You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

237 lines
8.1 KiB

/*
* 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.
*/
#ifndef ANDROID_FRAMEWORKS_ML_NN_COMMON_CPU_OPERATION_UTILS_H
#define ANDROID_FRAMEWORKS_ML_NN_COMMON_CPU_OPERATION_UTILS_H
#include <android-base/logging.h>
#include <tensorflow/lite/kernels/internal/types.h>
#include <algorithm>
#include <cmath>
#include <limits>
#include <vector>
#include "OperationsUtils.h"
namespace android {
namespace nn {
// The implementations in tflite/kernels/internal/ take a Dims<4> object
// even if the original tensors were not 4D.
inline tflite::Dims<4> convertShapeToDims(const Shape& shape) {
CHECK_LE(shape.dimensions.size(), 4u);
tflite::Dims<4> dims;
// The dimensions are reversed in Dims<4>.
for (int i = 0; i < 4; ++i) {
int src = static_cast<int>(shape.dimensions.size()) - i - 1;
if (src >= 0) {
dims.sizes[i] = static_cast<int>(getSizeOfDimension(shape, src));
} else {
dims.sizes[i] = 1;
}
}
dims.strides[0] = 1;
for (int i = 1; i < 4; i++) {
dims.strides[i] = dims.strides[i - 1] * dims.sizes[i - 1];
}
return dims;
}
inline tflite::RuntimeShape convertShapeToTflshape(const Shape& shape) {
std::vector<int32_t> tflShapeDim(shape.dimensions.begin(), shape.dimensions.end());
return tflite::RuntimeShape(tflShapeDim.size(), tflShapeDim.data());
}
inline void convertFloat16ToFloat32(const _Float16* input, std::vector<float>* output) {
CHECK(input != nullptr);
CHECK(output != nullptr);
for (int i = 0; i < output->size(); ++i) {
(*output)[i] = static_cast<float>(input[i]);
}
}
inline void convertFloat32ToFloat16(const std::vector<float>& input, _Float16* output) {
CHECK(output != nullptr);
for (int i = 0; i < input.size(); ++i) {
output[i] = input[i];
}
}
// Convert int8 quantized values to uint8 assuming that the scale is the same
// and the distance between offsets is 128.
inline void convertInt8ToUInt8(const int8_t* input, std::vector<uint8_t>* output) {
CHECK(input != nullptr);
CHECK(output != nullptr);
for (int i = 0; i < output->size(); ++i) {
(*output)[i] = static_cast<uint8_t>(static_cast<int32_t>(input[i]) + 128);
}
}
// Convert uint8 quantized values to int8 assuming that the scale is the same
// and the distance between offsets is 128.
inline void convertUInt8ToInt8(const std::vector<uint8_t>& input, int8_t* output) {
CHECK(output != nullptr);
for (int i = 0; i < input.size(); ++i) {
output[i] = static_cast<int8_t>(static_cast<int32_t>(input[i]) - 128);
}
}
template <typename T>
inline void convertQuantToFloat32(const T* input, float scale, int32_t zeroPoint,
std::vector<float>* output) {
CHECK(input != nullptr);
CHECK(output != nullptr);
for (int i = 0; i < output->size(); ++i) {
(*output)[i] = (static_cast<float>(input[i]) - zeroPoint) * scale;
}
}
template <typename T>
inline void convertFloat32ToQuant(const std::vector<float>& input, float scale, int32_t zeroPoint,
T* output) {
CHECK(output != nullptr);
for (int i = 0; i < input.size(); ++i) {
int32_t intVal = std::round(input[i] / scale + zeroPoint);
intVal = std::min<int32_t>(std::max<int32_t>(intVal, std::numeric_limits<T>::min()),
std::numeric_limits<T>::max());
output[i] = static_cast<T>(intVal);
}
}
template <typename T>
inline bool convertNchwToNhwc(const T* nchw, const Shape& nchwShape, std::vector<T>* nhwc,
Shape* nhwcShape) {
NN_RET_CHECK_EQ(getNumberOfDimensions(nchwShape), 4)
<< "Error converting a non-4-D tensor to NHWC layout";
*nhwcShape = nchwShape;
const auto& fromDim = nchwShape.dimensions;
nhwcShape->dimensions = {fromDim[0], fromDim[2], fromDim[3], fromDim[1]};
nhwc->resize(getNumberOfElements(nchwShape));
auto to = nhwc->data();
uint32_t spatialSize = fromDim[2] * fromDim[3];
for (uint32_t n = 0; n < fromDim[0]; n++) {
for (uint32_t hw = 0; hw < spatialSize; hw++) {
for (uint32_t c = 0; c < fromDim[1]; c++) {
uint32_t fromIndex = n * fromDim[1] * spatialSize + c * spatialSize + hw;
*to++ = nchw[fromIndex];
}
}
}
return true;
}
template <typename T>
inline bool convertNhwcToNchw(const std::vector<T>& nhwc, const Shape& nhwcShape, T* nchw) {
NN_RET_CHECK_EQ(getNumberOfDimensions(nhwcShape), 4)
<< "Error converting a non-4-D tensor to NCHW layout";
const auto& fromDim = nhwcShape.dimensions;
const auto from = nhwc.data();
uint32_t spatialSize = fromDim[1] * fromDim[2];
for (uint32_t n = 0; n < fromDim[0]; n++) {
for (uint32_t c = 0; c < fromDim[3]; c++) {
for (uint32_t hw = 0; hw < spatialSize; hw++) {
uint32_t fromIndex = n * spatialSize * fromDim[3] + hw * fromDim[3] + c;
*nchw++ = from[fromIndex];
}
}
}
return true;
}
template <typename T>
class InputWithLayout {
public:
InputWithLayout(bool useNchw) : mDataOriginal(nullptr), mUseNchw(useNchw) {}
bool initialize(const T* data, const Shape& shape) {
mDataOriginal = data;
mShape = shape;
if (mUseNchw) {
return convertNchwToNhwc(mDataOriginal, shape, &mDataNhwc, &mShape);
}
return true;
}
const T* getNhwcBuffer() { return mUseNchw ? mDataNhwc.data() : mDataOriginal; }
const Shape& getNhwcShape() { return mShape; }
private:
const T* mDataOriginal;
std::vector<T> mDataNhwc;
Shape mShape;
bool mUseNchw;
};
template <typename T>
class OutputWithLayout {
public:
OutputWithLayout(bool useNchw) : mDataOriginal(nullptr), mUseNchw(useNchw) {}
bool initialize(T* data, const Shape& shape) {
NN_RET_CHECK_EQ(getNumberOfDimensions(shape), 4);
mDataOriginal = data;
mShape = shape;
if (mUseNchw) {
const auto& dim = shape.dimensions;
mShape.dimensions = {dim[0], dim[2], dim[3], dim[1]};
mDataNhwc.resize(getNumberOfElements(shape));
}
return true;
}
T* getNhwcBuffer() { return mUseNchw ? mDataNhwc.data() : mDataOriginal; }
const Shape& getNhwcShape() { return mShape; }
bool commit() {
if (mUseNchw) {
return convertNhwcToNchw(mDataNhwc, mShape, mDataOriginal);
}
return true;
}
private:
T* mDataOriginal;
std::vector<T> mDataNhwc;
Shape mShape;
bool mUseNchw;
};
template <typename T>
inline void CalculateActivationRange(int32_t activation, const Shape& outputShape,
int32_t* outputActivationMin, int32_t* outputActivationMax);
template <>
inline void CalculateActivationRange<uint8_t>(int32_t activation, const Shape& outputShape,
int32_t* outputActivationMin,
int32_t* outputActivationMax) {
CalculateActivationRangeUint8(activation, outputShape, outputActivationMin,
outputActivationMax);
}
template <>
inline void CalculateActivationRange<int8_t>(int32_t activation, const Shape& outputShape,
int32_t* outputActivationMin,
int32_t* outputActivationMax) {
CalculateActivationRangeInt8(activation, outputShape, outputActivationMin, outputActivationMax);
}
} // namespace nn
} // namespace android
#endif // ANDROID_FRAMEWORKS_ML_NN_COMMON_CPU_OPERATION_UTILS_H