% rename from src/tensorflow/lite/micro/kernels/all_ops_resolver.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.cpp index ba725f6..9bfaf59 100644 --- a/src/tensorflow/lite/micro/kernels/all_ops_resolver.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -10,27 +11,25 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/kernels/all_ops_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.h" -#include "tensorflow/lite/micro/kernels/micro_ops.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_ops.h" namespace tflite { namespace ops { namespace micro { -// Register each supported op with: -// AddBuiltin(, , [min version], [max version]) AllOpsResolver::AllOpsResolver() { - AddBuiltin(BuiltinOperator_FULLY_CONNECTED, Register_FULLY_CONNECTED(), 1, 4); + AddBuiltin(BuiltinOperator_DEPTHWISE_CONV_2D, Register_DEPTHWISE_CONV_2D()); + AddBuiltin(BuiltinOperator_FULLY_CONNECTED, Register_FULLY_CONNECTED(), + /* min_version */ 1, + /* max_version */ 4); AddBuiltin(BuiltinOperator_MAX_POOL_2D, Register_MAX_POOL_2D()); - AddBuiltin(BuiltinOperator_SOFTMAX, Register_SOFTMAX(), 1, 2); + AddBuiltin(BuiltinOperator_SOFTMAX, Register_SOFTMAX()); AddBuiltin(BuiltinOperator_LOGISTIC, Register_LOGISTIC()); - AddBuiltin(BuiltinOperator_SVDF, Register_SVDF(), 1, 3); - AddBuiltin(BuiltinOperator_CONV_2D, Register_CONV_2D(), 1, 3); - AddBuiltin(BuiltinOperator_CONCATENATION, Register_CONCATENATION(), 1, 3); - AddBuiltin(BuiltinOperator_DEPTHWISE_CONV_2D, Register_DEPTHWISE_CONV_2D(), 1, - 3); - AddBuiltin(BuiltinOperator_AVERAGE_POOL_2D, Register_AVERAGE_POOL_2D(), 1, 2); + AddBuiltin(BuiltinOperator_SVDF, Register_SVDF()); + AddBuiltin(BuiltinOperator_CONV_2D, Register_CONV_2D()); + AddBuiltin(BuiltinOperator_AVERAGE_POOL_2D, Register_AVERAGE_POOL_2D()); AddBuiltin(BuiltinOperator_ABS, Register_ABS()); AddBuiltin(BuiltinOperator_SIN, Register_SIN()); AddBuiltin(BuiltinOperator_COS, Register_COS()); @@ -48,25 +47,24 @@ AllOpsResolver::AllOpsResolver() { AddBuiltin(BuiltinOperator_LOGICAL_AND, Register_LOGICAL_AND()); AddBuiltin(BuiltinOperator_LOGICAL_NOT, Register_LOGICAL_NOT()); AddBuiltin(BuiltinOperator_RESHAPE, Register_RESHAPE()); - AddBuiltin(BuiltinOperator_EQUAL, Register_EQUAL(), 1, 2); - AddBuiltin(BuiltinOperator_NOT_EQUAL, Register_NOT_EQUAL(), 1, 2); - AddBuiltin(BuiltinOperator_GREATER, Register_GREATER(), 1, 2); - AddBuiltin(BuiltinOperator_GREATER_EQUAL, Register_GREATER_EQUAL(), 1, 2); - AddBuiltin(BuiltinOperator_LESS, Register_LESS(), 1, 2); - AddBuiltin(BuiltinOperator_LESS_EQUAL, Register_LESS_EQUAL(), 1, 2); + AddBuiltin(BuiltinOperator_EQUAL, Register_EQUAL()); + AddBuiltin(BuiltinOperator_NOT_EQUAL, Register_NOT_EQUAL()); + AddBuiltin(BuiltinOperator_GREATER, Register_GREATER()); + AddBuiltin(BuiltinOperator_GREATER_EQUAL, Register_GREATER_EQUAL()); + AddBuiltin(BuiltinOperator_LESS, Register_LESS()); + AddBuiltin(BuiltinOperator_LESS_EQUAL, Register_LESS_EQUAL()); AddBuiltin(BuiltinOperator_CEIL, Register_CEIL()); AddBuiltin(BuiltinOperator_ROUND, Register_ROUND()); AddBuiltin(BuiltinOperator_STRIDED_SLICE, Register_STRIDED_SLICE()); - AddBuiltin(BuiltinOperator_PACK, Register_PACK(), 1, 2); - AddBuiltin(BuiltinOperator_PAD, Register_PAD(), 1, 2); - AddBuiltin(BuiltinOperator_PADV2, Register_PADV2(), 1, 2); - AddBuiltin(BuiltinOperator_SPLIT, Register_SPLIT(), 1, 3); - AddBuiltin(BuiltinOperator_UNPACK, Register_UNPACK(), 1, 2); + AddBuiltin(BuiltinOperator_PACK, Register_PACK()); + AddBuiltin(BuiltinOperator_SPLIT, Register_SPLIT(), + /* min_version */ 1, + /* max_version */ 3); + AddBuiltin(BuiltinOperator_UNPACK, Register_UNPACK()); AddBuiltin(BuiltinOperator_NEG, Register_NEG()); - AddBuiltin(BuiltinOperator_ADD, Register_ADD(), 1, 2); - AddBuiltin(BuiltinOperator_MUL, Register_MUL(), 1, 3); - AddBuiltin(BuiltinOperator_QUANTIZE, Register_QUANTIZE()); - AddBuiltin(BuiltinOperator_DEQUANTIZE, Register_DEQUANTIZE(), 1, 2); + AddBuiltin(BuiltinOperator_ADD, Register_ADD()); + AddBuiltin(BuiltinOperator_QUANTIZE, Register_QUANTIZE(), 1, 4); + AddBuiltin(BuiltinOperator_DEQUANTIZE, Register_DEQUANTIZE(), 1, 4); AddBuiltin(BuiltinOperator_RELU, Register_RELU()); AddBuiltin(BuiltinOperator_RELU6, Register_RELU6()); } @@ -74,3 +72,5 @@ AllOpsResolver::AllOpsResolver() { } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/all_ops_resolver.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.h similarity index 67% rename from src/tensorflow/lite/micro/kernels/all_ops_resolver.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.h index 26bb032..780c4a6 100644 --- a/src/tensorflow/lite/micro/kernels/all_ops_resolver.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/all_ops_resolver.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -9,11 +10,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_KERNELS_ALL_OPS_RESOLVER_H_ -#define TENSORFLOW_LITE_MICRO_KERNELS_ALL_OPS_RESOLVER_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_ALL_OPS_RESOLVER_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_ALL_OPS_RESOLVER_H_ -#include "tensorflow/lite/micro/compatibility.h" -#include "tensorflow/lite/micro/micro_mutable_op_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.h" namespace tflite { namespace ops { @@ -31,4 +32,6 @@ class AllOpsResolver : public MicroMutableOpResolver { } // namespace ops } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_KERNELS_ALL_OPS_RESOLVER_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_ALL_OPS_RESOLVER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/arg_min_max.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/arg_min_max.cpp similarity index 82% rename from src/tensorflow/lite/micro/kernels/arg_min_max.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/arg_min_max.cpp index f4d0021..fd0f73b 100644 --- a/src/tensorflow/lite/micro/kernels/arg_min_max.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/arg_min_max.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,13 +14,13 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/arg_min_max.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/arg_min_max.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/micro/kernels/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -104,19 +105,19 @@ TfLiteStatus ArgMaxEval(TfLiteContext* context, TfLiteNode* node) { } // namespace arg_min_max TfLiteRegistration* Register_ARG_MAX() { - static TfLiteRegistration r = {}; - r.prepare = arg_min_max::Prepare; - r.invoke = arg_min_max::ArgMaxEval; + static TfLiteRegistration r = {nullptr, nullptr, arg_min_max::Prepare, + arg_min_max::ArgMaxEval}; return &r; } TfLiteRegistration* Register_ARG_MIN() { - static TfLiteRegistration r = {}; - r.prepare = arg_min_max::Prepare; - r.invoke = arg_min_max::ArgMinEval; + static TfLiteRegistration r = {nullptr, nullptr, arg_min_max::Prepare, + arg_min_max::ArgMinEval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/ceil.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/ceil.cpp similarity index 81% rename from src/tensorflow/lite/micro/kernels/ceil.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/ceil.cpp index 9076019..8b56ecd 100644 --- a/src/tensorflow/lite/micro/kernels/ceil.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/ceil.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,11 +14,11 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/ceil.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/ceil.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -54,12 +55,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace ceil TfLiteRegistration* Register_CEIL() { - static TfLiteRegistration r = {}; - r.prepare = ceil::Prepare; - r.invoke = ceil::Eval; + static TfLiteRegistration r = {/*init=*/nullptr, + /*free=*/nullptr, ceil::Prepare, ceil::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/comparisons.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/comparisons.cpp similarity index 90% rename from src/tensorflow/lite/micro/kernels/comparisons.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/comparisons.cpp index fff056c..dac97c0 100644 --- a/src/tensorflow/lite/micro/kernels/comparisons.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/comparisons.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,12 +13,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/comparisons.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/comparisons.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -43,14 +44,12 @@ constexpr int kOutputTensor = 0; \ int32 input1_multiplier; \ int input1_shift; \ - QuantizeMultiplierSmallerThanOneExp( \ - static_cast(input1->params.scale), &input1_multiplier, \ - &input1_shift); \ + QuantizeMultiplierSmallerThanOneExp(input1->params.scale, \ + &input1_multiplier, &input1_shift); \ int32 input2_multiplier; \ int input2_shift; \ - QuantizeMultiplierSmallerThanOneExp( \ - static_cast(input2->params.scale), &input2_multiplier, \ - &input2_shift); \ + QuantizeMultiplierSmallerThanOneExp(input2->params.scale, \ + &input2_multiplier, &input2_shift); \ \ ComparisonParams op_params; \ op_params.left_shift = left_shift; \ @@ -300,41 +299,43 @@ TfLiteStatus LessEqualEval(TfLiteContext* context, TfLiteNode* node) { } // namespace comparisons TfLiteRegistration* Register_EQUAL() { - static TfLiteRegistration r = {}; - r.invoke = comparisons::EqualEval; + static TfLiteRegistration r = {nullptr, nullptr, nullptr, + comparisons::EqualEval}; return &r; } TfLiteRegistration* Register_NOT_EQUAL() { - static TfLiteRegistration r = {}; - r.invoke = comparisons::NotEqualEval; + static TfLiteRegistration r = {nullptr, nullptr, nullptr, + comparisons::NotEqualEval}; return &r; } TfLiteRegistration* Register_GREATER() { - static TfLiteRegistration r = {}; - r.invoke = comparisons::GreaterEval; + static TfLiteRegistration r = {nullptr, nullptr, nullptr, + comparisons::GreaterEval}; return &r; } TfLiteRegistration* Register_GREATER_EQUAL() { - static TfLiteRegistration r = {}; - r.invoke = comparisons::GreaterEqualEval; + static TfLiteRegistration r = {nullptr, nullptr, nullptr, + comparisons::GreaterEqualEval}; return &r; } TfLiteRegistration* Register_LESS() { - static TfLiteRegistration r = {}; - r.invoke = comparisons::LessEval; + static TfLiteRegistration r = {nullptr, nullptr, nullptr, + comparisons::LessEval}; return &r; } TfLiteRegistration* Register_LESS_EQUAL() { - static TfLiteRegistration r = {}; - r.invoke = comparisons::LessEqualEval; + static TfLiteRegistration r = {nullptr, nullptr, nullptr, + comparisons::LessEqualEval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/conv.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/conv.cpp similarity index 92% rename from src/tensorflow/lite/micro/kernels/conv.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/conv.cpp index 918b850..221cbd5 100644 --- a/src/tensorflow/lite/micro/kernels/conv.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/conv.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,16 +14,16 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/conv.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/conv.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/reference/integer_ops/conv.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/padding.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/padding.h" namespace tflite { namespace ops { @@ -37,6 +38,8 @@ constexpr int kMaxChannels = 256; // This file has 2 implementation of Conv. +const int kTensorNotAllocated = -1; + struct OpData { TfLitePaddingValues padding; // The scaling factor from input to output (aka the 'real multiplier') can @@ -160,8 +163,6 @@ void EvalQuantizedPerChannel(TfLiteContext* context, TfLiteNode* node, op_params.dilation_width_factor = params->dilation_width_factor; op_params.padding_values.height = data->padding.height; op_params.padding_values.width = data->padding.width; - op_params.quantized_activation_min = data->output_activation_min; - op_params.quantized_activation_max = data->output_activation_max; reference_integer_ops::ConvPerChannel( op_params, data->per_channel_output_multiplier, @@ -230,7 +231,6 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE(context, affine_quantization->zero_point); // Conv is quantized along dimension 0: // https://www.tensorflow.org/lite/performance/quantization_spec - TF_LITE_ENSURE_EQ(context, affine_quantization->quantized_dimension, 0); TF_LITE_ENSURE_EQ(context, filter->dims->data[0], affine_quantization->scale->size); TF_LITE_ENSURE_EQ(context, filter->dims->data[0], @@ -265,12 +265,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace conv TfLiteRegistration* Register_CONV_2D() { - static TfLiteRegistration r = {}; - r.prepare = conv::Prepare; - r.invoke = conv::Eval; + static TfLiteRegistration r = {conv::Init, conv::Free, conv::Prepare, + conv::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/depthwise_conv.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/depthwise_conv.cpp similarity index 90% rename from src/tensorflow/lite/micro/kernels/depthwise_conv.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/depthwise_conv.cpp index a6fd7c9..3dbcf2a 100644 --- a/src/tensorflow/lite/micro/kernels/depthwise_conv.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/depthwise_conv.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,17 +14,17 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h" -#include "tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/padding.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/padding.h" namespace tflite { namespace ops { @@ -35,7 +36,7 @@ constexpr int kInputTensor = 0; constexpr int kFilterTensor = 1; constexpr int kBiasTensor = 2; constexpr int kOutputTensor = 0; -constexpr int kMaxChannels = 256; +constexpr int kMaxChannels = 64; struct OpData { TfLitePaddingValues padding; @@ -116,8 +117,8 @@ void EvalFloat(TfLiteContext* context, TfLiteNode* node, op_params.padding_values.height = data->padding.height; op_params.stride_width = params->stride_width; op_params.stride_height = params->stride_height; - op_params.dilation_width_factor = params->dilation_width_factor; - op_params.dilation_height_factor = params->dilation_height_factor; + op_params.dilation_width_factor = 1; + op_params.dilation_height_factor = 1; op_params.depth_multiplier = params->depth_multiplier; op_params.float_activation_min = output_activation_min; op_params.float_activation_max = output_activation_max; @@ -174,8 +175,8 @@ void EvalQuantized(TfLiteContext* context, TfLiteNode* node, op_params.padding_values.height = data->padding.height; op_params.stride_width = params->stride_width; op_params.stride_height = params->stride_height; - op_params.dilation_width_factor = params->dilation_width_factor; - op_params.dilation_height_factor = params->dilation_height_factor; + op_params.dilation_width_factor = 1; + op_params.dilation_height_factor = 1; op_params.depth_multiplier = params->depth_multiplier; op_params.quantized_activation_min = data->output_activation_min; op_params.quantized_activation_max = data->output_activation_max; @@ -224,7 +225,6 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE(context, affine_quantization->zero_point); // Depthwise conv is quantized along dimension 3: // https://www.tensorflow.org/lite/performance/quantization_spec - TF_LITE_ENSURE_EQ(context, affine_quantization->quantized_dimension, 3); TF_LITE_ENSURE_EQ(context, filter->dims->data[3], affine_quantization->scale->size); TF_LITE_ENSURE_EQ(context, filter->dims->data[3], @@ -259,14 +259,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace depthwise_conv TfLiteRegistration* Register_DEPTHWISE_CONV_2D() { - static TfLiteRegistration r = {}; - r.init = depthwise_conv::Init; - r.free = depthwise_conv::Free; - r.prepare = depthwise_conv::Prepare; - r.invoke = depthwise_conv::Eval; + static TfLiteRegistration r = {depthwise_conv::Init, depthwise_conv::Free, + depthwise_conv::Prepare, depthwise_conv::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/dequantize.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/dequantize.cpp similarity index 80% rename from src/tensorflow/lite/micro/kernels/dequantize.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/dequantize.cpp index 1406be2..93f161c 100644 --- a/src/tensorflow/lite/micro/kernels/dequantize.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/dequantize.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,12 +14,12 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/dequantize.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/dequantize.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -46,7 +47,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { tflite::DequantizationParams op_params; op_params.zero_point = input->params.zero_point; - op_params.scale = static_cast(input->params.scale); + op_params.scale = input->params.scale; switch (input->type) { case kTfLiteUInt8: reference_ops::Dequantize( @@ -70,12 +71,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace dequantize TfLiteRegistration* Register_DEQUANTIZE() { - static TfLiteRegistration r = {}; - r.prepare = dequantize::Prepare; - r.invoke = dequantize::Eval; + static TfLiteRegistration r = {nullptr, nullptr, dequantize::Prepare, + dequantize::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/elementwise.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/elementwise.cpp similarity index 70% rename from src/tensorflow/lite/micro/kernels/elementwise.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/elementwise.cpp index cd77cec..12ac741 100644 --- a/src/tensorflow/lite/micro/kernels/elementwise.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/elementwise.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,9 @@ limitations under the License. #include -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -110,61 +111,71 @@ TfLiteStatus LogicalNotEval(TfLiteContext* context, TfLiteNode* node) { } // namespace elementwise TfLiteRegistration* Register_ABS() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::AbsEval; + static TfLiteRegistration r = { + /* init */ nullptr, /* free */ nullptr, + elementwise::GenericPrepare, + elementwise::AbsEval}; return &r; } TfLiteRegistration* Register_SIN() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::SinEval; + static TfLiteRegistration r = { + /* init */ nullptr, /* free */ nullptr, + elementwise::GenericPrepare, + elementwise::SinEval}; return &r; } TfLiteRegistration* Register_COS() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::CosEval; + static TfLiteRegistration r = { + /* init */ nullptr, /* free */ nullptr, + elementwise::GenericPrepare, + elementwise::CosEval}; return &r; } TfLiteRegistration* Register_LOG() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::LogEval; + static TfLiteRegistration r = { + /* init */ nullptr, /* free */ nullptr, + elementwise::GenericPrepare, + elementwise::LogEval}; return &r; } TfLiteRegistration* Register_SQRT() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::SqrtEval; + static TfLiteRegistration r = { + /* init */ nullptr, /* free */ nullptr, + elementwise::GenericPrepare, + elementwise::SqrtEval}; return &r; } TfLiteRegistration* Register_RSQRT() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::RsqrtEval; + static TfLiteRegistration r = { + /* init */ nullptr, /* free */ nullptr, + elementwise::GenericPrepare, + elementwise::RsqrtEval}; return &r; } TfLiteRegistration* Register_SQUARE() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::SquareEval; + static TfLiteRegistration r = { + /* init */ nullptr, /* free */ nullptr, + elementwise::GenericPrepare, + elementwise::SquareEval}; return &r; } TfLiteRegistration* Register_LOGICAL_NOT() { - static TfLiteRegistration r = {}; - r.prepare = elementwise::GenericPrepare; - r.invoke = elementwise::LogicalNotEval; + static TfLiteRegistration r = { + /*init=*/nullptr, /*free=*/nullptr, + elementwise::GenericPrepare, + elementwise::LogicalNotEval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/floor.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/floor.cpp similarity index 73% rename from src/tensorflow/lite/micro/kernels/floor.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/floor.cpp index 6082a86..525d3ba 100644 --- a/src/tensorflow/lite/micro/kernels/floor.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/floor.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,10 +14,10 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/reference/floor.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/floor.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -37,11 +38,14 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace floor TfLiteRegistration* Register_FLOOR() { - static TfLiteRegistration r = {}; - r.invoke = floor::Eval; + static TfLiteRegistration r = {/*init=*/nullptr, + /*free=*/nullptr, /*prepare=*/nullptr, + floor::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/fully_connected.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/fully_connected.cpp similarity index 90% rename from src/tensorflow/lite/micro/kernels/fully_connected.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/fully_connected.cpp index 8b23f02..9579de3 100644 --- a/src/tensorflow/lite/micro/kernels/fully_connected.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/fully_connected.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,15 +14,15 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/fully_connected.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/fully_connected.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -201,14 +202,14 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace fully_connected TfLiteRegistration* Register_FULLY_CONNECTED() { - static TfLiteRegistration r = {}; - r.init = fully_connected::Init; - r.free = fully_connected::Free; - r.prepare = fully_connected::Prepare; - r.invoke = fully_connected::Eval; + static TfLiteRegistration r = {fully_connected::Init, fully_connected::Free, + fully_connected::Prepare, + fully_connected::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/logical.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/logical.cpp similarity index 77% rename from src/tensorflow/lite/micro/kernels/logical.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/logical.cpp index 4b95bb4..84a96d9 100644 --- a/src/tensorflow/lite/micro/kernels/logical.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/logical.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,11 +13,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/reference/binary_function.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/binary_function.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -68,19 +69,22 @@ TfLiteStatus LogicalAndEval(TfLiteContext* context, TfLiteNode* node) { TfLiteRegistration* Register_LOGICAL_OR() { // Init, Free, Prepare, Eval are satisfying the Interface required by // TfLiteRegistration. - static TfLiteRegistration r = {}; - r.invoke = logical::LogicalOrEval; + static TfLiteRegistration r = {/* init */ nullptr, /* free */ nullptr, + /* prepare */ nullptr, logical::LogicalOrEval}; return &r; } TfLiteRegistration* Register_LOGICAL_AND() { // Init, Free, Prepare, Eval are satisfying the Interface required by // TfLiteRegistration. - static TfLiteRegistration r = {}; - r.invoke = logical::LogicalAndEval; + static TfLiteRegistration r = {/* init */ nullptr, /* free */ nullptr, + /* prepare */ nullptr, + logical::LogicalAndEval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/logistic.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/logistic.cpp similarity index 69% rename from src/tensorflow/lite/micro/kernels/logistic.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/logistic.cpp index 96cf6db..177e045 100644 --- a/src/tensorflow/lite/micro/kernels/logistic.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/logistic.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,15 +14,15 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/logistic.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/logistic.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -60,11 +61,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace activations TfLiteRegistration* Register_LOGISTIC() { - static TfLiteRegistration r = {}; - r.prepare = activations::Prepare; - r.invoke = activations::Eval; + static TfLiteRegistration r = {/*init=*/nullptr, + /*free=*/nullptr, activations::Prepare, + activations::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/maximum_minimum.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/maximum_minimum.cpp similarity index 76% rename from src/tensorflow/lite/micro/kernels/maximum_minimum.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/maximum_minimum.cpp index f49f87d..615a3d5 100644 --- a/src/tensorflow/lite/micro/kernels/maximum_minimum.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/maximum_minimum.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,15 +14,15 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/maximum_minimum.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/maximum_minimum.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -117,19 +118,27 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace maximum_minimum TfLiteRegistration* Register_MAXIMUM() { - static TfLiteRegistration r = {}; - r.invoke = maximum_minimum::Eval; + static TfLiteRegistration r = { + /* init */ nullptr, + /* free */ nullptr, + /* prepare */ nullptr, + maximum_minimum::Eval}; return &r; } TfLiteRegistration* Register_MINIMUM() { - static TfLiteRegistration r = {}; - r.invoke = maximum_minimum::Eval; + static TfLiteRegistration r = { + /* init */ nullptr, + /* free */ nullptr, + /* prepare */ nullptr, + maximum_minimum::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/micro_ops.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_ops.h similarity index 89% rename from src/tensorflow/lite/micro/kernels/micro_ops.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_ops.h index 35422ec..57eeebd 100644 --- a/src/tensorflow/lite/micro/kernels/micro_ops.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_ops.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,10 +13,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_KERNELS_MICRO_OPS_H_ -#define TENSORFLOW_LITE_MICRO_KERNELS_MICRO_OPS_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_MICRO_OPS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_MICRO_OPS_H_ -#include "tensorflow/lite/c/common.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" namespace tflite { namespace ops { @@ -36,7 +37,6 @@ TfLiteRegistration* Register_ARG_MIN(); TfLiteRegistration* Register_AVERAGE_POOL_2D(); TfLiteRegistration* Register_CEIL(); TfLiteRegistration* Register_CONV_2D(); -TfLiteRegistration* Register_CONCATENATION(); TfLiteRegistration* Register_COS(); TfLiteRegistration* Register_DEPTHWISE_CONV_2D(); TfLiteRegistration* Register_DEQUANTIZE(); @@ -55,12 +55,9 @@ TfLiteRegistration* Register_LOGISTIC(); TfLiteRegistration* Register_MAXIMUM(); TfLiteRegistration* Register_MAX_POOL_2D(); TfLiteRegistration* Register_MINIMUM(); -TfLiteRegistration* Register_MUL(); TfLiteRegistration* Register_NEG(); TfLiteRegistration* Register_NOT_EQUAL(); TfLiteRegistration* Register_PACK(); -TfLiteRegistration* Register_PAD(); -TfLiteRegistration* Register_PADV2(); TfLiteRegistration* Register_PRELU(); TfLiteRegistration* Register_QUANTIZE(); TfLiteRegistration* Register_RELU(); @@ -81,4 +78,6 @@ TfLiteRegistration* Register_UNPACK(); } // namespace ops } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_KERNELS_MICRO_OPS_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_MICRO_OPS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_utils.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_utils.h new file mode 100644 index 0000000..42394cf --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/micro_utils.h @@ -0,0 +1,40 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_MICRO_UTILS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_MICRO_UTILS_H_ +namespace tflite { +namespace ops { +namespace micro { + +// Same as gtl::Greater but defined here to reduce dependencies and +// binary size for micro environment. +struct Greater { + template + bool operator()(const T& x, const T& y) const { + return x > y; + } +}; + +struct Less { + template + bool operator()(const T& x, const T& y) const { + return x < y; + } +}; + +} // namespace micro +} // namespace ops +} // namespace tflite +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_KERNELS_MICRO_UTILS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/neg.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/neg.cpp similarity index 77% rename from src/tensorflow/lite/micro/kernels/neg.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/neg.cpp index 679b6f2..a96eca3 100644 --- a/src/tensorflow/lite/micro/kernels/neg.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/neg.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,11 +14,11 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/neg.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/neg.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -48,11 +49,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace neg TfLiteRegistration* Register_NEG() { - static TfLiteRegistration r = {}; - r.invoke = neg::Eval; + static TfLiteRegistration r = {/*init=*/nullptr, /*free=*/nullptr, + /*prepare=*/nullptr, neg::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/pack.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/pack.cpp similarity index 89% rename from src/tensorflow/lite/micro/kernels/pack.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/pack.cpp index 048f946..37ee32c 100644 --- a/src/tensorflow/lite/micro/kernels/pack.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/pack.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,10 +14,10 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -113,12 +114,12 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace pack TfLiteRegistration* Register_PACK() { - static TfLiteRegistration r = {}; - r.prepare = pack::Prepare; - r.invoke = pack::Eval; + static TfLiteRegistration r = {nullptr, nullptr, pack::Prepare, pack::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/pooling.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/pooling.cpp similarity index 69% rename from src/tensorflow/lite/micro/kernels/pooling.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/pooling.cpp index 696bbc7..9433a4c 100644 --- a/src/tensorflow/lite/micro/kernels/pooling.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/pooling.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,13 +13,13 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/pooling.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/pooling.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/padding.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/padding.h" namespace tflite { namespace ops { @@ -74,14 +75,12 @@ void AverageEvalFloat(const TfLiteContext* context, const TfLiteNode* node, GetTensorShape(output), GetTensorData(output)); } -void AverageEvalQuantized(TfLiteContext* context, const TfLiteNode* node, - const TfLitePoolParams* params, const OpData* data, - const TfLiteTensor* input, TfLiteTensor* output) { - assert(input->type == kTfLiteUInt8 || input->type == kTfLiteInt8); - +void AverageEvalUint8(const TfLiteContext* context, const TfLiteNode* node, + const TfLitePoolParams* params, const OpData* data, + const TfLiteTensor* input, TfLiteTensor* output) { int32_t activation_min, activation_max; - (void)CalculateActivationRangeQuantized(context, params->activation, output, - &activation_min, &activation_max); + CalculateActivationRangeUint8(params->activation, output, &activation_min, + &activation_max); PoolParams op_params; op_params.stride_height = params->stride_height; @@ -92,16 +91,30 @@ void AverageEvalQuantized(TfLiteContext* context, const TfLiteNode* node, op_params.padding_values.width = data->padding.width; op_params.quantized_activation_min = activation_min; op_params.quantized_activation_max = activation_max; + reference_ops::AveragePool( + op_params, GetTensorShape(input), GetTensorData(input), + GetTensorShape(output), GetTensorData(output)); +} - if (input->type == kTfLiteUInt8) { - reference_ops::AveragePool( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(output), GetTensorData(output)); - } else { - reference_integer_ops::AveragePool( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(output), GetTensorData(output)); - } +void AverageEvalInt8(const TfLiteContext* context, const TfLiteNode* node, + const TfLitePoolParams* params, const OpData* data, + const TfLiteTensor* input, TfLiteTensor* output) { + int32_t activation_min, activation_max; + CalculateActivationRangeInt8(params->activation, output, &activation_min, + &activation_max); + + PoolParams op_params; + op_params.stride_height = params->stride_height; + op_params.stride_width = params->stride_width; + op_params.filter_height = params->filter_height; + op_params.filter_width = params->filter_width; + op_params.padding_values.height = data->padding.height; + op_params.padding_values.width = data->padding.width; + op_params.quantized_activation_min = activation_min; + op_params.quantized_activation_max = activation_max; + reference_integer_ops::AveragePool( + op_params, GetTensorShape(input), GetTensorData(input), + GetTensorShape(output), GetTensorData(output)); } void MaxEvalFloat(TfLiteContext* context, TfLiteNode* node, @@ -125,14 +138,12 @@ void MaxEvalFloat(TfLiteContext* context, TfLiteNode* node, GetTensorData(output)); } -void MaxEvalQuantized(TfLiteContext* context, TfLiteNode* node, - TfLitePoolParams* params, OpData* data, - const TfLiteTensor* input, TfLiteTensor* output) { - assert(input->type == kTfLiteUInt8 || input->type == kTfLiteInt8); - +void MaxEvalQuantizedUInt8(TfLiteContext* context, TfLiteNode* node, + TfLitePoolParams* params, OpData* data, + const TfLiteTensor* input, TfLiteTensor* output) { int32_t activation_min, activation_max; - (void)CalculateActivationRangeQuantized(context, params->activation, output, - &activation_min, &activation_max); + CalculateActivationRangeUint8(params->activation, output, &activation_min, + &activation_max); tflite::PoolParams op_params; op_params.stride_height = params->stride_height; @@ -143,17 +154,11 @@ void MaxEvalQuantized(TfLiteContext* context, TfLiteNode* node, op_params.padding_values.width = data->padding.width; op_params.quantized_activation_min = activation_min; op_params.quantized_activation_max = activation_max; - - if (input->type == kTfLiteUInt8) { - reference_ops::MaxPool( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(output), GetTensorData(output)); - } else { - reference_integer_ops::MaxPool( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(output), GetTensorData(output)); - } + reference_ops::MaxPool(op_params, GetTensorShape(input), + GetTensorData(input), GetTensorShape(output), + GetTensorData(output)); } + } // namespace void* Init(TfLiteContext* context, const char* buffer, size_t length) { @@ -181,8 +186,10 @@ TfLiteStatus AverageEval(TfLiteContext* context, TfLiteNode* node) { AverageEvalFloat(context, node, params, &data, input, output); break; case kTfLiteUInt8: + AverageEvalUint8(context, node, params, &data, input, output); + break; case kTfLiteInt8: - AverageEvalQuantized(context, node, params, &data, input, output); + AverageEvalInt8(context, node, params, &data, input, output); break; default: context->ReportError(context, "Input type %s is not currently supported", @@ -206,8 +213,7 @@ TfLiteStatus MaxEval(TfLiteContext* context, TfLiteNode* node) { MaxEvalFloat(context, node, params, &data, input, output); break; case kTfLiteUInt8: - case kTfLiteInt8: - MaxEvalQuantized(context, node, params, &data, input, output); + MaxEvalQuantizedUInt8(context, node, params, &data, input, output); break; default: context->ReportError(context, "Type %s not currently supported.", @@ -220,23 +226,23 @@ TfLiteStatus MaxEval(TfLiteContext* context, TfLiteNode* node) { } // namespace pooling TfLiteRegistration* Register_AVERAGE_POOL_2D() { - static TfLiteRegistration r = {}; - r.init = pooling::Init; - r.free = pooling::Free; - r.prepare = pooling::Prepare; - r.invoke = pooling::AverageEval; + static TfLiteRegistration r = { + pooling::Init, + pooling::Free, + pooling::Prepare, + pooling::AverageEval, + }; return &r; } TfLiteRegistration* Register_MAX_POOL_2D() { - static TfLiteRegistration r = {}; - r.init = pooling::Init; - r.free = pooling::Free; - r.prepare = pooling::Prepare; - r.invoke = pooling::MaxEval; + static TfLiteRegistration r = {pooling::Init, pooling::Free, pooling::Prepare, + pooling::MaxEval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/prelu.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/prelu.cpp similarity index 84% rename from src/tensorflow/lite/micro/kernels/prelu.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/prelu.cpp index f35e784..27021f9 100644 --- a/src/tensorflow/lite/micro/kernels/prelu.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/prelu.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,12 +14,12 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/prelu.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/prelu.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -53,7 +54,7 @@ inline void BroadcastPrelu4DSlowFloat( auto in2_idx = SubscriptToIndex(desc2, b, y, x, c); auto in1_val = input1_data[in1_idx]; auto in2_val = input2_data[in2_idx]; - output_data[out_idx] = in1_val >= 0.0f ? in1_val : in1_val * in2_val; + output_data[out_idx] = in1_val >= 0.0 ? in1_val : in1_val * in2_val; } } } @@ -67,9 +68,8 @@ TfLiteStatus PreluEval(TfLiteContext* context, TfLiteNode* node) { int32_t output_multiplier = 0; int output_shift = 0; if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt16) { - double real_multiplier = static_cast(input->params.scale) * - static_cast(alpha->params.scale) / - static_cast(output->params.scale); + double real_multiplier = + input->params.scale * alpha->params.scale / output->params.scale; QuantizeMultiplierSmallerThanOneExp(real_multiplier, &output_multiplier, &output_shift); } @@ -105,12 +105,13 @@ TfLiteStatus PreluEval(TfLiteContext* context, TfLiteNode* node) { } // namespace activations TfLiteRegistration* Register_PRELU() { - static TfLiteRegistration r = {}; - r.prepare = activations::PreluPrepare; - r.invoke = activations::PreluEval; + static TfLiteRegistration r = {nullptr, nullptr, activations::PreluPrepare, + activations::PreluEval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/quantize.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/quantize.cpp similarity index 84% rename from src/tensorflow/lite/micro/kernels/quantize.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/quantize.cpp index 47c3e73..403ccc0 100644 --- a/src/tensorflow/lite/micro/kernels/quantize.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/quantize.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,12 +13,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/quantize.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/quantize.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -60,7 +61,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { tflite::QuantizationParams op_params; op_params.zero_point = output->params.zero_point; - op_params.scale = static_cast(output->params.scale); + op_params.scale = output->params.scale; switch (output->type) { case kTfLiteInt8: reference_ops::AffineQuantize( @@ -87,14 +88,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { // AffineQuantize takes scale and zero point and quantizes the float value to // quantized output, in int8 or uint8 format. TfLiteRegistration* Register_QUANTIZE() { - static TfLiteRegistration r = {}; - r.init = quantize::Init; - r.free = quantize::Free; - r.prepare = quantize::Prepare; - r.invoke = quantize::Eval; + static TfLiteRegistration r = {quantize::Init, quantize::Free, + quantize::Prepare, quantize::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/reshape.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/reshape.cpp similarity index 84% rename from src/tensorflow/lite/micro/kernels/reshape.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/reshape.cpp index d7a5a61..979e894 100644 --- a/src/tensorflow/lite/micro/kernels/reshape.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/reshape.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,11 +14,11 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -25,6 +26,7 @@ namespace micro { namespace reshape { constexpr int kInputTensor = 0; +constexpr int kShapeTensor = 1; constexpr int kOutputTensor = 0; TfLiteStatus ReshapeOutput(TfLiteContext* context, TfLiteNode* node) { @@ -79,7 +81,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { return kTfLiteError; } - for (size_t i = 0; i < input->bytes; ++i) { + for (int i = 0; i < input->bytes; ++i) { output->data.raw[i] = input->data.raw[i]; } return kTfLiteOk; @@ -88,12 +90,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace reshape TfLiteRegistration* Register_RESHAPE() { - static TfLiteRegistration r = {}; - r.prepare = reshape::Prepare; - r.invoke = reshape::Eval; + static TfLiteRegistration r = {nullptr, nullptr, reshape::Prepare, + reshape::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/round.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/round.cpp similarity index 81% rename from src/tensorflow/lite/micro/kernels/round.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/round.cpp index f4e0586..35b00e4 100644 --- a/src/tensorflow/lite/micro/kernels/round.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/round.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,11 +14,11 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/round.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -54,12 +55,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace round TfLiteRegistration* Register_ROUND() { - static TfLiteRegistration r = {}; - r.prepare = round::Prepare; - r.invoke = round::Eval; + static TfLiteRegistration r = {/*init=*/nullptr, + /*free=*/nullptr, round::Prepare, round::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/softmax.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/softmax.cpp similarity index 72% rename from src/tensorflow/lite/micro/kernels/softmax.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/softmax.cpp index dc9b8bd..5604a33 100644 --- a/src/tensorflow/lite/micro/kernels/softmax.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/softmax.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,16 +14,14 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/softmax.h" - -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/reference/integer_ops/softmax.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/softmax.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -42,19 +41,14 @@ TfLiteStatus CalculateSoftmaxOpData(TfLiteContext* context, TfLiteTensor* output, const TfLiteSoftmaxParams* params, OpData* data) { - if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8) { - if (input->type == kTfLiteUInt8) { - TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0); - } else { - TF_LITE_ENSURE_EQ(context, output->params.zero_point, -128); - } + if (input->type == kTfLiteUInt8) { + TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0); TF_LITE_ENSURE(context, output->params.scale == 1.f / 256); static const int kScaledDiffIntegerBits = 5; tflite::PreprocessSoftmaxScaling( - static_cast(params->beta), - static_cast(input->params.scale), kScaledDiffIntegerBits, + params->beta, input->params.scale, kScaledDiffIntegerBits, &data->input_multiplier, &data->input_left_shift); data->diff_min = -1.0 * tflite::CalculateInputRadius( kScaledDiffIntegerBits, data->input_left_shift); @@ -104,15 +98,9 @@ void Softmax1DQuantized(const TfLiteTensor* input, TfLiteTensor* output, op_params.input_multiplier = data->input_multiplier; op_params.input_left_shift = data->input_left_shift; op_params.diff_min = data->diff_min; - if (input->type == kTfLiteUInt8) { - tflite::reference_ops::Softmax(op_params, shape, - GetTensorData(input), shape, - GetTensorData(output)); - } else { - tflite::reference_integer_ops::Softmax(op_params, shape, - GetTensorData(input), shape, - GetTensorData(output)); - } + tflite::reference_ops::Softmax(op_params, shape, + GetTensorData(input), shape, + GetTensorData(output)); } void Softmax2DQuantized(const TfLiteTensor* input, TfLiteTensor* output, @@ -129,22 +117,16 @@ void Softmax2DQuantized(const TfLiteTensor* input, TfLiteTensor* output, op_params.input_multiplier = data->input_multiplier; op_params.input_left_shift = data->input_left_shift; op_params.diff_min = data->diff_min; - if (input->type == kTfLiteUInt8) { - tflite::reference_ops::Softmax(op_params, shape, - GetTensorData(input), shape, - GetTensorData(output)); - } else { - tflite::reference_integer_ops::Softmax(op_params, shape, - GetTensorData(input), shape, - GetTensorData(output)); - } + tflite::reference_ops::Softmax(op_params, shape, + GetTensorData(input), shape, + GetTensorData(output)); } // Takes a 4D tensor and perform softmax along the forth dimension. void Softmax4DFloat(const TfLiteTensor* input, TfLiteTensor* output, TfLiteSoftmaxParams* params) { SoftmaxParams op_params; - op_params.beta = static_cast(params->beta); + op_params.beta = params->beta; tflite::reference_ops::Softmax( op_params, GetTensorShape(input), GetTensorData(input), GetTensorShape(output), GetTensorData(output)); @@ -156,15 +138,9 @@ void Softmax4DQuantized(const TfLiteTensor* input, TfLiteTensor* output, op_params.input_multiplier = data->input_multiplier; op_params.input_left_shift = data->input_left_shift; op_params.diff_min = data->diff_min; - if (input->type == kTfLiteUInt8) { - tflite::reference_ops::Softmax( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(output), GetTensorData(output)); - } else { - tflite::reference_integer_ops::Softmax( - op_params, GetTensorShape(input), GetTensorData(input), - GetTensorShape(output), GetTensorData(output)); - } + tflite::reference_ops::Softmax( + op_params, GetTensorShape(input), GetTensorData(input), + GetTensorShape(output), GetTensorData(output)); } TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) { @@ -199,7 +175,6 @@ TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) { NumDimensions(input)); return kTfLiteError; } - case kTfLiteInt8: case kTfLiteUInt8: { if (NumDimensions(input) == 1) { Softmax1DQuantized(input, output, params, data); @@ -220,8 +195,7 @@ TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) { } default: context->ReportError( - context, - "Only float32, uint8_t and int8_t supported currently, got %d.", + context, "Only float32 and uint8_t supported currently, got %d.", input->type); return kTfLiteError; } @@ -229,14 +203,14 @@ TfLiteStatus SoftmaxEval(TfLiteContext* context, TfLiteNode* node) { } // namespace activations TfLiteRegistration* Register_SOFTMAX() { - static TfLiteRegistration r = {}; - r.init = activations::Init; - r.free = activations::Free; - r.prepare = activations::SoftmaxPrepare; - r.invoke = activations::SoftmaxEval; + static TfLiteRegistration r = {activations::Init, activations::Free, + activations::SoftmaxPrepare, + activations::SoftmaxEval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/split.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/split.cpp similarity index 90% rename from src/tensorflow/lite/micro/kernels/split.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/split.cpp index 3a1eb2e..a56abde 100644 --- a/src/tensorflow/lite/micro/kernels/split.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/split.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,10 +14,10 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -116,12 +117,12 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace split TfLiteRegistration* Register_SPLIT() { - static TfLiteRegistration r = {}; - r.prepare = split::Prepare; - r.invoke = split::Eval; + static TfLiteRegistration r = {nullptr, nullptr, split::Prepare, split::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/strided_slice.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/strided_slice.cpp similarity index 91% rename from src/tensorflow/lite/micro/kernels/strided_slice.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/strided_slice.cpp index 53a5107..99f99ef 100644 --- a/src/tensorflow/lite/micro/kernels/strided_slice.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/strided_slice.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,15 +13,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/internal/reference/strided_slice.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/reference/strided_slice.h" #include -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -169,12 +170,14 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace strided_slice TfLiteRegistration* Register_STRIDED_SLICE() { - static TfLiteRegistration r = {}; - r.prepare = strided_slice::Prepare; - r.invoke = strided_slice::Eval; + static TfLiteRegistration r = { + nullptr, nullptr, strided_slice::Prepare, + strided_slice::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/svdf.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/svdf.cpp similarity index 52% rename from src/tensorflow/lite/micro/kernels/svdf.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/svdf.cpp index 2f9883c..5e4b2f1 100644 --- a/src/tensorflow/lite/micro/kernels/svdf.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/svdf.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,15 +16,15 @@ limitations under the License. #include -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" -#include "tensorflow/lite/kernels/op_macros.h" -#include "tensorflow/lite/micro/kernels/activation_utils.h" -#include "tensorflow/lite/micro/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/activation_utils.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace ops { @@ -31,19 +32,6 @@ namespace micro { namespace svdf { namespace { -// These constants represent constants specific to the hotword "OK G" model. -// They exist until (b/132070898) is fixed. -constexpr int kScratchTensorMaxSize = 64; - -struct OpData { - int32 effective_scale_1_a; - int32 effective_scale_2_a; - // b versions of each scale are kept at int since the numbers are just the - // shift value - typically between [-32, 32]. - int effective_scale_1_b; - int effective_scale_2_b; -}; - /** * This version of SVDF is specific to TFLite Micro. It contains the following * differences between the TFLite version: @@ -55,6 +43,9 @@ struct OpData { * resizing. */ +// TODO(kreeger): upstream these reference methods into +// `lite/kernels/reference/svdf.h` + static inline void ApplyTimeWeightsBiasAndActivation( int batch_size, int memory_size, int num_filters, int num_units, int rank, const TfLiteTensor* weights_time, const TfLiteTensor* bias, @@ -82,7 +73,7 @@ static inline void ApplyTimeWeightsBiasAndActivation( // Initialize output with bias if provided. if (bias) { - // VectorBatchVectorAssign + // TODO(kreeger): doc me - VectorBatchVectorAssign const float* bias_data = GetTensorData(bias); float* output_data = GetTensorData(output); for (int i = 0; i < batch_size; ++i) { @@ -105,9 +96,10 @@ static inline void ApplyTimeWeightsBiasAndActivation( float* scratch_ptr_batch = GetTensorData(scratch) + b * num_filters; // Reduction sum vector + const float* input_vector_ptr = scratch_ptr_batch; for (int i = 0; i < num_units; ++i) { for (int j = 0; j < rank; j++) { - output_ptr_batch[i] += *scratch_ptr_batch++; + output_ptr_batch[i] += *input_vector_ptr++; } } } @@ -196,149 +188,91 @@ inline void EvalFloatSVDF(TfLiteContext* context, TfLiteNode* node, params->activation, activation_state, scratch, output); } -void EvalIntegerSVDF( - TfLiteContext* context, TfLiteNode* node, const TfLiteTensor* input_tensor, - const TfLiteTensor* weights_feature_tensor, - const TfLiteTensor* weights_time_tensor, const TfLiteTensor* bias_tensor, - const TfLiteSVDFParams* params, TfLiteTensor* activation_state_tensor, - TfLiteTensor* output_tensor, int32_t scale_1_a, int scale_1_b, - int32_t scale_2_a, int scale_2_b, int32_t input_zp, int32_t output_zp) { - const int n_rank = params->rank; - const int n_batch = input_tensor->dims->data[0]; - const int n_input = input_tensor->dims->data[1]; - const int n_filter = weights_feature_tensor->dims->data[0]; - const int n_unit = n_filter / n_rank; - const int n_memory = weights_time_tensor->dims->data[1]; - - // TODO(b/132070898): Move these temp variables to the new scratch buffer API - // when ready. - int32_t scratch_tensor[kScratchTensorMaxSize]; - int32_t scratch_output_tensor[kScratchTensorMaxSize]; - - // Rewrite last bit of state. - { - for (int b = 0; b < n_batch; ++b) { - int16_t* state_ptr_batch = - GetTensorData(activation_state_tensor) + - b * n_memory * n_filter; - for (int c = 0; c < n_filter; ++c) { - int16_t* state_ptr = state_ptr_batch + c * n_memory; - state_ptr[n_memory - 1] = 0; - } - } - } +inline void EvalHybridSVDF( + TfLiteContext* context, TfLiteNode* node, const TfLiteTensor* input, + const TfLiteTensor* weights_feature, const TfLiteTensor* weights_time, + const TfLiteTensor* bias, const TfLiteSVDFParams* params, + TfLiteTensor* scratch, TfLiteTensor* scaling_factors, + TfLiteTensor* input_quantized, TfLiteTensor* activation_state, + TfLiteTensor* output) { + const int rank = params->rank; + const int batch_size = input->dims->data[0]; + const int input_size = input->dims->data[1]; + const int num_filters = weights_feature->dims->data[0]; + const int num_units = num_filters / rank; + const int memory_size = weights_time->dims->data[1]; - // Feature matmul. - { - int16_t* state = GetTensorData(activation_state_tensor); - const int8_t* input = GetTensorData(input_tensor); - const int8_t* weight_feature = - GetTensorData(weights_feature_tensor); - const int32_t output_max = std::numeric_limits::max(); - const int32_t output_min = std::numeric_limits::min(); - int16_t* result_in_batch = state + (n_memory - 1); - for (int b = 0; b < n_batch; b++) { - const int8_t* matrix_ptr = weight_feature; - for (int r = 0; r < n_filter; r++) { - int32_t dot_prod = 0; - const int8_t* vector_in_batch = input + b * n_input; - for (int c = 0; c < n_input; c++) { - dot_prod += *matrix_ptr++ * (*vector_in_batch++ - input_zp); - } - dot_prod = - MultiplyByQuantizedMultiplier(dot_prod, scale_1_a, scale_1_b); - dot_prod = std::min(std::max(output_min, dot_prod), output_max); - *result_in_batch = dot_prod; - result_in_batch += n_memory; - } - } - } + // Initialize the pointer to input. + const float* input_ptr_batch = GetTensorData(input); - // Time. - { - for (int b = 0; b < n_batch; ++b) { - int32_t* scratch_ptr_batch = scratch_tensor + b * n_filter; - - // Perform batched vector dot product: - const int16_t* vector1_ptr = GetTensorData(weights_time_tensor); - const int16_t* vector2_ptr = - GetTensorData(activation_state_tensor) + - b * n_memory * n_filter; - - for (int i = 0; i < n_filter; i++) { - *scratch_ptr_batch = 0; - for (int j = 0; j < n_memory; j++) { - *scratch_ptr_batch += *vector1_ptr++ * *vector2_ptr++; - } - scratch_ptr_batch++; - } - } - } + int8_t* quantized_input_ptr_batch = GetTensorData(input_quantized); + const int8_t* weights_feature_ptr = GetTensorData(weights_feature); - // Reduce, add bias, rescale, activation. - { - // Add bias. - if (bias_tensor) { - // Vector batch assign: - const int32_t* bias_data = GetTensorData(bias_tensor); - for (int i = 0; i < n_batch; ++i) { - int32_t* output_ptr = scratch_output_tensor + i * n_unit; - const int32_t* bias_ptr = bias_data; - for (int j = 0; j < n_unit; ++j) { - *output_ptr++ = *bias_ptr++; - } - } - } else { - int32_t* output_ptr = scratch_output_tensor; - for (int i = 0; i < n_batch * n_unit; ++i) { - *output_ptr++ = 0; - } - } + // Initialize the pointer to storage for scaling factors. + float* scaling_factors_ptr = GetTensorData(scaling_factors); - // Reduce. - for (int b = 0; b < n_batch; ++b) { - int32_t* output_temp_ptr = scratch_output_tensor + b * n_unit; - int32_t* scratch_ptr_batch = scratch_tensor + b * n_filter; + // Initialize the weights scale. + const float weights_feature_scale = weights_feature->params.scale; - // Reduction sum vector - for (int i = 0; i < n_unit; ++i) { - for (int j = 0; j < n_rank; ++j) { - output_temp_ptr[i] += *scratch_ptr_batch++; - } - } + // Clear the activation (activation_state's leftmost column). + // TODO(ghodrat): Add a test which initialize activation_state with invalid + // values in the leftmost column and make sure it passes. + // TODO(kreeger): Use a port of tensor_utils when ready (b/140272187). + for (int b = 0; b < batch_size; ++b) { + float* state_ptr_batch = + GetTensorData(activation_state) + b * memory_size * num_filters; + for (int c = 0; c < num_filters; ++c) { + float* state_ptr = state_ptr_batch + c * memory_size; + state_ptr[memory_size - 1] = 0.0; } + } - // Rescale. - const int32_t output_max = std::numeric_limits::max(); - const int32_t output_min = std::numeric_limits::min(); - for (int i = 0; i < n_batch * n_unit; ++i) { - int32_t x1 = scratch_output_tensor[i]; - int32_t x2 = MultiplyByQuantizedMultiplier(x1, scale_2_a, scale_2_b); - int32_t x3 = x2 + output_zp; - int32_t x4 = std::min(std::max(output_min, x3), output_max); - GetTensorData(output_tensor)[i] = static_cast(x4); + // Determine if input pointer batch is a zero based vector: + bool is_zero_vector = true; + for (int i = 0; i < batch_size * input_size && is_zero_vector; ++i) { + if (input_ptr_batch[i] != 0.0f) { + is_zero_vector = false; } } - // Shift state. - { - for (int b = 0; b < n_batch; ++b) { - int16_t* state_ptr_batch = - GetTensorData(activation_state_tensor) + - b * n_memory * n_filter; - for (int f = 0; f < n_filter; ++f) { - // Shift the vector left: - int16_t* batch_ptr = state_ptr_batch; - int16_t* batch_start = state_ptr_batch + 1; - int16_t* batch_end = state_ptr_batch + n_memory; - while (batch_start != batch_end) { - *batch_ptr++ = *batch_start++; + if (!is_zero_vector) { + SignedSymmetricPerChannelQuantize(input_ptr_batch, input->dims, 0, + quantized_input_ptr_batch, + scaling_factors_ptr); + + // Quantize input from float to int8. + for (int b = 0; b < batch_size; ++b) { + scaling_factors_ptr[b] *= weights_feature_scale; + } + + // Compute conv1d(inputs, weights_feature). + // The rightmost column of activation_state is used to save the current + // cycle activation. This is achieved by starting at + // GetTensorData(activation_state)[memory_size - 1] and having the + // stride equal to memory_size. (Matrix batch vector multiply accumulate) + float* result = &GetTensorData(activation_state)[memory_size - 1]; + for (int i = 0; i < batch_size; + ++i, quantized_input_ptr_batch += input_size) { + const float batch_scaling_factor = scaling_factors_ptr[i]; + + // Get the address of the first row: + const int8_t* row_ptr = weights_feature_ptr; + for (int j = 0; j < num_filters; ++j, result += memory_size) { + // Initialize the dot product sum for the row to 0. + int32_t dotprod = 0; + for (int k = 0; k < input_size; ++k, ++row_ptr) { + dotprod += (*row_ptr) * (quantized_input_ptr_batch[k]); } - state_ptr_batch[n_memory - 1] = 0; - state_ptr_batch += n_memory; + *result += dotprod * batch_scaling_factor; } } } + + // TODO(alanchiao): can optimize hybrid case ~5% by unrolling loop in applying + // time weights so that the inner loop multiplies eight elements at a time. + ApplyTimeWeightsBiasAndActivation( + batch_size, memory_size, num_filters, num_units, rank, weights_time, bias, + params->activation, activation_state, scratch, output); } } // namespace @@ -370,7 +304,10 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { // [3] = Bias (optional), {1, num_units} // [4] = Activation State (variable), // {2, batch_size, memory_size * num_filters} - + // TODO(kreeger): Use input tensor as variable until scratch tensor allocation + // has been implemented (cl/263032056) + // TF_LITE_ENSURE_EQ(context, node->inputs->size, 5); + TF_LITE_ENSURE_EQ(context, node->inputs->size, 6); const TfLiteTensor* input = GetInput(context, node, kInputTensor); const TfLiteTensor* weights_feature = GetInput(context, node, kWeightsFeatureTensor); @@ -389,21 +326,10 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { const int num_units = num_filters / rank; const int memory_size = weights_time->dims->data[1]; - const bool is_full_integer = input->type == kTfLiteInt8; - // Validate Input Tensor: - TF_LITE_ENSURE(context, - input->type == kTfLiteFloat32 || input->type == kTfLiteInt8); + TF_LITE_ENSURE_EQ(context, input->type, kTfLiteFloat32); TF_LITE_ENSURE_EQ(context, NumDimensions(input), 2); - // Validate Tensor Output: - // [0] = float/int8, {2, batch_size, num_units} - TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - TF_LITE_ENSURE_EQ(context, NumDimensions(output), 2); - TF_LITE_ENSURE_EQ(context, output->dims->data[0], batch_size); - TF_LITE_ENSURE_EQ(context, output->dims->data[1], num_units); - // Validate Weights Feature Input Tensor: TF_LITE_ENSURE_EQ(context, NumDimensions(weights_feature), 2); TF_LITE_ENSURE_EQ(context, weights_feature->dims->data[1], input_size); @@ -416,73 +342,108 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { // Validate Optional Bias Input Tensor: if (bias) { TF_LITE_ENSURE_EQ(context, bias->dims->data[0], num_units); + TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteFloat32); } // Validate Activation State Input Tensor: + TF_LITE_ENSURE_EQ(context, activation_state->type, kTfLiteFloat32); TF_LITE_ENSURE_EQ(context, NumDimensions(activation_state), 2); TF_LITE_ENSURE_EQ(context, activation_state->dims->data[0], batch_size); TF_LITE_ENSURE_EQ(context, activation_state->dims->data[1], memory_size * num_filters); - if (is_full_integer) { - TF_LITE_ENSURE_EQ(context, node->inputs->size, 5); - - TF_LITE_ENSURE_EQ(context, weights_feature->type, kTfLiteInt8); - TF_LITE_ENSURE_EQ(context, weights_time->type, kTfLiteInt16); - - if (bias) { - TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteInt32); - } - - TF_LITE_ENSURE_EQ(context, activation_state->type, kTfLiteInt16); + // Validate shared Scratch Tensor (same for full float and hybrid): + // [0] = Holds dot-product of time-forward calculations in + // ApplyTimeWeightsBiasAndActivation(): + // float, {2, batch_size, num_filters} + // TODO(kreeger): Use input tensor as variable until scratch tensor allocation + // has been implemented (cl/263032056) + // TfLiteTensor* scratch_tensor = GetTemporary(context, node, 0); + TfLiteTensor* scratch_tensor = &context->tensors[node->inputs->data[5]]; + + TF_LITE_ENSURE_EQ(context, scratch_tensor->type, kTfLiteFloat32); + TF_LITE_ENSURE_EQ(context, NumDimensions(scratch_tensor), 2); + TF_LITE_ENSURE_EQ(context, scratch_tensor->dims->data[0], batch_size); + TF_LITE_ENSURE_EQ(context, scratch_tensor->dims->data[1], num_filters); + + // The weights are of consistent type, so it suffices to check one. + const bool is_hybrid_op = IsHybridOp(input, weights_feature); + // TODO(kreeger): Handle full quant svdf b/139435798 + if (is_hybrid_op) { + // Validate Input Tensor dtypes: + TF_LITE_ENSURE(context, weights_feature->type == kTfLiteUInt8 || + weights_feature->type == kTfLiteInt8); + TF_LITE_ENSURE(context, weights_time->type == kTfLiteUInt8 || + weights_time->type == kTfLiteInt8); // Validate Scratch Tensors: - // [0] = (shared - see float block below for usage) - // [1] = Output Temp, int8_t, {2, num_units, batch_size} - // TODO(b/132070898): Scratch values are used as stack variables in - // EvalIntegerSVDF(). - - // Validate output tensor: - TF_LITE_ENSURE_EQ(context, output->type, kTfLiteInt8); + // [0] = (shared - see above for usage) + // [1] = Input Quantized, int8_t/uint8_t, {2, batch_size, input_size} + // [2] = Scaling Factors, float, {1, batch_size} + // [3] = Float Weights Time, float, {2, num_filters, memory_size} + TF_LITE_ENSURE_EQ(context, node->temporaries->size, 4); + TfLiteTensor* scratch_input_quantized = GetTemporary(context, node, 1); + TfLiteTensor* scratch_scaling_factors = GetTemporary(context, node, 2); + TfLiteTensor* scratch_float_weights_time = GetTemporary(context, node, 3); + + // Validate Input Quantized Scratch Tensor: + TF_LITE_ENSURE(context, scratch_input_quantized->type == kTfLiteUInt8 || + scratch_input_quantized->type == kTfLiteInt8); + TF_LITE_ENSURE_EQ(context, scratch_input_quantized->dims->data[0], + batch_size); + TF_LITE_ENSURE_EQ(context, scratch_input_quantized->dims->data[1], + input_size); + + // Validate Scaling Factors Scratch Tensor: + TF_LITE_ENSURE_EQ(context, scratch_scaling_factors->type, kTfLiteFloat32); + TF_LITE_ENSURE_EQ(context, NumDimensions(scratch_scaling_factors), 1); + TF_LITE_ENSURE_EQ(context, scratch_scaling_factors->dims->data[0], + batch_size); + + // Validate Float Weights Time Scratch Tensor: + TF_LITE_ENSURE_EQ(context, scratch_float_weights_time->type, + kTfLiteFloat32); + TF_LITE_ENSURE_EQ(context, NumDimensions(scratch_float_weights_time), 2); + TF_LITE_ENSURE_EQ(context, scratch_float_weights_time->dims->data[0], + num_filters); + TF_LITE_ENSURE_EQ(context, scratch_float_weights_time->dims->data[1], + memory_size); + + // TfLite Micro has scratch tensors allocated at the time that Prepare() is + // called. Use this time to do a one-time de-quantization copy of + // the input values from the Weights Time tensor to the float weights time + // scratch tensor. + // TODO(kreeger): Consider doing this at model conversion time? + SymmetricDequantize(GetTensorData(weights_time), + NumElements(scratch_float_weights_time), + weights_time->params.scale, + GetTensorData(scratch_float_weights_time)); } else { - TF_LITE_ENSURE_EQ(context, node->inputs->size, 6); - // Validate Input Tensor dtypes: TF_LITE_ENSURE_EQ(context, weights_feature->type, kTfLiteFloat32); TF_LITE_ENSURE_EQ(context, weights_time->type, kTfLiteFloat32); - TF_LITE_ENSURE_EQ(context, activation_state->type, kTfLiteFloat32); - - if (bias) { - TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteFloat32); - } - - // Validate shared Scratch Tensor: - // [0] = Holds dot-product of time-forward calculations in - // ApplyTimeWeightsBiasAndActivation(): - // float/int32, {2, batch_size, num_filters} - // TODO(b/132070898): Use input tensor as variable until scratch tensor - // allocation has been implemented (b/132070898) TfLiteTensor* - // scratch_tensor = GetTemporary(context, node, 0); - TfLiteTensor* scratch_tensor = &context->tensors[node->inputs->data[5]]; - TF_LITE_ENSURE_EQ(context, scratch_tensor->type, kTfLiteFloat32); - - TF_LITE_ENSURE_EQ(context, NumDimensions(scratch_tensor), 2); - TF_LITE_ENSURE_EQ(context, scratch_tensor->dims->data[0], batch_size); - TF_LITE_ENSURE_EQ(context, scratch_tensor->dims->data[1], num_filters); // Full-float SVDF only uses the one shared scratch tensor (see above for // usage). - // TODO(b/132070898): Use input tensor as variable until scratch tensor - // allocation has been implemented. + // TODO(kreeger): Use input tensor as variable until scratch tensor + // allocation has been implemented (cl/263032056) // TF_LITE_ENSURE_EQ(context, node->temporaries->size, 1); - TF_LITE_ENSURE_EQ(context, output->type, kTfLiteFloat32); } + // Validate Tensor Output: + // [0] = float, {2, batch_size, num_units} + TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + TF_LITE_ENSURE_EQ(context, output->type, kTfLiteFloat32); + TF_LITE_ENSURE_EQ(context, NumDimensions(output), 2); + TF_LITE_ENSURE_EQ(context, output->dims->data[0], batch_size); + TF_LITE_ENSURE_EQ(context, output->dims->data[1], num_units); + return kTfLiteOk; } TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { - auto* params = reinterpret_cast(node->builtin_data); + const auto* params = reinterpret_cast(node->builtin_data); const TfLiteTensor* input = GetInput(context, node, kInputTensor); const TfLiteTensor* weights_feature = @@ -490,66 +451,39 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { const TfLiteTensor* weights_time = GetInput(context, node, kWeightsTimeTensor); const TfLiteTensor* bias = GetOptionalInputTensor(context, node, kBiasTensor); + + // TODO(kreeger): Use input tensor as variable until scratch tensor allocation + // has been implemented (cl/263032056) + // TfLiteTensor* scratch = GetTemporary(context, node, /*index=*/0); + TfLiteTensor* scratch = &context->tensors[node->inputs->data[5]]; + TfLiteTensor* activation_state = &context->tensors[node->inputs->data[kInputActivationStateTensor]]; TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - const bool is_full_integer = input->type == kTfLiteInt8; - switch (weights_feature->type) { case kTfLiteFloat32: { - // TODO(b/132070898): Use input tensor as variable until scratch tensor - // allocation has been implemented. TfLiteTensor* scratch = - // GetTemporary(context, node, /*index=*/0); - TfLiteTensor* scratch = &context->tensors[node->inputs->data[5]]; EvalFloatSVDF(context, node, input, weights_feature, weights_time, bias, params, scratch, activation_state, output); return kTfLiteOk; break; } + case kTfLiteUInt8: case kTfLiteInt8: { - if (is_full_integer) { - // TODO(b/132070898): Store these values in ::Prepare() instead of - // ::Eval(): - // Calculate effective scales. - OpData op_data; - auto* input_params = reinterpret_cast( - input->quantization.params); - auto* weights_feature_params = - reinterpret_cast( - weights_feature->quantization.params); - auto* state_params = reinterpret_cast( - activation_state->quantization.params); - auto* weight_time_params = reinterpret_cast( - weights_time->quantization.params); - auto* output_params = reinterpret_cast( - output->quantization.params); - const double effective_scale_1 = - static_cast(input_params->scale->data[0] * - weights_feature_params->scale->data[0] / - state_params->scale->data[0]); - const double effective_scale_2 = static_cast( - state_params->scale->data[0] * weight_time_params->scale->data[0] / - output_params->scale->data[0]); - QuantizeMultiplier(effective_scale_1, &op_data.effective_scale_1_a, - &op_data.effective_scale_1_b); - QuantizeMultiplier(effective_scale_2, &op_data.effective_scale_2_a, - &op_data.effective_scale_2_b); - - TF_LITE_ENSURE_EQ(context, params->activation, kTfLiteActRelu); - EvalIntegerSVDF( - context, node, input, weights_feature, weights_time, bias, params, - activation_state, output, op_data.effective_scale_1_a, - op_data.effective_scale_1_b, op_data.effective_scale_2_a, - op_data.effective_scale_2_b, input->params.zero_point, - output->params.zero_point); - return kTfLiteOk; - } + TfLiteTensor* scratch_input_quantized = GetTemporary(context, node, 1); + TfLiteTensor* scratch_scaling_factors = GetTemporary(context, node, 2); + TfLiteTensor* scratch_float_weights_time = GetTemporary(context, node, 3); + EvalHybridSVDF(context, node, input, weights_feature, + scratch_float_weights_time, bias, params, scratch, + scratch_scaling_factors, scratch_input_quantized, + activation_state, output); + return kTfLiteOk; break; } default: + // TODO(kreeger): Handle this case for full quant svdf b/139435798 context->ReportError(context, "Type %s not currently supported.", TfLiteTypeGetName(weights_feature->type)); return kTfLiteError; @@ -560,14 +494,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace svdf TfLiteRegistration* Register_SVDF() { - static TfLiteRegistration r = {}; - r.init = svdf::Init; - r.free = svdf::Free; - r.prepare = svdf::Prepare; - r.invoke = svdf::Eval; + static TfLiteRegistration r = {svdf::Init, svdf::Free, svdf::Prepare, + svdf::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/kernels/unpack.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/unpack.cpp similarity index 88% rename from src/tensorflow/lite/micro/kernels/unpack.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/unpack.cpp index 207570c..8f51d46 100644 --- a/src/tensorflow/lite/micro/kernels/unpack.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/kernels/unpack.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,10 +14,10 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" namespace tflite { namespace ops { @@ -106,12 +107,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { } // namespace unpack TfLiteRegistration* Register_UNPACK() { - static TfLiteRegistration r = {}; - r.prepare = unpack::Prepare; - r.invoke = unpack::Eval; + static TfLiteRegistration r = {nullptr, nullptr, unpack::Prepare, + unpack::Eval}; return &r; } } // namespace micro } // namespace ops } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/memory_helpers.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.cpp similarity index 93% rename from src/tensorflow/lite/micro/memory_helpers.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.cpp index 302f160..fce5af0 100644 --- a/src/tensorflow/lite/micro/memory_helpers.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,11 +14,11 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/memory_helpers.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.h" #include -#include "tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h" namespace tflite { @@ -93,3 +94,5 @@ TfLiteStatus BytesRequiredForTensor(const tflite::Tensor& flatbuffer_tensor, } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/memory_helpers.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.h similarity index 77% rename from src/tensorflow/lite/micro/memory_helpers.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.h index ef8205c..615e6ea 100644 --- a/src/tensorflow/lite/micro/memory_helpers.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,12 +13,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_MEMORY_HELPERS_H_ -#define TENSORFLOW_LITE_MICRO_MEMORY_HELPERS_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_HELPERS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_HELPERS_H_ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { @@ -41,4 +42,6 @@ TfLiteStatus BytesRequiredForTensor(const tflite::Tensor& flatbuffer_tensor, } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_MEMORY_HELPERS_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_HELPERS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.cpp similarity index 97% rename from src/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.cpp index d066ba8..764ce05 100644 --- a/src/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,7 +14,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/memory_planner/greedy_memory_planner.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.h" namespace tflite { @@ -240,19 +241,17 @@ void GreedyMemoryPlanner::CalculateOffsetsIfNeeded() { } } -size_t GreedyMemoryPlanner::GetMaximumMemorySize() { +int GreedyMemoryPlanner::GetMaximumMemorySize() { CalculateOffsetsIfNeeded(); if (buffer_count_ == 0) { return 0; } ListEntry* entry = &buffers_sorted_by_offset_[0]; - size_t max_size = 0; + int max_size = 0; while (entry) { BufferRequirements* requirements = &requirements_[entry->requirements_index]; - // TODO(b/148246793): Update all size and offset variables types from - // int to size_t - const size_t current_size = entry->offset + requirements->size; + const int current_size = entry->offset + requirements->size; if (current_size > max_size) { max_size = current_size; } @@ -386,3 +385,5 @@ bool GreedyMemoryPlanner::DoAnyBuffersOverlap(ErrorReporter* error_reporter) { } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.h similarity index 92% rename from src/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.h index f2c77ed..646d0e5 100644 --- a/src/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,11 +14,10 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ -#define TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ -#include "tensorflow/lite/micro/compatibility.h" -#include "tensorflow/lite/micro/memory_planner/memory_planner.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/memory_planner.h" namespace tflite { @@ -61,7 +61,7 @@ class GreedyMemoryPlanner : public MemoryPlanner { // Returns the high-water mark of used memory. This is the minimum size of a // memory arena you'd need to allocate to hold these buffers. - size_t GetMaximumMemorySize() override; + int GetMaximumMemorySize() override; // How many buffers have been recorded. int GetBufferCount() override; @@ -126,10 +126,10 @@ class GreedyMemoryPlanner : public MemoryPlanner { // Whether buffers have been added since the last plan was calculated. bool need_to_calculate_offsets_; - - TF_LITE_REMOVE_VIRTUAL_DELETE }; } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_GREEDY_MEMORY_PLANNER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/memory_planner/linear_memory_planner.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.cpp similarity index 88% rename from src/tensorflow/lite/micro/memory_planner/linear_memory_planner.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.cpp index 2e2e19a..65b7d03 100644 --- a/src/tensorflow/lite/micro/memory_planner/linear_memory_planner.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,7 +14,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/memory_planner/linear_memory_planner.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.h" namespace tflite { @@ -34,7 +35,7 @@ TfLiteStatus LinearMemoryPlanner::AddBuffer( return kTfLiteOk; } -size_t LinearMemoryPlanner::GetMaximumMemorySize() { return next_free_offset_; } +int LinearMemoryPlanner::GetMaximumMemorySize() { return next_free_offset_; } int LinearMemoryPlanner::GetBufferCount() { return current_buffer_count_; } @@ -50,3 +51,5 @@ TfLiteStatus LinearMemoryPlanner::GetOffsetForBuffer( } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.h new file mode 100644 index 0000000..df72e6e --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/linear_memory_planner.h @@ -0,0 +1,50 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_LINEAR_MEMORY_PLANNER_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_LINEAR_MEMORY_PLANNER_H_ + +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/memory_planner.h" + +namespace tflite { + +// The simplest possible memory planner that just lays out all buffers at +// increasing offsets without trying to reuse memory. +class LinearMemoryPlanner : public MemoryPlanner { + public: + LinearMemoryPlanner(); + ~LinearMemoryPlanner() override; + + TfLiteStatus AddBuffer(tflite::ErrorReporter* error_reporter, int size, + int first_time_used, int last_time_used) override; + + int GetMaximumMemorySize() override; + int GetBufferCount() override; + TfLiteStatus GetOffsetForBuffer(tflite::ErrorReporter* error_reporter, + int buffer_index, int* offset) override; + + private: + static constexpr int kMaxBufferCount = 1024; + int buffer_offsets_[kMaxBufferCount]; + int current_buffer_count_; + int next_free_offset_; +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_LINEAR_MEMORY_PLANNER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/memory_planner.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/memory_planner.h new file mode 100644 index 0000000..a97e835 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/memory_planner.h @@ -0,0 +1,74 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_MEMORY_PLANNER_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_MEMORY_PLANNER_H_ + +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" + +namespace tflite { + +// Interface class for planning the layout of memory buffers during the +// execution of a graph. +// It's designed to be used by a client that iterates in any order through the +// buffers it wants to lay out, and then calls the getter functions for +// information about the calculated layout. For example: +// +// SomeMemoryPlanner planner; +// planner.AddBuffer(reporter, 100, 0, 1); // Buffer 0 +// planner.AddBuffer(reporter, 50, 2, 3); // Buffer 1 +// planner.AddBuffer(reporter, 50, 2, 3); // Buffer 2 +// +// int offset0; +// TF_EXPECT_OK(planner.GetOffsetForBuffer(reporter, 0, &offset0)); +// int offset1; +// TF_EXPECT_OK(planner.GetOffsetForBuffer(reporter, 1, &offset1)); +// int offset2; +// TF_EXPECT_OK(planner.GetOffsetForBuffer(reporter, 2, &offset2)); +// const int arena_size_needed = planner.GetMaximumMemorySize(); +// +// The goal is for applications to be able to experiment with different layout +// strategies without changing their client code, by swapping out classes that +// implement this interface.= +class MemoryPlanner { + public: + MemoryPlanner() {} + virtual ~MemoryPlanner() {} + + // Pass information about a buffer's size and lifetime to the layout + // algorithm. The order this is called implicitly assigns an index to the + // result, so the buffer information that's passed into the N-th call of + // this method will be used as the buffer_index argument to + // GetOffsetForBuffer(). + virtual TfLiteStatus AddBuffer(tflite::ErrorReporter* error_reporter, + int size, int first_time_used, + int last_time_used) = 0; + + // The largest contguous block of memory that's needed to hold the layout. + virtual int GetMaximumMemorySize() = 0; + // How many buffers have been added to the planner. + virtual int GetBufferCount() = 0; + // Calculated layout offset for the N-th buffer added to the planner. + virtual TfLiteStatus GetOffsetForBuffer(tflite::ErrorReporter* error_reporter, + int buffer_index, int* offset) = 0; +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MEMORY_PLANNER_MEMORY_PLANNER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.cpp new file mode 100644 index 0000000..ea8929e --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.cpp @@ -0,0 +1,504 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.h" + +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_planner/greedy_memory_planner.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.h" + +namespace tflite { + +namespace { +// Used to hold information used during allocation calculations. +struct TensorInfo { + const tflite::Tensor* flatbuffer_tensor; + TfLiteTensor* runtime_tensor; + int first_created; + int last_used; + bool needs_allocating; +}; + +// We align tensor buffers to 16-byte boundaries, since this is a common +// requirement for SIMD extensions. +constexpr int kBufferAlignment = 16; +// For common data structures that doesn't need SIMD extensions. +constexpr int kDefaultAlignment = sizeof(int); + +class MicroBuiltinDataAllocator : public BuiltinDataAllocator { + public: + explicit MicroBuiltinDataAllocator(SimpleMemoryAllocator* memory_allocator) + : memory_allocator_(memory_allocator) {} + + void* Allocate(size_t size) override { + return memory_allocator_->AllocateFromTail(size, kDefaultAlignment); + } + void Deallocate(void* data) override { + // Do not deallocate, builtin data needs to be available for the life time + // of the model. + } + + private: + SimpleMemoryAllocator* memory_allocator_; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace + +MicroAllocator::MicroAllocator(TfLiteContext* context, const Model* model, + uint8_t* tensor_arena, size_t arena_size, + ErrorReporter* error_reporter) + : model_(model), + memory_allocator_(tensor_arena, arena_size), + error_reporter_(error_reporter), + context_(context), + arena_(tensor_arena), + arena_size_(arena_size) { + auto* subgraphs = model->subgraphs(); + if (subgraphs->size() != 1) { + error_reporter->Report("Only 1 subgraph is currently supported.\n"); + return; + } + subgraph_ = (*subgraphs)[0]; + tensors_ = subgraph_->tensors(); + operators_ = subgraph_->operators(); + + context_->tensors_size = tensors_->size(); + context_->tensors = + reinterpret_cast(memory_allocator_.AllocateFromTail( + sizeof(TfLiteTensor) * context_->tensors_size, kDefaultAlignment)); + + // Null all inputs so we can later perform a null check to avoid re-allocating + // registered pre-allocated inputs. + for (size_t i = 0; i < context_->tensors_size; ++i) { + context_->tensors[i].data.raw = nullptr; + } + + active_ = true; +} + +TfLiteStatus MicroAllocator::RegisterPreallocatedInput(uint8_t* buffer, + size_t input_index) { + if (buffer == nullptr || input_index < 0 || + input_index >= subgraph_->inputs()->size()) { + error_reporter_->Report("Invalid pre-allocated input %d provided.", + input_index); + return kTfLiteError; + } + const flatbuffers::Vector>* buffers = + model_->buffers(); + + const int tensor_index = subgraph_->inputs()->Get(input_index); + const auto* tensor = tensors_->Get(tensor_index); + return InitializeRuntimeTensor(*tensor, buffers, error_reporter_, + &context_->tensors[tensor_index], buffer); +} + +TfLiteStatus MicroAllocator::AllocateNodeAndRegistrations( + const OpResolver& op_resolver, + NodeAndRegistration** node_and_registrations) { + if (!active_) { + return kTfLiteError; + } + + auto* output = + reinterpret_cast(memory_allocator_.AllocateFromTail( + sizeof(NodeAndRegistration) * operators_->size(), kDefaultAlignment)); + if (output == nullptr) { + error_reporter_->Report( + "Failed to allocate memory for node_and_registrations."); + return kTfLiteError; + } + TfLiteStatus status = kTfLiteOk; + auto* opcodes = model_->operator_codes(); + MicroBuiltinDataAllocator builtin_data_allocator(&memory_allocator_); + for (size_t i = 0; i < operators_->size(); ++i) { + const auto* op = operators_->Get(i); + size_t index = op->opcode_index(); + if (index < 0 || index >= opcodes->size()) { + error_reporter_->Report("Missing registration for opcode_index %d\n", + index); + return kTfLiteError; + } + auto* opcode = (*opcodes)[index]; + status = GetRegistrationFromOpCode(opcode, op_resolver, error_reporter_, + &(output[i].registration)); + if (status != kTfLiteOk) { + error_reporter_->Report("Failed to get registration from op code % d\n ", + opcode); + return status; + } + const auto* registration = output[i].registration; + if (registration == nullptr) { + error_reporter_->Report("Skipping op for opcode_index %d\n", index); + return kTfLiteError; + } + BuiltinOperator op_type = + static_cast(registration->builtin_code); + + if (op_type != BuiltinOperator_CUSTOM && op->custom_options()) { + error_reporter_->Report( + "Unsupported behavior: found builtin operator %s with custom " + "options.\n", + EnumNameBuiltinOperator(op_type)); + return kTfLiteError; + } + + const char* custom_data = nullptr; + size_t custom_data_size = 0; + unsigned char* builtin_data = nullptr; + if (op->custom_options()) { + custom_data = reinterpret_cast(op->custom_options()->data()); + custom_data_size = op->custom_options()->size(); + } else { + TF_LITE_ENSURE_STATUS(ParseOpData(op, op_type, error_reporter_, + &builtin_data_allocator, + (void**)(&builtin_data))); + } + + // Disregard const qualifier to workaround with existing API. + TfLiteIntArray* inputs_array = const_cast( + reinterpret_cast(op->inputs())); + TfLiteIntArray* outputs_array = const_cast( + reinterpret_cast(op->outputs())); + + TfLiteNode* node = &(output[i].node); + node->inputs = inputs_array; + node->outputs = outputs_array; + // This is OK for now as temporary array is not in used. + // TODO(wangtz): Support scratch buffers. + node->temporaries = nullptr; + node->user_data = nullptr; // Will be filled in after `init` + node->builtin_data = reinterpret_cast(builtin_data); + node->custom_initial_data = custom_data; + node->custom_initial_data_size = custom_data_size; + node->delegate = nullptr; + } + *node_and_registrations = output; + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::FinishTensorAllocation() { + if (!active_) { + return kTfLiteError; + } + + const size_t tensors_size = tensors_->size(); + + const flatbuffers::Vector>* buffers = + model_->buffers(); + + // Initialize runtime tensors. + for (size_t i = 0; i < tensors_size; ++i) { + auto* runtime_tensor = &context_->tensors[i]; + auto* flatbuffer_tensor = tensors_->Get(i); + + // Preallocated inputs have already been set up earlier, so skip them. + const bool is_preallocated_input = (runtime_tensor->data.raw != nullptr); + if (!is_preallocated_input) { + TF_LITE_ENSURE_STATUS(InitializeRuntimeTensor(*flatbuffer_tensor, buffers, + error_reporter_, + runtime_tensor, nullptr)); + } + } + + // tensor_info is only used in this function. + auto tmp_allocator = memory_allocator_.CreateChildAllocator(); + TensorInfo* tensor_info = + reinterpret_cast(tmp_allocator.AllocateFromTail( + sizeof(TensorInfo) * tensors_size, sizeof(TensorInfo))); + + // Set up the runtime data structures for all tensors. + for (size_t i = 0; i < tensors_size; ++i) { + TensorInfo* current = &tensor_info[i]; + current->flatbuffer_tensor = &(*(tensors_->Get(i))); + current->runtime_tensor = &context_->tensors[i]; + const bool is_variable = current->flatbuffer_tensor->is_variable(); + if (is_variable) { + current->first_created = 0; + current->last_used = operators_->size(); + } else { + current->first_created = -1; + current->last_used = -1; + } + current->needs_allocating = false; + } + + // First go through the inputs and figure out if they need to be allocated. + for (size_t i = 0; i < subgraph_->inputs()->size(); ++i) { + const int tensor_index = subgraph_->inputs()->Get(i); + TensorInfo* current = &tensor_info[tensor_index]; + // Check for pre-allocated inputs. + current->needs_allocating = (current->runtime_tensor->data.raw == nullptr); + current->first_created = 0; + } + + // Mark all outputs as persistent to the end of the invocation. + for (size_t i = 0; i < subgraph_->outputs()->size(); ++i) { + const int tensor_index = subgraph_->outputs()->Get(i); + TensorInfo* current = &tensor_info[tensor_index]; + current->last_used = operators_->size() - 1; + } + + // Figure out when the first and last use of each tensor is. + for (int i = (operators_->size() - 1); i >= 0; --i) { + const auto* op = operators_->Get(i); + for (size_t n = 0; n < op->inputs()->size(); ++n) { + const int tensor_index = op->inputs()->Get(n); + TensorInfo* current = &tensor_info[tensor_index]; + if ((current->last_used == -1) || (current->last_used > i)) { + current->last_used = i; + } + } + for (size_t n = 0; n < op->outputs()->size(); ++n) { + const int tensor_index = op->outputs()->Get(n); + TensorInfo* current = &tensor_info[tensor_index]; + if ((current->first_created == -1) || (current->first_created < i)) { + current->first_created = i; + } + } + } + + // Work out which tensors need to be allocated. + for (size_t i = 0; i < tensors_->size(); ++i) { + TensorInfo* current = &tensor_info[i]; + const bool is_read_only = + (current->first_created == -1) && (current->last_used != -1); + const bool is_preallocated_input = + (current->runtime_tensor->data.raw != nullptr); + const bool has_partial_lifetime = + !is_read_only && + ((current->first_created == -1) || (current->last_used == -1)); + if (has_partial_lifetime) { + error_reporter_->Report( + "Logic error in memory planner, tensor %d has an invalid lifetime", + i); + return kTfLiteError; + } + if (!is_read_only && !is_preallocated_input) { + current->needs_allocating = true; + } + } + + uint8_t* aligned_arena = AlignPointerUp(arena_, kBufferAlignment); + const size_t alignment_loss = (aligned_arena - arena_); + + // Remaining arena size that memory planner can use for calculating offsets. + int remaining_arena_size = + arena_size_ - (tmp_allocator.GetDataSize() + alignment_loss); + GreedyMemoryPlanner planner(aligned_arena, remaining_arena_size); + + // Add the tensors to our allocation plan. + for (size_t i = 0; i < tensors_->size(); ++i) { + TensorInfo* current = &tensor_info[i]; + if (current->needs_allocating) { + size_t bytes_required; + size_t type_size; + TF_LITE_ENSURE_STATUS(BytesRequiredForTensor(*current->flatbuffer_tensor, + &bytes_required, &type_size, + error_reporter_)); + size_t aligned_bytes_required = + AlignSizeUp(bytes_required, kBufferAlignment); + TF_LITE_ENSURE_STATUS( + planner.AddBuffer(error_reporter_, aligned_bytes_required, + current->first_created, current->last_used)); + } + } + + // Actual size available for placing tensors. This includes memory held by the + // tensor info array, which will be released. + int actual_available_arena_size = + arena_size_ - (memory_allocator_.GetDataSize() + alignment_loss); + // Make sure we have enough room. + if (planner.GetMaximumMemorySize() > actual_available_arena_size) { + error_reporter_->Report( + "Arena size is too small for activation buffers. Needed %d but only %d " + "was available.", + planner.GetMaximumMemorySize(), remaining_arena_size); + return kTfLiteError; + } + + // Figure out the actual memory addresses for each buffer, based on the plan. + int planner_index = 0; + for (size_t i = 0; i < tensors_->size(); ++i) { + TensorInfo* current = &tensor_info[i]; + if (current->needs_allocating) { + int offset; + TF_LITE_ENSURE_STATUS( + planner.GetOffsetForBuffer(error_reporter_, planner_index, &offset)); + current->runtime_tensor->data.uint8 = aligned_arena + offset; + ++planner_index; + } + } + + // Copy default value for variable tensors. Note that this will overwrite + // the arena planner data so GetOffsetForBuffer will return wrong + // result. + for (size_t i = 0; i < tensors_->size(); ++i) { + TensorInfo* current = &tensor_info[i]; + // Set default value for variable tensors: + if (current->flatbuffer_tensor->is_variable()) { + if (current->runtime_tensor->data.uint8 == nullptr) { + error_reporter_->Report("Variable is not allocated"); + return kTfLiteError; + } + tflite::ResetVariableTensor(current->runtime_tensor); + } + } + + active_ = false; + return kTfLiteOk; +} + +TfLiteStatus MicroAllocator::InitializeRuntimeTensor( + const tflite::Tensor& flatbuffer_tensor, + const flatbuffers::Vector>* buffers, + ErrorReporter* error_reporter, TfLiteTensor* result, + uint8_t* preallocated_buffer) { + if (!active_) { + return kTfLiteError; + } + + // Make sure the serialized type is one we know how to deal with, and convert + // it from a flatbuffer enum into a constant used by the kernel C API. + TF_LITE_ENSURE_STATUS(ConvertTensorType(flatbuffer_tensor.type(), + &result->type, error_reporter)); + // Make sure we remember if the serialized tensor is designated as a variable. + result->is_variable = flatbuffer_tensor.is_variable(); + + // We need to figure out where the actual contents of this tensor are stored + // in memory. We'll check to see if there's a serialized buffer (pretty much + // the same as a constant op in TensorFlow) associated with this tensor first, + // and if there is update the runtime structure to point to its location in + // memory. + result->data.raw = nullptr; + result->bytes = 0; + // First see if there's any buffer information in the serialized tensor. + if (auto* buffer = (*buffers)[flatbuffer_tensor.buffer()]) { + // If we've found a buffer, does it have any data? + if (auto* array = buffer->data()) { + // If it has any data, is the data size larger than zero? + if (size_t array_size = array->size()) { + // We've found a buffer with valid data, so update the runtime tensor + // data structure to point to it. + result->data.raw = + const_cast(reinterpret_cast(array->data())); + // We set the data from a serialized buffer, so record tha. + result->allocation_type = kTfLiteMmapRo; + } + } + // TODO(petewarden): It's not clear in what circumstances we could have a + // buffer in the serialized tensor, but it doesn't have any data in it. Is + // that a validly-generated file, and if so what does it mean, or is it an + // error condition? It would be good to tighten up the specification to make + // it less ambiguous. + } + + // TODO(petewarden): Some of these paths aren't getting enough testing + // coverage, so we should figure out some tests that exercise them. + if (!result->data.raw) { + // The tensor contents haven't been set from a serialized buffer, so + // make a note that they will be allocated from memory. The actual + // allocation won't happen until later. + result->allocation_type = kTfLiteArenaRw; + if (preallocated_buffer != nullptr) { + // If the client is supplying memory for the contents of the tensor + // themselves, use it. + // TODO(petewarden): Should we store the fact this is a client-allocated + // buffer? + result->data.raw = reinterpret_cast(preallocated_buffer); + } + } + + // Figure out what the size in bytes of the buffer is and store it. + size_t type_size; + TF_LITE_ENSURE_STATUS(BytesRequiredForTensor( + flatbuffer_tensor, &result->bytes, &type_size, error_reporter)); + // Copy the shape of the tensor from the serialized data into the runtime + // form. We have to allocate memory for this. + result->dims = + reinterpret_cast(memory_allocator_.AllocateFromTail( + sizeof(int) * (flatbuffer_tensor.shape()->Length() + 1), + kDefaultAlignment)); + result->dims->size = flatbuffer_tensor.shape()->Length(); + for (size_t n = 0; n < flatbuffer_tensor.shape()->Length(); ++n) { + result->dims->data[n] = flatbuffer_tensor.shape()->Get(n); + } + // Copy the quantization information from the serialized data. + const auto* src_quantization = flatbuffer_tensor.quantization(); + if (src_quantization && src_quantization->scale() && + (src_quantization->scale()->size() > 0) && + src_quantization->zero_point() && + (src_quantization->zero_point()->size() > 0)) { + result->params.scale = src_quantization->scale()->Get(0); + // This magic handles issues with little-endianness. + for (unsigned int b = 0; b < sizeof(int64_t); ++b) + *(reinterpret_cast(&result->params.zero_point) + b) = + *(reinterpret_cast( + src_quantization->zero_point()->Data()) + + b); + result->params.zero_point = + flatbuffers::EndianScalar(result->params.zero_point); + + // Populate per-channel quantization params. + int channels = src_quantization->scale()->size(); + TfLiteAffineQuantization* quantization = + reinterpret_cast( + memory_allocator_.AllocateFromTail(sizeof(TfLiteAffineQuantization), + kDefaultAlignment)); + int* zero_point_array = + reinterpret_cast(memory_allocator_.AllocateFromTail( + channels * sizeof(int) + sizeof(int), kDefaultAlignment)); + int* scale_array = + reinterpret_cast(memory_allocator_.AllocateFromTail( + channels * sizeof(float) + sizeof(int), kDefaultAlignment)); + zero_point_array[0] = channels; + scale_array[0] = channels; + int* zero_point_data = &zero_point_array[1]; + float* scale_data = reinterpret_cast(&scale_array[1]); + for (int i = 0; i < channels; i++) { + zero_point_data[i] = src_quantization->zero_point()->Get(i); + scale_data[i] = src_quantization->scale()->Get(i); + } + quantization->scale = reinterpret_cast(scale_array); + quantization->zero_point = + reinterpret_cast(zero_point_array); + + result->quantization = {kTfLiteAffineQuantization, quantization}; + } + // Copy the name, if there is one. + if (flatbuffer_tensor.name()->c_str() != nullptr) { + result->name = flatbuffer_tensor.name()->c_str(); + } else { + result->name = ""; + } + // These aren't used by the micro flavor of TFL, so set them to defaults. + result->allocation = nullptr; + result->delegate = nullptr; + result->buffer_handle = 0; + result->data_is_stale = false; + return kTfLiteOk; +} + +} // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.h new file mode 100644 index 0000000..a770b17 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.h @@ -0,0 +1,96 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_ALLOCATOR_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_ALLOCATOR_H_ + +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" + +namespace tflite { + +typedef struct { + TfLiteNode node; + const TfLiteRegistration* registration; +} NodeAndRegistration; + +// Allocator responsible for allocating memory for all intermediate tensors +// necessary to invoke a model. +class MicroAllocator { + public: + // The lifetime of the model, tensor allocator and error reporter must be at + // least as long as that of the allocator object, since the allocator needs + // them to be accessible during its entire lifetime. + MicroAllocator(TfLiteContext* context, const Model* model, + uint8_t* tensor_arena, size_t arena_size, + ErrorReporter* error_reporter); + + // Specify a particular tensor as pre-allocated. This means that this tensor + // will internally point to the supplied buffer, and no new memory will be + // provided. The buffer must live at least as long as the allocator, since + // the buffer will be used every time an op is invoked which uses the + // specified tensor. Most commonly this is useful when a platform-provided + // DMA buffer is used as an input, and it is desirable to avoid unnecessarily + // allocating a new buffer and copying from the DMA buffer. The user must + // ensure the buffer is valid throughout each interpreter run, and is not + // prematurely overwritten. + TfLiteStatus RegisterPreallocatedInput(uint8_t* buffer, size_t input_index); + + // Sets up all of the data structure members for a runtime tensor based on the + // contents of a serialized tensor. This method doesn't allocate any memory, + // all allocations happen subsequently in AllocateTensors. + TfLiteStatus InitializeRuntimeTensor( + const tflite::Tensor& flatbuffer_tensor, + const flatbuffers::Vector>* buffers, + ErrorReporter* error_reporter, TfLiteTensor* result, + uint8_t* preallocated_buffer = nullptr); + + // Run through the model and allocate all necessary input, output and + // intermediate tensors except for those already provided via calls to + // registerPreallocatedInput. + // WARNING: doing any allocation after calling is method has the risk of + // corruption tensor data so this method is the last method to be called in + // this class. + TfLiteStatus FinishTensorAllocation(); + + // Run through the model to allocate nodes and registrations. We need to keep + // them for the entire life time of the model to allow persistent tensors. + // This method needs to be called before FinishTensorAllocation method. + TfLiteStatus AllocateNodeAndRegistrations( + const OpResolver& op_resolver, + NodeAndRegistration** node_and_registrations); + + private: + const Model* model_; + SimpleMemoryAllocator memory_allocator_; + ErrorReporter* error_reporter_; + TfLiteContext* context_; + uint8_t* arena_; + size_t arena_size_; + // Indicating if the allocator is ready for allocation. + bool active_ = false; + + const SubGraph* subgraph_; + const flatbuffers::Vector>* operators_; + const flatbuffers::Vector>* tensors_; +}; + +} // namespace tflite +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_ALLOCATOR_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_error_reporter.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.cpp similarity index 93% rename from src/tensorflow/lite/micro/micro_error_reporter.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.cpp index e1a6785..3091c1d 100644 --- a/src/tensorflow/lite/micro/micro_error_reporter.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,7 +14,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/micro_error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h" namespace tflite { namespace { @@ -66,3 +67,5 @@ int MicroErrorReporter::Report(const char* format, va_list args) { } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h new file mode 100644 index 0000000..0b1c6d3 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h @@ -0,0 +1,39 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_ERROR_REPORTER_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_ERROR_REPORTER_H_ + +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/debug_log_numbers.h" + +namespace tflite { + +class MicroErrorReporter : public ErrorReporter { + public: + ~MicroErrorReporter() {} + int Report(const char* format, va_list args) override; + + private: + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_ERROR_REPORTER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_interpreter.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.cpp similarity index 70% rename from src/tensorflow/lite/micro/micro_interpreter.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.cpp index f6f8127..10a8239 100644 --- a/src/tensorflow/lite/micro/micro_interpreter.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,16 +13,33 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/micro_interpreter.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/flatbuffer_conversions.h" -#include "tensorflow/lite/core/api/tensor_utils.h" -#include "tensorflow/lite/micro/compatibility.h" -#include "tensorflow/lite/micro/micro_optional_debug_tools.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h" namespace tflite { namespace { +const int kStackDataAllocatorSize = 128; +class StackDataAllocator : public BuiltinDataAllocator { + public: + void* Allocate(size_t size) override { + if (size > kStackDataAllocatorSize) { + return nullptr; + } else { + return data_; + } + } + void Deallocate(void* data) override { + // Do nothing. + } + + private: + uint8_t data_[kStackDataAllocatorSize]; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; const char* OpNameFromRegistration(const TfLiteRegistration* registration) { if (registration->builtin_code == BuiltinOperator_CUSTOM) { @@ -50,12 +68,11 @@ MicroInterpreter::MicroInterpreter(const Model* model, : model_(model), op_resolver_(op_resolver), error_reporter_(error_reporter), + context_(), allocator_(&context_, model_, tensor_arena, tensor_arena_size, error_reporter_), - tensors_allocated_(false), - tensors_prepared_(false) { - const flatbuffers::Vector>* subgraphs = - model->subgraphs(); + tensors_allocated_(false) { + auto* subgraphs = model->subgraphs(); if (subgraphs->size() != 1) { error_reporter->Report("Only 1 subgraph is currently supported.\n"); initialization_status_ = kTfLiteError; @@ -75,7 +92,7 @@ MicroInterpreter::MicroInterpreter(const Model* model, // NOTE: This requires that the flatbuffer is held in memory which can be // modified by this process. if (!FLATBUFFERS_LITTLEENDIAN) { - for (size_t t = 0; t < tensors_size(); ++t) { + for (int t = 0; t < tensors_size(); ++t) { TfLiteTensor* thisTensor = &context_.tensors[t]; if (thisTensor->allocation_type == kTfLiteMmapRo) CorrectTensorEndianness(thisTensor); @@ -122,6 +139,11 @@ void MicroInterpreter::CorrectTensorDataEndianness(T* data, int32_t size) { } } +TfLiteStatus MicroInterpreter::RegisterPreallocatedInput(uint8_t* buffer, + size_t input_index) { + return allocator_.RegisterPreallocatedInput(buffer, input_index); +} + TfLiteStatus MicroInterpreter::AllocateTensors() { TF_LITE_ENSURE_OK(&context_, allocator_.AllocateNodeAndRegistrations( op_resolver_, &node_and_registrations_)); @@ -156,30 +178,24 @@ TfLiteStatus MicroInterpreter::Invoke() { init_data = reinterpret_cast(node->builtin_data); init_data_size = 0; } - if (!tensors_prepared_ && registration->init) { + if (registration->init) { node->user_data = registration->init(&context_, init_data, init_data_size); } } - if (!tensors_prepared_) { - for (size_t i = 0; i < operators_->size(); ++i) { - auto* node = &(node_and_registrations_[i].node); - auto* registration = node_and_registrations_[i].registration; - if (registration->prepare) { - TfLiteStatus prepare_status = registration->prepare(&context_, node); - if (prepare_status != kTfLiteOk) { - error_reporter_->Report( - "Node %s (number %d) failed to prepare with status %d", - OpNameFromRegistration(registration), i, prepare_status); - return kTfLiteError; - } + for (size_t i = 0; i < operators_->size(); ++i) { + auto* node = &(node_and_registrations_[i].node); + auto* registration = node_and_registrations_[i].registration; + if (registration->prepare) { + TfLiteStatus prepare_status = registration->prepare(&context_, node); + if (prepare_status != kTfLiteOk) { + error_reporter_->Report( + "Node %s (number %d) failed to prepare with status %d", + OpNameFromRegistration(registration), i, prepare_status); + return kTfLiteError; } } -#ifdef TF_LITE_MICRO_TENSORS_PREPARED - // TODO(b/148085107): Turn this value on by default. - tensors_prepared_ = true; -#endif } for (size_t i = 0; i < operators_->size(); ++i) { @@ -241,20 +257,45 @@ TfLiteTensor* MicroInterpreter::tensor(size_t index) { return &context_.tensors[index]; } -TfLiteStatus MicroInterpreter::ResetVariableTensors() { - const size_t length = tensors_size(); - for (size_t i = 0; i < length; ++i) { - TfLiteTensor* cur_tensor = tensor(i); - if (cur_tensor->is_variable) { - TfLiteStatus status = tflite::ResetVariableTensor(cur_tensor); - if (status != kTfLiteOk) { - error_reporter_->Report("Failed to reset variable tensor at index: %d", - i); - return status; - } +struct pairTfLiteNodeAndRegistration MicroInterpreter::node_and_registration( + int node_index) { + TfLiteStatus status = kTfLiteOk; + struct pairTfLiteNodeAndRegistration tfNodeRegiPair; + auto opcodes = model_->operator_codes(); + { + const auto* op = operators_->Get(node_index); + size_t index = op->opcode_index(); + if (index < 0 || index >= opcodes->size()) { + error_reporter_->Report("Missing registration for opcode_index %d\n", + index); + } + auto opcode = (*opcodes)[index]; + const TfLiteRegistration* registration = nullptr; + status = GetRegistrationFromOpCode(opcode, op_resolver_, error_reporter_, + ®istration); + if (status != kTfLiteOk) { + error_reporter_->Report("Missing registration for opcode_index %d\n", + index); } + if (registration == nullptr) { + error_reporter_->Report("Skipping op for opcode_index %d\n", index); + } + + // Disregard const qualifier to workaround with existing API. + TfLiteIntArray* inputs_array = const_cast( + reinterpret_cast(op->inputs())); + TfLiteIntArray* outputs_array = const_cast( + reinterpret_cast(op->outputs())); + + TfLiteNode node; + node.inputs = inputs_array; + node.outputs = outputs_array; + tfNodeRegiPair.node = node; + tfNodeRegiPair.registration = registration; } - return kTfLiteOk; + return tfNodeRegiPair; } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_interpreter.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.h similarity index 56% rename from src/tensorflow/lite/micro/micro_interpreter.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.h index 941960a..6551bae 100644 --- a/src/tensorflow/lite/micro/micro_interpreter.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,16 +13,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_MICRO_INTERPRETER_H_ -#define TENSORFLOW_LITE_MICRO_MICRO_INTERPRETER_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_INTERPRETER_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_INTERPRETER_H_ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/core/api/op_resolver.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/micro/micro_allocator.h" -#include "tensorflow/lite/schema/schema_generated.h" -#include "tensorflow/lite/type_to_tflitetype.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_allocator.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { @@ -39,69 +39,41 @@ class MicroInterpreter { uint8_t* tensor_arena, size_t tensor_arena_size, ErrorReporter* error_reporter); - // Runs through the model and allocates all necessary input, output and - // intermediate tensors. + // Specify a particular tensor as pre-allocated. This means that this tensor + // will internally point to the supplied buffer, and no new memory will be + // provided. The buffer must live at least as long as the allocator, since + // the buffer will be used every time an op is invoked which uses the + // specified tensor. Most commonly this is useful when a platform-provided + // DMA buffer is used as an input, and it is desirable to avoid unnecessarily + // allocating a new buffer and copying from the DMA buffer. The user must + // ensure the buffer is valid throughout each interpreter run, and is not + // prematurely overwritten. + TfLiteStatus RegisterPreallocatedInput(uint8_t* buffer, size_t input_index); + + // Run through the model and allocate all necessary input, output and + // intermediate tensors except for those already provided via calls to + // registerPreallocatedInput. TfLiteStatus AllocateTensors(); TfLiteStatus Invoke(); size_t tensors_size() const { return context_.tensors_size; } TfLiteTensor* tensor(size_t tensor_index); - template - T* typed_tensor(int tensor_index) { - if (TfLiteTensor* tensor_ptr = tensor(tensor_index)) { - if (tensor_ptr->type == typeToTfLiteType()) { - return GetTensorData(tensor_ptr); - } - } - return nullptr; - } TfLiteTensor* input(size_t index); size_t inputs_size() const { return subgraph_->inputs()->Length(); } - const flatbuffers::Vector& inputs() const { - return *subgraph_->inputs(); - } - TfLiteTensor* input_tensor(size_t index) { return input(index); } - template - T* typed_input_tensor(int tensor_index) { - if (TfLiteTensor* tensor_ptr = input_tensor(tensor_index)) { - if (tensor_ptr->type == typeToTfLiteType()) { - return GetTensorData(tensor_ptr); - } - } - return nullptr; - } + const flatbuffers::Vector* inputs() { return subgraph_->inputs(); } TfLiteTensor* output(size_t index); size_t outputs_size() const { return subgraph_->outputs()->Length(); } - const flatbuffers::Vector& outputs() const { - return *subgraph_->outputs(); - } - TfLiteTensor* output_tensor(size_t index) { return output(index); } - template - T* typed_output_tensor(int tensor_index) { - if (TfLiteTensor* tensor_ptr = output_tensor(tensor_index)) { - if (tensor_ptr->type == typeToTfLiteType()) { - return GetTensorData(tensor_ptr); - } - } - return nullptr; - } - - // Reset all variable tensors to the default value. - TfLiteStatus ResetVariableTensors(); + const flatbuffers::Vector* outputs() { return subgraph_->outputs(); } TfLiteStatus initialization_status() const { return initialization_status_; } ErrorReporter* error_reporter() { return error_reporter_; } size_t operators_size() const { return operators_->size(); } - - // For debugging only. - const NodeAndRegistration node_and_registration(int node_index) const { - return node_and_registrations_[node_index]; - } + struct pairTfLiteNodeAndRegistration node_and_registration(int node_index); private: void CorrectTensorEndianness(TfLiteTensor* tensorCorr); @@ -114,10 +86,9 @@ class MicroInterpreter { const Model* model_; const OpResolver& op_resolver_; ErrorReporter* error_reporter_; - TfLiteContext context_ = {}; + TfLiteContext context_; MicroAllocator allocator_; bool tensors_allocated_; - bool tensors_prepared_; TfLiteStatus initialization_status_; const flatbuffers::Vector>* tensors_; @@ -128,4 +99,6 @@ class MicroInterpreter { } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_MICRO_INTERPRETER_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_INTERPRETER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.cpp new file mode 100644 index 0000000..5359ac0 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.cpp @@ -0,0 +1,83 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.h" + +namespace tflite { + +const TfLiteRegistration* MicroMutableOpResolver::FindOp( + tflite::BuiltinOperator op, int version) const { + for (int i = 0; i < registrations_len_; ++i) { + const TfLiteRegistration& registration = registrations_[i]; + if ((registration.builtin_code == op) && + (registration.version == version)) { + return ®istration; + } + } + return nullptr; +} + +const TfLiteRegistration* MicroMutableOpResolver::FindOp(const char* op, + int version) const { + for (int i = 0; i < registrations_len_; ++i) { + const TfLiteRegistration& registration = registrations_[i]; + if ((registration.builtin_code == BuiltinOperator_CUSTOM) && + (strcmp(registration.custom_name, op) == 0) && + (registration.version == version)) { + return ®istration; + } + } + return nullptr; +} + +void MicroMutableOpResolver::AddBuiltin(tflite::BuiltinOperator op, + TfLiteRegistration* registration, + int min_version, int max_version) { + for (int version = min_version; version <= max_version; ++version) { + if (registrations_len_ >= TFLITE_REGISTRATIONS_MAX) { + // TODO(petewarden) - Add error reporting hooks so we can report this! + return; + } + TfLiteRegistration* new_registration = ®istrations_[registrations_len_]; + registrations_len_ += 1; + + *new_registration = *registration; + new_registration->builtin_code = op; + new_registration->version = version; + } +} + +void MicroMutableOpResolver::AddCustom(const char* name, + TfLiteRegistration* registration, + int min_version, int max_version) { + for (int version = min_version; version <= max_version; ++version) { + if (registrations_len_ >= TFLITE_REGISTRATIONS_MAX) { + // TODO(petewarden) - Add error reporting hooks so we can report this! + return; + } + TfLiteRegistration* new_registration = ®istrations_[registrations_len_]; + registrations_len_ += 1; + + *new_registration = *registration; + new_registration->builtin_code = BuiltinOperator_CUSTOM; + new_registration->custom_name = name; + new_registration->version = version; + } +} + +} // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.h new file mode 100644 index 0000000..2580662 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_mutable_op_resolver.h @@ -0,0 +1,49 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ + +#include "tensorflow_esp32/tensorflow/lite/core/api/op_resolver.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/compatibility.h" + +#ifndef TFLITE_REGISTRATIONS_MAX +#define TFLITE_REGISTRATIONS_MAX (128) +#endif + +namespace tflite { + +class MicroMutableOpResolver : public OpResolver { + public: + const TfLiteRegistration* FindOp(tflite::BuiltinOperator op, + int version) const override; + const TfLiteRegistration* FindOp(const char* op, int version) const override; + void AddBuiltin(tflite::BuiltinOperator op, TfLiteRegistration* registration, + int min_version = 1, int max_version = 1); + void AddCustom(const char* name, TfLiteRegistration* registration, + int min_version = 1, int max_version = 1); + + private: + TfLiteRegistration registrations_[TFLITE_REGISTRATIONS_MAX]; + int registrations_len_ = 0; + + TF_LITE_REMOVE_VIRTUAL_DELETE +}; + +} // namespace tflite + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_MUTABLE_OP_RESOLVER_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_optional_debug_tools.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.cpp similarity index 76% rename from src/tensorflow/lite/micro/micro_optional_debug_tools.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.cpp index bc69eb5..fc008ef 100644 --- a/src/tensorflow/lite/micro/micro_optional_debug_tools.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,45 +13,37 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/micro_optional_debug_tools.h" +#include "Arduino.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.h" -// `cinttypes` requires `__STDC_FORMAT_MACROS` to be defined to expose `PRId32`. -#ifndef __STDC_FORMAT_MACROS -#define __STDC_FORMAT_MACROS -#endif - -#include - -#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { -namespace { std::vector flatbuffersVector2StdVector( - const flatbuffers::Vector& fVector) { + const flatbuffers::Vector* fVector) { std::vector stdVector; - stdVector.reserve(fVector.size()); - for (size_t i = 0; i < fVector.size(); i++) { - stdVector.push_back(fVector.Get(i)); + for (size_t i = 0; i < fVector->size(); i++) { + stdVector.push_back(fVector->Get(i)); } return stdVector; } void PrintIntVector(const std::vector& v) { for (const auto& it : v) { - printf(" %d", it); + Serial.printf(" %d", it); } - printf("\n"); + Serial.printf("\n"); } void PrintTfLiteIntVector(const TfLiteIntArray* v) { if (!v) { - printf(" (null)\n"); + Serial.printf(" (null)\n"); return; } for (int k = 0; k < v->size; k++) { - printf(" %d", v->data[k]); + Serial.printf(" %d", v->data[k]); } - printf("\n"); + Serial.printf("\n"); } const char* TensorTypeName(TfLiteType type) { @@ -96,47 +89,48 @@ const char* AllocTypeName(TfLiteAllocationType type) { } return "(invalid)"; } -} // namespace // Prints a dump of what tensors and what nodes are in the interpreter. void PrintInterpreterState(MicroInterpreter* interpreter) { - printf("Interpreter has %zu tensors and %zu nodes\n", + Serial.printf("Interpreter has %zu tensors and %zu nodes\n", interpreter->tensors_size(), interpreter->operators_size()); - printf("Inputs:"); + Serial.printf("Inputs:"); PrintIntVector(flatbuffersVector2StdVector(interpreter->inputs())); - printf("Outputs:"); + Serial.printf("Outputs:"); PrintIntVector(flatbuffersVector2StdVector(interpreter->outputs())); - printf("\n"); + Serial.printf("\n"); for (size_t tensor_index = 0; tensor_index < interpreter->tensors_size(); tensor_index++) { TfLiteTensor* tensor = interpreter->tensor(static_cast(tensor_index)); - printf("Tensor %3zu %-20s %10s %15s %10zu bytes (%4.1f MB) ", tensor_index, + Serial.printf("Tensor %3zu %-20s %10s %15s %10zu bytes (%4.1f MB) ", tensor_index, tensor->name, TensorTypeName(tensor->type), AllocTypeName(tensor->allocation_type), tensor->bytes, static_cast(tensor->bytes / (1 << 20))); PrintTfLiteIntVector(tensor->dims); } - printf("\n"); + Serial.printf("\n"); for (size_t node_index = 0; node_index < interpreter->operators_size(); node_index++) { - const NodeAndRegistration node_and_reg = + struct pairTfLiteNodeAndRegistration node_and_reg = interpreter->node_and_registration(static_cast(node_index)); const TfLiteNode& node = node_and_reg.node; const TfLiteRegistration* reg = node_and_reg.registration; if (reg->custom_name != nullptr) { - printf("Node %3zu Operator Custom Name %s\n", node_index, + Serial.printf("Node %3zu Operator Custom Name %s\n", node_index, reg->custom_name); } else { - printf("Node %3zu Operator Builtin Code %3" PRId32 " %s\n", node_index, + Serial.printf("Node %3zu Operator Builtin Code %3d %s\n", node_index, reg->builtin_code, EnumNamesBuiltinOperator()[reg->builtin_code]); } - printf(" Inputs:"); + Serial.printf(" Inputs:"); PrintTfLiteIntVector(node.inputs); - printf(" Outputs:"); + Serial.printf(" Outputs:"); PrintTfLiteIntVector(node.outputs); } } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_optional_debug_tools.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.h similarity index 63% rename from src/tensorflow/lite/micro/micro_optional_debug_tools.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.h index ae96b62..9404ab8 100644 --- a/src/tensorflow/lite/micro/micro_optional_debug_tools.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_optional_debug_tools.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -14,14 +15,29 @@ limitations under the License. ==============================================================================*/ // Optional debugging functionality. For small sized binaries, these are not // needed. -#ifndef TENSORFLOW_LITE_MICRO_MICRO_OPTIONAL_DEBUG_TOOLS_H_ -#define TENSORFLOW_LITE_MICRO_MICRO_OPTIONAL_DEBUG_TOOLS_H_ +#ifndef TENSORFLOW_LITE_MICRO_OPTIONAL_DEBUG_TOOLS_H_ +#define TENSORFLOW_LITE_MICRO_OPTIONAL_DEBUG_TOOLS_H_ -#include "tensorflow/lite/micro/micro_interpreter.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_interpreter.h" namespace tflite { // Prints a dump of what tensors and what nodes are in the interpreter. +class MicroInterpreter; void PrintInterpreterState(MicroInterpreter* interpreter); + +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus +struct pairTfLiteNodeAndRegistration { + TfLiteNode node; + const TfLiteRegistration* registration; +}; +#ifdef __cplusplus +} // extern "C" +#endif // __cplusplus + } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_MICRO_OPTIONAL_DEBUG_TOOLS_H_ +#endif // TENSORFLOW_LITE_MICRO_OPTIONAL_DEBUG_TOOLS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_utils.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.cpp similarity index 71% rename from src/tensorflow/lite/micro/micro_utils.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.cpp index fbd4a5e..b3919c7 100644 --- a/src/tensorflow/lite/micro/micro_utils.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,33 +14,26 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h" #include #include #include -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" namespace tflite { namespace { static const uint8_t kAsymmetricUInt8Min = 0; -static const uint8_t kAsymmetricUInt8Max = UINT8_MAX; +static const uint8_t kAsymmetricUInt8Max = 255; static const uint8_t kSymmetricUInt8Min = 1; -static const uint8_t kSymmetricUInt8Max = UINT8_MAX; -static const int8_t kAsymmetricInt8Min = INT8_MIN; -static const int8_t kAsymmetricInt8Max = INT8_MAX; +static const uint8_t kSymmetricUInt8Max = 255; +static const int8_t kAsymmetricInt8Min = -128; +static const int8_t kAsymmetricInt8Max = 127; static const int kSymmetricInt8Scale = kAsymmetricInt8Max; -static const int16_t kAsymmetricInt16Max = INT16_MAX; -static const int kSymmetricInt16Scale = kAsymmetricInt16Max; - -static const int32_t kAsymmetricInt32Max = INT32_MAX; -static const int kSymmetricInt32Scale = kAsymmetricInt32Max; - } // namespace int ElementCount(const TfLiteIntArray& dims) { @@ -82,7 +76,7 @@ int8_t FloatToAsymmetricQuantizedInt8(const float value, const float scale, } int8_t FloatToSymmetricQuantizedInt8(const float value, const float scale) { - return FloatToAsymmetricQuantizedInt8(value, scale, 0.0f); + return FloatToSymmetricQuantizedUInt8(value, scale) + kAsymmetricInt8Min; } int32_t FloatToSymmetricQuantizedInt32(const float value, const float scale) { @@ -137,24 +131,15 @@ void SignedSymmetricPerChannelQuantize(const float* values, int input_size = ElementCount(*dims); int channel_count = dims->data[quantized_dimension]; int per_channel_size = input_size / channel_count; - - int stride; - int channel_stride; - if (quantized_dimension == 0) { - stride = 1; - channel_stride = per_channel_size; - } else if (quantized_dimension == 3) { - stride = channel_count; - channel_stride = 1; - } else { - TF_LITE_FATAL("quantized dimension must be 0 or 3"); - } - - // Calculate scales for each channel. for (int channel = 0; channel < channel_count; channel++) { float min = 0; float max = 0; - + int stride = 1; + for (int i = 0; i < quantized_dimension; i++) { + stride *= dims->data[i]; + } + int channel_stride = per_channel_size / stride; + // Calculate scales for each channel. for (int i = 0; i < per_channel_size; i++) { int idx = channel * channel_stride + i * stride; min = fminf(min, values[idx]); @@ -193,47 +178,6 @@ void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims, } } -void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims, - int16_t* quantized_values, float* scaling_factor) { - int input_size = ElementCount(*dims); - - float min = 0; - float max = 0; - for (int i = 0; i < input_size; i++) { - min = fminf(min, values[i]); - max = fmaxf(max, values[i]); - } - *scaling_factor = fmaxf(fabs(min), fabs(max)) / kSymmetricInt16Scale; - for (int i = 0; i < input_size; i++) { - const int32_t quantized_value = - static_cast(roundf(values[i] / *scaling_factor)); - // Clamp: just in case some odd numeric offset. - quantized_values[i] = fminf(kSymmetricInt16Scale, - fmaxf(-kSymmetricInt16Scale, quantized_value)); - } -} - -void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims, - int32_t* quantized_values, float* scaling_factor) { - int input_size = ElementCount(*dims); - - float min = 0; - float max = 0; - for (int i = 0; i < input_size; i++) { - min = fminf(min, values[i]); - max = fmaxf(max, values[i]); - } - - *scaling_factor = fmaxf(fabs(min), fabs(max)) / kSymmetricInt32Scale; - for (int i = 0; i < input_size; i++) { - const int32_t quantized_value = - static_cast(roundf(values[i] / *scaling_factor)); - // Clamp: just in case some odd numeric offset. - quantized_values[i] = fminf(kSymmetricInt32Scale, - fmaxf(-kSymmetricInt32Scale, quantized_value)); - } -} - void SymmetricQuantize(const float* values, TfLiteIntArray* dims, uint8_t* quantized_values, float* scaling_factor) { SignedSymmetricQuantize(values, dims, @@ -250,3 +194,5 @@ void SymmetricDequantize(const int8_t* values, const int size, } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/micro_utils.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h similarity index 87% rename from src/tensorflow/lite/micro/micro_utils.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h index 42b33dc..4bdeea8 100644 --- a/src/tensorflow/lite/micro/micro_utils.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,12 +14,12 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_MICRO_UTILS_H_ -#define TENSORFLOW_LITE_MICRO_MICRO_UTILS_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_UTILS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_UTILS_H_ #include -#include "tensorflow/lite/c/common.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" namespace tflite { @@ -74,12 +75,6 @@ void SignedSymmetricPerChannelQuantize(const float* values, void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims, int8_t* quantized_values, float* scaling_factor); -void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims, - int16_t* quantized_values, float* scaling_factor); - -void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims, - int32_t* quantized_values, float* scaling_factor); - void SymmetricQuantize(const float* values, TfLiteIntArray* dims, uint8_t* quantized_values, float* scaling_factor); @@ -89,4 +84,6 @@ void SymmetricDequantize(const int8_t* values, const int size, } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_MICRO_UTILS_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_MICRO_UTILS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/simple_memory_allocator.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.cpp similarity index 71% rename from src/tensorflow/lite/micro/simple_memory_allocator.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.cpp index 89d6fd6..f3f9569 100644 --- a/src/tensorflow/lite/micro/simple_memory_allocator.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,25 +14,13 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/simple_memory_allocator.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.h" -#include - -#include "tensorflow/lite/core/api/flatbuffer_conversions.h" -#include "tensorflow/lite/micro/memory_helpers.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/flatbuffer_conversions.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/memory_helpers.h" namespace tflite { -SimpleMemoryAllocator* CreateInPlaceSimpleMemoryAllocator(uint8_t* buffer, - size_t buffer_size) { - SimpleMemoryAllocator tmp = SimpleMemoryAllocator(buffer, buffer_size); - SimpleMemoryAllocator* in_place_allocator = - reinterpret_cast(tmp.AllocateFromTail( - sizeof(SimpleMemoryAllocator), alignof(SimpleMemoryAllocator))); - *in_place_allocator = tmp; - return in_place_allocator; -} - uint8_t* SimpleMemoryAllocator::AllocateFromTail(size_t size, size_t alignment) { if (has_child_allocator_) { @@ -41,7 +30,7 @@ uint8_t* SimpleMemoryAllocator::AllocateFromTail(size_t size, uint8_t* previous_free = (data_ + data_size_max_) - data_size_; uint8_t* current_data = previous_free - size; uint8_t* aligned_result = AlignPointerDown(current_data, alignment); - std::ptrdiff_t aligned_size = (previous_free - aligned_result); + size_t aligned_size = (previous_free - aligned_result); if ((data_size_ + aligned_size) > data_size_max_) { // TODO(petewarden): Add error reporting beyond returning null! return nullptr; @@ -55,6 +44,7 @@ SimpleMemoryAllocator SimpleMemoryAllocator::CreateChildAllocator() { // is not what we expected. SimpleMemoryAllocator child = *this; child.parent_allocator_ = this; + // With C++ copy elision, &child should be available after return. has_child_allocator_ = true; return child; } @@ -67,3 +57,5 @@ SimpleMemoryAllocator::~SimpleMemoryAllocator() { } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/simple_memory_allocator.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.h similarity index 74% rename from src/tensorflow/lite/micro/simple_memory_allocator.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.h index e624d65..473aedb 100644 --- a/src/tensorflow/lite/micro/simple_memory_allocator.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/simple_memory_allocator.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,12 +14,12 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ -#define TENSORFLOW_LITE_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { @@ -35,9 +36,7 @@ class SimpleMemoryAllocator { // in ascending order. uint8_t* AllocateFromTail(size_t size, size_t alignment); - size_t GetDataSize() const { return data_size_; } - uint8_t* GetBuffer() const { return data_; } - size_t GetMaxBufferSize() const { return data_size_max_; } + int GetDataSize() const { return data_size_; } // Child allocator is something like a temporary allocator. Memory allocated // by the child allocator will be freed once the child allocator is @@ -52,7 +51,7 @@ class SimpleMemoryAllocator { ~SimpleMemoryAllocator(); private: - size_t data_size_ = 0; + int data_size_ = 0; size_t data_size_max_; uint8_t* data_; SimpleMemoryAllocator* parent_allocator_ = nullptr; @@ -60,11 +59,8 @@ class SimpleMemoryAllocator { bool has_child_allocator_ = false; }; -// Allocate a SimpleMemoryAllocator from the buffer and then return the pointer -// to this allocator. -SimpleMemoryAllocator* CreateInPlaceSimpleMemoryAllocator(uint8_t* buffer, - size_t buffer_size); - } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_SIMPLE_MEMORY_ALLOCATOR_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/test_helpers.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.cpp similarity index 65% rename from src/tensorflow/lite/micro/test_helpers.cpp rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.cpp index 6859a9a..5521e31 100644 --- a/src/tensorflow/lite/micro/test_helpers.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,11 +14,11 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/micro/test_helpers.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.h" -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/tensor_utils.h" -#include "tensorflow/lite/micro/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h" namespace tflite { namespace testing { @@ -47,7 +48,7 @@ class StackAllocator : public flatbuffers::Allocator { return *inst; } - static constexpr size_t kStackAllocatorSize = 4096; + static constexpr int kStackAllocatorSize = 4096; private: uint8_t data_backing_[kStackAllocatorSize]; @@ -63,7 +64,7 @@ flatbuffers::FlatBufferBuilder* BuilderInstance() { return inst; } -const Model* BuildSimpleMockModel() { +const Model* BuildMockModel() { using flatbuffers::Offset; flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); @@ -126,207 +127,24 @@ const Model* BuildSimpleMockModel() { return model; } -const Model* BuildComplexMockModel() { - using flatbuffers::Offset; - flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); - - constexpr size_t buffer_data_size = 1; - const uint8_t buffer_data_1[buffer_data_size] = {21}; - const uint8_t buffer_data_2[buffer_data_size] = {21}; - const uint8_t buffer_data_3[buffer_data_size] = {21}; - constexpr size_t buffers_size = 7; - const Offset buffers[buffers_size] = { - // Op 1 buffers: - CreateBuffer(*builder), - CreateBuffer(*builder), - CreateBuffer(*builder, - builder->CreateVector(buffer_data_1, buffer_data_size)), - // Op 2 buffers: - CreateBuffer(*builder), - CreateBuffer(*builder, - builder->CreateVector(buffer_data_2, buffer_data_size)), - // Op 3 buffers: - CreateBuffer(*builder), - CreateBuffer(*builder, - builder->CreateVector(buffer_data_3, buffer_data_size)), - }; - constexpr size_t tensor_shape_size = 1; - const int32_t tensor_shape[tensor_shape_size] = {1}; - - constexpr size_t tensors_size = 10; - const Offset tensors[tensors_size] = { - // Op 1 inputs: - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 0, builder->CreateString("test_input_tensor_1"), 0, - false /* is_variable */), - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 1, builder->CreateString("test_variable_tensor_1"), - 0, true /* is_variable */), - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_UINT8, 2, builder->CreateString("test_weight_tensor_1"), 0, - false /* is_variable */), - // Op 1 output / Op 2 input: - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 0, builder->CreateString("test_output_tensor_1"), 0, - false /* is_variable */), - // Op 2 inputs: - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 1, builder->CreateString("test_variable_tensor_2"), - 0, true /* is_variable */), - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_UINT8, 2, builder->CreateString("test_weight_tensor_2"), 0, - false /* is_variable */), - // Op 2 output / Op 3 input: - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 0, builder->CreateString("test_output_tensor_2"), 0, - false /* is_variable */), - // Op 3 inputs: - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 1, builder->CreateString("test_variable_tensor_3"), - 0, true /* is_variable */), - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_UINT8, 2, builder->CreateString("test_weight_tensor_3"), 0, - false /* is_variable */), - // Op 3 output: - CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 0, builder->CreateString("test_output_tensor_3"), 0, - false /* is_variable */), - }; - - constexpr size_t operators_size = 3; - Offset operators[operators_size]; - { - // Set Op 1 attributes: - constexpr size_t operator_inputs_size = 3; - const int32_t operator_inputs[operator_inputs_size] = {0, 1, 2}; - constexpr size_t operator_outputs_size = 1; - const int32_t operator_outputs[operator_outputs_size] = {3}; - - operators[0] = {CreateOperator( - *builder, 0, - builder->CreateVector(operator_inputs, operator_inputs_size), - builder->CreateVector(operator_outputs, operator_outputs_size), - BuiltinOptions_NONE)}; - } - - { - // Set Op 2 attributes - constexpr size_t operator_inputs_size = 3; - const int32_t operator_inputs[operator_inputs_size] = {3, 4, 5}; - constexpr size_t operator_outputs_size = 1; - const int32_t operator_outputs[operator_outputs_size] = {6}; - - operators[1] = {CreateOperator( - *builder, 0, - builder->CreateVector(operator_inputs, operator_inputs_size), - builder->CreateVector(operator_outputs, operator_outputs_size), - BuiltinOptions_NONE)}; - } - - { - // Set Op 3 attributes - constexpr size_t operator_inputs_size = 3; - const int32_t operator_inputs[operator_inputs_size] = {6, 7, 8}; - constexpr size_t operator_outputs_size = 1; - const int32_t operator_outputs[operator_outputs_size] = {9}; - - operators[2] = {CreateOperator( - *builder, 0, - builder->CreateVector(operator_inputs, operator_inputs_size), - builder->CreateVector(operator_outputs, operator_outputs_size), - BuiltinOptions_NONE)}; - } - - constexpr size_t inputs_size = 1; - const int32_t inputs[inputs_size] = {0}; - constexpr size_t outputs_size = 1; - const int32_t outputs[outputs_size] = {9}; - - constexpr size_t subgraphs_size = 1; - const Offset subgraphs[subgraphs_size] = { - CreateSubGraph(*builder, builder->CreateVector(tensors, tensors_size), - builder->CreateVector(inputs, inputs_size), - builder->CreateVector(outputs, outputs_size), - builder->CreateVector(operators, operators_size), - builder->CreateString("test_subgraph"))}; - - constexpr size_t operator_codes_size = 1; - const Offset operator_codes[operator_codes_size] = { - CreateOperatorCodeDirect(*builder, BuiltinOperator_CUSTOM, "mock_custom", - 0)}; - - const Offset model_offset = CreateModel( - *builder, 0, builder->CreateVector(operator_codes, operator_codes_size), - builder->CreateVector(subgraphs, subgraphs_size), - builder->CreateString("test_model"), - builder->CreateVector(buffers, buffers_size)); - - FinishModelBuffer(*builder, model_offset); - void* model_pointer = builder->GetBufferPointer(); - const Model* model = flatbuffers::GetRoot(model_pointer); - return model; -} - } // namespace -const Model* GetSimpleMockModel() { - static Model* model = nullptr; - if (!model) { - model = const_cast(BuildSimpleMockModel()); - } - return model; -} - -const Model* GetComplexMockModel() { +const Model* GetMockModel() { static Model* model = nullptr; if (!model) { - model = const_cast(BuildComplexMockModel()); + model = const_cast(BuildMockModel()); } return model; } -const Tensor* Create1dFlatbufferTensor(int size, bool is_variable) { - using flatbuffers::Offset; - flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); - constexpr size_t tensor_shape_size = 1; - const int32_t tensor_shape[tensor_shape_size] = {size}; - const Offset tensor_offset = CreateTensor( - *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 0, builder->CreateString("test_tensor"), 0, - is_variable); - builder->Finish(tensor_offset); - void* tensor_pointer = builder->GetBufferPointer(); - const Tensor* tensor = flatbuffers::GetRoot(tensor_pointer); - return tensor; -} - -const Tensor* CreateQuantizedFlatbufferTensor(int size) { +const Tensor* Create1dFlatbufferTensor(int size) { using flatbuffers::Offset; flatbuffers::FlatBufferBuilder* builder = BuilderInstance(); - const Offset quant_params = - CreateQuantizationParameters( - *builder, - /*min=*/builder->CreateVector({0.1f}), - /*max=*/builder->CreateVector({0.2f}), - /*scale=*/builder->CreateVector({0.3f}), - /*zero_point=*/builder->CreateVector({100ll})); - constexpr size_t tensor_shape_size = 1; const int32_t tensor_shape[tensor_shape_size] = {size}; const Offset tensor_offset = CreateTensor( *builder, builder->CreateVector(tensor_shape, tensor_shape_size), - TensorType_INT32, 0, builder->CreateString("test_tensor"), quant_params, - false); + TensorType_INT32, 0, builder->CreateString("test_tensor"), 0, false); builder->Finish(tensor_offset); void* tensor_pointer = builder->GetBufferPointer(); const Tensor* tensor = flatbuffers::GetRoot(tensor_pointer); @@ -490,18 +308,6 @@ TfLiteTensor CreateQuantizedTensor(const int8_t* data, TfLiteIntArray* dims, return result; } -TfLiteTensor CreateQuantizedTensor(const int16_t* data, TfLiteIntArray* dims, - float scale, int zero_point, - const char* name, bool is_variable) { - TfLiteTensor result = CreateTensor(dims, name, is_variable); - result.type = kTfLiteInt16; - result.data.i16 = const_cast(data); - result.params = {scale, zero_point}; - result.quantization = {kTfLiteAffineQuantization, nullptr}; - result.bytes = ElementCount(*dims) * sizeof(int16_t); - return result; -} - TfLiteTensor CreateQuantizedTensor(const float* input, int8_t* quantized, TfLiteIntArray* dims, float scale, int zero_point, const char* name, @@ -580,10 +386,6 @@ TfLiteTensor CreateSymmetricPerChannelQuantizedTensor( SignedSymmetricPerChannelQuantize(input, dims, quantized_dimension, quantized, &scales[1]); - for (int i = 0; i < channel_count; i++) { - zero_points[i + 1] = 0; - } - affine_quant->scale = FloatArrayFromFloats(scales); affine_quant->zero_point = IntArrayFromInts(zero_points); affine_quant->quantized_dimension = quantized_dimension; @@ -598,3 +400,5 @@ TfLiteTensor CreateSymmetricPerChannelQuantizedTensor( } // namespace testing } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/test_helpers.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.h similarity index 81% rename from src/tensorflow/lite/micro/test_helpers.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.h index 243f9fd..76fce02 100644 --- a/src/tensorflow/lite/micro/test_helpers.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -13,32 +14,23 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_TEST_HELPERS_H_ -#define TENSORFLOW_LITE_MICRO_TEST_HELPERS_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TEST_HELPERS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TEST_HELPERS_H_ // Useful functions for writing tests. -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/error_reporter.h" -#include "tensorflow/lite/schema/schema_generated.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/schema/schema_generated.h" namespace tflite { namespace testing { -// Returns a simple example flatbuffer TensorFlow Lite model. Contains 1 input, -// 1 layer of weights, 1 output Tensor, and 1 operator. -const Model* GetSimpleMockModel(); - -// Returns a flatbuffer TensorFlow Lite model with more inputs, variable -// tensors, and operators. -const Model* GetComplexMockModel(); +// Returns an example flatbuffer TensorFlow Lite model. +const Model* GetMockModel(); // Builds a one-dimensional flatbuffer tensor of the given size. -const Tensor* Create1dFlatbufferTensor(int size, bool is_variable = false); - -// Builds a one-dimensional flatbuffer tensor of the given size with -// quantization metadata. -const Tensor* CreateQuantizedFlatbufferTensor(int size); +const Tensor* Create1dFlatbufferTensor(int size); // Creates a one-dimensional tensor with no quantization metadata. const Tensor* CreateMissingQuantizationFlatbufferTensor(int size); @@ -88,10 +80,6 @@ TfLiteTensor CreateQuantizedTensor(const int8_t* data, TfLiteIntArray* dims, float scale, int zero_point, const char* name, bool is_variable = false); -TfLiteTensor CreateQuantizedTensor(const int16_t* data, TfLiteIntArray* dims, - float scale, int zero_point, - const char* name, bool is_variable = false); - TfLiteTensor CreateQuantizedTensor(const float* input, int8_t* quantized, TfLiteIntArray* dims, float scale, int zero_point, const char* name, @@ -118,4 +106,6 @@ TfLiteTensor CreateSymmetricPerChannelQuantizedTensor( } // namespace testing } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_TEST_HELPERS_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TEST_HELPERS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/testing/micro_test.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/testing/micro_test.h similarity index 95% rename from src/tensorflow/lite/micro/testing/micro_test.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/testing/micro_test.h index 695bc46..14bbece 100644 --- a/src/tensorflow/lite/micro/testing/micro_test.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/testing/micro_test.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -51,10 +52,10 @@ limitations under the License. // all on systems that struggle to run more conventional approaches, so use with // caution! -#ifndef TENSORFLOW_LITE_MICRO_TESTING_MICRO_TEST_H_ -#define TENSORFLOW_LITE_MICRO_TESTING_MICRO_TEST_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TESTING_MICRO_TEST_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TESTING_MICRO_TEST_H_ -#include "tensorflow/lite/micro/micro_error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h" namespace micro_test { extern int tests_passed; @@ -191,7 +192,7 @@ extern tflite::ErrorReporter* reporter; #define TF_LITE_MICRO_EXPECT_TRUE(x) \ do { \ - if (!(x)) { \ + if (!x) { \ micro_test::reporter->Report(#x " was not true failed at %s:%d", \ __FILE__, __LINE__); \ micro_test::did_test_fail = true; \ @@ -207,10 +208,6 @@ extern tflite::ErrorReporter* reporter; } \ } while (false) -#define TF_LITE_MICRO_FAIL(msg) \ - do { \ - micro_test::reporter->Report("FAIL: %s", msg, __FILE__, __LINE__); \ - micro_test::did_test_fail = true; \ - } while (false) +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TESTING_MICRO_TEST_H_ -#endif // TENSORFLOW_LITE_MICRO_TESTING_MICRO_TEST_H_ +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/micro/testing/test_utils.h b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/testing/test_utils.h similarity index 90% rename from src/tensorflow/lite/micro/testing/test_utils.h rename to src/tensorflow_esp32/tensorflow/lite/experimental/micro/testing/test_utils.h index 47535d5..d2c2634 100644 --- a/src/tensorflow/lite/micro/testing/test_utils.h +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/micro/testing/test_utils.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,19 +13,19 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#ifndef TENSORFLOW_LITE_MICRO_TESTING_TEST_UTILS_H_ -#define TENSORFLOW_LITE_MICRO_TESTING_TEST_UTILS_H_ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TESTING_TEST_UTILS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TESTING_TEST_UTILS_H_ #include #include #include #include -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/core/api/tensor_utils.h" -#include "tensorflow/lite/micro/micro_utils.h" -#include "tensorflow/lite/micro/test_helpers.h" -#include "tensorflow/lite/micro/testing/micro_test.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/core/api/tensor_utils.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_utils.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/test_helpers.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/testing/micro_test.h" namespace tflite { namespace testing { @@ -215,24 +216,6 @@ inline TfLiteTensor CreateQuantizedTensor(float* data, int8_t* quantized_data, return result; } -inline TfLiteTensor CreateQuantizedTensor(float* data, int16_t* quantized_data, - TfLiteIntArray* dims, - const char* name, - bool is_variable = false) { - TfLiteTensor result; - SignedSymmetricQuantize(data, dims, quantized_data, &result.params.scale); - result.data.i16 = quantized_data; - result.type = kTfLiteInt16; - result.dims = dims; - result.params.zero_point = 0; - result.allocation_type = kTfLiteMemNone; - result.bytes = ElementCount(*dims) * sizeof(int16_t); - result.allocation = nullptr; - result.name = name; - result.is_variable = is_variable; - return result; -} - inline TfLiteTensor CreateQuantized32Tensor(const int32_t* data, TfLiteIntArray* dims, const char* name, float scale, @@ -288,4 +271,6 @@ inline TfLiteTensor CreateTensor(std::initializer_list data, } // namespace testing } // namespace tflite -#endif // TENSORFLOW_LITE_MICRO_TESTING_TEST_UTILS_H_ +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICRO_TESTING_TEST_UTILS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h new file mode 100644 index 0000000..31455cc --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h @@ -0,0 +1,105 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_ + +#ifdef __cplusplus +#include + +extern "C" { +#endif + +static inline int CountLeadingZeros32Slow(uint64_t n) { + int zeroes = 28; + if (n >> 16) zeroes -= 16, n >>= 16; + if (n >> 8) zeroes -= 8, n >>= 8; + if (n >> 4) zeroes -= 4, n >>= 4; + return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[n] + zeroes; +} + +static inline int CountLeadingZeros32(uint32_t n) { +#if defined(_MSC_VER) + unsigned long result = 0; // NOLINT(runtime/int) + if (_BitScanReverse(&result, n)) { + return 31 - result; + } + return 32; +#elif defined(__GNUC__) + + // Handle 0 as a special case because __builtin_clz(0) is undefined. + if (n == 0) { + return 32; + } + return __builtin_clz(n); +#else + return CountLeadingZeros32Slow(n); +#endif +} + +static inline int MostSignificantBit32(uint32_t n) { + return 32 - CountLeadingZeros32(n); +} + +static inline int CountLeadingZeros64Slow(uint64_t n) { + int zeroes = 60; + if (n >> 32) zeroes -= 32, n >>= 32; + if (n >> 16) zeroes -= 16, n >>= 16; + if (n >> 8) zeroes -= 8, n >>= 8; + if (n >> 4) zeroes -= 4, n >>= 4; + return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[n] + zeroes; +} + +static inline int CountLeadingZeros64(uint64_t n) { +#if defined(_MSC_VER) && defined(_M_X64) + // MSVC does not have __builtin_clzll. Use _BitScanReverse64. + unsigned long result = 0; // NOLINT(runtime/int) + if (_BitScanReverse64(&result, n)) { + return 63 - result; + } + return 64; +#elif defined(_MSC_VER) + // MSVC does not have __builtin_clzll. Compose two calls to _BitScanReverse + unsigned long result = 0; // NOLINT(runtime/int) + if ((n >> 32) && _BitScanReverse(&result, n >> 32)) { + return 31 - result; + } + if (_BitScanReverse(&result, n)) { + return 63 - result; + } + return 64; +#elif defined(__GNUC__) + + // Handle 0 as a special case because __builtin_clzll(0) is undefined. + if (n == 0) { + return 64; + } + return __builtin_clzll(n); +#else + return CountLeadingZeros64Slow(n); +#endif +} + +static inline int MostSignificantBit64(uint64_t n) { + return 64 - CountLeadingZeros64(n); +} + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.cpp new file mode 100644 index 0000000..fd8970b --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.cpp @@ -0,0 +1,56 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h" + +#include + +#define FIXED_POINT 16 +#include "tensorflow_esp32/third_party/kissfft/kiss_fft.h" +#include "tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.h" + +void FftCompute(struct FftState* state, const int16_t* input, + int input_scale_shift) { + const size_t input_size = state->input_size; + const size_t fft_size = state->fft_size; + + int16_t* fft_input = state->input; + // First, scale the input by the given shift. + int i; + for (i = 0; i < input_size; ++i) { + *fft_input++ = (*input++) << input_scale_shift; + } + // Zero out whatever else remains in the top part of the input. + for (; i < fft_size; ++i) { + *fft_input++ = 0; + } + + // Apply the FFT. + kiss_fftr( + reinterpret_cast(state->scratch), + state->input, + reinterpret_cast(state->output)); +} + +void FftInit(struct FftState* state) { + // All the initialization is done in FftPopulateState() +} + +void FftReset(struct FftState* state) { + memset(state->input, 0, state->fft_size * sizeof(*state->input)); + memset(state->output, 0, (state->fft_size / 2 + 1) * sizeof(*state->output)); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h new file mode 100644 index 0000000..e9fe68a --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h @@ -0,0 +1,53 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_ + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct complex_int16_t { + int16_t real; + int16_t imag; +}; + +struct FftState { + int16_t* input; + struct complex_int16_t* output; + size_t fft_size; + size_t input_size; + void* scratch; + size_t scratch_size; +}; + +void FftCompute(struct FftState* state, const int16_t* input, + int input_scale_shift); + +void FftInit(struct FftState* state); + +void FftReset(struct FftState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp new file mode 100644 index 0000000..939202c --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.cpp @@ -0,0 +1,75 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.h" + +#include + +#define FIXED_POINT 16 +#include "tensorflow_esp32/third_party/kissfft/kiss_fft.h" +#include "tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.h" + +int FftPopulateState(struct FftState* state, size_t input_size) { + state->input_size = input_size; + state->fft_size = 1; + while (state->fft_size < state->input_size) { + state->fft_size <<= 1; + } + + state->input = reinterpret_cast( + malloc(state->fft_size * sizeof(*state->input))); + if (state->input == nullptr) { + fprintf(stderr, "Failed to alloc fft input buffer\n"); + return 0; + } + + state->output = reinterpret_cast( + malloc((state->fft_size / 2 + 1) * sizeof(*state->output) * 2)); + if (state->output == nullptr) { + fprintf(stderr, "Failed to alloc fft output buffer\n"); + return 0; + } + + // Ask kissfft how much memory it wants. + size_t scratch_size = 0; + kiss_fftr_cfg kfft_cfg = kiss_fftr_alloc( + state->fft_size, 0, nullptr, &scratch_size); + if (kfft_cfg != nullptr) { + fprintf(stderr, "Kiss memory sizing failed.\n"); + return 0; + } + state->scratch = malloc(scratch_size); + if (state->scratch == nullptr) { + fprintf(stderr, "Failed to alloc fft scratch buffer\n"); + return 0; + } + state->scratch_size = scratch_size; + // Let kissfft configure the scratch space we just allocated + kfft_cfg = kiss_fftr_alloc(state->fft_size, 0, + state->scratch, &scratch_size); + if (kfft_cfg != state->scratch) { + fprintf(stderr, "Kiss memory preallocation strategy failed.\n"); + return 0; + } + return 1; +} + +void FftFreeStateContents(struct FftState* state) { + free(state->input); + free(state->output); + free(state->scratch); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.h new file mode 100644 index 0000000..7e57cc8 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.h @@ -0,0 +1,37 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_ + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h" + +#ifdef __cplusplus +extern "C" { +#endif + +// Prepares and FFT for the given input size. +int FftPopulateState(struct FftState* state, size_t input_size); + +// Frees any allocated buffers. +void FftFreeStateContents(struct FftState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.c new file mode 100644 index 0000000..7ae2bef --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.c @@ -0,0 +1,137 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.h" + +#include + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +void FilterbankConvertFftComplexToEnergy(struct FilterbankState* state, + struct complex_int16_t* fft_output, + int32_t* energy) { + const int end_index = state->end_index; + int i; + energy += state->start_index; + fft_output += state->start_index; + for (i = state->start_index; i < end_index; ++i) { + const int32_t real = fft_output->real; + const int32_t imag = fft_output->imag; + fft_output++; + const uint32_t mag_squared = (real * real) + (imag * imag); + *energy++ = mag_squared; + } +} + +void FilterbankAccumulateChannels(struct FilterbankState* state, + const int32_t* energy) { + uint64_t* work = state->work; + uint64_t weight_accumulator = 0; + uint64_t unweight_accumulator = 0; + + const int16_t* channel_frequency_starts = state->channel_frequency_starts; + const int16_t* channel_weight_starts = state->channel_weight_starts; + const int16_t* channel_widths = state->channel_widths; + + int num_channels_plus_1 = state->num_channels + 1; + int i; + for (i = 0; i < num_channels_plus_1; ++i) { + const int32_t* magnitudes = energy + *channel_frequency_starts++; + const int16_t* weights = state->weights + *channel_weight_starts; + const int16_t* unweights = state->unweights + *channel_weight_starts++; + const int width = *channel_widths++; + int j; + for (j = 0; j < width; ++j) { + weight_accumulator += *weights++ * ((uint64_t)*magnitudes); + unweight_accumulator += *unweights++ * ((uint64_t)*magnitudes); + ++magnitudes; + } + *work++ = weight_accumulator; + weight_accumulator = unweight_accumulator; + unweight_accumulator = 0; + } +} + +static uint16_t Sqrt32(uint32_t num) { + if (num == 0) { + return 0; + } + uint32_t res = 0; + int max_bit_number = 32 - MostSignificantBit32(num); + max_bit_number |= 1; + uint32_t bit = 1U << (31 - max_bit_number); + int iterations = (31 - max_bit_number) / 2 + 1; + while (iterations--) { + if (num >= res + bit) { + num -= res + bit; + res = (res >> 1U) + bit; + } else { + res >>= 1U; + } + bit >>= 2U; + } + // Do rounding - if we have the bits. + if (num > res && res != 0xFFFF) { + ++res; + } + return res; +} + +static uint32_t Sqrt64(uint64_t num) { + // Take a shortcut and just use 32 bit operations if the upper word is all + // clear. This will cause a slight off by one issue for numbers close to 2^32, + // but it probably isn't going to matter (and gives us a big performance win). + if ((num >> 32) == 0) { + return Sqrt32((uint32_t)num); + } + uint64_t res = 0; + int max_bit_number = 64 - MostSignificantBit64(num); + max_bit_number |= 1; + uint64_t bit = 1ULL << (63 - max_bit_number); + int iterations = (63 - max_bit_number) / 2 + 1; + while (iterations--) { + if (num >= res + bit) { + num -= res + bit; + res = (res >> 1U) + bit; + } else { + res >>= 1U; + } + bit >>= 2U; + } + // Do rounding - if we have the bits. + if (num > res && res != 0xFFFFFFFFLL) { + ++res; + } + return res; +} + +uint32_t* FilterbankSqrt(struct FilterbankState* state, int scale_down_shift) { + const int num_channels = state->num_channels; + const int64_t* work = state->work + 1; + // Reuse the work buffer since we're fine clobbering it at this point to hold + // the output. + uint32_t* output = (uint32_t*)state->work; + int i; + for (i = 0; i < num_channels; ++i) { + *output++ = Sqrt64(*work++) >> scale_down_shift; + } + return (uint32_t*)state->work; +} + +void FilterbankReset(struct FilterbankState* state) { + memset(state->work, 0, (state->num_channels + 1) * sizeof(*state->work)); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.h new file mode 100644 index 0000000..9dbb693 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.h @@ -0,0 +1,66 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_ + +#include +#include + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h" + +#define kFilterbankBits 12 + +#ifdef __cplusplus +extern "C" { +#endif + +struct FilterbankState { + int num_channels; + int start_index; + int end_index; + int16_t* channel_frequency_starts; + int16_t* channel_weight_starts; + int16_t* channel_widths; + int16_t* weights; + int16_t* unweights; + uint64_t* work; +}; + +// Converts the relevant complex values of an FFT output into energy (the +// square magnitude). +void FilterbankConvertFftComplexToEnergy(struct FilterbankState* state, + struct complex_int16_t* fft_output, + int32_t* energy); + +// Computes the mel-scale filterbank on the given energy array. Output is cached +// internally - to fetch it, you need to call FilterbankSqrt. +void FilterbankAccumulateChannels(struct FilterbankState* state, + const int32_t* energy); + +// Applies an integer square root to the 64 bit intermediate values of the +// filterbank, and returns a pointer to them. Memory will be invalidated the +// next time FilterbankAccumulateChannels is called. +uint32_t* FilterbankSqrt(struct FilterbankState* state, int scale_down_shift); + +void FilterbankReset(struct FilterbankState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c new file mode 100644 index 0000000..382bb87 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c @@ -0,0 +1,223 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h" + +#include +#include +#include + +#define kFilterbankIndexAlignment 4 +#define kFilterbankChannelBlockSize 4 + +void FilterbankFillConfigWithDefaults(struct FilterbankConfig* config) { + config->num_channels = 32; + config->lower_band_limit = 125.0f; + config->upper_band_limit = 7500.0f; + config->output_scale_shift = 7; +} + +static float FreqToMel(float freq) { return 1127.0 * log1p(freq / 700.0); } + +static void CalculateCenterFrequencies(const int num_channels, + const float lower_frequency_limit, + const float upper_frequency_limit, + float* center_frequencies) { + assert(lower_frequency_limit >= 0.0f); + assert(upper_frequency_limit > lower_frequency_limit); + + const float mel_low = FreqToMel(lower_frequency_limit); + const float mel_hi = FreqToMel(upper_frequency_limit); + const float mel_span = mel_hi - mel_low; + const float mel_spacing = mel_span / ((float)num_channels); + int i; + for (i = 0; i < num_channels; ++i) { + center_frequencies[i] = mel_low + (mel_spacing * (i + 1)); + } +} + +static void QuantizeFilterbankWeights(const float float_weight, int16_t* weight, + int16_t* unweight) { + *weight = floor(float_weight * (1 << kFilterbankBits) + 0.5); + *unweight = floor((1.0 - float_weight) * (1 << kFilterbankBits) + 0.5); +} + +int FilterbankPopulateState(const struct FilterbankConfig* config, + struct FilterbankState* state, int sample_rate, + int spectrum_size) { + state->num_channels = config->num_channels; + const int num_channels_plus_1 = config->num_channels + 1; + + // How should we align things to index counts given the byte alignment? + const int index_alignment = + (kFilterbankIndexAlignment < sizeof(int16_t) + ? 1 + : kFilterbankIndexAlignment / sizeof(int16_t)); + + state->channel_frequency_starts = + malloc(num_channels_plus_1 * sizeof(*state->channel_frequency_starts)); + state->channel_weight_starts = + malloc(num_channels_plus_1 * sizeof(*state->channel_weight_starts)); + state->channel_widths = + malloc(num_channels_plus_1 * sizeof(*state->channel_widths)); + state->work = malloc(num_channels_plus_1 * sizeof(*state->work)); + + float* center_mel_freqs = + malloc(num_channels_plus_1 * sizeof(*center_mel_freqs)); + int16_t* actual_channel_starts = + malloc(num_channels_plus_1 * sizeof(*actual_channel_starts)); + int16_t* actual_channel_widths = + malloc(num_channels_plus_1 * sizeof(*actual_channel_widths)); + + if (state->channel_frequency_starts == NULL || + state->channel_weight_starts == NULL || state->channel_widths == NULL || + center_mel_freqs == NULL || actual_channel_starts == NULL || + actual_channel_widths == NULL) { + free(center_mel_freqs); + free(actual_channel_starts); + free(actual_channel_widths); + fprintf(stderr, "Failed to allocate channel buffers\n"); + return 0; + } + + CalculateCenterFrequencies(num_channels_plus_1, config->lower_band_limit, + config->upper_band_limit, center_mel_freqs); + + // Always exclude DC. + const float hz_per_sbin = 0.5 * sample_rate / ((float)spectrum_size - 1); + state->start_index = 1.5 + config->lower_band_limit / hz_per_sbin; + state->end_index = 0; // Initialized to zero here, but actually set below. + + // For each channel, we need to figure out what frequencies belong to it, and + // how much padding we need to add so that we can efficiently multiply the + // weights and unweights for accumulation. To simplify the multiplication + // logic, all channels will have some multiplication to do (even if there are + // no frequencies that accumulate to that channel) - they will be directed to + // a set of zero weights. + int chan_freq_index_start = state->start_index; + int weight_index_start = 0; + int needs_zeros = 0; + + int chan; + for (chan = 0; chan < num_channels_plus_1; ++chan) { + // Keep jumping frequencies until we overshoot the bound on this channel. + int freq_index = chan_freq_index_start; + while (FreqToMel((freq_index)*hz_per_sbin) <= center_mel_freqs[chan]) { + ++freq_index; + } + + const int width = freq_index - chan_freq_index_start; + actual_channel_starts[chan] = chan_freq_index_start; + actual_channel_widths[chan] = width; + + if (width == 0) { + // This channel doesn't actually get anything from the frequencies, it's + // always zero. We need then to insert some 'zero' weights into the + // output, and just redirect this channel to do a single multiplication at + // this point. For simplicity, the zeros are placed at the beginning of + // the weights arrays, so we have to go and update all the other + // weight_starts to reflect this shift (but only once). + state->channel_frequency_starts[chan] = 0; + state->channel_weight_starts[chan] = 0; + state->channel_widths[chan] = kFilterbankChannelBlockSize; + if (!needs_zeros) { + needs_zeros = 1; + int j; + for (j = 0; j < chan; ++j) { + state->channel_weight_starts[j] += kFilterbankChannelBlockSize; + } + weight_index_start += kFilterbankChannelBlockSize; + } + } else { + // How far back do we need to go to ensure that we have the proper + // alignment? + const int aligned_start = + (chan_freq_index_start / index_alignment) * index_alignment; + const int aligned_width = (chan_freq_index_start - aligned_start + width); + const int padded_width = + (((aligned_width - 1) / kFilterbankChannelBlockSize) + 1) * + kFilterbankChannelBlockSize; + + state->channel_frequency_starts[chan] = aligned_start; + state->channel_weight_starts[chan] = weight_index_start; + state->channel_widths[chan] = padded_width; + weight_index_start += padded_width; + } + chan_freq_index_start = freq_index; + } + + // Allocate the two arrays to store the weights - weight_index_start contains + // the index of what would be the next set of weights that we would need to + // add, so that's how many weights we need to allocate. + state->weights = calloc(weight_index_start, sizeof(*state->weights)); + state->unweights = calloc(weight_index_start, sizeof(*state->unweights)); + + // If the alloc failed, we also need to nuke the arrays. + if (state->weights == NULL || state->unweights == NULL) { + free(center_mel_freqs); + free(actual_channel_starts); + free(actual_channel_widths); + fprintf(stderr, "Failed to allocate weights or unweights\n"); + return 0; + } + + // Next pass, compute all the weights. Since everything has been memset to + // zero, we only need to fill in the weights that correspond to some frequency + // for a channel. + const float mel_low = FreqToMel(config->lower_band_limit); + for (chan = 0; chan < num_channels_plus_1; ++chan) { + int frequency = actual_channel_starts[chan]; + const int num_frequencies = actual_channel_widths[chan]; + const int frequency_offset = + frequency - state->channel_frequency_starts[chan]; + const int weight_start = state->channel_weight_starts[chan]; + const float denom_val = (chan == 0) ? mel_low : center_mel_freqs[chan - 1]; + + int j; + for (j = 0; j < num_frequencies; ++j, ++frequency) { + const float weight = + (center_mel_freqs[chan] - FreqToMel(frequency * hz_per_sbin)) / + (center_mel_freqs[chan] - denom_val); + + // Make the float into an integer for the weights (and unweights). + const int weight_index = weight_start + frequency_offset + j; + QuantizeFilterbankWeights(weight, state->weights + weight_index, + state->unweights + weight_index); + } + if (frequency > state->end_index) { + state->end_index = frequency; + } + } + + free(center_mel_freqs); + free(actual_channel_starts); + free(actual_channel_widths); + if (state->end_index >= spectrum_size) { + fprintf(stderr, "Filterbank end_index is above spectrum size.\n"); + return 0; + } + return 1; +} + +void FilterbankFreeStateContents(struct FilterbankState* state) { + free(state->channel_frequency_starts); + free(state->channel_weight_starts); + free(state->channel_widths); + free(state->weights); + free(state->unweights); + free(state->work); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h new file mode 100644 index 0000000..dea255c --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h @@ -0,0 +1,53 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_ + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct FilterbankConfig { + // number of frequency channel buckets for filterbank + int num_channels; + // maximum frequency to include + float upper_band_limit; + // minimum frequency to include + float lower_band_limit; + // unused + int output_scale_shift; +}; + +// Fills the frontendConfig with "sane" defaults. +void FilterbankFillConfigWithDefaults(struct FilterbankConfig* config); + +// Allocates any buffers. +int FilterbankPopulateState(const struct FilterbankConfig* config, + struct FilterbankState* state, int sample_rate, + int spectrum_size); + +// Frees any allocated buffers. +void FilterbankFreeStateContents(struct FilterbankState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.c new file mode 100644 index 0000000..8314f4f --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.c @@ -0,0 +1,75 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.h" + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +struct FrontendOutput FrontendProcessSamples(struct FrontendState* state, + const int16_t* samples, + size_t num_samples, + size_t* num_samples_read) { + struct FrontendOutput output; + output.values = NULL; + output.size = 0; + + // Try to apply the window - if it fails, return and wait for more data. + if (!WindowProcessSamples(&state->window, samples, num_samples, + num_samples_read)) { + return output; + } + + // Apply the FFT to the window's output (and scale it so that the fixed point + // FFT can have as much resolution as possible). + int input_shift = + 15 - MostSignificantBit32(state->window.max_abs_output_value); + FftCompute(&state->fft, state->window.output, input_shift); + + // We can re-ruse the fft's output buffer to hold the energy. + int32_t* energy = (int32_t*)state->fft.output; + + FilterbankConvertFftComplexToEnergy(&state->filterbank, state->fft.output, + energy); + + FilterbankAccumulateChannels(&state->filterbank, energy); + uint32_t* scaled_filterbank = FilterbankSqrt(&state->filterbank, input_shift); + + // Apply noise reduction. + NoiseReductionApply(&state->noise_reduction, scaled_filterbank); + + if (state->pcan_gain_control.enable_pcan) { + PcanGainControlApply(&state->pcan_gain_control, scaled_filterbank); + } + + // Apply the log and scale. + int correction_bits = + MostSignificantBit32(state->fft.fft_size) - 1 - (kFilterbankBits / 2); + uint16_t* logged_filterbank = + LogScaleApply(&state->log_scale, scaled_filterbank, + state->filterbank.num_channels, correction_bits); + + output.size = state->filterbank.num_channels; + output.values = logged_filterbank; + return output; +} + +void FrontendReset(struct FrontendState* state) { + WindowReset(&state->window); + FftReset(&state->fft); + FilterbankReset(&state->filterbank); + NoiseReductionReset(&state->noise_reduction); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.h new file mode 100644 index 0000000..a5b8e0f --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.h @@ -0,0 +1,67 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_ + +#include +#include + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct FrontendState { + struct WindowState window; + struct FftState fft; + struct FilterbankState filterbank; + struct NoiseReductionState noise_reduction; + struct PcanGainControlState pcan_gain_control; + struct LogScaleState log_scale; +}; + +struct FrontendOutput { + const uint16_t* values; + size_t size; +}; + +// Main entry point to processing frontend samples. Updates num_samples_read to +// contain the number of samples that have been consumed from the input array. +// Returns a struct containing the generated output. If not enough samples were +// added to generate a feature vector, the returned size will be 0 and the +// values pointer will be NULL. Note that the output pointer will be invalidated +// as soon as FrontendProcessSamples is called again, so copy the contents +// elsewhere if you need to use them later. +struct FrontendOutput FrontendProcessSamples(struct FrontendState* state, + const int16_t* samples, + size_t num_samples, + size_t* num_samples_read); + +void FrontendReset(struct FrontendState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c new file mode 100644 index 0000000..55126f7 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c @@ -0,0 +1,88 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h" + +#include +#include + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +void FrontendFillConfigWithDefaults(struct FrontendConfig* config) { + WindowFillConfigWithDefaults(&config->window); + FilterbankFillConfigWithDefaults(&config->filterbank); + NoiseReductionFillConfigWithDefaults(&config->noise_reduction); + PcanGainControlFillConfigWithDefaults(&config->pcan_gain_control); + LogScaleFillConfigWithDefaults(&config->log_scale); +} + +int FrontendPopulateState(const struct FrontendConfig* config, + struct FrontendState* state, int sample_rate) { + memset(state, 0, sizeof(*state)); + + if (!WindowPopulateState(&config->window, &state->window, sample_rate)) { + fprintf(stderr, "Failed to populate window state\n"); + return 0; + } + + if (!FftPopulateState(&state->fft, state->window.size)) { + fprintf(stderr, "Failed to populate fft state\n"); + return 0; + } + FftInit(&state->fft); + + if (!FilterbankPopulateState(&config->filterbank, &state->filterbank, + sample_rate, state->fft.fft_size / 2 + 1)) { + fprintf(stderr, "Failed to populate filterbank state\n"); + return 0; + } + + if (!NoiseReductionPopulateState(&config->noise_reduction, + &state->noise_reduction, + state->filterbank.num_channels)) { + fprintf(stderr, "Failed to populate noise reduction state\n"); + return 0; + } + + int input_correction_bits = + MostSignificantBit32(state->fft.fft_size) - 1 - (kFilterbankBits / 2); + if (!PcanGainControlPopulateState( + &config->pcan_gain_control, &state->pcan_gain_control, + state->noise_reduction.estimate, state->filterbank.num_channels, + state->noise_reduction.smoothing_bits, input_correction_bits)) { + fprintf(stderr, "Failed to populate pcan gain control state\n"); + return 0; + } + + if (!LogScalePopulateState(&config->log_scale, &state->log_scale)) { + fprintf(stderr, "Failed to populate log scale state\n"); + return 0; + } + + FrontendReset(state); + + // All good, return a true value. + return 1; +} + +void FrontendFreeStateContents(struct FrontendState* state) { + WindowFreeStateContents(&state->window); + FftFreeStateContents(&state->fft); + FilterbankFreeStateContents(&state->filterbank); + NoiseReductionFreeStateContents(&state->noise_reduction); + PcanGainControlFreeStateContents(&state->pcan_gain_control); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h new file mode 100644 index 0000000..598c6e3 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h @@ -0,0 +1,55 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_ + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/fft_util.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/frontend.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct FrontendConfig { + struct WindowConfig window; + struct FilterbankConfig filterbank; + struct NoiseReductionConfig noise_reduction; + struct PcanGainControlConfig pcan_gain_control; + struct LogScaleConfig log_scale; +}; + +// Fills the frontendConfig with "sane" defaults. +void FrontendFillConfigWithDefaults(struct FrontendConfig* config); + +// Allocates any buffers. +int FrontendPopulateState(const struct FrontendConfig* config, + struct FrontendState* state, int sample_rate); + +// Frees any allocated buffers. +void FrontendFreeStateContents(struct FrontendState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.c new file mode 100644 index 0000000..613dded --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.c @@ -0,0 +1,33 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.h" +const uint16_t kLogLut[] +#ifndef _MSC_VER + __attribute__((aligned(4))) +#endif // _MSV_VER + = {0, 224, 442, 654, 861, 1063, 1259, 1450, 1636, 1817, 1992, 2163, + 2329, 2490, 2646, 2797, 2944, 3087, 3224, 3358, 3487, 3611, 3732, 3848, + 3960, 4068, 4172, 4272, 4368, 4460, 4549, 4633, 4714, 4791, 4864, 4934, + 5001, 5063, 5123, 5178, 5231, 5280, 5326, 5368, 5408, 5444, 5477, 5507, + 5533, 5557, 5578, 5595, 5610, 5622, 5631, 5637, 5640, 5641, 5638, 5633, + 5626, 5615, 5602, 5586, 5568, 5547, 5524, 5498, 5470, 5439, 5406, 5370, + 5332, 5291, 5249, 5203, 5156, 5106, 5054, 5000, 4944, 4885, 4825, 4762, + 4697, 4630, 4561, 4490, 4416, 4341, 4264, 4184, 4103, 4020, 3935, 3848, + 3759, 3668, 3575, 3481, 3384, 3286, 3186, 3084, 2981, 2875, 2768, 2659, + 2549, 2437, 2323, 2207, 2090, 1971, 1851, 1729, 1605, 1480, 1353, 1224, + 1094, 963, 830, 695, 559, 421, 282, 142, 0, 0}; + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.h new file mode 100644 index 0000000..2baee0f --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.h @@ -0,0 +1,43 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_ + +#include + +#ifdef __cplusplus +extern "C" { +#endif + +// Number of segments in the log lookup table. The table will be kLogSegments+1 +// in length (with some padding). +#define kLogSegments 128 +#define kLogSegmentsLog2 7 + +// Scale used by lookup table. +#define kLogScale 65536 +#define kLogScaleLog2 16 +#define kLogCoeff 45426 + +extern const uint16_t kLogLut[]; + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.c new file mode 100644 index 0000000..fed741a --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.c @@ -0,0 +1,86 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.h" + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_lut.h" + +#define kuint16max 0x0000FFFF + +// The following functions implement integer logarithms of various sizes. The +// approximation is calculated according to method described in +// www.inti.gob.ar/electronicaeinformatica/instrumentacion/utic/ +// publicaciones/SPL2007/Log10-spl07.pdf +// It first calculates log2 of the input and then converts it to natural +// logarithm. + +static uint32_t Log2FractionPart(const uint32_t x, const uint32_t log2x) { + // Part 1 + int32_t frac = x - (1LL << log2x); + if (log2x < kLogScaleLog2) { + frac <<= kLogScaleLog2 - log2x; + } else { + frac >>= log2x - kLogScaleLog2; + } + // Part 2 + const uint32_t base_seg = frac >> (kLogScaleLog2 - kLogSegmentsLog2); + const uint32_t seg_unit = + (((uint32_t)1) << kLogScaleLog2) >> kLogSegmentsLog2; + + const int32_t c0 = kLogLut[base_seg]; + const int32_t c1 = kLogLut[base_seg + 1]; + const int32_t seg_base = seg_unit * base_seg; + const int32_t rel_pos = ((c1 - c0) * (frac - seg_base)) >> kLogScaleLog2; + return frac + c0 + rel_pos; +} + +static uint32_t Log(const uint32_t x, const uint32_t scale_shift) { + const uint32_t integer = MostSignificantBit32(x) - 1; + const uint32_t fraction = Log2FractionPart(x, integer); + const uint32_t log2 = (integer << kLogScaleLog2) + fraction; + const uint32_t round = kLogScale / 2; + const uint32_t loge = (((uint64_t)kLogCoeff) * log2 + round) >> kLogScaleLog2; + // Finally scale to our output scale + const uint32_t loge_scaled = ((loge << scale_shift) + round) >> kLogScaleLog2; + return loge_scaled; +} + +uint16_t* LogScaleApply(struct LogScaleState* state, uint32_t* signal, + int signal_size, int correction_bits) { + const int scale_shift = state->scale_shift; + uint16_t* output = (uint16_t*)signal; + uint16_t* ret = output; + int i; + for (i = 0; i < signal_size; ++i) { + uint32_t value = *signal++; + if (state->enable_log) { + if (correction_bits < 0) { + value >>= -correction_bits; + } else { + value <<= correction_bits; + } + if (value > 1) { + value = Log(value, scale_shift); + } else { + value = 0; + } + } + *output++ = (value < kuint16max) ? value : kuint16max; + } + return ret; +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.h new file mode 100644 index 0000000..770f88f --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.h @@ -0,0 +1,42 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_ + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct LogScaleState { + int enable_log; + int scale_shift; +}; + +// Applies a fixed point logarithm to the signal and converts it to 16 bit. Note +// that the signal array will be modified. +uint16_t* LogScaleApply(struct LogScaleState* state, uint32_t* signal, + int signal_size, int correction_bits); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c new file mode 100644 index 0000000..cac82a0 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c @@ -0,0 +1,30 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h" + +void LogScaleFillConfigWithDefaults(struct LogScaleConfig* config) { + config->enable_log = 1; + config->scale_shift = 6; +} + +int LogScalePopulateState(const struct LogScaleConfig* config, + struct LogScaleState* state) { + state->enable_log = config->enable_log; + state->scale_shift = config->scale_shift; + return 1; +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h new file mode 100644 index 0000000..6653536 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h @@ -0,0 +1,48 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_ + +#include +#include + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/log_scale.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct LogScaleConfig { + // set to false (0) to disable this module + int enable_log; + // scale results by 2^(scale_shift) + int scale_shift; +}; + +// Populates the LogScaleConfig with "sane" default values. +void LogScaleFillConfigWithDefaults(struct LogScaleConfig* config); + +// Allocates any buffers. +int LogScalePopulateState(const struct LogScaleConfig* config, + struct LogScaleState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c new file mode 100644 index 0000000..aeedcd5 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c @@ -0,0 +1,54 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h" + +#include + +void NoiseReductionApply(struct NoiseReductionState* state, uint32_t* signal) { + int i; + for (i = 0; i < state->num_channels; ++i) { + const uint32_t smoothing = + ((i & 1) == 0) ? state->even_smoothing : state->odd_smoothing; + const uint32_t one_minus_smoothing = (1 << kNoiseReductionBits) - smoothing; + + // Update the estimate of the noise. + const uint32_t signal_scaled_up = signal[i] << state->smoothing_bits; + uint32_t estimate = + (((uint64_t)signal_scaled_up * smoothing) + + ((uint64_t)state->estimate[i] * one_minus_smoothing)) >> + kNoiseReductionBits; + state->estimate[i] = estimate; + + // Make sure that we can't get a negative value for the signal - estimate. + if (estimate > signal_scaled_up) { + estimate = signal_scaled_up; + } + + const uint32_t floor = + ((uint64_t)signal[i] * state->min_signal_remaining) >> + kNoiseReductionBits; + const uint32_t subtracted = + (signal_scaled_up - estimate) >> state->smoothing_bits; + const uint32_t output = subtracted > floor ? subtracted : floor; + signal[i] = output; + } +} + +void NoiseReductionReset(struct NoiseReductionState* state) { + memset(state->estimate, 0, sizeof(*state->estimate) * state->num_channels); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h new file mode 100644 index 0000000..71eaa7e --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h @@ -0,0 +1,49 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_ + +#define kNoiseReductionBits 14 + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct NoiseReductionState { + int smoothing_bits; + uint16_t even_smoothing; + uint16_t odd_smoothing; + uint16_t min_signal_remaining; + int num_channels; + uint32_t* estimate; +}; + +// Removes stationary noise from each channel of the signal using a low pass +// filter. +void NoiseReductionApply(struct NoiseReductionState* state, uint32_t* signal); + +void NoiseReductionReset(struct NoiseReductionState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c new file mode 100644 index 0000000..baf8f93 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c @@ -0,0 +1,48 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h" + +#include + +void NoiseReductionFillConfigWithDefaults(struct NoiseReductionConfig* config) { + config->smoothing_bits = 10; + config->even_smoothing = 0.025; + config->odd_smoothing = 0.06; + config->min_signal_remaining = 0.05; +} + +int NoiseReductionPopulateState(const struct NoiseReductionConfig* config, + struct NoiseReductionState* state, + int num_channels) { + state->smoothing_bits = config->smoothing_bits; + state->odd_smoothing = config->odd_smoothing * (1 << kNoiseReductionBits); + state->even_smoothing = config->even_smoothing * (1 << kNoiseReductionBits); + state->min_signal_remaining = + config->min_signal_remaining * (1 << kNoiseReductionBits); + state->num_channels = num_channels; + state->estimate = calloc(state->num_channels, sizeof(*state->estimate)); + if (state->estimate == NULL) { + fprintf(stderr, "Failed to alloc estimate buffer\n"); + return 0; + } + return 1; +} + +void NoiseReductionFreeStateContents(struct NoiseReductionState* state) { + free(state->estimate); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h new file mode 100644 index 0000000..384785c --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h @@ -0,0 +1,53 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_ + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct NoiseReductionConfig { + // scale the signal up by 2^(smoothing_bits) before reduction + int smoothing_bits; + // smoothing coefficient for even-numbered channels + float even_smoothing; + // smoothing coefficient for odd-numbered channels + float odd_smoothing; + // fraction of signal to preserve (1.0 disables this module) + float min_signal_remaining; +}; + +// Populates the NoiseReductionConfig with "sane" default values. +void NoiseReductionFillConfigWithDefaults(struct NoiseReductionConfig* config); + +// Allocates any buffers. +int NoiseReductionPopulateState(const struct NoiseReductionConfig* config, + struct NoiseReductionState* state, + int num_channels); + +// Frees any allocated buffers. +void NoiseReductionFreeStateContents(struct NoiseReductionState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c new file mode 100644 index 0000000..70d1d85 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c @@ -0,0 +1,59 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h" + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/bits.h" + +int16_t WideDynamicFunction(const uint32_t x, const int16_t* lut) { + if (x <= 2) { + return lut[x]; + } + + const int16_t interval = MostSignificantBit32(x); + lut += 4 * interval - 6; + + const int16_t frac = + ((interval < 11) ? (x << (11 - interval)) : (x >> (interval - 11))) & + 0x3FF; + + int32_t result = ((int32_t)lut[2] * frac) >> 5; + result += ((int32_t)lut[1]) << 5; + result *= frac; + result = (result + (1 << 14)) >> 15; + result += lut[0]; + return (int16_t)result; +} + +uint32_t PcanShrink(const uint32_t x) { + if (x < (2 << kPcanSnrBits)) { + return (x * x) >> (2 + 2 * kPcanSnrBits - kPcanOutputBits); + } else { + return (x >> (kPcanSnrBits - kPcanOutputBits)) - (1 << kPcanOutputBits); + } +} + +void PcanGainControlApply(struct PcanGainControlState* state, + uint32_t* signal) { + int i; + for (i = 0; i < state->num_channels; ++i) { + const uint32_t gain = + WideDynamicFunction(state->noise_estimate[i], state->gain_lut); + const uint32_t snr = ((uint64_t)signal[i] * gain) >> state->snr_shift; + signal[i] = PcanShrink(snr); + } +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h new file mode 100644 index 0000000..79c8934 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h @@ -0,0 +1,49 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_ + +#include +#include + +#define kPcanSnrBits 12 +#define kPcanOutputBits 6 + +#ifdef __cplusplus +extern "C" { +#endif + +struct PcanGainControlState { + int enable_pcan; + uint32_t* noise_estimate; + int num_channels; + int16_t* gain_lut; + int32_t snr_shift; +}; + +int16_t WideDynamicFunction(const uint32_t x, const int16_t* lut); + +uint32_t PcanShrink(const uint32_t x); + +void PcanGainControlApply(struct PcanGainControlState* state, uint32_t* signal); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c new file mode 100644 index 0000000..699ec0d --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c @@ -0,0 +1,95 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h" + +#include +#include + +#define kint16max 0x00007FFF + +void PcanGainControlFillConfigWithDefaults( + struct PcanGainControlConfig* config) { + config->enable_pcan = 0; + config->strength = 0.95; + config->offset = 80.0; + config->gain_bits = 21; +} + +int16_t PcanGainLookupFunction(const struct PcanGainControlConfig* config, + int32_t input_bits, uint32_t x) { + const float x_as_float = ((float)x) / ((uint32_t)1 << input_bits); + const float gain_as_float = + ((uint32_t)1 << config->gain_bits) * + powf(x_as_float + config->offset, -config->strength); + + if (gain_as_float > kint16max) { + return kint16max; + } + return (int16_t)(gain_as_float + 0.5f); +} + +int PcanGainControlPopulateState(const struct PcanGainControlConfig* config, + struct PcanGainControlState* state, + uint32_t* noise_estimate, + const int num_channels, + const uint16_t smoothing_bits, + const int32_t input_correction_bits) { + state->enable_pcan = config->enable_pcan; + if (!state->enable_pcan) { + return 1; + } + state->noise_estimate = noise_estimate; + state->num_channels = num_channels; + state->gain_lut = malloc(kWideDynamicFunctionLUTSize * sizeof(int16_t)); + if (state->gain_lut == NULL) { + fprintf(stderr, "Failed to allocate gain LUT\n"); + return 0; + } + state->snr_shift = config->gain_bits - input_correction_bits - kPcanSnrBits; + + const int32_t input_bits = smoothing_bits - input_correction_bits; + state->gain_lut[0] = PcanGainLookupFunction(config, input_bits, 0); + state->gain_lut[1] = PcanGainLookupFunction(config, input_bits, 1); + state->gain_lut -= 6; + int interval; + for (interval = 2; interval <= kWideDynamicFunctionBits; ++interval) { + const uint32_t x0 = (uint32_t)1 << (interval - 1); + const uint32_t x1 = x0 + (x0 >> 1); + const uint32_t x2 = + (interval == kWideDynamicFunctionBits) ? x0 + (x0 - 1) : 2 * x0; + + const int16_t y0 = PcanGainLookupFunction(config, input_bits, x0); + const int16_t y1 = PcanGainLookupFunction(config, input_bits, x1); + const int16_t y2 = PcanGainLookupFunction(config, input_bits, x2); + + const int32_t diff1 = (int32_t)y1 - y0; + const int32_t diff2 = (int32_t)y2 - y0; + const int32_t a1 = 4 * diff1 - diff2; + const int32_t a2 = diff2 - a1; + + state->gain_lut[4 * interval] = y0; + state->gain_lut[4 * interval + 1] = (int16_t)a1; + state->gain_lut[4 * interval + 2] = (int16_t)a2; + } + state->gain_lut += 6; + return 1; +} + +void PcanGainControlFreeStateContents(struct PcanGainControlState* state) { + free(state->gain_lut); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h new file mode 100644 index 0000000..fff8bae --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h @@ -0,0 +1,60 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_ + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h" + +#define kWideDynamicFunctionBits 32 +#define kWideDynamicFunctionLUTSize (4 * kWideDynamicFunctionBits - 3) + +#ifdef __cplusplus +extern "C" { +#endif + +struct PcanGainControlConfig { + // set to false (0) to disable this module + int enable_pcan; + // gain normalization exponent (0.0 disables, 1.0 full strength) + float strength; + // positive value added in the normalization denominator + float offset; + // number of fractional bits in the gain + int gain_bits; +}; + +void PcanGainControlFillConfigWithDefaults( + struct PcanGainControlConfig* config); + +int16_t PcanGainLookupFunction(const struct PcanGainControlConfig* config, + int32_t input_bits, uint32_t x); + +int PcanGainControlPopulateState(const struct PcanGainControlConfig* config, + struct PcanGainControlState* state, + uint32_t* noise_estimate, + const int num_channels, + const uint16_t smoothing_bits, + const int32_t input_correction_bits); + +void PcanGainControlFreeStateContents(struct PcanGainControlState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.c new file mode 100644 index 0000000..56c0fe5 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.c @@ -0,0 +1,73 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.h" + +#include + +int WindowProcessSamples(struct WindowState* state, const int16_t* samples, + size_t num_samples, size_t* num_samples_read) { + const int size = state->size; + + // Copy samples from the samples buffer over to our local input. + size_t max_samples_to_copy = state->size - state->input_used; + if (max_samples_to_copy > num_samples) { + max_samples_to_copy = num_samples; + } + memcpy(state->input + state->input_used, samples, + max_samples_to_copy * sizeof(*samples)); + *num_samples_read = max_samples_to_copy; + state->input_used += max_samples_to_copy; + + if (state->input_used < state->size) { + // We don't have enough samples to compute a window. + return 0; + } + + // Apply the window to the input. + const int16_t* coefficients = state->coefficients; + const int16_t* input = state->input; + int16_t* output = state->output; + int i; + int16_t max_abs_output_value = 0; + for (i = 0; i < size; ++i) { + int16_t new_value = + (((int32_t)*input++) * *coefficients++) >> kFrontendWindowBits; + *output++ = new_value; + if (new_value < 0) { + new_value = -new_value; + } + if (new_value > max_abs_output_value) { + max_abs_output_value = new_value; + } + } + // Shuffle the input down by the step size, and update how much we have used. + memmove(state->input, state->input + state->step, + sizeof(*state->input) * (state->size - state->step)); + state->input_used -= state->step; + state->max_abs_output_value = max_abs_output_value; + + // Indicate that the output buffer is valid for the next stage. + return 1; +} + +void WindowReset(struct WindowState* state) { + memset(state->input, 0, state->size * sizeof(*state->input)); + memset(state->output, 0, state->size * sizeof(*state->output)); + state->input_used = 0; + state->max_abs_output_value = 0; +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.h new file mode 100644 index 0000000..669fcc4 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.h @@ -0,0 +1,52 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_ + +#include +#include + +#define kFrontendWindowBits 12 + +#ifdef __cplusplus +extern "C" { +#endif + +struct WindowState { + size_t size; + int16_t* coefficients; + size_t step; + + int16_t* input; + size_t input_used; + int16_t* output; + int16_t max_abs_output_value; +}; + +// Applies a window to the samples coming in, stepping forward at the given +// rate. +int WindowProcessSamples(struct WindowState* state, const int16_t* samples, + size_t num_samples, size_t* num_samples_read); + +void WindowReset(struct WindowState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.c b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.c new file mode 100644 index 0000000..85a50e7 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.c @@ -0,0 +1,76 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.h" + +#include +#include +#include +#include + +// Some platforms don't have M_PI +#ifndef M_PI +#define M_PI 3.14159265358979323846 +#endif + +void WindowFillConfigWithDefaults(struct WindowConfig* config) { + config->size_ms = 25; + config->step_size_ms = 10; +} + +int WindowPopulateState(const struct WindowConfig* config, + struct WindowState* state, int sample_rate) { + state->size = config->size_ms * sample_rate / 1000; + state->step = config->step_size_ms * sample_rate / 1000; + + state->coefficients = malloc(state->size * sizeof(*state->coefficients)); + if (state->coefficients == NULL) { + fprintf(stderr, "Failed to allocate window coefficients\n"); + return 0; + } + + // Populate the window values. + const float arg = M_PI * 2.0 / ((float)state->size); + int i; + for (i = 0; i < state->size; ++i) { + float float_value = 0.5 - (0.5 * cos(arg * (i + 0.5))); + // Scale it to fixed point and round it. + state->coefficients[i] = + floor(float_value * (1 << kFrontendWindowBits) + 0.5); + } + + state->input_used = 0; + state->input = malloc(state->size * sizeof(*state->input)); + if (state->input == NULL) { + fprintf(stderr, "Failed to allocate window input\n"); + return 0; + } + + state->output = malloc(state->size * sizeof(*state->output)); + if (state->output == NULL) { + fprintf(stderr, "Failed to allocate window output\n"); + return 0; + } + + return 1; +} + +void WindowFreeStateContents(struct WindowState* state) { + free(state->coefficients); + free(state->input); + free(state->output); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.h b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.h new file mode 100644 index 0000000..543021f --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window_util.h @@ -0,0 +1,48 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_ +#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_ + +#include "tensorflow_esp32/tensorflow/lite/experimental/microfrontend/lib/window.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct WindowConfig { + // length of window frame in milliseconds + size_t size_ms; + // length of step for next frame in milliseconds + size_t step_size_ms; +}; + +// Populates the WindowConfig with "sane" default values. +void WindowFillConfigWithDefaults(struct WindowConfig* config); + +// Allocates any buffers. +int WindowPopulateState(const struct WindowConfig* config, + struct WindowState* state, int sample_rate); + +// Frees any allocated buffers. +void WindowFreeStateContents(struct WindowState* state); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/common.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/common.h similarity index 90% rename from src/tensorflow/lite/kernels/internal/common.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/common.h index 8b4ecfa..72b098f 100644 --- a/src/tensorflow/lite/kernels/internal/common.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/common.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -21,9 +22,9 @@ limitations under the License. #endif #endif -#include "fixedpoint/fixedpoint.h" -#include "tensorflow/lite/kernels/internal/optimized/neon_check.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/optimized/neon_check.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -432,23 +433,12 @@ inline int32 GetReciprocal(int32 x, int x_integer_digits, inline void GetInvSqrtQuantizedMultiplierExp(int32 input, int reverse_shift, int32* output_inv_sqrt, int* output_shift) { - TFLITE_DCHECK_GE(input, 0); - if (input <= 1) { - // Handle the input value 1 separately to avoid overflow in that case - // in the general computation below (b/143972021). Also handle 0 as if it - // were a 1. 0 is an invalid input here (divide by zero) and 1 is a valid - // but rare/unrealistic input value. We can expect both to occur in some - // incompletely trained models, but probably not in fully trained models. - *output_inv_sqrt = std::numeric_limits::max(); - *output_shift = 0; - return; - } - TFLITE_DCHECK_GT(input, 1); *output_shift = 11; while (input >= (1 << 29)) { input /= 4; ++*output_shift; } + TFLITE_DCHECK_GT(input, 0); const unsigned max_left_shift_bits = CountLeadingZeros(static_cast(input)) - 1; const unsigned max_left_shift_bit_pairs = max_left_shift_bits / 2; @@ -581,18 +571,6 @@ inline void NdArrayDescsForElementwiseBroadcast(const Dims& input0_dims, } } -// Copies dims to desc, calculating strides. -template -inline void CopyDimsToDesc(const RuntimeShape& input_shape, - NdArrayDesc* desc_out) { - int desc_stride = 1; - for (int i = N - 1; i >= 0; --i) { - desc_out->extents[i] = input_shape.Dims(i); - desc_out->strides[i] = desc_stride; - desc_stride *= input_shape.Dims(i); - } -} - template inline void NdArrayDescsForElementwiseBroadcast( const RuntimeShape& input0_shape, const RuntimeShape& input1_shape, @@ -604,8 +582,16 @@ inline void NdArrayDescsForElementwiseBroadcast( auto extended_input1_shape = RuntimeShape::ExtendedShape(N, input1_shape); // Copy dims to desc, calculating strides. - CopyDimsToDesc(extended_input0_shape, desc0_out); - CopyDimsToDesc(extended_input1_shape, desc1_out); + int desc0_stride = 1; + int desc1_stride = 1; + for (int i = N - 1; i >= 0; --i) { + desc0_out->extents[i] = extended_input0_shape.Dims(i); + desc0_out->strides[i] = desc0_stride; + desc0_stride *= extended_input0_shape.Dims(i); + desc1_out->extents[i] = extended_input1_shape.Dims(i); + desc1_out->strides[i] = desc1_stride; + desc1_stride *= extended_input1_shape.Dims(i); + } // Walk over each dimension. If the extents are equal do nothing. // Otherwise, set the desc with extent 1 to have extent equal to the other and @@ -626,57 +612,6 @@ inline void NdArrayDescsForElementwiseBroadcast( } } -template -inline void NdArrayDescsForElementwiseBroadcast( - const RuntimeShape& input0_shape, const RuntimeShape& input1_shape, - const RuntimeShape& input2_shape, NdArrayDesc* desc0_out, - NdArrayDesc* desc1_out, NdArrayDesc* desc2_out) { - TFLITE_DCHECK(desc0_out != nullptr); - TFLITE_DCHECK(desc1_out != nullptr); - TFLITE_DCHECK(desc2_out != nullptr); - - auto extended_input0_shape = RuntimeShape::ExtendedShape(N, input0_shape); - auto extended_input1_shape = RuntimeShape::ExtendedShape(N, input1_shape); - auto extended_input2_shape = RuntimeShape::ExtendedShape(N, input2_shape); - - // Copy dims to desc, calculating strides. - CopyDimsToDesc(extended_input0_shape, desc0_out); - CopyDimsToDesc(extended_input1_shape, desc1_out); - CopyDimsToDesc(extended_input2_shape, desc2_out); - - // Walk over each dimension. If the extents are equal do nothing. - // Otherwise, set the desc with extent 1 to have extent equal to the other and - // stride 0. - for (int i = 0; i < N; ++i) { - const int extent0 = extended_input0_shape.Dims(i); - const int extent1 = extended_input1_shape.Dims(i); - const int extent2 = extended_input2_shape.Dims(i); - - int extent = extent0; - if (extent1 != 1) extent = extent1; - if (extent2 != 1) extent = extent2; - - TFLITE_DCHECK(extent0 == 1 || extent0 == extent); - TFLITE_DCHECK(extent1 == 1 || extent1 == extent); - TFLITE_DCHECK(extent2 == 1 || extent2 == extent); - - if (!(extent0 == extent1 && extent1 == extent2)) { - if (extent0 == 1) { - desc0_out->strides[i] = 0; - desc0_out->extents[i] = extent; - } - if (extent1 == 1) { - desc1_out->strides[i] = 0; - desc1_out->extents[i] = extent; - } - if (extent2 == 1) { - desc2_out->strides[i] = 0; - desc2_out->extents[i] = extent; - } - } - } -} - // Copied from gemmlowp::RoundDown when we dropped direct dependency on // gemmlowp. // @@ -779,3 +714,5 @@ void optimized_ops_prefetch_write_l1_keep(const T* ptr) { } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_COMMON_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/compatibility.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h similarity index 96% rename from src/tensorflow/lite/kernels/internal/compatibility.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h index 2093062..978337f 100644 --- a/src/tensorflow/lite/kernels/internal/compatibility.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,7 +18,7 @@ limitations under the License. #include -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" #ifndef TFLITE_DCHECK #define TFLITE_DCHECK(condition) (condition) ? (void)0 : TFLITE_ASSERT_FALSE @@ -81,7 +82,6 @@ using int8 = std::int8_t; using uint8 = std::uint8_t; using int16 = std::int16_t; using uint16 = std::uint16_t; -#define int32 foo using int32 = std::int32_t; using uint32 = std::uint32_t; @@ -109,3 +109,5 @@ using uint32 = std::uint32_t; #endif #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/optimized/neon_check.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/optimized/neon_check.h similarity index 96% rename from src/tensorflow/lite/kernels/internal/optimized/neon_check.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/optimized/neon_check.h index a72af90..e953bd6 100644 --- a/src/tensorflow/lite/kernels/internal/optimized/neon_check.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/optimized/neon_check.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -44,3 +45,5 @@ limitations under the License. #endif // defined(USE_NEON) #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/quantization_util.cpp b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.cpp similarity index 94% rename from src/tensorflow/lite/kernels/internal/quantization_util.cpp rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.cpp index 2411b64..1e32cc1 100644 --- a/src/tensorflow/lite/kernels/internal/quantization_util.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,9 +18,9 @@ limitations under the License. #include #include -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" namespace tflite { @@ -183,11 +184,11 @@ double DoubleFromFractionAndShift(int64_t fraction, int shift) { // Detect NaNs and infinities. if (shift == std::numeric_limits::max()) { if (fraction == 0) { - return std::numeric_limits::quiet_NaN(); + return NAN; } else if (fraction > 0) { - return std::numeric_limits::infinity(); + return INFINITY; } else { - return -std::numeric_limits::infinity(); + return -INFINITY; } } @@ -229,7 +230,7 @@ double IntegerDoubleMultiply(double a, double b) { // Detect NaNs and infinities. if (a_shift == std::numeric_limits::max() || (b_shift == std::numeric_limits::max())) { - return std::numeric_limits::quiet_NaN(); + return NAN; } const int result_shift = a_shift + b_shift + 1; const int64_t result_fraction = (a_fraction * b_fraction) >> 32; @@ -288,11 +289,8 @@ void PreprocessSoftmaxScaling(double beta, double input_scale, input_beta_real_multiplier = (1ll << 31) - 1.0; } #else // TFLITE_EMULATE_FLOAT -// @Eloquent std::min and std::max produces error on some compilers, replace with #defines -#define min(a, b) ((a) < (b)) ? (a) : (b) -#define max(a, b) ((a) > (b)) ? (a) : (b) - - const double input_beta_real_multiplier = min(beta * input_scale * (1 << (31 - input_integer_bits)), (1ll << 31) - 1.0); + const double input_beta_real_multiplier = std::min( + beta * input_scale * (1 << (31 - input_integer_bits)), (1ll << 31) - 1.0); #endif // TFLITE_EMULATE_FLOAT QuantizeMultiplierGreaterThanOne(input_beta_real_multiplier, @@ -364,7 +362,7 @@ void FakeQuantizeArray(const float nudged_scale, const float nudged_min, for (int i = 0; i < size; i++) { const float src_val = input_data[i]; - const float clamped = min(nudged_max, max(nudged_min, src_val)); + const float clamped = std::min(nudged_max, std::max(nudged_min, src_val)); const float clamped_shifted = clamped - nudged_min; const float dst_val = TfLiteRound(clamped_shifted * inv_nudged_scale) * nudged_scale + @@ -382,7 +380,7 @@ bool CheckedLog2(const float x, int* log2_result) { const float x_log2_fracpart = x_log2 - x_log2_rounded; *log2_result = static_cast(x_log2_rounded); - return std::abs(x_log2_fracpart) < 1e-3f; + return std::abs(x_log2_fracpart) < 1e-3; } void QuantizeMultiplierArray(const double* effective_scales, size_t size, @@ -395,3 +393,5 @@ void QuantizeMultiplierArray(const double* effective_scales, size_t size, } } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h new file mode 100644 index 0000000..6a39111 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h @@ -0,0 +1,295 @@ +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_ + +#include +#include +#include + +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +// Given the min and max values of a float array, return +// reasonable quantization parameters to use for this array. +template +QuantizationParams ChooseQuantizationParams(double rmin, double rmax, + bool narrow_range) { + const T qmin = std::numeric_limits::min() + (narrow_range ? 1 : 0); + const T qmax = std::numeric_limits::max(); + const double qmin_double = qmin; + const double qmax_double = qmax; + // 0 should always be a representable value. Let's assume that the initial + // min,max range contains 0. + TFLITE_CHECK_LE(rmin, 0.); + TFLITE_CHECK_GE(rmax, 0.); + if (rmin == rmax) { + // Special case where the min,max range is a point. Should be {0}. + TFLITE_CHECK_EQ(rmin, 0.); + TFLITE_CHECK_EQ(rmax, 0.); + QuantizationParams quantization_params; + quantization_params.zero_point = 0; + quantization_params.scale = 0.; + return quantization_params; + } + + // General case. + // + // First determine the scale. + const double scale = (rmax - rmin) / (qmax_double - qmin_double); + + // Zero-point computation. + // First the initial floating-point computation. The zero-point can be + // determined from solving an affine equation for any known pair + // (real value, corresponding quantized value). + // We know two such pairs: (rmin, qmin) and (rmax, qmax). + // The arithmetic error on the zero point computed from either pair + // will be roughly machine_epsilon * (sum of absolute values of terms) + // so we want to use the variant that adds the smaller terms. + const double zero_point_from_min = qmin_double - rmin / scale; + const double zero_point_from_max = qmax_double - rmax / scale; + const double zero_point_from_min_error = + std::abs(qmin_double) + std::abs(rmin / scale); + const double zero_point_from_max_error = + std::abs(qmax_double) + std::abs(rmax / scale); + + const double zero_point_double = + zero_point_from_min_error < zero_point_from_max_error + ? zero_point_from_min + : zero_point_from_max; + + // Now we need to nudge the zero point to be an integer + // (our zero points are integer, and this is motivated by the requirement + // to be able to represent the real value "0" exactly as a quantized value, + // which is required in multiple places, for example in Im2col with SAME + // padding). + T nudged_zero_point = 0; + if (zero_point_double < qmin_double) { + nudged_zero_point = qmin; + } else if (zero_point_double > qmax_double) { + nudged_zero_point = qmax; + } else { + nudged_zero_point = static_cast(round(zero_point_double)); + } + // The zero point should always be in the range of quantized value, + // [qmin, qmax]. + TFLITE_CHECK_GE(nudged_zero_point, qmin); + TFLITE_CHECK_LE(nudged_zero_point, qmax); + + // Finally, store the result nudged quantization params. + QuantizationParams quantization_params; + quantization_params.zero_point = nudged_zero_point; + quantization_params.scale = scale; + return quantization_params; +} + +template +QuantizationParams ChooseQuantizationParams(double rmin, double rmax) { + return ChooseQuantizationParams(rmin, rmax, false); +} + +// Converts a floating-point number to an integer. For all inputs x where +// static_cast(x) is legal according to the C++ standard, the result +// is identical to that cast (i.e. the result is x with its fractional part +// truncated whenever that is representable as IntOut). +// +// static_cast would cause undefined behavior for the following cases, which +// have well-defined behavior for this function: +// +// 1. If x is NaN, the result is zero. +// +// 2. If the truncated form of x is above the representable range of IntOut, +// the result is std::numeric_limits::max(). +// +// 3. If the truncated form of x is below the representable range of IntOut, +// the result is std::numeric_limits::min(). +// +// Note that cases #2 and #3 cover infinities as well as finite numbers. +// +// The range of FloatIn must include the range of IntOut, otherwise +// the results are undefined. +// TODO(sfeuz): Replace by absl::SafeCast once available. +template +IntOut SafeCast(FloatIn x) { + static_assert(!std::numeric_limits::is_integer, + "FloatIn is integer"); + static_assert(std::numeric_limits::is_integer, + "IntOut is not integer"); + static_assert(std::numeric_limits::radix == 2, "IntOut is base 2"); + + // Special case NaN, for which the logic below doesn't work. + if (std::isnan(x)) { + return 0; + } + + // Negative values all clip to zero for unsigned results. + if (!std::numeric_limits::is_signed && x < 0) { + return 0; + } + + // Handle infinities. + if (std::isinf(x)) { + return x < 0 ? std::numeric_limits::min() + : std::numeric_limits::max(); + } + + // Set exp such that x == f * 2^exp for some f with |f| in [0.5, 1.0), + // unless x is zero in which case exp == 0. Note that this implies that the + // magnitude of x is strictly less than 2^exp. + int exp = 0; + std::frexp(x, &exp); + + // Let N be the number of non-sign bits in the representation of IntOut. If + // the magnitude of x is strictly less than 2^N, the truncated version of x + // is representable as IntOut. The only representable integer for which this + // is not the case is kMin for signed types (i.e. -2^N), but that is covered + // by the fall-through below. + if (exp <= std::numeric_limits::digits) { + return x; + } + + // Handle numbers with magnitude >= 2^N. + return x < 0 ? std::numeric_limits::min() + : std::numeric_limits::max(); +} + +// Decompose a double multiplier into a Q0.31 int32 representation of its +// significand, and shift representation of NEGATIVE its exponent --- +// this is intended as a RIGHT-shift. +// +// Restricted to the case where the multiplier < 1 (and non-negative). +void QuantizeMultiplierSmallerThanOneExp(double double_multiplier, + int32_t* quantized_multiplier, + int* left_shift); + +// Decompose a double multiplier into a Q0.31 int32 representation of its +// significand, and shift representation of its exponent. +// +// Restricted to the case where the multiplier > 1. +void QuantizeMultiplierGreaterThanOne(double double_multiplier, + int32_t* quantized_multiplier, + int* left_shift); + +// Decompose a double multiplier into a Q0.31 int32 representation of its +// significand, and shift representation of its exponent. +// +// Handles an arbitrary positive multiplier. The 'shift' output-value is +// basically the 'floating-point exponent' of the multiplier: +// Negative for a right-shift (when the multiplier is <1), positive for a +// left-shift (when the multiplier is >1) +void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier, + int* shift); + +// Splits a double input value into a returned fraction, and a shift value from +// the exponent, using only bitwise and integer operations to support +// microcontrollers and other environments without floating-point support. +// +// This is designed to be a replacement for how std::frexp() is used within the +// QuantizeMultiplier() function, and so has a different signature than the +// standard version, returning a 64-bit integer rather than a double. This +// result has a maximum value of 1<<31, with the fraction expressed as a +// proportion of that maximum. +// +// std::frexp() returns NaNs and infinities unmodified, but since we're +// returning integers that can't represent those values, instead we return +// a shift of std::numeric_limits::max() for all bad numbers, with an int64 +// result of 0 for NaNs, std:numeric_limits::max() for +INFINITY, and +// std::numeric_limits::min() for -INFINITY. Denormalized inputs will +// result in return values that end up truncating some bits at the end, +// reflecting the loss of precision inherent in denormalization. +int64_t IntegerFrExp(double input, int* shift); + +// Converts an integer fraction in the format produced by IntegerFrExp (where +// 0x40000000 is 1.0) and an exponent shift (between -1022 and +1022) into an +// IEEE binary64 double format result. The implementation uses only integer and +// bitwise operators, so no floating point hardware support or emulation is +// needed. This is here so quantized operations can run non-time-critical +// preparation calculations on microcontrollers and other platforms without +// float support. +double DoubleFromFractionAndShift(int64_t fraction, int shift); + +// Performs a multiplication of two numbers in double format, using only integer +// and bitwise instructions. This is aimed at supporting housekeeping functions +// for quantized operations on microcontrollers without floating-point hardware. +double IntegerDoubleMultiply(double a, double b); + +// Returns -1 if a is less than b, 0 if a and b are equal, and +1 if a is +// greater than b. It is implemented using only integer and logical instructions +// so that it can be easily run on microcontrollers for quantized operations. +int IntegerDoubleCompare(double a, double b); + +// This first creates a multiplier in a double equivalent of +// Q(input_integer_bits).(31-input_integer_bits) representation, with extra +// precision in the double's fractional bits. It then splits the result into +// significand and exponent. +void PreprocessSoftmaxScaling(double beta, double input_scale, + int input_integer_bits, + int32_t* quantized_multiplier, int* left_shift); +// Like PreprocessSoftmaxScaling, but inverse scaling factors also calculated. +void PreprocessLogSoftmaxScalingExp(double beta, double input_scale, + int input_integer_bits, + int32_t* quantized_multiplier, + int* left_shift, + int32_t* reverse_scaling_divisor, + int* reverse_scaling_left_shift); +// Calculate the largest input that will result in a within-bounds intermediate +// result within MultiplyByQuantizedMultiplierGreaterThanOne. In other words, +// it must not overflow before we reduce the value by multiplication by the +// input multiplier. The negative radius is used as the minimum difference in +// Softmax. +int CalculateInputRadius(int input_integer_bits, int input_left_shift, + int total_signed_bits = 31); + +// Nudges a min/max quantization range to ensure zero is zero. +// Gymnastics with nudged zero point is to ensure that real zero maps to +// an integer, which is required for e.g. zero-padding in convolutional layers. +// Outputs nudged_min, nudged_max, nudged_scale. +void NudgeQuantizationRange(const float min, const float max, + const int quant_min, const int quant_max, + float* nudged_min, float* nudged_max, + float* nudged_scale); + +// Fake quantizes (quantizes and dequantizes) input_data using the scale, +// nudged_min, and nudged_max from NudgeQuantizationRange. This matches the code +// in TensorFlow's FakeQuantizeWithMinMaxVarsFunctor. +void FakeQuantizeArray(const float nudged_scale, const float nudged_min, + const float nudged_max, const float* input_data, + float* output_data, const float size); + +// If x is approximately a power of two (with any positive or negative +// exponent), stores that exponent (i.e. log2(x)) in *log2_result, otherwise +// returns false. +bool CheckedLog2(const float x, int* log2_result); + +// Decomposes an array of double multipliers into a Q0.31 int32 representation +// of its significand, and shift representation of its exponent. +// +// Handles an arbitrary multiplier. The 'shift' output-value is +// basically the 'floating-point exponent' of the multiplier: +// Negative for a right-shift (when the multiplier is <1), positive for a +// left-shift (when the multiplier is >1) +void QuantizeMultiplierArray(const double* effective_scales, size_t size, + int32_t* effective_scale_significand, + int* effective_shift); + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/add.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/add.h similarity index 98% rename from src/tensorflow/lite/kernels/internal/reference/add.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/add.h index d0c4091..70c749e 100644 --- a/src/tensorflow/lite/kernels/internal/reference/add.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/add.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,8 +16,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_ -#include "fixedpoint/fixedpoint.h" -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" namespace tflite { @@ -417,3 +418,5 @@ inline void BroadcastAddFivefold(const ArithmeticParams& unswitched_params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/arg_min_max.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/arg_min_max.h new file mode 100644 index 0000000..a40f827 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/arg_min_max.h @@ -0,0 +1,71 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_ + +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +template +void ArgMinMax(const RuntimeShape& input1_shape, const T1* input1_data, + const T3* input2_data, const RuntimeShape& output_shape, + T2* output_data, const Cmp& cmp) { + TFLITE_DCHECK_GT(input1_shape.DimensionsCount(), 0); + TFLITE_DCHECK_EQ(input1_shape.DimensionsCount() - 1, + output_shape.DimensionsCount()); + int axis = input2_data[0]; + if (axis < 0) { + axis += input1_shape.DimensionsCount(); + } + const int axis_size = input1_shape.Dims(axis); + + int outer_size = 1; + for (int i = 0; i < axis; ++i) { + TFLITE_DCHECK_EQ(input1_shape.Dims(i), output_shape.Dims(i)); + outer_size *= input1_shape.Dims(i); + } + + int inner_size = 1; + const int dims_count = input1_shape.DimensionsCount(); + for (int i = axis + 1; i < dims_count; ++i) { + TFLITE_DCHECK_EQ(input1_shape.Dims(i), output_shape.Dims(i - 1)); + inner_size *= input1_shape.Dims(i); + } + for (int outer = 0; outer < outer_size; ++outer) { + for (int inner = 0; inner < inner_size; ++inner) { + auto min_max_value = input1_data[outer * axis_size * inner_size + inner]; + T2 min_max_index = 0; + for (int i = 1; i < axis_size; ++i) { + const auto& curr_value = + input1_data[(outer * axis_size + i) * inner_size + inner]; + if (cmp(curr_value, min_max_value)) { + min_max_value = curr_value; + min_max_index = static_cast(i); + } + } + output_data[outer * inner_size + inner] = min_max_index; + } + } +} +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/binary_function.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/binary_function.h similarity index 92% rename from src/tensorflow/lite/kernels/internal/reference/binary_function.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/binary_function.h index 82095af..839ee52 100644 --- a/src/tensorflow/lite/kernels/internal/reference/binary_function.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/binary_function.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,9 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -82,3 +83,5 @@ inline void BinaryFunction(const RuntimeShape& input1_shape, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/ceil.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/ceil.h new file mode 100644 index 0000000..67651eb --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/ceil.h @@ -0,0 +1,40 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_ + +#include + +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline void Ceil(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; ++i) { + output_data[i] = std::ceil(input_data[i]); + } +} + +} // namespace reference_ops +} // namespace tflite +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/comparisons.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/comparisons.h similarity index 98% rename from src/tensorflow/lite/kernels/internal/reference/comparisons.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/comparisons.h index 19a968e..6a43b66 100644 --- a/src/tensorflow/lite/kernels/internal/reference/comparisons.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/comparisons.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,8 +16,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -265,3 +266,5 @@ TFLITE_COMPARISON_OP(LessEqual); } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/conv.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/conv.h similarity index 98% rename from src/tensorflow/lite/kernels/internal/reference/conv.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/conv.h index 55dd869..6a81732 100644 --- a/src/tensorflow/lite/kernels/internal/reference/conv.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/conv.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,8 +16,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ -#include "tensorflow/lite/kernels/internal/types.h" -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" @@ -260,3 +261,5 @@ inline void HybridConvPerChannel( #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h new file mode 100644 index 0000000..eb91e18 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h @@ -0,0 +1,103 @@ +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_ + +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { +namespace reference_ops { + +inline void DepthwiseConv( + const DepthwiseParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& filter_shape, + const float* filter_data, const RuntimeShape& bias_shape, + const float* bias_data, const RuntimeShape& output_shape, + float* output_data) { + const int stride_width = params.stride_width; + const int stride_height = params.stride_height; + const int dilation_width_factor = params.dilation_width_factor; + const int dilation_height_factor = params.dilation_height_factor; + const int pad_width = params.padding_values.width; + const int pad_height = params.padding_values.height; + const int depth_multiplier = params.depth_multiplier; + const float output_activation_min = params.float_activation_min; + const float output_activation_max = params.float_activation_max; + TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4); + TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); + + const int batches = MatchingDim(input_shape, 0, output_shape, 0); + const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3); + const int input_height = input_shape.Dims(1); + const int input_width = input_shape.Dims(2); + const int input_depth = input_shape.Dims(3); + const int filter_height = filter_shape.Dims(1); + const int filter_width = filter_shape.Dims(2); + const int output_height = output_shape.Dims(1); + const int output_width = output_shape.Dims(2); + TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier); + TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth); + + for (int b = 0; b < batches; ++b) { + for (int out_y = 0; out_y < output_height; ++out_y) { + for (int out_x = 0; out_x < output_width; ++out_x) { + for (int ic = 0; ic < input_depth; ++ic) { + for (int m = 0; m < depth_multiplier; m++) { + const int oc = m + ic * depth_multiplier; + const int in_x_origin = (out_x * stride_width) - pad_width; + const int in_y_origin = (out_y * stride_height) - pad_height; + float total = 0.f; + for (int filter_y = 0; filter_y < filter_height; ++filter_y) { + for (int filter_x = 0; filter_x < filter_width; ++filter_x) { + const int in_x = in_x_origin + dilation_width_factor * filter_x; + const int in_y = + in_y_origin + dilation_height_factor * filter_y; + // If the location is outside the bounds of the input image, + // use zero as a default value. + if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) && + (in_y < input_height)) { + float input_value = + input_data[Offset(input_shape, b, in_y, in_x, ic)]; + float filter_value = filter_data[Offset( + filter_shape, 0, filter_y, filter_x, oc)]; + total += (input_value * filter_value); + } + } + } + float bias_value = 0.0f; + if (bias_data) { + bias_value = bias_data[oc]; + } + output_data[Offset(output_shape, b, out_y, out_x, oc)] = + ActivationFunctionWithMinMax(total + bias_value, + output_activation_min, + output_activation_max); + } + } + } + } + } +} + +} // end namespace reference_ops +} // end namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h similarity index 96% rename from src/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h index c38f374..c476c1a 100644 --- a/src/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,10 +18,10 @@ limitations under the License. #include -#include "fixedpoint/fixedpoint.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -193,3 +194,5 @@ inline void DepthwiseConv( } // end namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/dequantize.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/dequantize.h similarity index 88% rename from src/tensorflow/lite/kernels/internal/reference/dequantize.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/dequantize.h index 1040001..1079ff0 100644 --- a/src/tensorflow/lite/kernels/internal/reference/dequantize.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/dequantize.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,8 +16,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -41,3 +42,5 @@ inline void Dequantize(const tflite::DequantizationParams& op_params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/floor.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/floor.h new file mode 100644 index 0000000..c2ffb10 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/floor.h @@ -0,0 +1,42 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_ + +#include + +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline void Floor(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; i++) { + int offset = i; + output_data[offset] = std::floor(input_data[offset]); + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/fully_connected.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/fully_connected.h similarity index 97% rename from src/tensorflow/lite/kernels/internal/reference/fully_connected.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/fully_connected.h index 51c1def..1eb6f03 100644 --- a/src/tensorflow/lite/kernels/internal/reference/fully_connected.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/fully_connected.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,10 +16,10 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/round.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { namespace reference_ops { @@ -319,3 +320,5 @@ inline void ShuffledFullyConnected( } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/integer_ops/add.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/add.h similarity index 97% rename from src/tensorflow/lite/kernels/internal/reference/integer_ops/add.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/add.h index 69b42e0..1f974e6 100644 --- a/src/tensorflow/lite/kernels/internal/reference/integer_ops/add.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/add.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,8 +18,8 @@ limitations under the License. #include -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { namespace reference_integer_ops { @@ -141,3 +142,5 @@ inline void BroadcastAdd4DSlow(const ArithmeticParams& params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h similarity index 95% rename from src/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h index 4b101f7..e514751 100644 --- a/src/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_ -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" namespace tflite { namespace reference_integer_ops { @@ -39,8 +40,8 @@ inline void ConvPerChannel( const int32 output_offset = params.output_offset; // Set min and max value of the output. - const int32 output_activation_min = params.quantized_activation_min; - const int32 output_activation_max = params.quantized_activation_max; + const int32 output_activation_min = std::numeric_limits::min(); + const int32 output_activation_max = std::numeric_limits::max(); // Sanity check. TFLITE_DCHECK_LE(output_activation_min, output_activation_max); @@ -126,3 +127,5 @@ inline void ConvPerChannel( } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h similarity index 98% rename from src/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h index a21f837..38e8be5 100644 --- a/src/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_ -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" namespace tflite { namespace reference_integer_ops { @@ -122,3 +123,5 @@ inline void DepthwiseConvPerChannel( } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h similarity index 95% rename from src/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h index 6431384..e52362c 100644 --- a/src/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_ -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" namespace tflite { namespace reference_integer_ops { @@ -66,3 +67,5 @@ inline void FullyConnected( } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h similarity index 98% rename from src/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h index 2762bec..e306383 100644 --- a/src/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -16,7 +17,7 @@ limitations under the License. #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_ #include -#include "tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" namespace tflite { namespace reference_integer_ops { @@ -139,3 +140,5 @@ inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/logistic.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/logistic.h similarity index 69% rename from src/tensorflow/lite/kernels/internal/reference/logistic.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/logistic.h index 29fd97d..1eee800 100644 --- a/src/tensorflow/lite/kernels/internal/reference/logistic.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/logistic.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,43 +16,23 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_ -#include - -#include "fixedpoint/fixedpoint.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/round.h" -#include "tensorflow/lite/kernels/internal/types.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace reference_ops { inline void Logistic(const RuntimeShape& input_shape, const float* input_data, const RuntimeShape& output_shape, float* output_data) { - const float cutoff_upper = 16.619047164916992188f; - const float cutoff_lower = -9.f; - const int flat_size = MatchingFlatSize(input_shape, output_shape); - // Rational for using approximation in reference kernel. - // 0. This approximation gives enough precision for float. - // 1. This works around an issue on an embedded chipset where exp() does not - // return correctly as expected - exp(x) should return inf when overflown - // not 1.701417 IEEE 754 defines representation for inf. - // 2. This will speed up calculation and is matching the behavior in the - // optimized kernels. (check the definition of scalar_logistic_op) - for (int i = 0; i < flat_size; i++) { float val = input_data[i]; - float result; - if (val > cutoff_upper) { - result = 1.0f; - } else if (val < cutoff_lower) { - result = std::exp(val); - } else { - result = 1.f / (1.f + std::exp(-val)); - } + float result = 1.f / (1.f + std::exp(-val)); output_data[i] = result; } } @@ -88,3 +69,5 @@ inline void Logistic(const LogisticParams& params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/maximum_minimum.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/maximum_minimum.h similarity index 92% rename from src/tensorflow/lite/kernels/internal/reference/maximum_minimum.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/maximum_minimum.h index 480069a..d3b81ec 100644 --- a/src/tensorflow/lite/kernels/internal/reference/maximum_minimum.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/maximum_minimum.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,8 +16,8 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { namespace reference_ops { @@ -59,3 +60,5 @@ void MaximumMinimumBroadcast4DSlow(const RuntimeShape& unextended_input1_shape, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/neg.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/neg.h new file mode 100644 index 0000000..ef5561e --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/neg.h @@ -0,0 +1,40 @@ +#if defined(ESP32) +/* Copyright 2019 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_ + +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +template +inline void Negate(const RuntimeShape& input_shape, const T* input_data, + const RuntimeShape& output_shape, T* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + + for (int i = 0; i < flat_size; ++i) { + output_data[i] = -input_data[i]; + } +} + +} // namespace reference_ops +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/pooling.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/pooling.h similarity index 97% rename from src/tensorflow/lite/kernels/internal/reference/pooling.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/pooling.h index 2cb2347..840db76 100644 --- a/src/tensorflow/lite/kernels/internal/reference/pooling.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/pooling.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,10 +16,10 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/round.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { namespace reference_ops { @@ -294,3 +295,5 @@ inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/prelu.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/prelu.h similarity index 92% rename from src/tensorflow/lite/kernels/internal/reference/prelu.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/prelu.h index adbbf66..d8fb80d 100644 --- a/src/tensorflow/lite/kernels/internal/reference/prelu.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/prelu.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,9 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -75,3 +76,5 @@ inline void BroadcastPrelu4DSlow(const PreluParams& params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h similarity index 97% rename from src/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h index 8e1a6c8..796f6d1 100644 --- a/src/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_ -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -132,3 +133,5 @@ inline bool ProcessBroadcastShapes(const RuntimeShape& shape0, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/quantize.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/quantize.h similarity index 83% rename from src/tensorflow/lite/kernels/internal/reference/quantize.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/quantize.h index 807eccb..bc092c5 100644 --- a/src/tensorflow/lite/kernels/internal/reference/quantize.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/quantize.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,9 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/round.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -36,9 +37,7 @@ inline void AffineQuantize(const tflite::QuantizationParams& op_params, for (int i = 0; i < flat_size; i++) { const float val = input_data[i]; - int32 unclamped = - static_cast(TfLiteRound(val / static_cast(scale))) + - zero_point; + int32 unclamped = static_cast(TfLiteRound(val / scale)) + zero_point; int32 clamped = std::min(std::max(unclamped, min_val), max_val); output_data[i] = clamped; } @@ -48,3 +47,5 @@ inline void AffineQuantize(const tflite::QuantizationParams& op_params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/round.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/round.h new file mode 100644 index 0000000..2bdae13 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/round.h @@ -0,0 +1,54 @@ +#if defined(ESP32) +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_ + +#include + +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +namespace reference_ops { + +inline float RoundToNearest(float value) { + auto floor_val = std::floor(value); + auto diff = value - floor_val; + if ((diff < 0.5f) || + ((diff == 0.5f) && (static_cast(floor_val) % 2 == 0))) { + return floor_val; + } else { + return floor_val = floor_val + 1.0f; + } +} + +inline void Round(const RuntimeShape& input_shape, const float* input_data, + const RuntimeShape& output_shape, float* output_data) { + const int flat_size = MatchingFlatSize(input_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + // Note that this implementation matches that of tensorFlow tf.round + // and corresponds to the bankers rounding method. + // cfenv (for fesetround) is not yet supported universally on Android, so + // using a work around. + output_data[i] = RoundToNearest(input_data[i]); + } +} + +} // namespace reference_ops +} // namespace tflite +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/softmax.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/softmax.h similarity index 90% rename from src/tensorflow/lite/kernels/internal/reference/softmax.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/softmax.h index 790f4d2..e7e91c1 100644 --- a/src/tensorflow/lite/kernels/internal/reference/softmax.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/softmax.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,12 +16,12 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_ -#include "fixedpoint/fixedpoint.h" -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/round.h" -#include "tensorflow/lite/kernels/internal/types.h" -#include "tensorflow/lite/kernels/op_macros.h" +#include "tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/op_macros.h" namespace tflite { namespace reference_ops { @@ -43,20 +44,16 @@ inline void Softmax(const SoftmaxParams& params, max = std::max(max, input_data[i * depth + c]); } - // TODO(b/148114827): Improve this code. // Compute sum. float sum = 0.f; for (int c = 0; c < depth; ++c) { - sum += std::exp(static_cast(input_data[i * depth + c] - max) * - params.beta); + sum += std::exp((input_data[i * depth + c] - max) * params.beta); } // Compute result. for (int c = 0; c < depth; ++c) { output_data[i * depth + c] = - std::exp(static_cast(input_data[i * depth + c] - max) * - params.beta) / - static_cast(sum); + std::exp((input_data[i * depth + c] - max) * params.beta) / sum; } } } @@ -171,3 +168,5 @@ inline void Softmax(const float* in, const int input_size, const int batch_size, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/reference/strided_slice.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/strided_slice.h similarity index 92% rename from src/tensorflow/lite/kernels/internal/reference/strided_slice.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/strided_slice.h index 921c49e..8486a73 100644 --- a/src/tensorflow/lite/kernels/internal/reference/strided_slice.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/reference/strided_slice.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,9 +16,9 @@ limitations under the License. #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_ #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_ -#include "tensorflow/lite/kernels/internal/common.h" -#include "tensorflow/lite/kernels/internal/strided_slice_logic.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/common.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/strided_slice_logic.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { @@ -78,3 +79,5 @@ inline void StridedSlice(const tflite::StridedSliceParams& op_params, } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/round.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/round.h similarity index 96% rename from src/tensorflow/lite/kernels/internal/round.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/round.h index d102d37..d403311 100644 --- a/src/tensorflow/lite/kernels/internal/round.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/round.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -38,3 +39,5 @@ inline T TfLiteRound(const T x) { } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_ROUND_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/strided_slice_logic.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/strided_slice_logic.h similarity index 97% rename from src/tensorflow/lite/kernels/internal/strided_slice_logic.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/strided_slice_logic.h index 3022ac7..7de5f28 100644 --- a/src/tensorflow/lite/kernels/internal/strided_slice_logic.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/strided_slice_logic.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -18,8 +19,8 @@ limitations under the License. #include #include -#include "tensorflow/lite/kernels/internal/compatibility.h" -#include "tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" namespace tflite { namespace strided_slice { @@ -196,3 +197,5 @@ inline tflite::StridedSliceParams BuildStridedSliceParams( } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_STRIDED_SLICE_LOGIC_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/tensor.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/tensor.h similarity index 93% rename from src/tensorflow/lite/kernels/internal/tensor.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/tensor.h index 0005bf3..39fb024 100644 --- a/src/tensorflow/lite/kernels/internal/tensor.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/tensor.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -18,10 +19,10 @@ limitations under the License. #include #include -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/kernels/internal/tensor_ctypes.h" -#include "tensorflow/lite/kernels/internal/types.h" -#include "tensorflow/lite/string_util.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" +#include "tensorflow_esp32/tensorflow/lite/string_util.h" namespace tflite { @@ -142,3 +143,5 @@ class SequentialTensorWriter { } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h new file mode 100644 index 0000000..3111b6b --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/tensor_ctypes.h @@ -0,0 +1,50 @@ +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_ +#define TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_ + +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/types.h" + +namespace tflite { + +template +inline T* GetTensorData(TfLiteTensor* tensor) { + return tensor != nullptr ? reinterpret_cast(tensor->data.raw) : nullptr; +} + +template +inline const T* GetTensorData(const TfLiteTensor* tensor) { + return tensor != nullptr ? reinterpret_cast(tensor->data.raw) + : nullptr; +} + +inline RuntimeShape GetTensorShape(const TfLiteTensor* tensor) { + if (tensor == nullptr) { + return RuntimeShape(); + } + + TfLiteIntArray* dims = tensor->dims; + const int dims_size = dims->size; + const int32_t* dims_data = reinterpret_cast(dims->data); + return RuntimeShape(dims_size, dims_data); +} + +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/internal/types.h b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/types.h similarity index 98% rename from src/tensorflow/lite/kernels/internal/types.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/internal/types.h index ad51f29..f2ab6ca 100644 --- a/src/tensorflow/lite/kernels/internal/types.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/internal/types.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -20,7 +21,7 @@ limitations under the License. #include #include -#include "tensorflow/lite/kernels/internal/compatibility.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/compatibility.h" namespace tflite { @@ -880,9 +881,6 @@ struct FullyConnectedParams { // float activation params. float float_activation_min; float float_activation_max; - // Mark the operands as cacheable if they are unchanging, e.g. weights. - bool lhs_cacheable; - bool rhs_cacheable; FullyConnectedWeightsFormat weights_format; }; @@ -991,10 +989,6 @@ struct ReshapeParams { struct ResizeBilinearParams { bool align_corners; - // half_pixel_centers assumes pixels are of half the actual dimensions, and - // yields more accurate resizes. Corresponds to the same argument for the - // original TensorFlow op in TF2.0. - bool half_pixel_centers; }; struct ResizeNearestNeighborParams { @@ -1047,11 +1041,11 @@ struct SqueezeParams { struct StridedSliceParams { int8 start_indices_count; - int32 start_indices[4]; + int16 start_indices[4]; int8 stop_indices_count; - int32 stop_indices[4]; + int16 stop_indices[4]; int8 strides_count; - int32 strides[4]; + int16 strides[4]; int16 begin_mask; int16 ellipsis_mask; @@ -1113,3 +1107,5 @@ inline void GetActivationParams(const P& params, float* min, float* max) { } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_INTERNAL_TYPES_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/kernel_util.cpp b/src/tensorflow_esp32/tensorflow/lite/kernels/kernel_util.cpp similarity index 68% rename from src/tensorflow/lite/kernels/kernel_util.cpp rename to src/tensorflow_esp32/tensorflow/lite/kernels/kernel_util.cpp index f9c2352..2321b6c 100644 --- a/src/tensorflow/lite/kernels/kernel_util.cpp +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/kernel_util.cpp @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,39 +13,23 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/lite/kernels/kernel_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h" #include #include #include -#include "tensorflow/lite/kernels/internal/quantization_util.h" -#include "tensorflow/lite/kernels/internal/round.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/quantization_util.h" +#include "tensorflow_esp32/tensorflow/lite/kernels/internal/round.h" namespace tflite { -// Per-axis TfLiteStatus PopulateConvolutionQuantizationParams( TfLiteContext* context, const TfLiteTensor* input, const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output, const TfLiteFusedActivation& activation, int32_t* multiplier, int* shift, int32_t* output_activation_min, int32_t* output_activation_max, int32_t* per_channel_multiplier, int* per_channel_shift) { - const auto* affine_quantization = - reinterpret_cast(filter->quantization.params); - return PopulateConvolutionQuantizationParams( - context, input, filter, bias, output, activation, multiplier, shift, - output_activation_min, output_activation_max, per_channel_multiplier, - per_channel_shift, affine_quantization->scale->size); -} - -// Per-axis & per-tensor -TfLiteStatus PopulateConvolutionQuantizationParams( - TfLiteContext* context, const TfLiteTensor* input, - const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output, - const TfLiteFusedActivation& activation, int32_t* multiplier, int* shift, - int32_t* output_activation_min, int32_t* output_activation_max, - int32_t* per_channel_multiplier, int* per_channel_shift, int num_channels) { TF_LITE_ENSURE_EQ(context, input->quantization.type, kTfLiteAffineQuantization); TF_LITE_ENSURE_EQ(context, filter->quantization.type, @@ -65,29 +50,26 @@ TfLiteStatus PopulateConvolutionQuantizationParams( // Currently only Int8 is supported for per channel quantization. TF_LITE_ENSURE_EQ(context, input->type, kTfLiteInt8); TF_LITE_ENSURE_EQ(context, filter->type, kTfLiteInt8); - TF_LITE_ENSURE_EQ(context, affine_quantization->scale->size, num_channels); TF_LITE_ENSURE_EQ( - context, num_channels, + context, affine_quantization->scale->size, filter->dims->data[affine_quantization->quantized_dimension]); } // Populate multiplier and shift using affine quantization. + const int num_channels = affine_quantization->scale->size; const float input_scale = input->params.scale; const float output_scale = output->params.scale; const float* filter_scales = affine_quantization->scale->data; for (int i = 0; i < num_channels; ++i) { - // If per-tensor quantization parameter is specified, broadcast it along the - // quantization dimension (channels_out). - const float scale = is_per_channel ? filter_scales[i] : filter_scales[0]; - const double filter_scale = static_cast(scale); + const double filter_scale = static_cast(filter_scales[i]); const double effective_output_scale = static_cast(input_scale) * filter_scale / static_cast(output_scale); int32_t significand; - int channel_shift; - QuantizeMultiplier(effective_output_scale, &significand, &channel_shift); + int shift; + QuantizeMultiplier(effective_output_scale, &significand, &shift); per_channel_multiplier[i] = significand; - per_channel_shift[i] = channel_shift; + per_channel_shift[i] = shift; } // Populate scalar quantization parameters. @@ -103,11 +85,8 @@ TfLiteStatus PopulateConvolutionQuantizationParams( // Populate quantization parameteters with multiplier and shift. QuantizeMultiplier(real_multiplier, multiplier, &exponent); *shift = -exponent; - } - if (input->type == kTfLiteInt8 || input->type == kTfLiteUInt8) { - TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( - context, activation, output, output_activation_min, - output_activation_max)); + CalculateActivationRangeUint8(activation, output, output_activation_min, + output_activation_max); } return kTfLiteOk; } @@ -118,12 +97,11 @@ TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context, const TfLiteTensor* bias, TfLiteTensor* output, double* multiplier) { - const double input_product_scale = static_cast(input->params.scale) * - static_cast(filter->params.scale); + const double input_product_scale = input->params.scale * filter->params.scale; // TODO(ahentz): The following conditions must be guaranteed by the training // pipeline. if (bias) { - const double bias_scale = static_cast(bias->params.scale); + const double bias_scale = bias->params.scale; TF_LITE_ENSURE(context, std::abs(input_product_scale - bias_scale) <= 1e-6 * std::min(input_product_scale, bias_scale)); @@ -137,10 +115,9 @@ TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context, const TfLiteTensor* filter, TfLiteTensor* output, double* multiplier) { - const double input_product_scale = - static_cast(input->params.scale * filter->params.scale); + const double input_product_scale = input->params.scale * filter->params.scale; TF_LITE_ENSURE(context, input_product_scale >= 0); - *multiplier = input_product_scale / static_cast(output->params.scale); + *multiplier = input_product_scale / output->params.scale; return kTfLiteOk; } @@ -198,6 +175,26 @@ TfLiteStatus CalculateActivationRangeQuantized(TfLiteContext* context, return kTfLiteOk; } +void CalculateActivationRangeUint8(TfLiteFusedActivation activation, + TfLiteTensor* output, int32_t* act_min, + int32_t* act_max) { + const int32_t qmin = std::numeric_limits::min(); + const int32_t qmax = std::numeric_limits::max(); + + CalculateActivationRangeQuantizedImpl(activation, qmin, qmax, output, act_min, + act_max); +} + +void CalculateActivationRangeInt8(TfLiteFusedActivation activation, + TfLiteTensor* output, int32_t* act_min, + int32_t* act_max) { + const int32_t qmin = std::numeric_limits::min(); + const int32_t qmax = std::numeric_limits::max(); + + CalculateActivationRangeQuantizedImpl(activation, qmin, qmax, output, act_min, + act_max); +} + bool HaveSameShapes(const TfLiteTensor* input1, const TfLiteTensor* input2) { return TfLiteIntArrayEqual(input1->dims, input2->dims); } @@ -209,9 +206,9 @@ TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, const TfLiteTensor* input1, const TfLiteTensor* input2, TfLiteIntArray** output_shape) { - int dims1 = NumDimensions(input1); - int dims2 = NumDimensions(input2); - int out_dims = std::max(dims1, dims2); + int64_t dims1 = NumDimensions(input1); + int64_t dims2 = NumDimensions(input2); + int64_t out_dims = std::max(dims1, dims2); if (NumElements(input1) == 0) { *output_shape = TfLiteIntArrayCopy(input1->dims); return kTfLiteOk; @@ -219,39 +216,16 @@ TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, std::unique_ptr shape( TfLiteIntArrayCreate(out_dims), TfLiteIntArrayFree); for (int i = 0; i < out_dims; ++i) { - int d1 = i >= dims1 ? 1 : SizeOfDimension(input1, dims1 - i - 1); - int d2 = i >= dims2 ? 1 : SizeOfDimension(input2, dims2 - i - 1); + int64_t d1 = i >= dims1 ? 1 : SizeOfDimension(input1, dims1 - i - 1); + int64_t d2 = i >= dims2 ? 1 : SizeOfDimension(input2, dims2 - i - 1); TF_LITE_ENSURE(context, d1 == d2 || d1 == 1 || d2 == 1); shape->data[out_dims - i - 1] = std::max(d1, d2); } *output_shape = shape.release(); return kTfLiteOk; } - -TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, - const TfLiteTensor* input1, - const TfLiteTensor* input2, - const TfLiteTensor* input3, - TfLiteIntArray** output_shape) { - int dims1 = NumDimensions(input1); - int dims2 = NumDimensions(input2); - int dims3 = NumDimensions(input3); - int out_dims = std::max(std::max(dims1, dims2), dims3); - std::unique_ptr shape( - TfLiteIntArrayCreate(out_dims), TfLiteIntArrayFree); - for (int i = 0; i < out_dims; ++i) { - int d1 = i >= dims1 ? 1 : SizeOfDimension(input1, dims1 - i - 1); - int d2 = i >= dims2 ? 1 : SizeOfDimension(input2, dims2 - i - 1); - int d3 = i >= dims3 ? 1 : SizeOfDimension(input3, dims3 - i - 1); - int max_value = std::max(std::max(d1, d2), d3); - TF_LITE_ENSURE(context, d1 == 1 || d1 == max_value); - TF_LITE_ENSURE(context, d2 == 1 || d2 == max_value); - TF_LITE_ENSURE(context, d3 == 1 || d3 == max_value); - shape->data[out_dims - i - 1] = max_value; - } - *output_shape = shape.release(); - return kTfLiteOk; -} #endif // TF_LITE_STATIC_MEMORY } // namespace tflite + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/kernel_util.h b/src/tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h similarity index 83% rename from src/tensorflow/lite/kernels/kernel_util.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h index ad068dd..7084aeb 100644 --- a/src/tensorflow/lite/kernels/kernel_util.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/kernel_util.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -18,9 +19,9 @@ limitations under the License. #include #include -#include "flatbuffers/flatbuffers.h" -#include "tensorflow/lite/c/builtin_op_data.h" -#include "tensorflow/lite/c/common.h" +#include "tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" namespace tflite { @@ -33,9 +34,6 @@ inline const TfLiteTensor* GetInput(TfLiteContext* context, return &context ->tensors[flatbuffers::EndianScalar(node->inputs->data[index])]; } -// Note: You must check if result is not null: -// TfLiteTensor* my_tensor = GetVariableInput(context, node, kMyTensorIdx); -// TF_LITE_ENSURE(context, my_tensor != nullptr); inline TfLiteTensor* GetVariableInput(TfLiteContext* context, const TfLiteNode* node, int index) { TfLiteTensor* tensor = @@ -77,8 +75,7 @@ inline int64_t NumElements(const TfLiteTensor* t) { inline const TfLiteTensor* GetOptionalInputTensor(TfLiteContext* context, const TfLiteNode* node, int index) { - const bool use_tensor = index < node->inputs->size && - node->inputs->data[index] != kTfLiteOptionalTensor; + const bool use_tensor = node->inputs->data[index] != kOptionalTensor; if (use_tensor) { return &context ->tensors[flatbuffers::EndianScalar(node->inputs->data[index])]; @@ -120,13 +117,6 @@ TfLiteStatus PopulateConvolutionQuantizationParams( int32_t* output_activation_min, int32_t* output_activation_max, int32_t* per_channel_multiplier, int* per_channel_shift); -TfLiteStatus PopulateConvolutionQuantizationParams( - TfLiteContext* context, const TfLiteTensor* input, - const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output, - const TfLiteFusedActivation& activation, int32_t* multiplier, int* shift, - int32_t* output_activation_min, int32_t* output_activation_max, - int32_t* per_channel_multiplier, int* per_channel_shift, int num_channels); - // Calculates the multiplication factor for a quantized convolution (or // quantized depthwise convolution) involving the given tensors. Returns an // error if the scales of the tensors are not compatible. @@ -150,7 +140,12 @@ TfLiteStatus CalculateActivationRangeQuantized(TfLiteContext* context, TfLiteTensor* output, int32_t* act_min, int32_t* act_max); - +void CalculateActivationRangeUint8(TfLiteFusedActivation activation, + TfLiteTensor* output, int32_t* act_min, + int32_t* act_max); +void CalculateActivationRangeInt8(TfLiteFusedActivation activation, + TfLiteTensor* output, int32_t* act_min, + int32_t* act_max); // Calculates the useful range of an activation layer given its activation // tensor.a template @@ -174,20 +169,14 @@ void CalculateActivationRange(TfLiteFusedActivation activation, // Return true if the given tensors have the same shape. bool HaveSameShapes(const TfLiteTensor* input1, const TfLiteTensor* input2); -// Calculates the output_shape that is necessary for element-wise operations +// Calculate the output_shape that is necessary for element-wise operations // with broadcasting involving the two input tensors. TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, const TfLiteTensor* input1, const TfLiteTensor* input2, TfLiteIntArray** output_shape); - -// Calculates the output_shape that is necessary for element-wise operations -// with broadcasting involving the three input tensors. -TfLiteStatus CalculateShapeForBroadcast(TfLiteContext* context, - const TfLiteTensor* input1, - const TfLiteTensor* input2, - const TfLiteTensor* input3, - TfLiteIntArray** output_shape); } // namespace tflite #endif // TENSORFLOW_LITE_KERNELS_KERNEL_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/kernels/op_macros.h b/src/tensorflow_esp32/tensorflow/lite/kernels/op_macros.h similarity index 93% rename from src/tensorflow/lite/kernels/op_macros.h rename to src/tensorflow_esp32/tensorflow/lite/kernels/op_macros.h index 4420800..f7bb756 100644 --- a/src/tensorflow/lite/kernels/op_macros.h +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/op_macros.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -19,7 +20,7 @@ limitations under the License. // non-portable function. #ifdef TF_LITE_MCU_DEBUG_LOG -#include "tensorflow/lite/micro/micro_error_reporter.h" +#include "tensorflow_esp32/tensorflow/lite/experimental/micro/micro_error_reporter.h" #define DEBUG_LOG(x) \ do { \ @@ -73,3 +74,5 @@ inline void InfiniteLoop() { } while (0) #endif // TENSORFLOW_LITE_KERNELS_OP_MACROS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/kernels/padding.h b/src/tensorflow_esp32/tensorflow/lite/kernels/padding.h new file mode 100644 index 0000000..a7a97f8 --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/kernels/padding.h @@ -0,0 +1,83 @@ +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_KERNELS_PADDING_H_ +#define TENSORFLOW_LITE_KERNELS_PADDING_H_ + +#include "tensorflow_esp32/tensorflow/lite/c/builtin_op_data.h" + +namespace tflite { + +// TODO(renjieliu): Migrate others to use ComputePaddingWithLeftover. +inline int ComputePadding(int stride, int dilation_rate, int in_size, + int filter_size, int out_size) { + int effective_filter_size = (filter_size - 1) * dilation_rate + 1; + int padding = ((out_size - 1) * stride + effective_filter_size - in_size) / 2; + return padding > 0 ? padding : 0; +} + +// It's not guaranteed that padding is symmetric. It's important to keep +// offset for algorithms need all paddings. +inline int ComputePaddingWithOffset(int stride, int dilation_rate, int in_size, + int filter_size, int out_size, + int* offset) { + int effective_filter_size = (filter_size - 1) * dilation_rate + 1; + int total_padding = + ((out_size - 1) * stride + effective_filter_size - in_size); + total_padding = total_padding > 0 ? total_padding : 0; + *offset = total_padding % 2; + return total_padding / 2; +} + +// Matching GetWindowedOutputSize in TensorFlow. +inline int ComputeOutSize(TfLitePadding padding, int image_size, + int filter_size, int stride, int dilation_rate = 1) { + int effective_filter_size = (filter_size - 1) * dilation_rate + 1; + switch (padding) { + case kTfLitePaddingSame: + return (image_size + stride - 1) / stride; + case kTfLitePaddingValid: + return (image_size + stride - effective_filter_size) / stride; + default: + return 0; + } +} + +inline TfLitePaddingValues ComputePaddingHeightWidth( + int stride_height, int stride_width, int dilation_rate_height, + int dilation_rate_width, int in_height, int in_width, int filter_height, + int filter_width, TfLitePadding padding, int* out_height, int* out_width) { + *out_width = ComputeOutSize(padding, in_width, filter_width, stride_width, + dilation_rate_width); + *out_height = ComputeOutSize(padding, in_height, filter_height, stride_height, + dilation_rate_height); + + TfLitePaddingValues padding_values; + int offset = 0; + padding_values.height = + ComputePaddingWithOffset(stride_height, dilation_rate_height, in_height, + filter_height, *out_height, &offset); + padding_values.height_offset = offset; + padding_values.width = + ComputePaddingWithOffset(stride_width, dilation_rate_width, in_width, + filter_width, *out_width, &offset); + padding_values.width_offset = offset; + return padding_values; +} +} // namespace tflite + +#endif // TENSORFLOW_LITE_KERNELS_PADDING_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/schema/schema_generated.h b/src/tensorflow_esp32/tensorflow/lite/schema/schema_generated.h old mode 100755 new mode 100644 similarity index 95% rename from src/tensorflow/lite/schema/schema_generated.h rename to src/tensorflow_esp32/tensorflow/lite/schema/schema_generated.h index 5daa078..cf42ef3 --- a/src/tensorflow/lite/schema/schema_generated.h +++ b/src/tensorflow_esp32/tensorflow/lite/schema/schema_generated.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2018 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -18,7 +19,7 @@ limitations under the License. #ifndef FLATBUFFERS_GENERATED_SCHEMA_TFLITE_H_ #define FLATBUFFERS_GENERATED_SCHEMA_TFLITE_H_ -#include "flatbuffers/flatbuffers.h" +#include "tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/flatbuffers.h" namespace tflite { @@ -28,12 +29,6 @@ struct CustomQuantizationT; struct QuantizationParameters; struct QuantizationParametersT; -struct DimensionMetadata; -struct DimensionMetadataT; - -struct SparsityParameters; -struct SparsityParametersT; - struct Tensor; struct TensorT; @@ -328,15 +323,6 @@ struct NonMaxSuppressionV5OptionsT; struct ScatterNdOptions; struct ScatterNdOptionsT; -struct SelectV2Options; -struct SelectV2OptionsT; - -struct DensifyOptions; -struct DensifyOptionsT; - -struct SegmentSumOptions; -struct SegmentSumOptionsT; - struct OperatorCode; struct OperatorCodeT; @@ -492,36 +478,6 @@ struct QuantizationDetailsUnion { bool VerifyQuantizationDetails(flatbuffers::Verifier &verifier, const void *obj, QuantizationDetails type); bool VerifyQuantizationDetailsVector(flatbuffers::Verifier &verifier, const flatbuffers::Vector> *values, const flatbuffers::Vector *types); -enum DimensionType { - DimensionType_DENSE = 0, - DimensionType_SPARSE_CSR = 1, - DimensionType_MIN = DimensionType_DENSE, - DimensionType_MAX = DimensionType_SPARSE_CSR -}; - -inline const DimensionType (&EnumValuesDimensionType())[2] { - static const DimensionType values[] = { - DimensionType_DENSE, - DimensionType_SPARSE_CSR - }; - return values; -} - -inline const char * const *EnumNamesDimensionType() { - static const char * const names[] = { - "DENSE", - "SPARSE_CSR", - nullptr - }; - return names; -} - -inline const char *EnumNameDimensionType(DimensionType e) { - if (e < DimensionType_DENSE || e > DimensionType_SPARSE_CSR) return ""; - const size_t index = static_cast(e); - return EnumNamesDimensionType()[index]; -} - enum BuiltinOperator { BuiltinOperator_ADD = 0, BuiltinOperator_AVERAGE_POOL_2D = 1, @@ -646,14 +602,11 @@ enum BuiltinOperator { BuiltinOperator_NON_MAX_SUPPRESSION_V4 = 120, BuiltinOperator_NON_MAX_SUPPRESSION_V5 = 121, BuiltinOperator_SCATTER_ND = 122, - BuiltinOperator_SELECT_V2 = 123, - BuiltinOperator_DENSIFY = 124, - BuiltinOperator_SEGMENT_SUM = 125, BuiltinOperator_MIN = BuiltinOperator_ADD, - BuiltinOperator_MAX = BuiltinOperator_SEGMENT_SUM + BuiltinOperator_MAX = BuiltinOperator_SCATTER_ND }; -inline const BuiltinOperator (&EnumValuesBuiltinOperator())[126] { +inline const BuiltinOperator (&EnumValuesBuiltinOperator())[123] { static const BuiltinOperator values[] = { BuiltinOperator_ADD, BuiltinOperator_AVERAGE_POOL_2D, @@ -777,10 +730,7 @@ inline const BuiltinOperator (&EnumValuesBuiltinOperator())[126] { BuiltinOperator_WHILE, BuiltinOperator_NON_MAX_SUPPRESSION_V4, BuiltinOperator_NON_MAX_SUPPRESSION_V5, - BuiltinOperator_SCATTER_ND, - BuiltinOperator_SELECT_V2, - BuiltinOperator_DENSIFY, - BuiltinOperator_SEGMENT_SUM + BuiltinOperator_SCATTER_ND }; return values; } @@ -910,16 +860,13 @@ inline const char * const *EnumNamesBuiltinOperator() { "NON_MAX_SUPPRESSION_V4", "NON_MAX_SUPPRESSION_V5", "SCATTER_ND", - "SELECT_V2", - "DENSIFY", - "SEGMENT_SUM", nullptr }; return names; } inline const char *EnumNameBuiltinOperator(BuiltinOperator e) { - if (e < BuiltinOperator_ADD || e > BuiltinOperator_SEGMENT_SUM) return ""; + if (e < BuiltinOperator_ADD || e > BuiltinOperator_SCATTER_ND) return ""; const size_t index = static_cast(e); return EnumNamesBuiltinOperator()[index]; } @@ -1023,14 +970,11 @@ enum BuiltinOptions { BuiltinOptions_NonMaxSuppressionV4Options = 95, BuiltinOptions_NonMaxSuppressionV5Options = 96, BuiltinOptions_ScatterNdOptions = 97, - BuiltinOptions_SelectV2Options = 98, - BuiltinOptions_DensifyOptions = 99, - BuiltinOptions_SegmentSumOptions = 100, BuiltinOptions_MIN = BuiltinOptions_NONE, - BuiltinOptions_MAX = BuiltinOptions_SegmentSumOptions + BuiltinOptions_MAX = BuiltinOptions_ScatterNdOptions }; -inline const BuiltinOptions (&EnumValuesBuiltinOptions())[101] { +inline const BuiltinOptions (&EnumValuesBuiltinOptions())[98] { static const BuiltinOptions values[] = { BuiltinOptions_NONE, BuiltinOptions_Conv2DOptions, @@ -1129,10 +1073,7 @@ inline const BuiltinOptions (&EnumValuesBuiltinOptions())[101] { BuiltinOptions_DepthToSpaceOptions, BuiltinOptions_NonMaxSuppressionV4Options, BuiltinOptions_NonMaxSuppressionV5Options, - BuiltinOptions_ScatterNdOptions, - BuiltinOptions_SelectV2Options, - BuiltinOptions_DensifyOptions, - BuiltinOptions_SegmentSumOptions + BuiltinOptions_ScatterNdOptions }; return values; } @@ -1237,16 +1178,13 @@ inline const char * const *EnumNamesBuiltinOptions() { "NonMaxSuppressionV4Options", "NonMaxSuppressionV5Options", "ScatterNdOptions", - "SelectV2Options", - "DensifyOptions", - "SegmentSumOptions", nullptr }; return names; } inline const char *EnumNameBuiltinOptions(BuiltinOptions e) { - if (e < BuiltinOptions_NONE || e > BuiltinOptions_SegmentSumOptions) return ""; + if (e < BuiltinOptions_NONE || e > BuiltinOptions_ScatterNdOptions) return ""; const size_t index = static_cast(e); return EnumNamesBuiltinOptions()[index]; } @@ -1643,18 +1581,6 @@ template<> struct BuiltinOptionsTraits { static const BuiltinOptions enum_value = BuiltinOptions_ScatterNdOptions; }; -template<> struct BuiltinOptionsTraits { - static const BuiltinOptions enum_value = BuiltinOptions_SelectV2Options; -}; - -template<> struct BuiltinOptionsTraits { - static const BuiltinOptions enum_value = BuiltinOptions_DensifyOptions; -}; - -template<> struct BuiltinOptionsTraits { - static const BuiltinOptions enum_value = BuiltinOptions_SegmentSumOptions; -}; - struct BuiltinOptionsUnion { BuiltinOptions type; void *value; @@ -2463,30 +2389,6 @@ struct BuiltinOptionsUnion { return type == BuiltinOptions_ScatterNdOptions ? reinterpret_cast(value) : nullptr; } - SelectV2OptionsT *AsSelectV2Options() { - return type == BuiltinOptions_SelectV2Options ? - reinterpret_cast(value) : nullptr; - } - const SelectV2OptionsT *AsSelectV2Options() const { - return type == BuiltinOptions_SelectV2Options ? - reinterpret_cast(value) : nullptr; - } - DensifyOptionsT *AsDensifyOptions() { - return type == BuiltinOptions_DensifyOptions ? - reinterpret_cast(value) : nullptr; - } - const DensifyOptionsT *AsDensifyOptions() const { - return type == BuiltinOptions_DensifyOptions ? - reinterpret_cast(value) : nullptr; - } - SegmentSumOptionsT *AsSegmentSumOptions() { - return type == BuiltinOptions_SegmentSumOptions ? - reinterpret_cast(value) : nullptr; - } - const SegmentSumOptionsT *AsSegmentSumOptions() const { - return type == BuiltinOptions_SegmentSumOptions ? - reinterpret_cast(value) : nullptr; - } }; bool VerifyBuiltinOptions(flatbuffers::Verifier &verifier, const void *obj, BuiltinOptions type); @@ -2966,206 +2868,6 @@ inline flatbuffers::Offset CreateQuantizationParametersD flatbuffers::Offset CreateQuantizationParameters(flatbuffers::FlatBufferBuilder &_fbb, const QuantizationParametersT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); -struct DimensionMetadataT : public flatbuffers::NativeTable { - typedef DimensionMetadata TableType; - DimensionType format; - int32_t dense_size; - std::vector array_segments; - std::vector array_indices; - DimensionMetadataT() - : format(DimensionType_DENSE), - dense_size(0) { - } -}; - -struct DimensionMetadata FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef DimensionMetadataT NativeTableType; - enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_FORMAT = 4, - VT_DENSE_SIZE = 6, - VT_ARRAY_SEGMENTS = 8, - VT_ARRAY_INDICES = 10 - }; - DimensionType format() const { - return static_cast(GetField(VT_FORMAT, 0)); - } - int32_t dense_size() const { - return GetField(VT_DENSE_SIZE, 0); - } - const flatbuffers::Vector *array_segments() const { - return GetPointer *>(VT_ARRAY_SEGMENTS); - } - const flatbuffers::Vector *array_indices() const { - return GetPointer *>(VT_ARRAY_INDICES); - } - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - VerifyField(verifier, VT_FORMAT) && - VerifyField(verifier, VT_DENSE_SIZE) && - VerifyOffset(verifier, VT_ARRAY_SEGMENTS) && - verifier.VerifyVector(array_segments()) && - VerifyOffset(verifier, VT_ARRAY_INDICES) && - verifier.VerifyVector(array_indices()) && - verifier.EndTable(); - } - DimensionMetadataT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; - void UnPackTo(DimensionMetadataT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; - static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); -}; - -struct DimensionMetadataBuilder { - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_format(DimensionType format) { - fbb_.AddElement(DimensionMetadata::VT_FORMAT, static_cast(format), 0); - } - void add_dense_size(int32_t dense_size) { - fbb_.AddElement(DimensionMetadata::VT_DENSE_SIZE, dense_size, 0); - } - void add_array_segments(flatbuffers::Offset> array_segments) { - fbb_.AddOffset(DimensionMetadata::VT_ARRAY_SEGMENTS, array_segments); - } - void add_array_indices(flatbuffers::Offset> array_indices) { - fbb_.AddOffset(DimensionMetadata::VT_ARRAY_INDICES, array_indices); - } - explicit DimensionMetadataBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - DimensionMetadataBuilder &operator=(const DimensionMetadataBuilder &); - flatbuffers::Offset Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset(end); - return o; - } -}; - -inline flatbuffers::Offset CreateDimensionMetadata( - flatbuffers::FlatBufferBuilder &_fbb, - DimensionType format = DimensionType_DENSE, - int32_t dense_size = 0, - flatbuffers::Offset> array_segments = 0, - flatbuffers::Offset> array_indices = 0) { - DimensionMetadataBuilder builder_(_fbb); - builder_.add_array_indices(array_indices); - builder_.add_array_segments(array_segments); - builder_.add_dense_size(dense_size); - builder_.add_format(format); - return builder_.Finish(); -} - -inline flatbuffers::Offset CreateDimensionMetadataDirect( - flatbuffers::FlatBufferBuilder &_fbb, - DimensionType format = DimensionType_DENSE, - int32_t dense_size = 0, - const std::vector *array_segments = nullptr, - const std::vector *array_indices = nullptr) { - auto array_segments__ = array_segments ? _fbb.CreateVector(*array_segments) : 0; - auto array_indices__ = array_indices ? _fbb.CreateVector(*array_indices) : 0; - return tflite::CreateDimensionMetadata( - _fbb, - format, - dense_size, - array_segments__, - array_indices__); -} - -flatbuffers::Offset CreateDimensionMetadata(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); - -struct SparsityParametersT : public flatbuffers::NativeTable { - typedef SparsityParameters TableType; - std::vector traversal_order; - std::vector block_map; - std::vector> dim_metadata; - SparsityParametersT() { - } -}; - -struct SparsityParameters FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef SparsityParametersT NativeTableType; - enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_TRAVERSAL_ORDER = 4, - VT_BLOCK_MAP = 6, - VT_DIM_METADATA = 8 - }; - const flatbuffers::Vector *traversal_order() const { - return GetPointer *>(VT_TRAVERSAL_ORDER); - } - const flatbuffers::Vector *block_map() const { - return GetPointer *>(VT_BLOCK_MAP); - } - const flatbuffers::Vector> *dim_metadata() const { - return GetPointer> *>(VT_DIM_METADATA); - } - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - VerifyOffset(verifier, VT_TRAVERSAL_ORDER) && - verifier.VerifyVector(traversal_order()) && - VerifyOffset(verifier, VT_BLOCK_MAP) && - verifier.VerifyVector(block_map()) && - VerifyOffset(verifier, VT_DIM_METADATA) && - verifier.VerifyVector(dim_metadata()) && - verifier.VerifyVectorOfTables(dim_metadata()) && - verifier.EndTable(); - } - SparsityParametersT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; - void UnPackTo(SparsityParametersT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; - static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); -}; - -struct SparsityParametersBuilder { - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - void add_traversal_order(flatbuffers::Offset> traversal_order) { - fbb_.AddOffset(SparsityParameters::VT_TRAVERSAL_ORDER, traversal_order); - } - void add_block_map(flatbuffers::Offset> block_map) { - fbb_.AddOffset(SparsityParameters::VT_BLOCK_MAP, block_map); - } - void add_dim_metadata(flatbuffers::Offset>> dim_metadata) { - fbb_.AddOffset(SparsityParameters::VT_DIM_METADATA, dim_metadata); - } - explicit SparsityParametersBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - SparsityParametersBuilder &operator=(const SparsityParametersBuilder &); - flatbuffers::Offset Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset(end); - return o; - } -}; - -inline flatbuffers::Offset CreateSparsityParameters( - flatbuffers::FlatBufferBuilder &_fbb, - flatbuffers::Offset> traversal_order = 0, - flatbuffers::Offset> block_map = 0, - flatbuffers::Offset>> dim_metadata = 0) { - SparsityParametersBuilder builder_(_fbb); - builder_.add_dim_metadata(dim_metadata); - builder_.add_block_map(block_map); - builder_.add_traversal_order(traversal_order); - return builder_.Finish(); -} - -inline flatbuffers::Offset CreateSparsityParametersDirect( - flatbuffers::FlatBufferBuilder &_fbb, - const std::vector *traversal_order = nullptr, - const std::vector *block_map = nullptr, - const std::vector> *dim_metadata = nullptr) { - auto traversal_order__ = traversal_order ? _fbb.CreateVector(*traversal_order) : 0; - auto block_map__ = block_map ? _fbb.CreateVector(*block_map) : 0; - auto dim_metadata__ = dim_metadata ? _fbb.CreateVector>(*dim_metadata) : 0; - return tflite::CreateSparsityParameters( - _fbb, - traversal_order__, - block_map__, - dim_metadata__); -} - -flatbuffers::Offset CreateSparsityParameters(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); - struct TensorT : public flatbuffers::NativeTable { typedef Tensor TableType; std::vector shape; @@ -3174,7 +2876,6 @@ struct TensorT : public flatbuffers::NativeTable { std::string name; std::unique_ptr quantization; bool is_variable; - std::unique_ptr sparsity; TensorT() : type(TensorType_FLOAT32), buffer(0), @@ -3190,8 +2891,7 @@ struct Tensor FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { VT_BUFFER = 8, VT_NAME = 10, VT_QUANTIZATION = 12, - VT_IS_VARIABLE = 14, - VT_SPARSITY = 16 + VT_IS_VARIABLE = 14 }; const flatbuffers::Vector *shape() const { return GetPointer *>(VT_SHAPE); @@ -3211,9 +2911,6 @@ struct Tensor FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { bool is_variable() const { return GetField(VT_IS_VARIABLE, 0) != 0; } - const SparsityParameters *sparsity() const { - return GetPointer(VT_SPARSITY); - } bool Verify(flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyOffset(verifier, VT_SHAPE) && @@ -3225,8 +2922,6 @@ struct Tensor FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { VerifyOffset(verifier, VT_QUANTIZATION) && verifier.VerifyTable(quantization()) && VerifyField(verifier, VT_IS_VARIABLE) && - VerifyOffset(verifier, VT_SPARSITY) && - verifier.VerifyTable(sparsity()) && verifier.EndTable(); } TensorT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; @@ -3255,9 +2950,6 @@ struct TensorBuilder { void add_is_variable(bool is_variable) { fbb_.AddElement(Tensor::VT_IS_VARIABLE, static_cast(is_variable), 0); } - void add_sparsity(flatbuffers::Offset sparsity) { - fbb_.AddOffset(Tensor::VT_SPARSITY, sparsity); - } explicit TensorBuilder(flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); @@ -3277,10 +2969,8 @@ inline flatbuffers::Offset CreateTensor( uint32_t buffer = 0, flatbuffers::Offset name = 0, flatbuffers::Offset quantization = 0, - bool is_variable = false, - flatbuffers::Offset sparsity = 0) { + bool is_variable = false) { TensorBuilder builder_(_fbb); - builder_.add_sparsity(sparsity); builder_.add_quantization(quantization); builder_.add_name(name); builder_.add_buffer(buffer); @@ -3297,8 +2987,7 @@ inline flatbuffers::Offset CreateTensorDirect( uint32_t buffer = 0, const char *name = nullptr, flatbuffers::Offset quantization = 0, - bool is_variable = false, - flatbuffers::Offset sparsity = 0) { + bool is_variable = false) { auto shape__ = shape ? _fbb.CreateVector(*shape) : 0; auto name__ = name ? _fbb.CreateString(name) : 0; return tflite::CreateTensor( @@ -3308,8 +2997,7 @@ inline flatbuffers::Offset CreateTensorDirect( buffer, name__, quantization, - is_variable, - sparsity); + is_variable); } flatbuffers::Offset CreateTensor(flatbuffers::FlatBufferBuilder &_fbb, const TensorT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); @@ -4813,29 +4501,22 @@ flatbuffers::Offset CreateBidirectionalSequenc struct ResizeBilinearOptionsT : public flatbuffers::NativeTable { typedef ResizeBilinearOptions TableType; bool align_corners; - bool half_pixel_centers; ResizeBilinearOptionsT() - : align_corners(false), - half_pixel_centers(false) { + : align_corners(false) { } }; struct ResizeBilinearOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { typedef ResizeBilinearOptionsT NativeTableType; enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE { - VT_ALIGN_CORNERS = 8, - VT_HALF_PIXEL_CENTERS = 10 + VT_ALIGN_CORNERS = 8 }; bool align_corners() const { return GetField(VT_ALIGN_CORNERS, 0) != 0; } - bool half_pixel_centers() const { - return GetField(VT_HALF_PIXEL_CENTERS, 0) != 0; - } bool Verify(flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyField(verifier, VT_ALIGN_CORNERS) && - VerifyField(verifier, VT_HALF_PIXEL_CENTERS) && verifier.EndTable(); } ResizeBilinearOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; @@ -4849,9 +4530,6 @@ struct ResizeBilinearOptionsBuilder { void add_align_corners(bool align_corners) { fbb_.AddElement(ResizeBilinearOptions::VT_ALIGN_CORNERS, static_cast(align_corners), 0); } - void add_half_pixel_centers(bool half_pixel_centers) { - fbb_.AddElement(ResizeBilinearOptions::VT_HALF_PIXEL_CENTERS, static_cast(half_pixel_centers), 0); - } explicit ResizeBilinearOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); @@ -4866,10 +4544,8 @@ struct ResizeBilinearOptionsBuilder { inline flatbuffers::Offset CreateResizeBilinearOptions( flatbuffers::FlatBufferBuilder &_fbb, - bool align_corners = false, - bool half_pixel_centers = false) { + bool align_corners = false) { ResizeBilinearOptionsBuilder builder_(_fbb); - builder_.add_half_pixel_centers(half_pixel_centers); builder_.add_align_corners(align_corners); return builder_.Finish(); } @@ -8612,126 +8288,6 @@ inline flatbuffers::Offset CreateScatterNdOptions( flatbuffers::Offset CreateScatterNdOptions(flatbuffers::FlatBufferBuilder &_fbb, const ScatterNdOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); -struct SelectV2OptionsT : public flatbuffers::NativeTable { - typedef SelectV2Options TableType; - SelectV2OptionsT() { - } -}; - -struct SelectV2Options FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef SelectV2OptionsT NativeTableType; - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - verifier.EndTable(); - } - SelectV2OptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; - void UnPackTo(SelectV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; - static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); -}; - -struct SelectV2OptionsBuilder { - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - explicit SelectV2OptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - SelectV2OptionsBuilder &operator=(const SelectV2OptionsBuilder &); - flatbuffers::Offset Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset(end); - return o; - } -}; - -inline flatbuffers::Offset CreateSelectV2Options( - flatbuffers::FlatBufferBuilder &_fbb) { - SelectV2OptionsBuilder builder_(_fbb); - return builder_.Finish(); -} - -flatbuffers::Offset CreateSelectV2Options(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); - -struct DensifyOptionsT : public flatbuffers::NativeTable { - typedef DensifyOptions TableType; - DensifyOptionsT() { - } -}; - -struct DensifyOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef DensifyOptionsT NativeTableType; - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - verifier.EndTable(); - } - DensifyOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; - void UnPackTo(DensifyOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; - static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); -}; - -struct DensifyOptionsBuilder { - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - explicit DensifyOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - DensifyOptionsBuilder &operator=(const DensifyOptionsBuilder &); - flatbuffers::Offset Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset(end); - return o; - } -}; - -inline flatbuffers::Offset CreateDensifyOptions( - flatbuffers::FlatBufferBuilder &_fbb) { - DensifyOptionsBuilder builder_(_fbb); - return builder_.Finish(); -} - -flatbuffers::Offset CreateDensifyOptions(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); - -struct SegmentSumOptionsT : public flatbuffers::NativeTable { - typedef SegmentSumOptions TableType; - SegmentSumOptionsT() { - } -}; - -struct SegmentSumOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { - typedef SegmentSumOptionsT NativeTableType; - bool Verify(flatbuffers::Verifier &verifier) const { - return VerifyTableStart(verifier) && - verifier.EndTable(); - } - SegmentSumOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; - void UnPackTo(SegmentSumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; - static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); -}; - -struct SegmentSumOptionsBuilder { - flatbuffers::FlatBufferBuilder &fbb_; - flatbuffers::uoffset_t start_; - explicit SegmentSumOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) - : fbb_(_fbb) { - start_ = fbb_.StartTable(); - } - SegmentSumOptionsBuilder &operator=(const SegmentSumOptionsBuilder &); - flatbuffers::Offset Finish() { - const auto end = fbb_.EndTable(start_); - auto o = flatbuffers::Offset(end); - return o; - } -}; - -inline flatbuffers::Offset CreateSegmentSumOptions( - flatbuffers::FlatBufferBuilder &_fbb) { - SegmentSumOptionsBuilder builder_(_fbb); - return builder_.Finish(); -} - -flatbuffers::Offset CreateSegmentSumOptions(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); - struct OperatorCodeT : public flatbuffers::NativeTable { typedef OperatorCode TableType; BuiltinOperator builtin_code; @@ -9159,15 +8715,6 @@ struct Operator FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { const ScatterNdOptions *builtin_options_as_ScatterNdOptions() const { return builtin_options_type() == BuiltinOptions_ScatterNdOptions ? static_cast(builtin_options()) : nullptr; } - const SelectV2Options *builtin_options_as_SelectV2Options() const { - return builtin_options_type() == BuiltinOptions_SelectV2Options ? static_cast(builtin_options()) : nullptr; - } - const DensifyOptions *builtin_options_as_DensifyOptions() const { - return builtin_options_type() == BuiltinOptions_DensifyOptions ? static_cast(builtin_options()) : nullptr; - } - const SegmentSumOptions *builtin_options_as_SegmentSumOptions() const { - return builtin_options_type() == BuiltinOptions_SegmentSumOptions ? static_cast(builtin_options()) : nullptr; - } const flatbuffers::Vector *custom_options() const { return GetPointer *>(VT_CUSTOM_OPTIONS); } @@ -9592,18 +9139,6 @@ template<> inline const ScatterNdOptions *Operator::builtin_options_as inline const SelectV2Options *Operator::builtin_options_as() const { - return builtin_options_as_SelectV2Options(); -} - -template<> inline const DensifyOptions *Operator::builtin_options_as() const { - return builtin_options_as_DensifyOptions(); -} - -template<> inline const SegmentSumOptions *Operator::builtin_options_as() const { - return builtin_options_as_SegmentSumOptions(); -} - struct OperatorBuilder { flatbuffers::FlatBufferBuilder &fbb_; flatbuffers::uoffset_t start_; @@ -10192,73 +9727,6 @@ inline flatbuffers::Offset CreateQuantizationParameters( _quantized_dimension); } -inline DimensionMetadataT *DimensionMetadata::UnPack(const flatbuffers::resolver_function_t *_resolver) const { - auto _o = new DimensionMetadataT(); - UnPackTo(_o, _resolver); - return _o; -} - -inline void DimensionMetadata::UnPackTo(DimensionMetadataT *_o, const flatbuffers::resolver_function_t *_resolver) const { - (void)_o; - (void)_resolver; - { auto _e = format(); _o->format = _e; }; - { auto _e = dense_size(); _o->dense_size = _e; }; - { auto _e = array_segments(); if (_e) { _o->array_segments.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->array_segments[_i] = _e->Get(_i); } } }; - { auto _e = array_indices(); if (_e) { _o->array_indices.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->array_indices[_i] = _e->Get(_i); } } }; -} - -inline flatbuffers::Offset DimensionMetadata::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT* _o, const flatbuffers::rehasher_function_t *_rehasher) { - return CreateDimensionMetadata(_fbb, _o, _rehasher); -} - -inline flatbuffers::Offset CreateDimensionMetadata(flatbuffers::FlatBufferBuilder &_fbb, const DimensionMetadataT *_o, const flatbuffers::rehasher_function_t *_rehasher) { - (void)_rehasher; - (void)_o; - struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DimensionMetadataT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; - auto _format = _o->format; - auto _dense_size = _o->dense_size; - auto _array_segments = _o->array_segments.size() ? _fbb.CreateVector(_o->array_segments) : 0; - auto _array_indices = _o->array_indices.size() ? _fbb.CreateVector(_o->array_indices) : 0; - return tflite::CreateDimensionMetadata( - _fbb, - _format, - _dense_size, - _array_segments, - _array_indices); -} - -inline SparsityParametersT *SparsityParameters::UnPack(const flatbuffers::resolver_function_t *_resolver) const { - auto _o = new SparsityParametersT(); - UnPackTo(_o, _resolver); - return _o; -} - -inline void SparsityParameters::UnPackTo(SparsityParametersT *_o, const flatbuffers::resolver_function_t *_resolver) const { - (void)_o; - (void)_resolver; - { auto _e = traversal_order(); if (_e) { _o->traversal_order.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->traversal_order[_i] = _e->Get(_i); } } }; - { auto _e = block_map(); if (_e) { _o->block_map.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->block_map[_i] = _e->Get(_i); } } }; - { auto _e = dim_metadata(); if (_e) { _o->dim_metadata.resize(_e->size()); for (flatbuffers::uoffset_t _i = 0; _i < _e->size(); _i++) { _o->dim_metadata[_i] = std::unique_ptr(_e->Get(_i)->UnPack(_resolver)); } } }; -} - -inline flatbuffers::Offset SparsityParameters::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT* _o, const flatbuffers::rehasher_function_t *_rehasher) { - return CreateSparsityParameters(_fbb, _o, _rehasher); -} - -inline flatbuffers::Offset CreateSparsityParameters(flatbuffers::FlatBufferBuilder &_fbb, const SparsityParametersT *_o, const flatbuffers::rehasher_function_t *_rehasher) { - (void)_rehasher; - (void)_o; - struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SparsityParametersT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; - auto _traversal_order = _o->traversal_order.size() ? _fbb.CreateVector(_o->traversal_order) : 0; - auto _block_map = _o->block_map.size() ? _fbb.CreateVector(_o->block_map) : 0; - auto _dim_metadata = _o->dim_metadata.size() ? _fbb.CreateVector> (_o->dim_metadata.size(), [](size_t i, _VectorArgs *__va) { return CreateDimensionMetadata(*__va->__fbb, __va->__o->dim_metadata[i].get(), __va->__rehasher); }, &_va ) : 0; - return tflite::CreateSparsityParameters( - _fbb, - _traversal_order, - _block_map, - _dim_metadata); -} - inline TensorT *Tensor::UnPack(const flatbuffers::resolver_function_t *_resolver) const { auto _o = new TensorT(); UnPackTo(_o, _resolver); @@ -10274,7 +9742,6 @@ inline void Tensor::UnPackTo(TensorT *_o, const flatbuffers::resolver_function_t { auto _e = name(); if (_e) _o->name = _e->str(); }; { auto _e = quantization(); if (_e) _o->quantization = std::unique_ptr(_e->UnPack(_resolver)); }; { auto _e = is_variable(); _o->is_variable = _e; }; - { auto _e = sparsity(); if (_e) _o->sparsity = std::unique_ptr(_e->UnPack(_resolver)); }; } inline flatbuffers::Offset Tensor::Pack(flatbuffers::FlatBufferBuilder &_fbb, const TensorT* _o, const flatbuffers::rehasher_function_t *_rehasher) { @@ -10291,7 +9758,6 @@ inline flatbuffers::Offset CreateTensor(flatbuffers::FlatBufferBuilder & auto _name = _o->name.empty() ? 0 : _fbb.CreateString(_o->name); auto _quantization = _o->quantization ? CreateQuantizationParameters(_fbb, _o->quantization.get(), _rehasher) : 0; auto _is_variable = _o->is_variable; - auto _sparsity = _o->sparsity ? CreateSparsityParameters(_fbb, _o->sparsity.get(), _rehasher) : 0; return tflite::CreateTensor( _fbb, _shape, @@ -10299,8 +9765,7 @@ inline flatbuffers::Offset CreateTensor(flatbuffers::FlatBufferBuilder & _buffer, _name, _quantization, - _is_variable, - _sparsity); + _is_variable); } inline Conv2DOptionsT *Conv2DOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { @@ -10921,7 +10386,6 @@ inline void ResizeBilinearOptions::UnPackTo(ResizeBilinearOptionsT *_o, const fl (void)_o; (void)_resolver; { auto _e = align_corners(); _o->align_corners = _e; }; - { auto _e = half_pixel_centers(); _o->half_pixel_centers = _e; }; } inline flatbuffers::Offset ResizeBilinearOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ResizeBilinearOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { @@ -10933,11 +10397,9 @@ inline flatbuffers::Offset CreateResizeBilinearOptions(fl (void)_o; struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ResizeBilinearOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; auto _align_corners = _o->align_corners; - auto _half_pixel_centers = _o->half_pixel_centers; return tflite::CreateResizeBilinearOptions( _fbb, - _align_corners, - _half_pixel_centers); + _align_corners); } inline ResizeNearestNeighborOptionsT *ResizeNearestNeighborOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { @@ -12855,75 +12317,6 @@ inline flatbuffers::Offset CreateScatterNdOptions(flatbuffers: _fbb); } -inline SelectV2OptionsT *SelectV2Options::UnPack(const flatbuffers::resolver_function_t *_resolver) const { - auto _o = new SelectV2OptionsT(); - UnPackTo(_o, _resolver); - return _o; -} - -inline void SelectV2Options::UnPackTo(SelectV2OptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { - (void)_o; - (void)_resolver; -} - -inline flatbuffers::Offset SelectV2Options::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { - return CreateSelectV2Options(_fbb, _o, _rehasher); -} - -inline flatbuffers::Offset CreateSelectV2Options(flatbuffers::FlatBufferBuilder &_fbb, const SelectV2OptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { - (void)_rehasher; - (void)_o; - struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SelectV2OptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; - return tflite::CreateSelectV2Options( - _fbb); -} - -inline DensifyOptionsT *DensifyOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { - auto _o = new DensifyOptionsT(); - UnPackTo(_o, _resolver); - return _o; -} - -inline void DensifyOptions::UnPackTo(DensifyOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { - (void)_o; - (void)_resolver; -} - -inline flatbuffers::Offset DensifyOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { - return CreateDensifyOptions(_fbb, _o, _rehasher); -} - -inline flatbuffers::Offset CreateDensifyOptions(flatbuffers::FlatBufferBuilder &_fbb, const DensifyOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { - (void)_rehasher; - (void)_o; - struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const DensifyOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; - return tflite::CreateDensifyOptions( - _fbb); -} - -inline SegmentSumOptionsT *SegmentSumOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { - auto _o = new SegmentSumOptionsT(); - UnPackTo(_o, _resolver); - return _o; -} - -inline void SegmentSumOptions::UnPackTo(SegmentSumOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { - (void)_o; - (void)_resolver; -} - -inline flatbuffers::Offset SegmentSumOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { - return CreateSegmentSumOptions(_fbb, _o, _rehasher); -} - -inline flatbuffers::Offset CreateSegmentSumOptions(flatbuffers::FlatBufferBuilder &_fbb, const SegmentSumOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { - (void)_rehasher; - (void)_o; - struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const SegmentSumOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; - return tflite::CreateSegmentSumOptions( - _fbb); -} - inline OperatorCodeT *OperatorCode::UnPack(const flatbuffers::resolver_function_t *_resolver) const { auto _o = new OperatorCodeT(); UnPackTo(_o, _resolver); @@ -13152,7 +12545,7 @@ inline bool VerifyQuantizationDetails(flatbuffers::Verifier &verifier, const voi auto ptr = reinterpret_cast(obj); return verifier.VerifyTable(ptr); } - default: return true; + default: return false; } } @@ -13605,19 +12998,7 @@ inline bool VerifyBuiltinOptions(flatbuffers::Verifier &verifier, const void *ob auto ptr = reinterpret_cast(obj); return verifier.VerifyTable(ptr); } - case BuiltinOptions_SelectV2Options: { - auto ptr = reinterpret_cast(obj); - return verifier.VerifyTable(ptr); - } - case BuiltinOptions_DensifyOptions: { - auto ptr = reinterpret_cast(obj); - return verifier.VerifyTable(ptr); - } - case BuiltinOptions_SegmentSumOptions: { - auto ptr = reinterpret_cast(obj); - return verifier.VerifyTable(ptr); - } - default: return true; + default: return false; } } @@ -14023,18 +13404,6 @@ inline void *BuiltinOptionsUnion::UnPack(const void *obj, BuiltinOptions type, c auto ptr = reinterpret_cast(obj); return ptr->UnPack(resolver); } - case BuiltinOptions_SelectV2Options: { - auto ptr = reinterpret_cast(obj); - return ptr->UnPack(resolver); - } - case BuiltinOptions_DensifyOptions: { - auto ptr = reinterpret_cast(obj); - return ptr->UnPack(resolver); - } - case BuiltinOptions_SegmentSumOptions: { - auto ptr = reinterpret_cast(obj); - return ptr->UnPack(resolver); - } default: return nullptr; } } @@ -14429,18 +13798,6 @@ inline flatbuffers::Offset BuiltinOptionsUnion::Pack(flatbuffers::FlatBuff auto ptr = reinterpret_cast(value); return CreateScatterNdOptions(_fbb, ptr, _rehasher).Union(); } - case BuiltinOptions_SelectV2Options: { - auto ptr = reinterpret_cast(value); - return CreateSelectV2Options(_fbb, ptr, _rehasher).Union(); - } - case BuiltinOptions_DensifyOptions: { - auto ptr = reinterpret_cast(value); - return CreateDensifyOptions(_fbb, ptr, _rehasher).Union(); - } - case BuiltinOptions_SegmentSumOptions: { - auto ptr = reinterpret_cast(value); - return CreateSegmentSumOptions(_fbb, ptr, _rehasher).Union(); - } default: return 0; } } @@ -14835,18 +14192,6 @@ inline BuiltinOptionsUnion::BuiltinOptionsUnion(const BuiltinOptionsUnion &u) FL value = new ScatterNdOptionsT(*reinterpret_cast(u.value)); break; } - case BuiltinOptions_SelectV2Options: { - value = new SelectV2OptionsT(*reinterpret_cast(u.value)); - break; - } - case BuiltinOptions_DensifyOptions: { - value = new DensifyOptionsT(*reinterpret_cast(u.value)); - break; - } - case BuiltinOptions_SegmentSumOptions: { - value = new SegmentSumOptionsT(*reinterpret_cast(u.value)); - break; - } default: break; } @@ -15339,21 +14684,6 @@ inline void BuiltinOptionsUnion::Reset() { delete ptr; break; } - case BuiltinOptions_SelectV2Options: { - auto ptr = reinterpret_cast(value); - delete ptr; - break; - } - case BuiltinOptions_DensifyOptions: { - auto ptr = reinterpret_cast(value); - delete ptr; - break; - } - case BuiltinOptions_SegmentSumOptions: { - auto ptr = reinterpret_cast(value); - delete ptr; - break; - } default: break; } value = nullptr; @@ -15412,3 +14742,5 @@ inline std::unique_ptr UnPackModel( } // namespace tflite #endif // FLATBUFFERS_GENERATED_SCHEMA_TFLITE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/string_type.h b/src/tensorflow_esp32/tensorflow/lite/string_type.h similarity index 94% rename from src/tensorflow/lite/string_type.h rename to src/tensorflow_esp32/tensorflow/lite/string_type.h index f5a7f83..97df246 100644 --- a/src/tensorflow/lite/string_type.h +++ b/src/tensorflow_esp32/tensorflow/lite/string_type.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -25,3 +26,5 @@ using std::string; } // namespace tflite #endif // TENSORFLOW_LITE_STRING_TYPE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow/lite/string_util.h b/src/tensorflow_esp32/tensorflow/lite/string_util.h similarity index 94% rename from src/tensorflow/lite/string_util.h rename to src/tensorflow_esp32/tensorflow/lite/string_util.h index 779b1e1..44c6b02 100644 --- a/src/tensorflow/lite/string_util.h +++ b/src/tensorflow_esp32/tensorflow/lite/string_util.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright 2017 The TensorFlow Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); @@ -22,7 +23,7 @@ limitations under the License. // Example of a string tensor: // [ // 2, 0, 0, 0, # 2 strings. -// 16, 0, 0, 0, # 0-th string starts from index 16. +// 16, 0, 0, 0, # 0-th string starts from index 12. // 18, 0, 0, 0, # 1-st string starts from index 18. // 18, 0, 0, 0, # total length of array. // 'A', 'B', # 0-th string [16..17]: "AB" @@ -42,8 +43,8 @@ limitations under the License. #include -#include "tensorflow/lite/c/common.h" -#include "tensorflow/lite/string_type.h" +#include "tensorflow_esp32/tensorflow/lite/c/c_api_internal.h" +#include "tensorflow_esp32/tensorflow/lite/string_type.h" namespace tflite { @@ -101,3 +102,5 @@ StringRef GetString(const TfLiteTensor* tensor, int string_index); } // namespace tflite #endif // TENSORFLOW_LITE_STRING_UTIL_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/tensorflow/lite/version.h b/src/tensorflow_esp32/tensorflow/lite/version.h new file mode 100644 index 0000000..28a35fd --- /dev/null +++ b/src/tensorflow_esp32/tensorflow/lite/version.h @@ -0,0 +1,32 @@ +#if defined(ESP32) +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_LITE_VERSION_H_ +#define TENSORFLOW_LITE_VERSION_H_ + +#include "tensorflow_esp32/tensorflow/core/public/version.h" + +// The version number of the Schema. Ideally all changes will be backward +// compatible. If that ever changes, we must ensure that version is the first +// entry in the new tflite root so that we can see that version is not 1. +#define TFLITE_SCHEMA_VERSION (3) + +// TensorFlow Lite Runtime version. +// This value is currently shared with that of TensorFlow. +#define TFLITE_VERSION_STRING TF_VERSION_STRING + +#endif // TENSORFLOW_LITE_VERSION_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/third_party/flatbuffers/LICENSE.txt b/src/tensorflow_esp32/third_party/flatbuffers/LICENSE.txt new file mode 100644 index 0000000..a4c5efd --- /dev/null +++ b/src/tensorflow_esp32/third_party/flatbuffers/LICENSE.txt @@ -0,0 +1,202 @@ + + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright 2014 Google Inc. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. diff --git a/src/flatbuffers/base.h b/src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/base.h similarity index 98% rename from src/flatbuffers/base.h rename to src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/base.h index d7faa7d..4ca8caf 100644 --- a/src/flatbuffers/base.h +++ b/src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/base.h @@ -1,3 +1,4 @@ +#if defined(ESP32) #ifndef FLATBUFFERS_BASE_H_ #define FLATBUFFERS_BASE_H_ @@ -33,9 +34,6 @@ #include -#include - - #include #include #include @@ -51,7 +49,7 @@ #include #endif -#include "flatbuffers/stl_emulation.h" +#include "tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/stl_emulation.h" // Note the __clang__ check is needed, because clang presents itself // as an older GNUC compiler (4.2). @@ -375,3 +373,5 @@ inline size_t PaddingBytes(size_t buf_size, size_t scalar_size) { } // namespace flatbuffers #endif // FLATBUFFERS_BASE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/flatbuffers/flatbuffers.h b/src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/flatbuffers.h similarity index 99% rename from src/flatbuffers/flatbuffers.h rename to src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/flatbuffers.h index a1a95f0..d525ba1 100644 --- a/src/flatbuffers/flatbuffers.h +++ b/src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/flatbuffers.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* * Copyright 2014 Google Inc. All rights reserved. * @@ -17,7 +18,7 @@ #ifndef FLATBUFFERS_H_ #define FLATBUFFERS_H_ -#include "flatbuffers/base.h" +#include "tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/base.h" #if defined(FLATBUFFERS_NAN_DEFAULTS) #include @@ -2611,3 +2612,5 @@ volatile __attribute__((weak)) const char *flatbuffer_version_string = // clang-format on #endif // FLATBUFFERS_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/flatbuffers/stl_emulation.h b/src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/stl_emulation.h similarity index 99% rename from src/flatbuffers/stl_emulation.h rename to src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/stl_emulation.h index 6f6e766..ab59317 100644 --- a/src/flatbuffers/stl_emulation.h +++ b/src/tensorflow_esp32/third_party/flatbuffers/include/flatbuffers/stl_emulation.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* * Copyright 2017 Google Inc. All rights reserved. * @@ -273,3 +274,5 @@ inline void vector_emplace_back(std::vector *vector, V &&data) { } // namespace flatbuffers #endif // FLATBUFFERS_STL_EMULATION_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/third_party/gemmlowp/LICENSE b/src/tensorflow_esp32/third_party/gemmlowp/LICENSE new file mode 100644 index 0000000..d645695 --- /dev/null +++ b/src/tensorflow_esp32/third_party/gemmlowp/LICENSE @@ -0,0 +1,202 @@ + + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. diff --git a/src/fixedpoint/fixedpoint.h b/src/tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h similarity index 99% rename from src/fixedpoint/fixedpoint.h rename to src/tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h index 51b5aff..2654ab2 100644 --- a/src/fixedpoint/fixedpoint.h +++ b/src/tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint.h @@ -1,3 +1,4 @@ +#if defined(ESP32) // Copyright 2015 The Gemmlowp Authors. All Rights Reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); @@ -898,3 +899,5 @@ FixedPoint logistic(FixedPoint a) { #endif #endif // GEMMLOWP_INTERNAL_FIXEDPOINT_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/fixedpoint/fixedpoint_sse.h b/src/tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint_sse.h similarity index 99% rename from src/fixedpoint/fixedpoint_sse.h rename to src/tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint_sse.h index a1fae32..8990c6f 100644 --- a/src/fixedpoint/fixedpoint_sse.h +++ b/src/tensorflow_esp32/third_party/gemmlowp/fixedpoint/fixedpoint_sse.h @@ -1,3 +1,4 @@ +#if defined(ESP32) // Copyright 2015 Google Inc. All Rights Reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); @@ -382,3 +383,5 @@ inline int16x8_m128i SaturatingAdd(int16x8_m128i a, int16x8_m128i b) { } // end namespace gemmlowp #endif // GEMMLOWP_INTERNAL_FIXEDPOINT_SSE_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/internal/detect_platform.h b/src/tensorflow_esp32/third_party/gemmlowp/internal/detect_platform.h similarity index 98% rename from src/internal/detect_platform.h rename to src/tensorflow_esp32/third_party/gemmlowp/internal/detect_platform.h index 6f06d19..463c8d3 100644 --- a/src/internal/detect_platform.h +++ b/src/tensorflow_esp32/third_party/gemmlowp/internal/detect_platform.h @@ -1,3 +1,4 @@ +#if defined(ESP32) // Copyright 2018 The Gemmlowp Authors. All Rights Reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); @@ -164,3 +165,5 @@ #endif #endif // GEMMLOWP_INTERNAL_DETECT_PLATFORM_H_ + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/third_party/kissfft/COPYING b/src/tensorflow_esp32/third_party/kissfft/COPYING new file mode 100644 index 0000000..2fc6685 --- /dev/null +++ b/src/tensorflow_esp32/third_party/kissfft/COPYING @@ -0,0 +1,11 @@ +Copyright (c) 2003-2010 Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/src/kissfft/_kiss_fft_guts.h b/src/tensorflow_esp32/third_party/kissfft/_kiss_fft_guts.h similarity index 99% rename from src/kissfft/_kiss_fft_guts.h rename to src/tensorflow_esp32/third_party/kissfft/_kiss_fft_guts.h index ba66144..428af36 100644 --- a/src/kissfft/_kiss_fft_guts.h +++ b/src/tensorflow_esp32/third_party/kissfft/_kiss_fft_guts.h @@ -1,3 +1,4 @@ +#if defined(ESP32) /* Copyright (c) 2003-2010, Mark Borgerding @@ -162,3 +163,5 @@ struct kiss_fft_state{ #define KISS_FFT_TMP_ALLOC(nbytes) KISS_FFT_MALLOC(nbytes) #define KISS_FFT_TMP_FREE(ptr) KISS_FFT_FREE(ptr) #endif + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/third_party/kissfft/kiss_fft.c b/src/tensorflow_esp32/third_party/kissfft/kiss_fft.c new file mode 100644 index 0000000..071a0ea --- /dev/null +++ b/src/tensorflow_esp32/third_party/kissfft/kiss_fft.c @@ -0,0 +1,411 @@ +#if defined(ESP32) +/* +Copyright (c) 2003-2010, Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + + +#include "_kiss_fft_guts.h" +/* The guts header contains all the multiplication and addition macros that are defined for + fixed or floating point complex numbers. It also delares the kf_ internal functions. + */ + +static void kf_bfly2( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m + ) +{ + kiss_fft_cpx * Fout2; + kiss_fft_cpx * tw1 = st->twiddles; + kiss_fft_cpx t; + Fout2 = Fout + m; + do{ + C_FIXDIV(*Fout,2); C_FIXDIV(*Fout2,2); + + C_MUL (t, *Fout2 , *tw1); + tw1 += fstride; + C_SUB( *Fout2 , *Fout , t ); + C_ADDTO( *Fout , t ); + ++Fout2; + ++Fout; + }while (--m); +} + +static void kf_bfly4( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + const size_t m + ) +{ + kiss_fft_cpx *tw1,*tw2,*tw3; + kiss_fft_cpx scratch[6]; + size_t k=m; + const size_t m2=2*m; + const size_t m3=3*m; + + + tw3 = tw2 = tw1 = st->twiddles; + + do { + C_FIXDIV(*Fout,4); C_FIXDIV(Fout[m],4); C_FIXDIV(Fout[m2],4); C_FIXDIV(Fout[m3],4); + + C_MUL(scratch[0],Fout[m] , *tw1 ); + C_MUL(scratch[1],Fout[m2] , *tw2 ); + C_MUL(scratch[2],Fout[m3] , *tw3 ); + + C_SUB( scratch[5] , *Fout, scratch[1] ); + C_ADDTO(*Fout, scratch[1]); + C_ADD( scratch[3] , scratch[0] , scratch[2] ); + C_SUB( scratch[4] , scratch[0] , scratch[2] ); + C_SUB( Fout[m2], *Fout, scratch[3] ); + tw1 += fstride; + tw2 += fstride*2; + tw3 += fstride*3; + C_ADDTO( *Fout , scratch[3] ); + + if(st->inverse) { + Fout[m].r = scratch[5].r - scratch[4].i; + Fout[m].i = scratch[5].i + scratch[4].r; + Fout[m3].r = scratch[5].r + scratch[4].i; + Fout[m3].i = scratch[5].i - scratch[4].r; + }else{ + Fout[m].r = scratch[5].r + scratch[4].i; + Fout[m].i = scratch[5].i - scratch[4].r; + Fout[m3].r = scratch[5].r - scratch[4].i; + Fout[m3].i = scratch[5].i + scratch[4].r; + } + ++Fout; + }while(--k); +} + +static void kf_bfly3( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + size_t m + ) +{ + size_t k=m; + const size_t m2 = 2*m; + kiss_fft_cpx *tw1,*tw2; + kiss_fft_cpx scratch[5]; + kiss_fft_cpx epi3; + epi3 = st->twiddles[fstride*m]; + + tw1=tw2=st->twiddles; + + do{ + C_FIXDIV(*Fout,3); C_FIXDIV(Fout[m],3); C_FIXDIV(Fout[m2],3); + + C_MUL(scratch[1],Fout[m] , *tw1); + C_MUL(scratch[2],Fout[m2] , *tw2); + + C_ADD(scratch[3],scratch[1],scratch[2]); + C_SUB(scratch[0],scratch[1],scratch[2]); + tw1 += fstride; + tw2 += fstride*2; + + Fout[m].r = Fout->r - HALF_OF(scratch[3].r); + Fout[m].i = Fout->i - HALF_OF(scratch[3].i); + + C_MULBYSCALAR( scratch[0] , epi3.i ); + + C_ADDTO(*Fout,scratch[3]); + + Fout[m2].r = Fout[m].r + scratch[0].i; + Fout[m2].i = Fout[m].i - scratch[0].r; + + Fout[m].r -= scratch[0].i; + Fout[m].i += scratch[0].r; + + ++Fout; + }while(--k); +} + +static void kf_bfly5( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m + ) +{ + kiss_fft_cpx *Fout0,*Fout1,*Fout2,*Fout3,*Fout4; + int u; + kiss_fft_cpx scratch[13]; + kiss_fft_cpx * twiddles = st->twiddles; + kiss_fft_cpx *tw; + kiss_fft_cpx ya,yb; + ya = twiddles[fstride*m]; + yb = twiddles[fstride*2*m]; + + Fout0=Fout; + Fout1=Fout0+m; + Fout2=Fout0+2*m; + Fout3=Fout0+3*m; + Fout4=Fout0+4*m; + + tw=st->twiddles; + for ( u=0; ur += scratch[7].r + scratch[8].r; + Fout0->i += scratch[7].i + scratch[8].i; + + scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r); + scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r); + + scratch[6].r = S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i); + scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i); + + C_SUB(*Fout1,scratch[5],scratch[6]); + C_ADD(*Fout4,scratch[5],scratch[6]); + + scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r); + scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r); + scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i); + scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i); + + C_ADD(*Fout2,scratch[11],scratch[12]); + C_SUB(*Fout3,scratch[11],scratch[12]); + + ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4; + } +} + +/* perform the butterfly for one stage of a mixed radix FFT */ +static void kf_bfly_generic( + kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_cfg st, + int m, + int p + ) +{ + int u,k,q1,q; + kiss_fft_cpx * twiddles = st->twiddles; + kiss_fft_cpx t; + int Norig = st->nfft; + + kiss_fft_cpx * scratch = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC(sizeof(kiss_fft_cpx)*p); + + for ( u=0; u=Norig) twidx-=Norig; + C_MUL(t,scratch[q] , twiddles[twidx] ); + C_ADDTO( Fout[ k ] ,t); + } + k += m; + } + } + KISS_FFT_TMP_FREE(scratch); +} + +static +void kf_work( + kiss_fft_cpx * Fout, + const kiss_fft_cpx * f, + const size_t fstride, + int in_stride, + int * factors, + const kiss_fft_cfg st + ) +{ + kiss_fft_cpx * Fout_beg=Fout; + const int p=*factors++; /* the radix */ + const int m=*factors++; /* stage's fft length/p */ + const kiss_fft_cpx * Fout_end = Fout + p*m; + +#ifdef _OPENMP + // use openmp extensions at the + // top-level (not recursive) + if (fstride==1 && p<=5) + { + int k; + + // execute the p different work units in different threads +# pragma omp parallel for + for (k=0;k floor_sqrt) + p = n; /* no more factors, skip to end */ + } + n /= p; + *facbuf++ = p; + *facbuf++ = n; + } while (n > 1); +} + +/* + * + * User-callable function to allocate all necessary storage space for the fft. + * + * The return value is a contiguous block of memory, allocated with malloc. As such, + * It can be freed with free(), rather than a kiss_fft-specific function. + * */ +kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem ) +{ + kiss_fft_cfg st=NULL; + size_t memneeded = sizeof(struct kiss_fft_state) + + sizeof(kiss_fft_cpx)*(nfft-1); /* twiddle factors*/ + + if ( lenmem==NULL ) { + st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded ); + }else{ + if (mem != NULL && *lenmem >= memneeded) + st = (kiss_fft_cfg)mem; + *lenmem = memneeded; + } + if (st) { + int i; + st->nfft=nfft; + st->inverse = inverse_fft; + + for (i=0;iinverse) + phase *= -1; + kf_cexp(st->twiddles+i, phase ); + } + + kf_factor(nfft,st->factors); + } + return st; +} + + +void kiss_fft_stride(kiss_fft_cfg st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int in_stride) +{ + if (fin == fout) { + //NOTE: this is not really an in-place FFT algorithm. + //It just performs an out-of-place FFT into a temp buffer + kiss_fft_cpx * tmpbuf = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC( sizeof(kiss_fft_cpx)*st->nfft); + kf_work(tmpbuf,fin,1,in_stride, st->factors,st); + memcpy(fout,tmpbuf,sizeof(kiss_fft_cpx)*st->nfft); + KISS_FFT_TMP_FREE(tmpbuf); + }else{ + kf_work( fout, fin, 1,in_stride, st->factors,st ); + } +} + +void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout) +{ + kiss_fft_stride(cfg,fin,fout,1); +} + + +void kiss_fft_cleanup(void) +{ + // nothing needed any more +} + +int kiss_fft_next_fast_size(int n) +{ + while(1) { + int m=n; + while ( (m%2) == 0 ) m/=2; + while ( (m%3) == 0 ) m/=3; + while ( (m%5) == 0 ) m/=5; + if (m<=1) + break; /* n is completely factorable by twos, threes, and fives */ + n++; + } + return n; +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/kissfft/kiss_fft.h b/src/tensorflow_esp32/third_party/kissfft/kiss_fft.h similarity index 97% rename from src/kissfft/kiss_fft.h rename to src/tensorflow_esp32/third_party/kissfft/kiss_fft.h index ee8f221..d6365a0 100644 --- a/src/kissfft/kiss_fft.h +++ b/src/tensorflow_esp32/third_party/kissfft/kiss_fft.h @@ -1,3 +1,4 @@ +#if defined(ESP32) #ifndef KISS_FFT_H #define KISS_FFT_H @@ -5,6 +6,7 @@ #include #include #include +#include #ifdef __cplusplus extern "C" { @@ -128,3 +130,5 @@ int kiss_fft_next_fast_size(int n); #endif #endif + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.c b/src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.c new file mode 100644 index 0000000..13452a6 --- /dev/null +++ b/src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.c @@ -0,0 +1,162 @@ +#if defined(ESP32) +/* +Copyright (c) 2003-2004, Mark Borgerding + +All rights reserved. + +Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. + * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#include "kiss_fftr.h" +#include "../_kiss_fft_guts.h" + +struct kiss_fftr_state{ + kiss_fft_cfg substate; + kiss_fft_cpx * tmpbuf; + kiss_fft_cpx * super_twiddles; +#ifdef USE_SIMD + void * pad; +#endif +}; + +kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem) +{ + int i; + kiss_fftr_cfg st = NULL; + size_t subsize, memneeded; + + if (nfft & 1) { + /* fprintf(stderr,"Real FFT optimization must be even.\n"); */ + return NULL; + } + nfft >>= 1; + + kiss_fft_alloc (nfft, inverse_fft, NULL, &subsize); + memneeded = sizeof(struct kiss_fftr_state) + subsize + sizeof(kiss_fft_cpx) * ( nfft * 3 / 2); + + if (lenmem == NULL) { + st = (kiss_fftr_cfg) KISS_FFT_MALLOC (memneeded); + } else { + if (*lenmem >= memneeded) + st = (kiss_fftr_cfg) mem; + *lenmem = memneeded; + } + if (!st) + return NULL; + + st->substate = (kiss_fft_cfg) (st + 1); /*just beyond kiss_fftr_state struct */ + st->tmpbuf = (kiss_fft_cpx *) (((char *) st->substate) + subsize); + st->super_twiddles = st->tmpbuf + nfft; + kiss_fft_alloc(nfft, inverse_fft, st->substate, &subsize); + + for (i = 0; i < nfft/2; ++i) { + double phase = + -3.14159265358979323846264338327 * ((double) (i+1) / nfft + .5); + if (inverse_fft) + phase *= -1; + kf_cexp (st->super_twiddles+i,phase); + } + return st; +} + +void kiss_fftr(kiss_fftr_cfg st,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata) +{ + /* input buffer timedata is stored row-wise */ + int k,ncfft; + kiss_fft_cpx fpnk,fpk,f1k,f2k,tw,tdc; + + if ( st->substate->inverse) { + /* fprintf(stderr,"kiss fft usage error: improper alloc\n"); */ + return; /* exit(1); */ + } + + ncfft = st->substate->nfft; + + /*perform the parallel fft of two real signals packed in real,imag*/ + kiss_fft( st->substate , (const kiss_fft_cpx*)timedata, st->tmpbuf ); + /* The real part of the DC element of the frequency spectrum in st->tmpbuf + * contains the sum of the even-numbered elements of the input time sequence + * The imag part is the sum of the odd-numbered elements + * + * The sum of tdc.r and tdc.i is the sum of the input time sequence. + * yielding DC of input time sequence + * The difference of tdc.r - tdc.i is the sum of the input (dot product) [1,-1,1,-1... + * yielding Nyquist bin of input time sequence + */ + + tdc.r = st->tmpbuf[0].r; + tdc.i = st->tmpbuf[0].i; + C_FIXDIV(tdc,2); + CHECK_OVERFLOW_OP(tdc.r ,+, tdc.i); + CHECK_OVERFLOW_OP(tdc.r ,-, tdc.i); + freqdata[0].r = tdc.r + tdc.i; + freqdata[ncfft].r = tdc.r - tdc.i; +#ifdef USE_SIMD + freqdata[ncfft].i = freqdata[0].i = _mm_set1_ps(0); +#else + freqdata[ncfft].i = freqdata[0].i = 0; +#endif + + for ( k=1;k <= ncfft/2 ; ++k ) { + fpk = st->tmpbuf[k]; + fpnk.r = st->tmpbuf[ncfft-k].r; + fpnk.i = - st->tmpbuf[ncfft-k].i; + C_FIXDIV(fpk,2); + C_FIXDIV(fpnk,2); + + C_ADD( f1k, fpk , fpnk ); + C_SUB( f2k, fpk , fpnk ); + C_MUL( tw , f2k , st->super_twiddles[k-1]); + + freqdata[k].r = HALF_OF(f1k.r + tw.r); + freqdata[k].i = HALF_OF(f1k.i + tw.i); + freqdata[ncfft-k].r = HALF_OF(f1k.r - tw.r); + freqdata[ncfft-k].i = HALF_OF(tw.i - f1k.i); + } +} + +void kiss_fftri(kiss_fftr_cfg st,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata) +{ + /* input buffer timedata is stored row-wise */ + int k, ncfft; + + if (st->substate->inverse == 0) { + /* fprintf (stderr, "kiss fft usage error: improper alloc\n"); */ + return; /* exit (1); */ + } + + ncfft = st->substate->nfft; + + st->tmpbuf[0].r = freqdata[0].r + freqdata[ncfft].r; + st->tmpbuf[0].i = freqdata[0].r - freqdata[ncfft].r; + C_FIXDIV(st->tmpbuf[0],2); + + for (k = 1; k <= ncfft / 2; ++k) { + kiss_fft_cpx fk, fnkc, fek, fok, tmp; + fk = freqdata[k]; + fnkc.r = freqdata[ncfft - k].r; + fnkc.i = -freqdata[ncfft - k].i; + C_FIXDIV( fk , 2 ); + C_FIXDIV( fnkc , 2 ); + + C_ADD (fek, fk, fnkc); + C_SUB (tmp, fk, fnkc); + C_MUL (fok, tmp, st->super_twiddles[k-1]); + C_ADD (st->tmpbuf[k], fek, fok); + C_SUB (st->tmpbuf[ncfft - k], fek, fok); +#ifdef USE_SIMD + st->tmpbuf[ncfft - k].i *= _mm_set1_ps(-1.0); +#else + st->tmpbuf[ncfft - k].i *= -1; +#endif + } + kiss_fft (st->substate, st->tmpbuf, (kiss_fft_cpx *) timedata); +} + +#endif // end of #if defined(ESP32) \ No newline at end of file diff --git a/src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.h b/src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.h new file mode 100644 index 0000000..6f4fd68 --- /dev/null +++ b/src/tensorflow_esp32/third_party/kissfft/tools/kiss_fftr.h @@ -0,0 +1,49 @@ +#if defined(ESP32) +#ifndef KISS_FTR_H +#define KISS_FTR_H + +#include "tensorflow_esp32/third_party/kissfft/kiss_fft.h" +#ifdef __cplusplus +extern "C" { +#endif + + +/* + + Real optimized version can save about 45% cpu time vs. complex fft of a real seq. + + + + */ + +typedef struct kiss_fftr_state *kiss_fftr_cfg; + + +kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem, size_t * lenmem); +/* + nfft must be even + + If you don't care to allocate space, use mem = lenmem = NULL +*/ + + +void kiss_fftr(kiss_fftr_cfg cfg,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata); +/* + input timedata has nfft scalar points + output freqdata has nfft/2+1 complex points +*/ + +void kiss_fftri(kiss_fftr_cfg cfg,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata); +/* + input freqdata has nfft/2+1 complex points + output timedata has nfft scalar points +*/ + +#define kiss_fftr_free free + +#ifdef __cplusplus +} +#endif +#endif + +#endif // end of #if defined(ESP32) \ No newline at end of file