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/ConvolutionWinograd3D.cpp

401 lines
19 KiB
C++
Raw Blame History

//
// ConvolutionWinograd3D.cpp
// MNN
//
// Created by MNN on 2018/09/23.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "backend/cpu/compute/ConvolutionWinograd3D.hpp"
#include "backend/cpu/CPUBackend.hpp"
#include <math.h>
#include "backend/cpu/compute/CommonOptFunction.h"
#include "core/Concurrency.h"
#include "backend/cpu/compute/ConvOpt.h"
#include "core/Macro.h"
#include "core/TensorUtils.hpp"
#include "math/WingoradGenerater.hpp"
#ifdef MNN_USE_NEON
#include <arm_neon.h>
#endif
#define CONVOLUTION_WINOGRAD_MAX_UNIT 8
#define CONVOLUTION_WINOGRAD_MIN_UNIT 2
using namespace MNN::Math;
//#define MNN_WINOGRAD_PRINT_REDUCE_RATE
namespace MNN {
ConvolutionWinograd3D::ConvolutionWinograd3D(const Convolution3DCommon *convOp, const Tensor *input, const Tensor *output,
Backend *b, const float *originWeight, size_t originWeightSize,
const float *bias, size_t biasSize, int unit) : Execution(b), mUnit(unit) {
for (int32_t kernel: *(convOp->kernels())) {
mKernels.push_back(kernel);
}
MNN_ASSERT(mKernels[1] == mKernels[2]);
mPadMode = convOp->padMode();
if (mPadMode != PadMode_SAME) {
for (int32_t pad: *(convOp->pads())) {
mPads.push_back(pad);
}
}
mPostFunction = CPUConvolution3D::getPostFunction(convOp);
const int inputChannel = convOp->inputCount(), outputChannel = convOp->outputCount();
const int kernelDepth = mKernels[0], kernelSize = mKernels[1], alpha = unit + kernelSize - 1, alpha2 = alpha * alpha;
mAlpha = alpha;
mSourceTransform = WinogradFunction::chooseSourceTransform(alpha, alpha);
mDestTransform = WinogradFunction::chooseDestTransform(alpha, unit);
mWeight.reset(Tensor::createDevice<float>({ALIGN_UP4(inputChannel) * ALIGN_UP4(outputChannel) * kernelDepth * alpha2}));
mBias.reset(Tensor::createDevice<float>({ALIGN_UP4((int)biasSize)}));
bool valid = b->onAcquireBuffer(mWeight.get(), Backend::STATIC);
valid = valid && b->onAcquireBuffer(mBias.get(), Backend::STATIC);
if (!valid) {
return;
}
memset(mBias->host<float>(), 0, mBias->size());
memcpy(mBias->host<float>(), bias, biasSize * sizeof(float));
WinogradGenerater generator(unit, kernelSize);
const int srcDepthStep = inputChannel * outputChannel * kernelSize * kernelSize;
const int dstDepthStep = ALIGN_UP4(inputChannel) * ALIGN_UP4(outputChannel) * alpha2;
std::shared_ptr<Tensor> srcWeight, transWeight;
for (int d = 0; d < kernelDepth; ++d) {
srcWeight.reset(Tensor::create<float>({outputChannel, inputChannel, kernelSize, kernelSize}, (void*)(originWeight + d * srcDepthStep)));
transWeight.reset(Tensor::create<float>({alpha2, UP_DIV(outputChannel, 4), UP_DIV(inputChannel, 4), 4, 4},
(void*)(mWeight->host<float>() + d * dstDepthStep)));
generator.transformWeight(transWeight.get(), srcWeight.get());
}
}
ConvolutionWinograd3D::~ConvolutionWinograd3D() {
if (nullptr != mBias) {
backend()->onReleaseBuffer(mBias.get(), Backend::STATIC);
}
if (nullptr != mWeight) {
backend()->onReleaseBuffer(mWeight.get(), Backend::STATIC);
}
}
ErrorCode ConvolutionWinograd3D::onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
auto input = inputs[0];
auto output = outputs[0];
const int oc = output->length(1), od = output->length(2);
const int ic = input->length(1), id = input->length(2);
const int threadNumber = ((CPUBackend*)backend())->threadNumber();
const int alpha2 = mAlpha * mAlpha;
if (mPadMode == PadMode_SAME) {
mPads.clear();
for (int i = 0; i < 3; ++i) {
int inputNeeded = output->length(i + 2) - 1 + mKernels[i];
mPads.push_back((inputNeeded - input->length(i + 2)) / 2);
}
}
mSourceBuffer.reset(Tensor::createDevice<float>({threadNumber, id, alpha2, UP_DIV(ic, 4), CONVOLUTION_TILED_NUMBER, 4}));
mDestBuffer.reset(Tensor::createDevice<float>({threadNumber, od + 1, alpha2, UP_DIV(oc, 4), CONVOLUTION_TILED_NUMBER, 4}));
mTempBuffer.reset(Tensor::createDevice<float>({threadNumber, 2, alpha2, 4}));
bool succ = backend()->onAcquireBuffer(mSourceBuffer.get(), Backend::DYNAMIC);
succ = succ && backend()->onAcquireBuffer(mDestBuffer.get(), Backend::DYNAMIC);
succ = succ && backend()->onAcquireBuffer(mTempBuffer.get(), Backend::DYNAMIC);
if (!succ) {
return OUT_OF_MEMORY;
}
backend()->onReleaseBuffer(mSourceBuffer.get(), Backend::DYNAMIC);
backend()->onReleaseBuffer(mDestBuffer.get(), Backend::DYNAMIC);
backend()->onReleaseBuffer(mTempBuffer.get(), Backend::DYNAMIC);
return NO_ERROR;
}
ErrorCode ConvolutionWinograd3D::onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
auto input = inputs[0];
auto output = outputs[0];
const int dstUnit = mUnit, srcUnit = mAlpha, srcUnit2 = srcUnit * srcUnit;
const int outputWidth = output->length(4), outputHeight = output->length(3), outputDepth = output->length(2);
const int inputWidth = input->length(4), inputHeight = input->length(3), inputDepth = input->length(2);
const int wUnit = UP_DIV(outputWidth, dstUnit), hUnit = UP_DIV(outputHeight, dstUnit);
const int ic_4 = UP_DIV(input->length(1), 4), dc_4 = UP_DIV(output->length(1), 4);
const int padY = mPads[1], padX = mPads[2], padDepth = mPads[0], kernelDepth = mKernels[0];
const int totalCount = wUnit * hUnit, tileCount = UP_DIV(totalCount, CONVOLUTION_TILED_NUMBER);
auto postFunction = mPostFunction;
const int threadNumber = std::max(((CPUBackend *)backend())->threadNumber(), 1);
auto sourceTransformFunc = [=](int xIndex, int xC, const float* srcOrigin, float* dstOrigin, float* midBuffer0, float* midBuffer1) {
int sourceZStep = inputDepth * inputWidth * inputHeight * 4;
int dstZStep = xC * 4;
int unitStep = ic_4 * xC * 4;
for (int xi = 0; xi < xC; ++xi) {
const int index = xIndex + xi, wIndex = index % wUnit, hIndex = index / wUnit;
const int srcX = wIndex * dstUnit - padX, srcY = hIndex * dstUnit - padY;
const int sx = ALIMAX(0, srcX) - srcX, ex = ALIMIN(srcX + srcUnit, inputWidth) - srcX;
const int sy = ALIMAX(0, srcY) - srcY, ey = ALIMIN(srcY + srcUnit, inputHeight) - srcY;
const int count = 4 * (ex - sx);
auto dst_x = dstOrigin + 4 * xi;
auto srcStart = srcOrigin + (srcX + srcY * inputWidth) * 4;
if (ey - sy < srcUnit) {
memset(midBuffer1, 0, srcUnit2 * 4 * sizeof(float));
}
if (ex - sx == srcUnit) {
for (int z = 0; z < ic_4; ++z) {
auto srcZ = srcStart + z * sourceZStep;
auto dstZ = dst_x + z * dstZStep;
for (int d = 0; d < inputDepth; ++d) {
auto src_depth = srcZ + d * inputWidth * inputHeight * 4;
auto dst_depth = dstZ + d * srcUnit2 * ic_4 * xC * 4;
// Transform
for (int i = sy; i < ey; ++i) {
mSourceTransform(src_depth + 4 * i * inputWidth, midBuffer1 + 4 * i, 4, 4 * srcUnit);
}
for (int i = 0; i < srcUnit; ++i) {
mSourceTransform(midBuffer1 + 4 * i * srcUnit, dst_depth + i * unitStep, 4,
unitStep * srcUnit);
}
}
}
} else {
memset(midBuffer0, 0, srcUnit2 * 4 * sizeof(float));
for (int z = 0; z < ic_4; ++z) {
// Extract
auto srcZ = srcStart + z * sourceZStep;
auto dstZ = dst_x + z * dstZStep;
for (int d = 0; d < inputDepth; ++d) {
auto src_depth = srcZ + d * inputWidth * inputHeight * 4;
auto dst_depth = dstZ + d * srcUnit2 * ic_4 * xC * 4;
if (count > 0) {
for (int yy = sy; yy < ey; ++yy) {
auto dst_yy = midBuffer0 + yy * srcUnit * 4 + sx * 4;
auto src_yy = src_depth + 4 * inputWidth * yy + sx * 4;
memcpy(dst_yy, src_yy, count * sizeof(float));
}
}
// Transform
for (int i = sy; i < ey; ++i) {
mSourceTransform(midBuffer0 + 4 * i * srcUnit, midBuffer1 + 4 * i, 4, 4 * srcUnit);
}
for (int i = 0; i < srcUnit; ++i) {
mSourceTransform(midBuffer1 + 4 * i * srcUnit, dst_depth + i * unitStep, 4,
unitStep * srcUnit);
}
}
}
}
}
};
auto destTransformFunc = [=](int xIndex, int xC, const float* srcOrigin, float* dstOrigin, float* midBuffer0, float* midBuffer1) {
int dstZStep = outputDepth * outputHeight * outputWidth * 4;
int srcZStep = xC * 4;
int unitStep = dc_4 * xC * 4;
for (int xi = 0; xi < xC; ++xi) {
const int index = xIndex + xi, wIndex = index % wUnit, hIndex = index / wUnit;
auto srcXi = srcOrigin + 4 * xi;
const int dstX = wIndex * dstUnit, dstY = hIndex * dstUnit;
auto dstStart = dstOrigin + 4 * (dstX + dstY * outputWidth);
const int ey = ALIMIN(dstY + dstUnit, outputHeight) - dstY;
const int ex = ALIMIN(dstX + dstUnit, outputWidth) - dstX;
const int count = ex * 4;
if (ex == dstUnit) {
for (int z = 0; z < dc_4; ++z) {
auto dstZAddr = dstStart + z * dstZStep;
auto srcZ = srcXi + z * srcZStep;
for (int d = 0; d < outputDepth; ++d) {
auto dst_depth = dstZAddr + d * outputHeight * outputWidth * 4;
auto src_depth = srcZ + d * srcUnit2 * dc_4 * xC * 4;
for (int i = 0; i < srcUnit; ++i) {
mDestTransform(src_depth + i * unitStep, midBuffer0 + i * dstUnit * 4,
srcUnit * unitStep, 4);
}
for (int i = 0; i < ey; ++i) {
auto dstAddr = dst_depth + i * 4 * outputWidth;
mDestTransform(midBuffer0 + i * 4, dstAddr, 4 * dstUnit, 4);
}
}
}
} else {
for (int z = 0; z < dc_4; ++z) {
auto dstZAddr = dstStart + z * dstZStep;
auto srcZ = srcXi + z * srcZStep;
for (int d = 0; d < outputDepth; ++d) {
auto dst_depth = dstZAddr + d * outputHeight * outputWidth * 4;
auto src_depth = srcZ + d * srcUnit2 * dc_4 * xC * 4;
for (int i = 0; i < srcUnit; ++i) {
mDestTransform(src_depth + i * unitStep, midBuffer0 + i * dstUnit * 4,
srcUnit * unitStep, 4);
}
for (int i = 0; i < ey; ++i) {
mDestTransform(midBuffer0 + i * 4, midBuffer1 + i * dstUnit * 4, 4 * dstUnit, 4);
}
for (int yy = 0; yy < ey; ++yy) {
auto dstYAddr = dst_depth + yy * 4 * outputWidth;
auto srcYAddr = midBuffer1 + yy * 4 * dstUnit;
memcpy(dstYAddr, srcYAddr, count * sizeof(float));
}
}
}
}
}
};
auto gemmFunc = [=](int xC, int start, int end, const float* srcOrigin, const float* weight, float* dstOrigin) {
float* tempDst = dstOrigin + outputDepth * srcUnit2 * dc_4 * xC * 4;
const int element = (end - start) * dc_4 * xC * 4, offset = start * dc_4 * xC * 4;
for (int od = 0; od < outputDepth; ++od) {
bool add = false;
float* _dstOrigin = dstOrigin + (od * srcUnit2 + start) * dc_4 * xC * 4;
const int srcD = od - padDepth, kdStart = -ALIMIN(srcD, 0), kdEnd = kernelDepth - ALIMAX(srcD + kernelDepth - inputDepth, 0);
for (int kd = kdStart; kd < kdEnd; ++kd) {
const float* _srcOrigin = srcOrigin + (kd + srcD) * srcUnit2 * ic_4 * xC * 4;
const float* _weight = weight + kd * srcUnit2 * dc_4 * ic_4 * 16;
for (int i = start; i < end; ++i) {
if (xC == CONVOLUTION_TILED_NUMBER) {
MNNGemmFloatUnit(tempDst + i * dc_4 * xC * 4, _srcOrigin + i * ic_4 * 4 * xC,
_weight + i * 16 * ic_4 * dc_4, ic_4, xC * 4, dc_4, 0);
} else {
MNNGemmFloatCommon_4(tempDst + i * dc_4 * xC * 4, _srcOrigin + i * ic_4 * 4 * xC,
_weight + (i * dc_4) * ic_4 * 16, ic_4, xC * 4, dc_4, xC, 0);
}
}
if (add) {
MNNMatrixAdd(_dstOrigin, _dstOrigin, tempDst + offset, element / 4, 0, 0, 0, 1);
} else {
memcpy(_dstOrigin, tempDst + offset, element * sizeof(float));
}
add = true;
}
}
};
auto gemmConcurrencyFunc = [=, &gemmFunc](int xC, const float* _srcOrigin, const float* weight, float* _dstOrigin) {
MNN_CONCURRENCY_BEGIN(tId, threadNumber) {
const int step = UP_DIV(srcUnit2, threadNumber);
gemmFunc(xC, tId * step, ALIMIN((tId + 1) * step, srcUnit2), _srcOrigin, weight, _dstOrigin);
}
MNN_CONCURRENCY_END()
};
auto tFunction = [&](const int tId, const int tileStart, const int tileStep, const int tileEnd, const float* srcOrigin, float* dstOrigin) {
auto _srcOrigin = mSourceBuffer->host<float>() + tId * mSourceBuffer->stride(0);
auto _dstOrigin = mDestBuffer->host<float>() + tId * mDestBuffer->stride(0);
auto midBuffer0 = mTempBuffer->host<float>() + tId * mTempBuffer->stride(0);
auto midBuffer1 = midBuffer0 + mTempBuffer->stride(1);
for (int tIndex = (int)tId; tIndex < tileCount; tIndex += threadNumber) {
int xIndex = (int)tIndex * CONVOLUTION_TILED_NUMBER;
int xReamin = totalCount - xIndex;
int xC = xReamin > CONVOLUTION_TILED_NUMBER ? CONVOLUTION_TILED_NUMBER : xReamin;
sourceTransformFunc(xIndex, xC, srcOrigin, _srcOrigin, midBuffer0, midBuffer1);
if (threadNumber != tileStep) {
gemmConcurrencyFunc(xC, _srcOrigin, mWeight->host<float>(), _dstOrigin);
} else {
gemmFunc(xC, 0, srcUnit2, _srcOrigin, mWeight->host<float>(), _dstOrigin);
}
destTransformFunc(xIndex, xC, _dstOrigin, dstOrigin, midBuffer0, midBuffer1);
}
};
for (int batchIndex = 0; batchIndex < input->batch(); ++batchIndex) {
auto srcOrigin = input->host<float>() + batchIndex * input->stride(0);
auto dstOrigin = output->host<float>() + batchIndex * output->stride(0);
if (tileCount >= threadNumber) {
MNN_CONCURRENCY_BEGIN(tId, threadNumber) {
tFunction((int)tId, (int)tId, threadNumber, tileCount / threadNumber * threadNumber, srcOrigin, dstOrigin);
}
MNN_CONCURRENCY_END();
}
if (tileCount % threadNumber != 0) {
tFunction(0, tileCount / threadNumber * threadNumber, 1, tileCount, srcOrigin, dstOrigin);
}
MNN_CONCURRENCY_BEGIN(tId, threadNumber) {
int channelStep = UP_DIV(dc_4, threadNumber);
int channelStart = channelStep * tId, channelNum = ALIMIN(channelStep * (tId + 1), dc_4) - channelStart;
if (channelNum > 0) {
postFunction(dstOrigin + channelStart * outputHeight * outputWidth * outputDepth * 4, mBias->host<float>() + 4 * channelStart, outputWidth * outputHeight * outputDepth, channelNum);
}
}
MNN_CONCURRENCY_END();
}
return NO_ERROR;
}
int ConvolutionWinograd3D::bestWinogradUnit(const Convolution3DCommon *common, const Tensor *inputTensor,
const Tensor *outputTensor, int threadNumber) {
const int ow = outputTensor->length(4), oh = outputTensor->length(3), oc = outputTensor->length(1);
int unit2 = UP_DIV(ow * oh, CONVOLUTION_TILED_NUMBER * threadNumber);
int maxUnit = (int)::sqrtf((float)unit2);
maxUnit = std::min(maxUnit, CONVOLUTION_WINOGRAD_MAX_UNIT);
maxUnit = std::max(maxUnit, CONVOLUTION_WINOGRAD_MIN_UNIT);
int ic = inputTensor->channel();
auto kernelSize = (*common->kernels())[1];
int unit = CONVOLUTION_WINOGRAD_MIN_UNIT;
float maxRate = 0.0f;
float originCost = (float)ow * oh * (float)ic * oc * kernelSize * kernelSize;
static std::set<int> supportSu{4, 8};
for (int u = CONVOLUTION_WINOGRAD_MIN_UNIT; u <= maxUnit; ++u) {
float su = (float)(u + kernelSize - 1);
if (supportSu.find(su) == supportSu.end()) {
continue;
}
if (nullptr == WinogradFunction::chooseDestTransform((int)su, u)) {
continue;
}
/*Let F(6,3) be choosed when it can speed up from F(2,3) than 0.6*/
float penalty = (su * su) / (float)(kernelSize * kernelSize) * 0.12f;
float winogradCost =
(2 * su * su * su * ic + su * su * ic * oc + 2 * su * u * u * oc) * (UP_DIV(ow, u) * UP_DIV(oh, u));
float reduceRate = originCost / winogradCost - penalty;
// MNN_PRINT("ow=%d, oh=%d, %f, %f, winograd unit:%d\n", ow, oh, winogradCost, reduceRate, u);
if (reduceRate > maxRate) {
maxRate = reduceRate;
unit = u;
}
}
if (maxRate < 1.0f) {
return 0;
}
return unit;
}
bool ConvolutionWinograd3D::canUseWinograd(const Convolution3DCommon *common) {
std::vector<int> kernels;
for (int kernel: *(common->kernels())) {
if (kernel <= 1) {
return false;
}
kernels.push_back(kernel);
}
if (kernels[1] != kernels[2]) {
return false;
}
for (int dialate: *(common->dilates())) {
if (dialate != 1) {
return false;
}
}
for (int stride: *(common->strides())) {
if (stride != 1) {
return false;
}
}
return true;
}
} // namespace MNN
Reference in New Issue View Git Blame Copy Permalink