/*
* Copyright (C) 2017 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
// Contains the implementation of the operations.
#define LOG_TAG "Operations"
#include "CpuOperationUtils.h"
#include "Operations.h"
#include "tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h"
#include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
#include "Tracing.h"
namespace android {
namespace nn {
bool copyData(const void* inputData, const Shape& inputShape, void* outputData,
const Shape& outputShape) {
NNTRACE_COMP("copyData");
size_t count = nonExtensionOperandSizeOfData(inputShape.type, inputShape.dimensions);
memcpy(outputData, inputData, count);
return true;
}
template <typename T>
bool depthToSpaceGeneric(const T* inputData, const Shape& inputShape, int32_t blockSize,
T* outputData, const Shape& outputShape) {
NNTRACE_COMP("optimized_ops::DepthToSpace");
tflite::optimized_ops::DepthToSpace(inputData, convertShapeToDims(inputShape), blockSize,
outputData, convertShapeToDims(outputShape));
return true;
}
template bool depthToSpaceGeneric<float>(const float* inputData, const Shape& inputShape,
int32_t blockSize, float* outputData,
const Shape& outputShape);
template bool depthToSpaceGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
int32_t blockSize, _Float16* outputData,
const Shape& outputShape);
template bool depthToSpaceGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
int32_t blockSize, uint8_t* outputData,
const Shape& outputShape);
template <typename T>
bool spaceToDepthGeneric(const T* inputData, const Shape& inputShape, int32_t blockSize,
T* outputData, const Shape& outputShape) {
NNTRACE_COMP("optimized_ops::SpaceToDepth");
tflite::optimized_ops::SpaceToDepth(inputData, convertShapeToDims(inputShape), blockSize,
outputData, convertShapeToDims(outputShape));
return true;
}
template bool spaceToDepthGeneric<float>(const float* inputData, const Shape& inputShape,
int32_t blockSize, float* outputData,
const Shape& outputShape);
template bool spaceToDepthGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
int32_t blockSize, _Float16* outputData,
const Shape& outputShape);
template bool spaceToDepthGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
int32_t blockSize, uint8_t* outputData,
const Shape& outputShape);
template <typename T>
bool padGeneric(const T* inputData, const Shape& inputShape, const int32_t* paddings, T padValue,
T* outputData, const Shape& outputShape) {
NNTRACE_TRANS("padGeneric");
// Based on
// http://google3/third_party/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h?l=6194&rcl=213557260
// TFLite runtime calls are currently fixed at 4 dimensions. Copy inputs so
// we can pad them to 4 dims (yes, we are "padding the padding").
int32_t numInputDims = static_cast<int32_t>(getNumberOfDimensions(inputShape));
NN_OPS_CHECK(numInputDims <= 4);
std::vector<int> leftPaddings(4 - numInputDims, 0);
std::vector<int> rightPaddings(4 - numInputDims, 0);
for (int32_t i = 0; i < numInputDims; ++i) {
leftPaddings.push_back(paddings[i * 2]);
rightPaddings.push_back(paddings[i * 2 + 1]);
}
const int leftBPadding = leftPaddings[0];
const int leftHPadding = leftPaddings[1];
const int leftWPadding = leftPaddings[2];
const int leftDPadding = leftPaddings[3];
const int rightBPadding = rightPaddings[0];
const int rightHPadding = rightPaddings[1];
const int rightWPadding = rightPaddings[2];
const int rightDPadding = rightPaddings[3];
const auto extInputShape =
tflite::RuntimeShape::ExtendedShape(4, convertShapeToTflshape(inputShape));
const auto extOutputShape =
tflite::RuntimeShape::ExtendedShape(4, convertShapeToTflshape(outputShape));
const int outputBatch = extOutputShape.Dims(0);
const int outputHeight = extOutputShape.Dims(1);
const int outputWidth = extOutputShape.Dims(2);
const int outputDepth = extOutputShape.Dims(3);
const int inputDepth = extInputShape.Dims(3);
NNTRACE_COMP_SWITCH("padGeneric");
if (leftBPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData, padValue, leftBPadding * outputHeight * outputWidth * outputDepth);
}
for (int outB = leftBPadding; outB < outputBatch - rightBPadding; ++outB) {
if (leftHPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData + tflite::Offset(extOutputShape, outB, 0, 0, 0), padValue,
leftHPadding * outputWidth * outputDepth);
}
for (int outH = leftHPadding; outH < outputHeight - rightHPadding; ++outH) {
if (leftWPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData + tflite::Offset(extOutputShape, outB, outH, 0, 0), padValue,
leftWPadding * outputDepth);
}
for (int outW = leftWPadding; outW < outputWidth - rightWPadding; ++outW) {
if (leftDPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData + tflite::Offset(extOutputShape, outB, outH, outW, 0),
padValue, leftDPadding);
}
T* out =
outputData + tflite::Offset(extOutputShape, outB, outH, outW, leftDPadding);
const T* in =
inputData + tflite::Offset(extInputShape, outB - leftBPadding,
outH - leftHPadding, outW - leftWPadding, 0);
memcpy(out, in, inputDepth * sizeof(T));
if (rightDPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData + tflite::Offset(extOutputShape, outB, outH, outW,
outputDepth - rightDPadding),
padValue, rightDPadding);
}
}
if (rightWPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData + tflite::Offset(extOutputShape, outB, outH,
outputWidth - rightWPadding, 0),
padValue, rightWPadding * outputDepth);
}
}
if (rightHPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData + tflite::Offset(extOutputShape, outB, outputHeight - rightHPadding,
0, 0),
padValue, rightHPadding * outputWidth * outputDepth);
}
}
if (rightBPadding != 0) {
tflite::optimized_ops::TypedMemset<T>(
outputData + tflite::Offset(extOutputShape, outputBatch - rightBPadding, 0, 0, 0),
padValue, rightBPadding * outputHeight * outputWidth * outputDepth);
}
return true;
}
template bool padGeneric<float>(const float* inputData, const Shape& inputShape,
const int32_t* paddings, float padValue, float* outputData,
const Shape& outputShape);
template bool padGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
const int32_t* paddings, _Float16 padValue, _Float16* outputData,
const Shape& outputShape);
template bool padGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
const int32_t* paddings, uint8_t padValue, uint8_t* outputData,
const Shape& outputShape);
template <typename T>
bool batchToSpaceGeneric(const T* inputData, const Shape& inputShape, const int32_t* blockSize,
T* outputData, const Shape& outputShape) {
// Needed by low level implementation, but not really used.
tflite::Dims<4> blockSizeDim, cropsDim;
const int32 crops[4] = {0, 0, 0, 0};
NNTRACE_COMP("optimized_ops::BatchToSpaceND");
tflite::optimized_ops::BatchToSpaceND(inputData, convertShapeToDims(inputShape), blockSize,
blockSizeDim, crops, cropsDim, outputData,
convertShapeToDims(outputShape));
return true;
}
template bool batchToSpaceGeneric<float>(const float* inputData, const Shape& inputShape,
const int32_t* blockSize, float* outputData,
const Shape& outputShape);
template bool batchToSpaceGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
const int32_t* blockSize, _Float16* outputData,
const Shape& outputShape);
template bool batchToSpaceGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
const int32_t* blockSize, uint8_t* outputData,
const Shape& outputShape);
template <typename T>
bool spaceToBatchGeneric(const T* inputData, const Shape& inputShape, const int32_t* blockSize,
const int32_t* padding, const Shape& paddingShape, T* outputData,
const Shape& outputShape) {
// Needed by low level implementation, but not really used.
tflite::RuntimeShape blockSizeDim;
NNTRACE_COMP("optimized_ops::SpaceToBatchND");
tflite::optimized_ops::SpaceToBatchND(
{.output_offset = outputShape.offset}, convertShapeToTflshape(inputShape), inputData,
blockSizeDim, blockSize, convertShapeToTflshape(paddingShape), padding,
convertShapeToTflshape(outputShape), outputData);
return true;
}
template bool spaceToBatchGeneric<float>(const float* inputData, const Shape& inputShape,
const int32_t* blockSize, const int32_t* padding,
const Shape& paddingShape, float* outputData,
const Shape& outputShape);
template bool spaceToBatchGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
const int32_t* blockSize, const int32_t* padding,
const Shape& paddingShape, _Float16* outputData,
const Shape& outputShape);
template bool spaceToBatchGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
const int32_t* blockSize, const int32_t* padding,
const Shape& paddingShape, uint8_t* outputData,
const Shape& outputShape);
} // namespace nn
} // namespace android