root
/
MNN
mirror of https://github.com/alibaba/MNN.git
1
0
Fork 0
Code Issues Actions 1 Packages Projects Releases Wiki Activity
MNN/source/backend/cpu/CPUConvolution.cpp

305 lines
12 KiB
C++
Raw Blame History

//
// CPUConvolution.cpp
// MNN
//
// Created by MNN on 2018/07/15.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "backend/cpu/CPUConvolution.hpp"
#include <math.h>
#include "backend/cpu/compute/CommonOptFunction.h"
#include "core/Macro.h"
#include "core/TensorUtils.hpp"
#include <limits>
#include "backend/cpu/compute/ConvolutionFloatFactory.h"
//#define MNN_OPEN_TIME_TRACE
#include <MNN/AutoTime.hpp>
#include "core/ConvolutionCommon.hpp"
#include "backend/cpu/compute/ConvInt8Winograd.hpp"
#include "backend/cpu/compute/ConvInt8TiledExecutor.hpp"
#ifdef MNN_USE_ONEDNN
#include "backend/cpu/OneDNNConvInt8.hpp"
#endif
namespace MNN {
bool CPUConvolution::Resource::copyBiasAlign(const float* bias, int outputCount) {
auto core = static_cast<CPUBackend*>(backend)->functions();
int bytes = core->bytes;
int unit = core->pack;
auto alignOutput = UP_DIV(outputCount, unit) * unit;
int remain = alignOutput - outputCount;
mBias.reset(Tensor::createDevice<uint8_t>(std::vector<int>{alignOutput * bytes}));
bool success = backend->onAcquireBuffer(mBias.get(), Backend::STATIC);
if (!success) {
MNN_ERROR("Error for alloc memory for Alloc Bias\n");
return false;;
}
if (bytes < 4) {
core->MNNFp32ToLowp(bias, mBias->host<int16_t>(), outputCount);
} else {
::memcpy(mBias->host<float>(), bias, outputCount * bytes);
}
if (remain > 0) {
::memset(mBias->host<uint8_t>() + outputCount * bytes, 0, remain * bytes);
}
return true;
}
void CPUConvolution::ResourceInt8::updateInputOutputScale(std::vector<float> inputQuantInfo, std::vector<float> outputQuantInfo) {
std::call_once(flag, [&](){
// new scales and zero points
float inputScale = inputQuantInfo[0];
float outputScale = outputQuantInfo[0];
float inputZeroPoint = inputQuantInfo[1];
float outputZeroPoint = outputQuantInfo[1];
if (inputScale == 0.f || outputScale == 0.f) {
return;
}
if (mInputScale == inputScale && mOutputScale == outputScale) {
return;
}
auto scalePtr = mScaleFloat->host<float>();
auto biasPtr = mBiasInt32->host<int>();
int size = mScaleFloat->elementSize();
float is = mInputScale / inputScale;
float os = mOutputScale / outputScale;
const int kernelNum = mInt8WeightKernelSum.size();
// compute remains used in asymmetric quant
std::vector<int> remainsCorrection;
for (int i = 0; i < kernelNum; i++) {
int temp = (int(inputZeroPoint) - mInputZeroPoint) * mInt8WeightKernelSum[i];
remainsCorrection.emplace_back(temp);
}
for (int i = kernelNum; i < size; i++) {
remainsCorrection.emplace_back(0);
}
for (int i = 0; i < size; i++) {
// compute outputZeroPointFused in asymmetric quant
int correction1 = static_cast<int32_t>(mOutputZeroPoint / scalePtr[i]);
scalePtr[i] = scalePtr[i] * os / is;
int correction2 = static_cast<int32_t>(outputZeroPoint / scalePtr[i]);
int outputZeroPointFusedCorrection = correction2 - correction1;
#ifdef MNN_USE_SSE
if (offsets.empty()) {
biasPtr[i] = biasPtr[i] - remainsCorrection[i] + outputZeroPointFusedCorrection;
biasPtr[i] = static_cast<int32_t>(biasPtr[i] * is);
} else {
biasPtr[i] = biasPtr[i] - offsets[i];
biasPtr[i] = biasPtr[i] - remainsCorrection[i] + outputZeroPointFusedCorrection;
biasPtr[i] = static_cast<int32_t>(biasPtr[i] * is + offsets[i]);
}
#else
biasPtr[i] = biasPtr[i] - remainsCorrection[i] + outputZeroPointFusedCorrection;
biasPtr[i] = static_cast<int32_t>(biasPtr[i] * is);
#endif
}
mInputScale = inputScale;
mOutputScale = outputScale;
mInputZeroPoint = int8_t(inputZeroPoint);
mOutputZeroPoint = int8_t(outputZeroPoint);
mClampMin = int8_t(outputQuantInfo[2]);
mClampMax = int8_t(outputQuantInfo[3]);
});
}
CPUConvolution::ResourceInt8::~ResourceInt8() {
if(mWeightInt8 != nullptr) {
backend->onReleaseBuffer(mWeightInt8.get(), Backend::STATIC);
}
if(mBiasInt32 != nullptr){
backend->onReleaseBuffer(mBiasInt32.get(), Backend::STATIC);
}
if(mScaleFloat != nullptr){
backend->onReleaseBuffer(mScaleFloat.get(), Backend::STATIC);
}
}
std::shared_ptr<CPUConvolution::ResourceInt8> CPUConvolution::makeResourceInt8(Backend* backend, const MNN::Convolution2D *convParam,
std::vector<float> inputQuantInfo, std::vector<float> outputQuantInfo) {
if (inputQuantInfo.empty() && outputQuantInfo.empty()) {
inputQuantInfo = {0.f, 0.f, -127, 127};
outputQuantInfo = {0.f, 0.f, -127, 127};
}
auto core = static_cast<CPUBackend*>(backend)->int8Functions();
int UNIT, SRC_UNIT, DST_XUNIT;
core->MNNGetGemmUnit(&UNIT, &SRC_UNIT, &DST_XUNIT);
std::shared_ptr<CPUConvolution::ResourceInt8> resource(new ResourceInt8);
resource->backend = backend;
resource->mInputScale = inputQuantInfo[0];
resource->mOutputScale = outputQuantInfo[0];
const auto convCommon = convParam->common();
const auto kernelCount = convCommon->kernelY() * convCommon->kernelX();
const auto group = convParam->common()->group();
const auto srcCount = convCommon->inputCount();
const auto outputCount = convCommon->outputCount();
const auto outputChannleUp4 = UP_DIV(outputCount, UNIT) * UNIT;
resource->mActBits = convParam->symmetricQuan()->nbits();
resource->mWeightInt8.reset(Tensor::createDevice<int8_t>({group, outputCount / group, srcCount / group, kernelCount}));
resource->mBiasInt32.reset(Tensor::createDevice<int32_t>({outputChannleUp4}));
resource->mScaleFloat.reset(Tensor::createDevice<float>({outputChannleUp4}));
auto allocRes = backend->onAcquireBuffer(resource->mWeightInt8.get(), Backend::STATIC);
allocRes &= backend->onAcquireBuffer(resource->mBiasInt32.get(), Backend::STATIC);
allocRes &= backend->onAcquireBuffer(resource->mScaleFloat.get(), Backend::STATIC);
if (!allocRes) {
return nullptr;
}
auto biasPtr = resource->mBiasInt32->host<int32_t>();
memset(biasPtr, 0, outputChannleUp4 * sizeof(int32_t));
auto scalePtr = resource->mScaleFloat->host<float>();
memset(scalePtr, 0, outputChannleUp4 * sizeof(float));
const int8_t* weightSrc = nullptr;
std::shared_ptr<ConvolutionCommon::Int8Common> quanCommon;
if (!ConvolutionCommon::getConvInt8Parameters(convParam, quanCommon, weightSrc, scalePtr, biasPtr,
inputQuantInfo[0], outputQuantInfo[0],
convParam->symmetricQuan()->zeroPoint(),
convParam->symmetricQuan()->outputZeroPoint())) {
return nullptr;
}
const int kernelNum = outputCount;
int kernelChannel = srcCount;
if ((srcCount == outputCount) && (group == srcCount)) {
kernelChannel = 1; // depthwise
}
const int kernelSize = kernelChannel * kernelCount;
for (int i = 0; i < kernelNum; i++) {
int temp = 0;
int offset = i * kernelSize;
for (int j = 0; j < kernelSize; j++) {
temp += int(weightSrc[offset + j]);
}
resource->mInt8WeightKernelSum.emplace_back(temp);
}
#ifdef MNN_USE_SSE
resource->offsets.resize(outputCount);
// For SSE use uint8_t, int8_t -> uint8_t, x + 128 -> x', x * w + b = (x' - 128) * w + b = x' * w + (-128 * w) + b
for (int x = 0; x < outputCount; ++x) {
int offset = resource->mInt8WeightKernelSum[x] * (-128);
resource->offsets[x] = offset;
biasPtr[x] = biasPtr[x] + offset;
}
#endif
auto weightDst = resource->mWeightInt8->host<int8_t>();
memcpy(weightDst, weightSrc, resource->mWeightInt8->size());
resource->mInputZeroPoint = convParam->symmetricQuan()->zeroPoint();
resource->mOutputZeroPoint = convParam->symmetricQuan()->outputZeroPoint();
resource->mClampMin = convParam->symmetricQuan()->clampMin();
resource->mClampMax = convParam->symmetricQuan()->clampMax();
resource->mRelu = convCommon->relu() || convCommon->relu6();
return resource;
}
CPUConvolution::CPUConvolution(const Convolution2DCommon *convOp, Backend *b) : MNN::Execution(b), mCommon(convOp) {
// Do nothing
}
std::vector<float> CPUConvolution::getPostParameters() const {
std::vector<float> postParameters = {
1.0f,
1.0f,
-std::numeric_limits<float>().max(),
std::numeric_limits<float>().max(),
};
if (mCommon->relu()) {
postParameters[2] = 0.0f;
}
if (mCommon->relu6()) {
postParameters[2] = 0.0f;
postParameters[3] = 6.0f;
}
return postParameters;
}
int CPUConvolution::reorderWeightSize(int depth, int outputCount, int kernelSize, int unitDepth, int unitOC) {
return UP_DIV(outputCount, unitOC) * UP_DIV(depth, unitDepth) * kernelSize * unitDepth * unitOC;
}
template<typename T>
void CPUConvolution::reorderWeightSlow(T* dest, const T* source, size_t depth, size_t outputCount, size_t kernelSize,
size_t unitDepth, size_t unitOC, bool transpose) {
memset(dest, 0, reorderWeightSize(depth, outputCount, kernelSize, unitDepth, unitOC) * sizeof(T));
for (int dz = 0; dz < outputCount; ++dz) {
auto dz_unit = dz / unitOC;
auto mx = dz % unitOC;
auto dst_dz = dest + dz_unit * UP_DIV(depth, unitDepth) * kernelSize * unitDepth * unitOC;
for (int sz = 0; sz < depth; ++sz) {
auto sz_unit = sz / unitDepth;
auto my = sz % unitDepth;
auto dst_sz = dst_dz + sz_unit * kernelSize * unitDepth * unitOC;
auto src = source + kernelSize * (sz + dz * depth);
for (int ki = 0; ki < kernelSize; ++ki) {
auto dst_i = dst_sz + ki * unitDepth * unitOC;
if (transpose) {
dst_i[unitDepth * mx + my] = src[ki];
} else {
dst_i[unitOC * my + mx] = src[ki];
}
}
}
}
}
template void CPUConvolution::reorderWeightSlow<int8_t>(int8_t*, const int8_t*, size_t, size_t, size_t, size_t, size_t, bool);
ErrorCode CPUConvolution::onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
auto input = inputs[0];
auto output = outputs[0];
auto pad = ConvolutionCommon::convolutionPad(input, output, mCommon);
mPadY = pad.second;
mPadX = pad.first;
return NO_ERROR;
}
class ConvolutionFactory : public CPUBackend::Creator {
public:
virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
const MNN::Op *op, Backend *backend) const override {
return ConvolutionFloatFactory::create(inputs, outputs, op, backend);
}
};
class CPUConvInt8Creator : public CPUBackend::Creator {
public:
virtual Execution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
const MNN::Op* op, Backend* backend) const override {
std::vector<float> inputQuantInfo;
std::vector<float> outputQuantInfo;
if (inputs.size() > 0) {
inputQuantInfo = TensorUtils::getQuantInfo(inputs[0]);
outputQuantInfo = TensorUtils::getQuantInfo(outputs[0]);
}
auto convOp = op->main_as_Convolution2D();
#ifdef MNN_USE_ONEDNN
return OneDNNConvInt8::create(backend, convOp, inputs, outputs);
#endif
/*int nbit = 6;
auto quantAttr = new QuantAttr;
quantAttr->min = -(1<<(nbit-1))+1;
quantAttr->max = (1<<(nbit-1))-1;
TensorUtils::getDescribe(inputs[0])->quantAttr.reset(quantAttr);*/
auto res = CPUConvolution::makeResourceInt8(backend, convOp, inputQuantInfo, outputQuantInfo);
if (!inputs.empty()) {
std::vector<ConvInt8Winograd::UnitAttr> unitAttrs;
if (ConvInt8Winograd::bestWinogradUnit(convOp, inputs[0], res->mWeightInt8.get(), outputs[0], backend, unitAttrs)) {
return new ConvInt8Winograd(backend, convOp, res, unitAttrs);
}
}
return new ConvInt8TiledExecutor(backend, convOp, res);
}
};
REGISTER_CPU_OP_CREATOR(ConvolutionFactory, OpType_Convolution);
REGISTER_CPU_OP_CREATOR(CPUConvInt8Creator, OpType_ConvInt8);
} // namespace MNN
Reference in New Issue View Git Blame Copy Permalink