Porting to new autotiler

MFCC_params_LARGE.h

@@ -5,12 +5,14 @@
 #define	N_FRAME                      49
 #define	N_FFT                      1024
 #define	USE_RADIX_4                   0
-#define	USE_HIGH_PREC                 0
+#define	DATA_TYPE                     0
 #define	USE_POWER                     1
+#define	OUT_FFT                       0
 #define	MFCC_BANK_CNT                40
 #define	FMIN                       20.0
 #define	FMAX                     4000.0
 #define	MFCC_COEFF_CNT              494
 #define	N_DCT                        40
 #define	OUTPUT_MELSPECTROGRAM         0
 #define	LIFTER_COEFF                  0
+#define	USE_DB                        0

MFCC_params_MEDIUM.h

@@ -5,8 +5,9 @@
 #define	N_FRAME                      49
 #define	N_FFT                      1024
 #define	USE_RADIX_4                   0
-#define	USE_HIGH_PREC                 0
+#define	DATA_TYPE                     0
 #define	USE_POWER                     1
+#define	OUT_FFT                       0
 #define	MFCC_BANK_CNT                40
 #define	FMIN                       20.0
 #define	FMAX                     4000.0

MFCC_params_SMALL.h

@@ -5,12 +5,14 @@
 #define	N_FRAME                      49
 #define	N_FFT                      1024
 #define	USE_RADIX_4                   0
-#define	USE_HIGH_PREC                 0
+#define	DATA_TYPE                     0
 #define	USE_POWER                     1
+#define	OUT_FFT                       0
 #define	MFCC_BANK_CNT                40
 #define	FMIN                       20.0
 #define	FMAX                     4000.0
 #define	MFCC_COEFF_CNT              494
 #define	N_DCT                        40
 #define	OUTPUT_MELSPECTROGRAM         0
 #define	LIFTER_COEFF                  0
+#define	USE_DB                        0

MFCCmodel.c

@@ -31,7 +31,7 @@ int main(int argc, char **argv)
     MFCCConfiguration(51200);
     // Load FIR basic kernels
     LoadMFCCLibrary();
-    MFCC_Generator("MFCC",        0, N_FRAME, FRAME_SIZE, FRAME_STEP, N_FFT, MFCC_COEFF_CNT, N_DCT, PREEMP_FACTOR, LIFTER_COEFF, USE_RADIX_4, USE_POWER, DATA_TYPE, MFCC_BANK_CNT, 0);
-  //MFCC_Generator("MFCC_single", 0,       1, FRAME_SIZE, FRAME_STEP, N_FFT, MFCC_COEFF_CNT, N_DCT, PREEMP_FACTOR, LIFTER_COEFF, USE_RADIX_4, USE_POWER, DATA_TYPE, MFCC_BANK_CNT, 0);
+    MFCC_Generator("MFCC",        0, N_FRAME, FRAME_SIZE, FRAME_STEP, N_FFT, MFCC_COEFF_CNT, N_DCT, PREEMP_FACTOR, LIFTER_COEFF, USE_RADIX_4, USE_POWER, DATA_TYPE, MFCC_BANK_CNT, 0, 0);
+  //MFCC_Generator("MFCC_single", 0,       1, FRAME_SIZE, FRAME_STEP, N_FFT, MFCC_COEFF_CNT, N_DCT, PREEMP_FACTOR, LIFTER_COEFF, USE_RADIX_4, USE_POWER, DATA_TYPE, MFCC_BANK_CNT, 0, 0);
     GenerateTilingCode();
 }

Makefile

@@ -38,7 +38,6 @@ else
 		ifeq ($(LARGE), 1)
 			MODEL_PREFIX = KWS_ds_cnn_l_quant
 			APP_CFLAGS += -DLARGE
-			LARGE_OPT = nodeoption 1 PARALLELFEATURES 0
 		else
 $(error You must set to 1 one of SMALL, MEDIUM, LARGE to select a network)
 		endif
@@ -56,9 +55,17 @@ READFS_FILES=$(realpath $(MODEL_TENSORS))
 QUANT_BITS=8
 BUILD_DIR=BUILD
 
+#MODEL_NE16 ?= 0
 NNTOOL_SCRIPT_PARAMETRIC=model/nntool_script_params
-NNTOOL_EXTRA_FLAGS = -q
 MODEL_SUFFIX = _$(QUANT_BITS)BIT
+ifeq ($(MODEL_NE16), 1)
+	NNTOOL_SCRIPT_PARAMETRIC=model/nntool_script_ne16_params
+	MODEL_SUFFIX = _NE16
+endif
+ifeq ($(DUMP_TENSORS), 1)
+	NNTOOL_SET_GRAPH_DUMP = set graph_dump_tensor 7
+endif
+NNTOOL_EXTRA_FLAGS = -q
 MODEL_BUILD = BUILD_MODEL_$(QUANT_BITS)BIT
 
 CLUSTER_STACK_SIZE=4096

common/model_decl.mk

@@ -6,20 +6,7 @@
 
 MODEL_SUFFIX?=
 
-MODEL_PREFIX?=
-
-# The training of the model is slightly different depending on
-# the quantization. This is because in 8 bit mode we used signed
-# 8 bit so the input to the model needs to be shifted 1 bit
-ifeq ($(QUANT_BITS),8)
-  TRAIN_SUFFIX=_8BIT
-else
-  ifeq ($(QUANT_BITS),16)
-    TRAIN_SUFFIX=_16BIT
-  else
-    $(error Dont know how to build with this bit width)
-  endif
-endif
+MODEL_PREFIX?=GapFlow
 
 MODEL_PYTHON=python3
 
@@ -29,7 +16,11 @@ MODEL_COMMON_INC ?= $(GAP_SDK_HOME)/libs/gap_lib/include
 MODEL_COMMON_SRC ?= $(GAP_SDK_HOME)/libs/gap_lib/img_io
 MODEL_COMMON_SRC_FILES ?= ImgIO.c
 MODEL_COMMON_SRCS = $(realpath $(addprefix $(MODEL_COMMON_SRC)/,$(MODEL_COMMON_SRC_FILES)))
+MODEL_HEADERS = $(MODEL_COMMON)/headers
+MODEL_TRAIN = model/train.py
 MODEL_BUILD = BUILD_MODEL$(MODEL_SUFFIX)
+MODEL_TRAIN_BUILD = BUILD_TRAIN$(TRAIN_SUFFIX)
+MODEL_H5 = $(MODEL_TRAIN_BUILD)/$(MODEL_PREFIX).h5
 
 MODEL_TFLITE = $(MODEL_BUILD)/$(MODEL_PREFIX).tflite
 
@@ -44,10 +35,6 @@ MODEL_GEN_C = $(addsuffix .c, $(MODEL_GEN))
 MODEL_GEN_CLEAN = $(MODEL_GEN_C) $(addsuffix .h, $(MODEL_GEN))
 MODEL_GEN_EXE = $(MODEL_BUILD)/GenTile
 
-ifdef MODEL_QUANTIZED
-  NNTOOL_EXTRA_FLAGS = -q
-endif
-
 MODEL_GENFLAGS_EXTRA =
 
 EXTRA_GENERATOR_SRC =
@@ -60,61 +47,8 @@ IMAGES = images
 RM=rm -f
 
 NNTOOL=nntool
-
-NNTOOL_PATH = $(GAP_SDK_HOME)/tools/nntool
-NNTOOL_KERNEL_PATH = $(NNTOOL_PATH)/autotiler/kernels
-NNTOOL_GENERATOR_PATH = $(NNTOOL_PATH)/autotiler/generators
-# hram - HyperBus RAM
-# qspiram - Quad SPI RA
-MODEL_L3_EXEC=hram
-# hflash - HyperBus Flash
-# qpsiflash - Quad SPI Flash
-MODEL_L3_CONST=hflash
-
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_BiasReLULinear_BasicKernels.c
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_BasicKernels.c
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DP_BasicKernels.c
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_BasicKernels.c
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_DP_BasicKernels.c
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Pooling_BasicKernels.c
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_MatAlgebra.c
-MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_SoftMax.c
-MODEL_LIB_POW2 += $(NNTOOL_KERNEL_PATH)/norm_transpose.c
-MODEL_LIB_INCLUDE_POW2 = -I$(TILER_CNN_KERNEL_PATH) -I$(NNTOOL_KERNEL_PATH)
-MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generator_Util.c
-MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generators.c
-MODEL_GEN_POW2 += $(NNTOOL_GENERATOR_PATH)/nntool_extra_generators.c
-MODEL_GEN_INCLUDE_POW2 = -I$(TILER_CNN_GENERATOR_PATH) -I$(NNTOOL_GENERATOR_PATH)
-
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_Activation_SQ8.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_Bias_Linear_SQ8.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_Conv_SQ8.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_Pooling_SQ8.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_Conv_DW_SQ8.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_MatAlgebra_SQ8.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_SoftMax_SQ8.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/CNN_AT_Misc.c
-MODEL_LIB_SQ8 += $(TILER_CNN_KERNEL_PATH_SQ8)/RNN_SQ8.c
-MODEL_LIB_SQ8 += $(NNTOOL_KERNEL_PATH)/norm_transpose.c
-MODEL_LIB_INCLUDE_SQ8 = -I$(TILER_CNN_KERNEL_PATH) -I$(TILER_CNN_KERNEL_PATH_SQ8) -I$(NNTOOL_KERNEL_PATH)
-MODEL_GEN_SQ8 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generator_Util.c
-MODEL_GEN_SQ8 += $(TILER_CNN_GENERATOR_PATH_SQ8)/CNN_Generators_SQ8.c
-MODEL_GEN_SQ8 += $(TILER_CNN_GENERATOR_PATH_SQ8)/RNN_Generators_SQ8.c
-MODEL_GEN_SQ8 += $(NNTOOL_GENERATOR_PATH)/nntool_extra_generators.c
-MODEL_GEN_INCLUDE_SQ8 = -I$(TILER_CNN_GENERATOR_PATH) -I$(TILER_CNN_GENERATOR_PATH_SQ8) -I$(NNTOOL_GENERATOR_PATH)
-
-
 MODEL_SIZE_CFLAGS = -DAT_INPUT_HEIGHT=$(AT_INPUT_HEIGHT) -DAT_INPUT_WIDTH=$(AT_INPUT_WIDTH) -DAT_INPUT_COLORS=$(AT_INPUT_COLORS)
 
-ifdef MODEL_SQ8
-  CNN_GEN = $(MODEL_GEN_SQ8)
-  CNN_GEN_INCLUDE = $(MODEL_GEN_INCLUDE_SQ8)
-  CNN_LIB = $(MODEL_LIB_SQ8)
-  CNN_LIB_INCLUDE = $(MODEL_LIB_INCLUDE_SQ8)
-else
-  CNN_GEN = $(MODEL_GEN_POW2)
-  CNN_GEN_INCLUDE = $(MODEL_GEN_INCLUDE_POW2)
-  CNN_LIB = $(MODEL_LIB_POW2)
-  CNN_LIB_INCLUDE = $(MODEL_LIB_INCLUDE_POW2)
-endif
+include $(RULES_DIR)/at_common_decl.mk
 $(info GEN ... $(CNN_GEN))
+

common/model_rules.mk

@@ -13,21 +13,10 @@ else
   MODEL_TRAIN_FLAGS =
 endif
 
-ifdef MODEL_SQ8
-  CNN_GEN = $(MODEL_GEN_SQ8)
-  CNN_GEN_INCLUDE = $(MODEL_GEN_INCLUDE_SQ8)
-  CNN_LIB = $(MODEL_LIB_SQ8)
-  CNN_LIB_INCLUDE = $(MODEL_LIB_INCLUDE_SQ8)
-else
-  CNN_GEN = $(MODEL_GEN_POW2)
-  CNN_GEN_INCLUDE = $(MODEL_GEN_INCLUDE_POW2)
-  CNN_LIB = $(MODEL_LIB_POW2)
-  CNN_LIB_INCLUDE = $(MODEL_LIB_INCLUDE_POW2)
-endif
-
 USE_DISP=1
+
 ifdef USE_DISP
-  SDL_FLAGS= -lSDL2 -lSDL2_ttf
+  SDL_FLAGS= -lSDL2 -lSDL2_ttf -DAT_DISPLAY
 else
   SDL_FLAGS=
 endif
@@ -51,13 +40,12 @@ $(MODEL_BUILD):
 $(MODEL_TFLITE): $(TRAINED_TFLITE_MODEL) | $(MODEL_BUILD)
 	cp $< $@
 
-# Creates an Autotiler Model file by running the commands in the script
+# Creates an NNTOOL state file by running the commands in the script
 # These commands could be run interactively
 # The commands:
 # 	Adjust the model to match AutoTiler tensor order
 #	Fuse nodes together to match fused AutoTiler generators
-#	Quantize the graph if not already done with tflite quantization
-#	Generate the Autotiler model code
+#	Save the graph state files
 
 # Runs NNTOOL with its state file to generate the autotiler model code
 $(MODEL_BUILD)/$(MODEL_SRC): $(MODEL_TFLITE) | $(MODEL_BUILD) $(SAMPLES)
@@ -71,7 +59,7 @@ nntool_gen: $(MODEL_BUILD)/$(MODEL_SRC)
 # Build the code generator from the model code
 $(MODEL_GEN_EXE): $(CNN_GEN) $(MODEL_BUILD)/$(MODEL_SRC) $(EXTRA_GENERATOR_SRC) 
 	echo "COMPILING AUTOTILER MODEL"
-	gcc -g -o $(MODEL_GEN_EXE) -I. -I$(TILER_INC) -I$(TILER_EMU_INC) $(CNN_GEN_INCLUDE) $(CNN_LIB_INCLUDE) $? $(TILER_LIB) $(SDL_FLAGS)
+	gcc -g -o $(MODEL_GEN_EXE) -I. -I$(TILER_INC) -I$(TILER_EMU_INC) $(CNN_GEN_INCLUDE) $(CNN_LIB_INCLUDE) $^ $(TILER_LIB) $(SDL_FLAGS) 
 
 compile_model: $(MODEL_GEN_EXE)
 

emul.mk

@@ -30,7 +30,7 @@ else
 		ifeq ($(LARGE), 1)
 			MODEL_PREFIX = KWS_ds_cnn_l_quant
 			CFLAGS += -DLARGE
-			LARGE_OPT = nodeoption 1 PARALLELFEATURES 0
+			#LARGE_OPT = nodeoption 1 PARALLELFEATURES 0
 		else
 $(error You must set to 1 one of SMALL, MEDIUM, LARGE to select a network)
 		endif

main_with_mfcc.c

@@ -176,13 +176,6 @@ static void Runkws()
 
   if(rec_digit>1)
     printf("Recognized: %s\twith confidence: %d\n", LABELS[rec_digit], highest);
-
-#ifdef PERF
-    if (rec_digit!=8){
-        printf("App didn't recognize ON with %s test sample\n", WavName);
-        pmsis_exit(-1);
-    }
-#endif
 }
 
 
@@ -378,6 +371,10 @@ while(1)
         printf("%45s: Cycles: %10d, Operations: %10d, Operations/Cycle: %f\n", "Total", TotalCycles, TotalOper, ((float) TotalOper)/ TotalCycles);
         printf("\n");
         break;
+        if (rec_digit!=8){
+            printf("App didn't recognize ON with %s test sample\n", WavName);
+            pmsis_exit(-1);
+        }
     }
     #endif  /* PERF */
     AT_L1_FREE(0, __PREFIX(_L1_Memory), L1_SIZE);

mfcc_model.mk

@@ -51,18 +51,19 @@ endif
 $(MFCCBUILD_DIR):
 	mkdir $(MFCCBUILD_DIR)
 
-# Build the code generator from the model code
-$(MFCC_MODEL_GEN): $(MFCCBUILD_DIR)
-	gcc -g -o $(MFCC_MODEL_GEN) -I. -I$(MFCC_GENERATOR) -I$(TILER_INC) -I$(TILER_EMU_INC) MFCCmodel.c $(MFCC_SRCG) $(TILER_LIB) $(TABLE_CFLAGS) $(SIZE_DEF)
-
-gen_lut: $(MFCCBUILD_DIR)
+$(MFCCBUILD_DIR)/LUT.def: $(MFCCBUILD_DIR)
 	python3 $(MFCC_GENERATOR)/GenLUT.py --fft_lut_file $(MFCCBUILD_DIR)/LUT.def --mfcc_bf_lut_file $(MFCCBUILD_DIR)/MFCC_FB.def                \
 									   --sample_rate 16000 --frame_size $(FRAME_SIZE) --frame_step $(FRAME_STEP) --n_frame $(AT_INPUT_HEIGHT) \
 									   --n_fft 1024 --n_dct 40 --mfcc_bank_cnt 40 --fmin 20 --fmax 4000 --mfcc_bank_cnt 40 --preempfactor 0.0 \
 									   --use_tf_mfcc --save_params_header MFCC_params_$(NN_SIZE).h $(EXTRA_FLAGS)
 
+# Build the code generator from the model code
+$(MFCC_MODEL_GEN): $(MFCCBUILD_DIR)/LUT.def $(MFCCBUILD_DIR)
+	gcc -g -o $(MFCC_MODEL_GEN) -I. -I$(MFCC_GENERATOR) -I$(TILER_INC) -I$(TILER_EMU_INC) MFCCmodel.c $(MFCC_SRCG) $(TILER_LIB) $(TABLE_CFLAGS) $(SIZE_DEF)
+
+
 # Run the code generator  kernel code
-$(MFCCBUILD_DIR)/MFCCKernels.c: gen_lut $(MFCC_MODEL_GEN)
+$(MFCCBUILD_DIR)/MFCCKernels.c: $(MFCC_MODEL_GEN)
 	$(MFCC_MODEL_GEN) -o $(MFCCBUILD_DIR) -c $(MFCCBUILD_DIR) $(MODEL_GEN_EXTRA_FLAGS)
 
 mfcc_model: $(MFCCBUILD_DIR)/MFCCKernels.c