Skip to content

Commit

Permalink
feat(compression): implement tensor decompression in op conv
Browse files Browse the repository at this point in the history 
Implement tensor decompression in op conv. Extend tests to
validate operation on compressed tensors.

BUG=part of tensorflow#2636
  • Loading branch information
@ddavis-2015 @rkuester
ddavis-2015 authored and rkuester committed Dec 15, 2024
1 parent b2f2718 commit 7a38395
Show file tree
Hide file tree
Showing 11 changed files with 749 additions and 134 deletions.
51 changes: 49 additions & 2 deletions tensorflow/lite/micro/kernels/conv.cc
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
/* Copyright 2023 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 Down Expand Up @@ -45,15 +45,35 @@ TfLiteStatus ConvEval(TfLiteContext* context, TfLiteNode* node) {
TFLITE_DCHECK(node->user_data != nullptr);
const auto& data = *(static_cast<const OpDataConv*>(node->user_data));

#ifdef USE_TFLM_COMPRESSION

MicroContext* micro_context = GetMicroContext(context);

const CompressionTensorData* weights_comp_td =
micro_context->GetTensorCompressionData(node, kConvWeightsTensor);
const CompressionTensorData* bias_comp_td =
micro_context->GetTensorCompressionData(node, kConvBiasTensor);

#endif // USE_TFLM_COMPRESSION

switch (input->type) { // Already know in/out types are same.
case kTfLiteFloat32: {
tflite::reference_ops::Conv(
ConvParamsFloat(params, data), tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<float>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<float>(micro_context, filter,
weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<float>(
micro_context, bias, bias_comp_td, data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<float>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<float>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<float>(output),
tflite::micro::GetTensorShape(nullptr), nullptr);
Expand All @@ -67,9 +87,18 @@ TfLiteStatus ConvEval(TfLiteContext* context, TfLiteNode* node) {
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(micro_context, filter,
weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int32_t>(
micro_context, bias, bias_comp_td, data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<std::int32_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output));
} else if (bias->type == kTfLiteInt64) {
Expand All @@ -79,9 +108,18 @@ TfLiteStatus ConvEval(TfLiteContext* context, TfLiteNode* node) {
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(micro_context, filter,
weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetTensorData<int64_t>(
micro_context, bias, bias_comp_td, data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<std::int64_t>(bias),
tflite::micro::GetTensorData<std::int64_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output));
} else {
Expand Down Expand Up @@ -119,9 +157,18 @@ TfLiteStatus ConvEval(TfLiteContext* context, TfLiteNode* node) {
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int8_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(micro_context, filter,
weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int32_t>(
micro_context, bias, bias_comp_td, data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int32_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int8_t>(output));
break;
Expand Down
10 changes: 9 additions & 1 deletion tensorflow/lite/micro/kernels/conv.h
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
/* Copyright 2023 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 Down Expand Up @@ -49,6 +49,14 @@ struct OpDataConv {
// A buffer used to store unpacked filter values. This is used if the source
// tensor is of n-bit precision that cannot be easily processed by kernels.
int filter_buffer_index;

#ifdef USE_TFLM_COMPRESSION

// scratch buffers for compressed tensors
int weights_scratch_index;
int bias_scratch_index;

#endif // USE_TFLM_COMPRESSION
};

extern const int kConvInputTensor;
Expand Down
19 changes: 18 additions & 1 deletion tensorflow/lite/micro/kernels/conv_common.cc
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
/* Copyright 2023 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 Down Expand Up @@ -209,6 +209,23 @@ TfLiteStatus ConvPrepare(TfLiteContext* context, TfLiteNode* node) {
&data->filter_buffer_index);
}

#ifdef USE_TFLM_COMPRESSION

// Compression scratch buffers.
// These will only be allocated if the tensor is compressed.
if (micro_context->IsTensorCompressed(node, kConvWeightsTensor) &&
filter->type == kTfLiteInt4) {
MicroPrintf("Compression not supported with INT4 tensors");
return kTfLiteError;
}
data->weights_scratch_index =
micro_context->AllocateDecompressionScratchBuffer(node,
kConvWeightsTensor);
data->bias_scratch_index =
micro_context->AllocateDecompressionScratchBuffer(node, kConvBiasTensor);

#endif // USE_TFLM_COMPRESSION

micro_context->DeallocateTempTfLiteTensor(filter);
micro_context->DeallocateTempTfLiteTensor(input);
micro_context->DeallocateTempTfLiteTensor(output);
Expand Down
Loading

0 comments on commit 7a38395

Please sign in to comment.