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/compute/IdstConvolutionInt8.cpp

280 lines
12 KiB
C++
Raw Blame History

//
// IdstConvolutionInt8.cpp
// MNN
//
// Created by MNN on 2018/07/16.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "IdstConvolutionInt8.hpp"
#include "ConvInt8TiledExecutor.hpp"
#include "ConvolutionTiledExecutor.hpp"
#include "CommonOptFunction.h"
#include "core/Concurrency.h"
#include "core/BufferAllocator.hpp"
#include "ConvOpt.h"
#include "ConvolutionIntFactory.hpp"
#include "core/Macro.h"
#include "core/TensorUtils.hpp"
#include "Int8FunctionsOpt.h"
#define MNN_OPEN_TIME_TRACE
#include <MNN/AutoTime.hpp>
#ifdef MNN_USE_NEON
#include <arm_neon.h>
#endif
#ifdef MNN_USE_SSE
extern "C" {
void MNNInt8ToUInt8(void* ptr, int count);
}
#endif
#define QUANT_INFO_BYTES 4
namespace MNN {
IdstConvolutionInt8::IdstConvolutionInt8(const Convolution2DCommon* convOp, Backend* b,
const ConvolutionCommon::Int8Common* common, const float* bias,
size_t biasSize) : MNN::CPUConvolution(convOp, b) {
auto core = static_cast<CPUBackend*>(b)->int8Functions();
int UNIT, SRC_UNIT, DST_XUNIT;
core->MNNGetGemmUnit(&UNIT, &SRC_UNIT, &DST_XUNIT);
int PackUnit = static_cast<CPUBackend*>(b)->functions()->pack;
int ocUp4 = ROUND_UP(biasSize, PackUnit);
int ocUpHp = ROUND_UP(biasSize, UNIT);
mBias.reset(ocUp4);
mBias.clear();
auto biasDest = mBias.get();
mAMin = common->quan->aMin();
mAMax = common->quan->aMaxOrBits();
mQuanScale = common->quan->quantScale();
// The postTreat will contain scale_bias and biasRelu, so the bias will be add twice
for (int i = 0; i < biasSize; ++i) {
biasDest[i] = bias[i] * 0.5f;
}
int outputCount = (int)biasSize;
mQuan = common->quan;
MNN_ASSERT(nullptr != mQuan);
mAlpha.reset(ROUND_UP(common->alpha.size(), PackUnit));
mAlpha.clear();
::memcpy(mAlpha.get(), common->alpha.get(), common->alpha.size() * sizeof(float));
auto weightLength = common->weight.size();
mSrcCount = (int)weightLength / mCommon->kernelX() / mCommon->kernelY() / outputCount;
auto kx = mCommon->kernelX();
auto ky = mCommon->kernelY();
auto kernelCount = kx * ky;
auto srcCount = mSrcCount;
std::vector<int> shape;
shape = {1, UP_DIV(outputCount, UNIT), UP_DIV(srcCount, SRC_UNIT) * kernelCount, UNIT, SRC_UNIT};
mFakeBias.reset(Tensor::createDevice<float>({ocUpHp}));
int weightlen = shape[0] * shape[1] * shape[2] * shape[3] * shape[4];
int quantlen = 2 * ocUpHp * QUANT_INFO_BYTES;
mWeight.reset(Tensor::createDevice<int8_t>({weightlen + quantlen}));
mValid = b->onAcquireBuffer(mWeight.get(), Backend::STATIC);
mValid &= b->onAcquireBuffer(mFakeBias.get(), Backend::STATIC);
if (!mValid) {
MNN_ERROR("Memory not enough\n");
return;
}
AutoStorage<uint8_t> weightReordered(weightlen);
AutoStorage<float> fakeWeightScaleBias(2 * ocUp4);
if (weightReordered.get() == nullptr || fakeWeightScaleBias.get() == nullptr) {
MNN_ERROR("Memory not enough\n");
return;
}
int32_t info[6] = {1, outputCount, srcCount, kernelCount, UNIT, SRC_UNIT};
ConvInt8TiledExecutor::reorderWeight(weightReordered.get(), (uint8_t*)common->weight.get(), info);
::memset(mFakeBias->host<float>(), 0, mFakeBias->size());
auto ptr = (float*)fakeWeightScaleBias.get();
::memset(ptr, 0, 2 * ocUp4 * 4);
for (int i = 0; i < ocUp4; ++i) {
ptr[i] = 1.f;
}
#ifdef MNN_USE_SSE
for (int oz = 0; oz < outputCount; ++oz) {
auto srcZ = common->weight.get() + oz * kernelCount * srcCount;
int32_t offset = 0;
for (int i = 0; i < kernelCount * srcCount; ++i) {
offset += srcZ[i] * (-128);
}
mFakeBias->host<float>()[oz] = static_cast<float>(offset) * 1.f;
}
#endif
int32_t params[6] = {shape[0], shape[1], shape[2], shape[3], shape[4], ocUp4};
ConvInt8TiledExecutor::packWeightAndQuantInfo(mWeight->host<int8_t>(), (int8_t*)weightReordered.get(), (int8_t*)fakeWeightScaleBias.get(), params, QUANT_INFO_BYTES);
}
IdstConvolutionInt8::~IdstConvolutionInt8() {
// Do nothing
}
ErrorCode IdstConvolutionInt8::onResize(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
auto core = static_cast<CPUBackend*>(backend())->int8Functions();
int UNIT, SRC_UNIT, DST_XUNIT;
core->MNNGetGemmUnit(&UNIT, &SRC_UNIT, &DST_XUNIT);
int PackUnit = static_cast<CPUBackend*>(backend())->functions()->pack;
CPUConvolution::onResize(inputs, outputs);
ConvolutionTiledExecutor::setIm2ColParameter(mIm2ColParamter, mCommon, inputs[0], outputs[0], mPadX, mPadY, static_cast<CPUBackend*>(backend())->functions(), core);
auto ow = mIm2ColParamter.ow;
auto oh = mIm2ColParamter.oh;
int tileCount = UP_DIV(ow * oh, DST_XUNIT);
auto outputCountUnit = UP_DIV(outputs[0]->channel(), PackUnit);
int number = std::max(((CPUBackend*)backend())->threadNumber(), 1);
number = std::min(number, tileCount);
TensorUtils::copyShape(inputs[0], &mSrcCopyBuffer, true);
mSrcCopyBuffer.buffer().dim[0].extent = 1;
mSrcCopyBuffer.buffer().type = halide_type_of<int8_t>();
TensorUtils::setLinearLayout(&mSrcCopyBuffer);
mTempBuffer.buffer().type = halide_type_of<int8_t>();
mTempBuffer.buffer().dimensions = 3;
mTempBuffer.buffer().dim[0].extent = number;
mTempBuffer.buffer().dim[1].extent = DST_XUNIT;
mTempBuffer.buffer().dim[2].extent = mIm2ColParamter.kernelCountUnit * SRC_UNIT;
TensorUtils::setLinearLayout(&mTempBuffer);
bool success = backend()->onAcquireBuffer(&mSrcCopyBuffer, Backend::DYNAMIC);
success &= backend()->onAcquireBuffer(&mTempBuffer, Backend::DYNAMIC);
if (!success) {
return OUT_OF_MEMORY;
}
auto bufferAlloc = static_cast<CPUBackend*>(backend())->getBufferAllocator();
auto blitInfoSize = ConvolutionTiledExecutor::computeBlitInfoSize(DST_XUNIT, mIm2ColParamter.ow, mIm2ColParamter.kernelX * mIm2ColParamter.kernelY, number);
mBlitInfo = bufferAlloc->alloc(blitInfoSize.first);
if (mBlitInfo.invalid()) {
return OUT_OF_MEMORY;
}
bufferAlloc->free(mBlitInfo);
mBlitInfoStride = blitInfoSize.second;
backend()->onReleaseBuffer(&mSrcCopyBuffer, Backend::DYNAMIC);
backend()->onReleaseBuffer(&mTempBuffer, Backend::DYNAMIC);
mPostParameters = getPostParameters();
return NO_ERROR;
}
ErrorCode IdstConvolutionInt8::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
auto coreFloat = static_cast<CPUBackend*>(backend())->functions();
auto coreInt = static_cast<CPUBackend*>(backend())->int8Functions();
int UNIT__, SRC_UNIT, DST_XUNIT;
coreInt->MNNGetGemmUnit(&UNIT__, &SRC_UNIT, &DST_XUNIT);
int PackUnit = static_cast<CPUBackend*>(backend())->functions()->pack;
auto gemmKernel = coreInt->Int8GemmKernel;
if (SRC_UNIT > PackUnit) {
memset(mTempBuffer.host<int8_t>(), 0, mTempBuffer.size());
}
// AUTOTIME;
auto input = inputs[0];
auto output = outputs[0];
auto weightOrigin = mWeight->host<int8_t>();
auto dstZStep = mIm2ColParamter.ow * mIm2ColParamter.oh * PackUnit * input->batch();
int threadNumber = 1;
auto blitProc = coreInt->MNNPackC4Int8ForMatMul_A;
int batch = input->batch();
int width = mIm2ColParamter.ow;
int height = mIm2ColParamter.oh;
auto ocC4 = UP_DIV(output->channel(), PackUnit);
auto kernelCountUnit = mIm2ColParamter.kernelCountUnit;
int count = width * height;
float quantScale[] = {
mQuanScale,
mQuanScale,
mQuanScale,
mQuanScale
};
float zeroPoint = 0;
std::vector<float> fakeScale(ocC4 * PackUnit, 1.0f);
QuanPostTreatParameters quanParam;
quanParam.biasFloat = mFakeBias->host<float>();
quanParam.scale = fakeScale.data();
quanParam.useInt8 = 0;
float fp32minmax[2] = {-std::numeric_limits<float>().max(), std::numeric_limits<float>().max()};
quanParam.fp32minmax = fp32minmax;
std::vector<float> fakeSrcKernleSum(DST_XUNIT, 0.f);
quanParam.srcKernelSum = fakeSrcKernleSum.data();
std::vector<float> fakeInputScale(DST_XUNIT, 1.f);
quanParam.inputScale = fakeInputScale.data();
std::vector<float> fakeWeightKernelsSum(ROUND_UP(output->channel(), UNIT__), 0.f);
quanParam.weightKernelSum = fakeWeightKernelsSum.data();
quanParam.inputBias = nullptr;
quanParam.blockNum = 1;
// MNN_PRINT("%s, %d, %d, %d,%d->%d,%d\n", layer->layer.layerId, layer->kernelSize[0], layer->kernelSize[1],
// input->d1, input->d2, output->d1, output->d2);
auto bn = static_cast<CPUBackend*>(backend());
int inputTotalSize = bn->getTensorSize(&mSrcCopyBuffer, true);
int8_t* srcCopy = mSrcCopyBuffer.host<int8_t>();
const int col_buffer_size = mIm2ColParamter.kernelCountUnit * DST_XUNIT * SRC_UNIT * sizeof(int8_t);
for (int batchIndex = 0; batchIndex < batch; ++batchIndex) {
auto srcOrigin = input->host<float>() + input->stride(0) * batchIndex;
auto dstOrigin = output->host<float>() + output->stride(0) * batchIndex;
MNNFloat2Int8(srcOrigin, srcCopy, inputTotalSize / 4, &mQuanScale, mAMin, mAMax, &zeroPoint, 0);
int tileCount = UP_DIV(count, DST_XUNIT);
threadNumber = std::max(((CPUBackend*)backend())->threadNumber(), 1);
threadNumber = std::min(threadNumber, tileCount);
auto outputOrigin = output->host<float>() + batchIndex * output->stride(0);
auto threadFunction = [&](int tId) {
auto colAddr = mTempBuffer.host<int8_t>() + tId * mTempBuffer.buffer().dim[0].stride;
auto srcPtr = (int8_t const **)(mBlitInfo.ptr() + tId * mBlitInfoStride.first);
auto el = (int32_t *)(srcPtr + mBlitInfoStride.second);
int32_t info[5];
info[1] = mIm2ColParamter.iw * mIm2ColParamter.ih;
info[2] = DST_XUNIT;
info[3] = mIm2ColParamter.strideX;
for (int tIndex = (int)tId; tIndex < tileCount; tIndex += threadNumber) {
int xIndexStart = tIndex * DST_XUNIT;
int realDstCount = ALIMIN(count - xIndexStart, DST_XUNIT);
auto res = ConvolutionTiledExecutor::turnIm2ColToBlitInfo((const float**)srcPtr, el, xIndexStart, realDstCount, mIm2ColParamter, (const uint8_t*)srcCopy, sizeof(int8_t));
int number = res.first;
bool needZero = res.second;
if (needZero) {
::memset(colAddr, zeroPoint, col_buffer_size);
}
info[0] = number;
info[4] = realDstCount;
if (number > 0) {
blitProc(colAddr, srcPtr, info, el);
}
auto outputInTile = outputOrigin + xIndexStart * PackUnit;
// GEMM
#ifdef MNN_USE_SSE
const int col_buffer_size = mIm2ColParamter.kernelCountUnit * DST_XUNIT * SRC_UNIT;
MNNInt8ToUInt8(colAddr, col_buffer_size);
#endif
gemmKernel((int8_t*)outputInTile, colAddr, weightOrigin, kernelCountUnit, dstZStep * sizeof(float), ocC4, &quanParam, realDstCount);
}
};
MNN_CONCURRENCY_BEGIN(tId, threadNumber) {
threadFunction((int)tId);
}
MNN_CONCURRENCY_END();
threadNumber = std::max(((CPUBackend*)backend())->threadNumber(), 1);
threadNumber = std::min(threadNumber, ocC4);
MNN_CONCURRENCY_BEGIN(tId, threadNumber) {
for (int z = (int)tId; z < ocC4; z += threadNumber) {
coreFloat->MNNScaleAndAddBias(dstOrigin + z * dstZStep, dstOrigin + z * dstZStep, mBias.get() + PackUnit * z,
mAlpha.get() + PackUnit * z, width * height, 1);
coreFloat->MNNAxByClampBroadcastUnit(dstOrigin + z * dstZStep, dstOrigin + z * dstZStep, mBias.get() + PackUnit * z, width * height, 0, 0, 1, mPostParameters.data());
}
}
MNN_CONCURRENCY_END();
}
return NO_ERROR;
}
} // namespace MNN
Reference in New Issue View Git Blame Copy Permalink