Skip to content

Commit

Permalink
CMSIS-NN Min Max int8 support (#2753)
Browse files Browse the repository at this point in the history 
* Moves common functions to new maximum_minimum.h
  * Creates cmsis-nn/maximum_minimum.cc

BUG=#2752
Change-Id: Ifbb3fedf53043b2f8d4c48d73c2ca44c7f0f87ca
  • Loading branch information
@ArmRyan
ArmRyan authored Feb 5, 2025
1 parent 9b79b9f commit 740cef3
Show file tree
Hide file tree
Showing 5 changed files with 367 additions and 58 deletions.
1 change: 1 addition & 0 deletions tensorflow/lite/micro/kernels/BUILD
View file
Original file line number Diff line number Diff line change
Expand Up @@ -333,6 +333,7 @@ tflm_kernel_cc_library(
"logistic.h",
"lstm_eval.h",
"lstm_shared.h",
"maximum_minimum.h",
"micro_ops.h",
"mul.h",
"pad.h",
Expand Down
247 changes: 247 additions & 0 deletions tensorflow/lite/micro/kernels/cmsis_nn/maximum_minimum.cc
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,247 @@
/* Copyright 2024 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/kernels/maximum_minimum.h"

#include "Include/arm_nnfunctions.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/lite/micro/kernels/kernel_util.h"
#include "tensorflow/lite/micro/micro_log.h"

namespace tflite {

namespace {

cmsis_nn_dims FillVariableShape(int32_t rank, int32_t* tensor_dims) {
if (rank == 4) {
return {tensor_dims[0], tensor_dims[1], tensor_dims[2], tensor_dims[3]};
} else if (rank == 3) {
return {1, tensor_dims[0], tensor_dims[1], tensor_dims[2]};
} else if (rank == 2) {
return {1, 1, tensor_dims[0], tensor_dims[1]};
} else {
return {1, 1, 1, 1};
}
}

TfLiteStatus EvalMaximum(TfLiteContext* context, TfLiteNode* node) {
OpContext op_context(context, node);
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 input_1_shape = tflite::micro::GetTensorShape(input1);
RuntimeShape input_2_shape = tflite::micro::GetTensorShape(input2);
RuntimeShape output_shape = tflite::micro::GetTensorShape(output);

cmsis_nn_dims input_1_dims = FillVariableShape(
input_1_shape.DimensionsCount(), input_1_shape.DimsData());
cmsis_nn_dims input_2_dims = FillVariableShape(
input_2_shape.DimensionsCount(), input_2_shape.DimsData());
cmsis_nn_dims output_dims = FillVariableShape(output_shape.DimensionsCount(),
output_shape.DimsData());

switch (op_context.output->type) {
case kTfLiteInt8:
cmsis_nn_context ctx;
ctx.buf = nullptr;
ctx.size = 0;

arm_maximum_s8(
&ctx, tflite::micro::GetTensorData<int8_t>(input1), &input_1_dims,
tflite::micro::GetTensorData<int8_t>(input2), &input_2_dims,
tflite::micro::GetTensorData<int8_t>(output), &output_dims);
break;
case kTfLiteFloat32:
TFLiteOperation<float, MaximumOp>(context, node, op_context);
break;
case kTfLiteInt16:
TFLiteOperation<int16_t, MaximumOp>(context, node, op_context);
break;
case kTfLiteInt32:
TFLiteOperation<int32_t, MaximumOp>(context, node, op_context);
break;
case kTfLiteInt64:
TFLiteOperation<int64_t, MaximumOp>(context, node, op_context);
break;
default:
MicroPrintf("Type %s (%d) is not supported by Maximum/Minimum.",
TfLiteTypeGetName(op_context.output->type),
op_context.output->type);
return kTfLiteError;
}
return kTfLiteOk;
}

TfLiteStatus EvalMaximumInt8(TfLiteContext* context, TfLiteNode* node) {
OpContext op_context(context, node);
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 input_1_shape = tflite::micro::GetTensorShape(input1);
RuntimeShape input_2_shape = tflite::micro::GetTensorShape(input2);
RuntimeShape output_shape = tflite::micro::GetTensorShape(output);

cmsis_nn_dims input_1_dims = FillVariableShape(
input_1_shape.DimensionsCount(), input_1_shape.DimsData());
cmsis_nn_dims input_2_dims = FillVariableShape(
input_2_shape.DimensionsCount(), input_2_shape.DimsData());
cmsis_nn_dims output_dims = FillVariableShape(output_shape.DimensionsCount(),
output_shape.DimsData());

switch (op_context.output->type) {
case kTfLiteInt8:
cmsis_nn_context ctx;
ctx.buf = nullptr;
ctx.size = 0;

arm_maximum_s8(
&ctx, tflite::micro::GetTensorData<int8_t>(input1), &input_1_dims,
tflite::micro::GetTensorData<int8_t>(input2), &input_2_dims,
tflite::micro::GetTensorData<int8_t>(output), &output_dims);
break;
default:
MicroPrintf("Type %s (%d) is not supported by Maximum Int8 Registration.",
TfLiteTypeGetName(op_context.output->type),
op_context.output->type);
return kTfLiteError;
}
return kTfLiteOk;
}

TfLiteStatus EvalMinimum(TfLiteContext* context, TfLiteNode* node) {
OpContext op_context(context, node);
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 input_1_shape = tflite::micro::GetTensorShape(input1);
RuntimeShape input_2_shape = tflite::micro::GetTensorShape(input2);
RuntimeShape output_shape = tflite::micro::GetTensorShape(output);

cmsis_nn_dims input_1_dims = FillVariableShape(
input_1_shape.DimensionsCount(), input_1_shape.DimsData());
cmsis_nn_dims input_2_dims = FillVariableShape(
input_2_shape.DimensionsCount(), input_2_shape.DimsData());
cmsis_nn_dims output_dims = FillVariableShape(output_shape.DimensionsCount(),
output_shape.DimsData());

switch (op_context.output->type) {
case kTfLiteInt8:
cmsis_nn_context ctx;
ctx.buf = nullptr;
ctx.size = 0;

arm_minimum_s8(
&ctx, tflite::micro::GetTensorData<int8_t>(input1), &input_1_dims,
tflite::micro::GetTensorData<int8_t>(input2), &input_2_dims,
tflite::micro::GetTensorData<int8_t>(output), &output_dims);
break;
case kTfLiteFloat32:
TFLiteOperation<float, MinimumOp>(context, node, op_context);
break;
case kTfLiteInt16:
TFLiteOperation<int16_t, MinimumOp>(context, node, op_context);
break;
case kTfLiteInt32:
TFLiteOperation<int32_t, MinimumOp>(context, node, op_context);
break;
case kTfLiteInt64:
TFLiteOperation<int64_t, MinimumOp>(context, node, op_context);
break;
default:
MicroPrintf("Type %s (%d) is not supported by Maximum/Minimum.",
TfLiteTypeGetName(op_context.output->type),
op_context.output->type);
return kTfLiteError;
}
return kTfLiteOk;
}

TfLiteStatus EvalMinimumInt8(TfLiteContext* context, TfLiteNode* node) {
OpContext op_context(context, node);
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 input_1_shape = tflite::micro::GetTensorShape(input1);
RuntimeShape input_2_shape = tflite::micro::GetTensorShape(input2);
RuntimeShape output_shape = tflite::micro::GetTensorShape(output);

cmsis_nn_dims input_1_dims = FillVariableShape(
input_1_shape.DimensionsCount(), input_1_shape.DimsData());
cmsis_nn_dims input_2_dims = FillVariableShape(
input_2_shape.DimensionsCount(), input_2_shape.DimsData());
cmsis_nn_dims output_dims = FillVariableShape(output_shape.DimensionsCount(),
output_shape.DimsData());

switch (op_context.output->type) {
case kTfLiteInt8:
cmsis_nn_context ctx;
ctx.buf = nullptr;
ctx.size = 0;

arm_minimum_s8(
&ctx, tflite::micro::GetTensorData<int8_t>(input1), &input_1_dims,
tflite::micro::GetTensorData<int8_t>(input2), &input_2_dims,
tflite::micro::GetTensorData<int8_t>(output), &output_dims);
break;
default:
MicroPrintf("Type %s (%d) is not supported by Minimum Int8 registration.",
TfLiteTypeGetName(op_context.output->type),
op_context.output->type);
return kTfLiteError;
}
return kTfLiteOk;
}

} // namespace

TFLMRegistration Register_MAXIMUM() {
return tflite::micro::RegisterOp(nullptr, nullptr, EvalMaximum);
}

TFLMRegistration Register_MINIMUM() {
return tflite::micro::RegisterOp(nullptr, nullptr, EvalMinimum);
}

TFLMRegistration Register_MAXIMUM_INT8() {
return tflite::micro::RegisterOp(nullptr, nullptr, EvalMaximumInt8);
}

TFLMRegistration Register_MINIMUM_INT8() {
return tflite::micro::RegisterOp(nullptr, nullptr, EvalMinimumInt8);
}

} // namespace tflite
50 changes: 2 additions & 48 deletions tensorflow/lite/micro/kernels/maximum_minimum.cc
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
/* Copyright 2022 The TensorFlow Authors. All Rights Reserved.
/* Copyright 2024 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.
Expand All @@ -23,59 +23,13 @@ limitations under the License.
#include "tensorflow/lite/kernels/kernel_util.h"
#include "tensorflow/lite/kernels/op_macros.h"
#include "tensorflow/lite/micro/kernels/kernel_util.h"
#include "tensorflow/lite/micro/kernels/maximum_minimum.h"
#include "tensorflow/lite/micro/micro_log.h"

namespace tflite {

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 <typename data_type>
static data_type op(data_type el1, data_type el2) {
return el1 > el2 ? el1 : el2;
}
};

struct MinimumOp {
template <typename data_type>
static data_type op(data_type el1, data_type el2) {
return el1 < el2 ? el1 : el2;
}
};

template <typename data_type, typename op_type>
void TFLiteOperation(TfLiteContext* context, TfLiteNode* node,
const OpContext& op_context) {
reference_ops::MaximumMinimumBroadcastSlow(
tflite::micro::GetTensorShape(op_context.input1),
tflite::micro::GetTensorData<data_type>(op_context.input1),
tflite::micro::GetTensorShape(op_context.input2),
tflite::micro::GetTensorData<data_type>(op_context.input2),
tflite::micro::GetTensorShape(op_context.output),
tflite::micro::GetTensorData<data_type>(op_context.output),
op_type::template op<data_type>);
}

template <KernelType kernel_type, typename OpType>
TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
OpContext op_context(context, node);
Expand Down
Loading

0 comments on commit 740cef3

Please sign in to comment.