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/Convolution3D3x3.cpp

433 lines
19 KiB
C++
Raw Blame History

//
// Convolution3D3x3.cpp
// MNN
//
// Created by MNN on 2019/09/18.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "../CPUBackend.hpp"
#include "backend/cpu/compute/Convolution3D3x3.hpp"
#include <MNN/AutoTime.hpp>
#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/Vec4.hpp"
using namespace MNN::Math;
typedef Vec4 float4;
#define SOURCE_BLOCK 64
#define WEIGHT_BLOCK 256
#define SOURCE_BLOCK_VEC 16
#define SRC_BLOCK_UNIT 3
#define SRC_BLOCK_UNIT2 9
#define BLOCK_UNIT 4
#define BLOCK_UNIT2 16
#define SOURCE_BLOCK 64
#define BLOCK_UNIT2 16
namespace MNN {
static void sourceTransform(const float* srcBlock, float* dstStart, size_t step) {
auto _x = (float*)srcBlock;
float4 m00;
float4 m01;
float4 m02;
float4 m03;
float4 m10;
float4 m11;
float4 m12;
float4 m13;
float4 m20;
float4 m21;
float4 m22;
float4 m23;
float4 m30;
float4 m31;
float4 m32;
float4 m33;
auto _y = dstStart;
m00 = Vec4::load(_x + 4 * 0) - Vec4::load(_x + 4 * 8);
m01 = Vec4::load(_x + 4 * 1) - Vec4::load(_x + 4 * 9);
m02 = Vec4::load(_x + 4 * 2) - Vec4::load(_x + 4 * 10);
m03 = Vec4::load(_x + 4 * 3) - Vec4::load(_x + 4 * 11);
m10 = Vec4::load(_x + 4 * 4) + Vec4::load(_x + 4 * 8);
m11 = Vec4::load(_x + 4 * 5) + Vec4::load(_x + 4 * 9);
m12 = Vec4::load(_x + 4 * 6) + Vec4::load(_x + 4 * 10);
m13 = Vec4::load(_x + 4 * 7) + Vec4::load(_x + 4 * 11);
m20 = Vec4::load(_x + 4 * 8) - Vec4::load(_x + 4 * 4);
m21 = Vec4::load(_x + 4 * 9) - Vec4::load(_x + 4 * 5);
m22 = Vec4::load(_x + 4 * 10) - Vec4::load(_x + 4 * 6);
m23 = Vec4::load(_x + 4 * 11) - Vec4::load(_x + 4 * 7);
m30 = Vec4::load(_x + 4 * 12) - Vec4::load(_x + 4 * 4);
m31 = Vec4::load(_x + 4 * 13) - Vec4::load(_x + 4 * 5);
m32 = Vec4::load(_x + 4 * 14) - Vec4::load(_x + 4 * 6);
m33 = Vec4::load(_x + 4 * 15) - Vec4::load(_x + 4 * 7);
Vec4::save(_y + step * 0, m00 - m02);
Vec4::save(_y + step * 1, m01 + m02);
Vec4::save(_y + step * 2, m02 - m01);
Vec4::save(_y + step * 3, m03 - m01);
Vec4::save(_y + step * 4, m10 - m12);
Vec4::save(_y + step * 5, m11 + m12);
Vec4::save(_y + step * 6, m12 - m11);
Vec4::save(_y + step * 7, m13 - m11);
Vec4::save(_y + step * 8, m20 - m22);
Vec4::save(_y + step * 9, m21 + m22);
Vec4::save(_y + step * 10, m22 - m21);
Vec4::save(_y + step * 11, m23 - m21);
Vec4::save(_y + step * 12, m30 - m32);
Vec4::save(_y + step * 13, m31 + m32);
Vec4::save(_y + step * 14, m32 - m31);
Vec4::save(_y + step * 15, m33 - m31);
}
static void destTransform(const float* srcZ, float* dstBlock, size_t step) {
auto yy = dstBlock;
float4 m00;
float4 m01;
float4 m02;
float4 m03;
float4 m10;
float4 m11;
float4 m12;
float4 m13;
auto x = srcZ;
m00 = Vec4::load(x + step * 0) + Vec4::load(x + step * 4) + Vec4::load(x + step * 8);
m01 = Vec4::load(x + step * 1) + Vec4::load(x + step * 5) + Vec4::load(x + step * 9);
m02 = Vec4::load(x + step * 2) + Vec4::load(x + step * 6) + Vec4::load(x + step * 10);
m03 = Vec4::load(x + step * 3) + Vec4::load(x + step * 7) + Vec4::load(x + step * 11);
m10 = Vec4::load(x + step * 4) - Vec4::load(x + step * 8) + Vec4::load(x + step * 12);
m11 = Vec4::load(x + step * 5) - Vec4::load(x + step * 9) + Vec4::load(x + step * 13);
m12 = Vec4::load(x + step * 6) - Vec4::load(x + step * 10) + Vec4::load(x + step * 14);
m13 = Vec4::load(x + step * 7) - Vec4::load(x + step * 11) + Vec4::load(x + step * 15);
Vec4::save(yy + 4 * 0, m00 + m01 + m02);
Vec4::save(yy + 4 * 1, m01 - m02 + m03);
Vec4::save(yy + 4 * 2, m10 + m11 + m12);
Vec4::save(yy + 4 * 3, m11 - m12 + m13);
}
static void kernelTransform(float* reorderedWeight, const float* srcWeight, int srcCount, int outputCount) {
float weight[BLOCK_UNIT2];
int srcDepthD4 = UP_DIV((int)srcCount, 4);
int dstDepthD4 = UP_DIV((int)outputCount, 4);
for (int dz = 0; dz < outputCount; ++dz) {
auto dz_4 = dz / BLOCK_UNIT;
auto mx = dz % BLOCK_UNIT;
auto dst_dz = reorderedWeight + dz_4 * srcDepthD4 * 16;
for (int sz = 0; sz < srcCount; ++sz) {
auto sz_4 = sz / BLOCK_UNIT;
auto my = sz % BLOCK_UNIT;
auto dst_sz = dst_dz + sz_4 * BLOCK_UNIT2;
auto src = srcWeight + SRC_BLOCK_UNIT2 * (sz + dz * srcCount);
auto dst = weight;
float* k = (float*)src;
float m00;
float m01;
float m02;
float m10;
float m11;
float m12;
float m20;
float m21;
float m22;
float m30;
float m31;
float m32;
m00 = k[0];
m01 = k[1];
m02 = k[2];
m10 = 0.500000 * k[0] + 0.500000 * k[3] + 0.500000 * k[6];
m11 = 0.500000 * k[1] + 0.500000 * k[4] + 0.500000 * k[7];
m12 = 0.500000 * k[2] + 0.500000 * k[5] + 0.500000 * k[8];
m20 = 0.500000 * k[0] + -0.500000 * k[3] + 0.500000 * k[6];
m21 = 0.500000 * k[1] + -0.500000 * k[4] + 0.500000 * k[7];
m22 = 0.500000 * k[2] + -0.500000 * k[5] + 0.500000 * k[8];
m30 = 0 + k[6];
m31 = 0 + k[7];
m32 = 0 + k[8];
k = dst;
k[0] = m00;
k[1] = 0.500000 * m00 + 0.500000 * m01 + 0.500000 * m02;
k[2] = 0.500000 * m00 + -0.500000 * m01 + 0.500000 * m02;
k[3] = 0 + m02;
k[4] = m10;
k[5] = 0.500000 * m10 + 0.500000 * m11 + 0.500000 * m12;
k[6] = 0.500000 * m10 + -0.500000 * m11 + 0.500000 * m12;
k[7] = 0 + m12;
k[8] = m20;
k[9] = 0.500000 * m20 + 0.500000 * m21 + 0.500000 * m22;
k[10] = 0.500000 * m20 + -0.500000 * m21 + 0.500000 * m22;
k[11] = 0 + m22;
k[12] = m30;
k[13] = 0.500000 * m30 + 0.500000 * m31 + 0.500000 * m32;
k[14] = 0.500000 * m30 + -0.500000 * m31 + 0.500000 * m32;
k[15] = 0 + m32;
for (int ki = 0; ki < BLOCK_UNIT2; ++ki) {
auto dst_i = dst_sz + ki * srcDepthD4 * dstDepthD4 * 16;
dst_i[4 * my + mx] = weight[ki];
}
}
}
}
Convolution3D3x3::Convolution3D3x3(const Convolution3DCommon* convOp, Backend *b, const float* originWeight,
int originWeightSize, const float* bias, int biasSize) : Execution(b) {
mPadMode = convOp->padMode();
if (mPadMode != PadMode_SAME) {
for (int32_t pad: *(convOp->pads())) {
mPads.push_back(pad);
}
}
mKernelDepth = (*(convOp->kernels()))[0];
mPostFunction = CPUConvolution3D::getPostFunction(convOp);
int inputChannel = convOp->inputCount();
int outputChannel = convOp->outputCount();
mWeight.reset(Tensor::createDevice<float>({ALIGN_UP4(inputChannel) * ALIGN_UP4(outputChannel) * mKernelDepth * BLOCK_UNIT2}));
mBias.reset(Tensor::createDevice<float>({ALIGN_UP4((int)biasSize)}));
bool valid = backend()->onAcquireBuffer(mWeight.get(), Backend::STATIC);
valid = valid && backend()->onAcquireBuffer(mBias.get(), Backend::STATIC);
if (!valid) {
return;
}
memset(mBias->host<float>(), 0, mBias->size());
memcpy(mBias->host<float>(), bias, biasSize * sizeof(float));
if (inputChannel % 4 != 0 || outputChannel % 4 != 0) {
memset(mWeight->host<float>(), 0, mWeight->size());
}
const int srcDepthStep = inputChannel * outputChannel * 9;
const int dstDepthStep = ALIGN_UP4(inputChannel) * ALIGN_UP4(outputChannel) * BLOCK_UNIT2;
for (int d = 0; d < mKernelDepth; ++d) {
kernelTransform(mWeight->host<float>() + d * dstDepthStep, originWeight + d * srcDepthStep, inputChannel, outputChannel);
}
}
Convolution3D3x3::~Convolution3D3x3() {
MNN_ASSERT(nullptr != mWeight);
MNN_ASSERT(nullptr != mBias);
if (nullptr != mBias) {
backend()->onReleaseBuffer(mBias.get(), Backend::STATIC);
}
if (nullptr != mWeight) {
backend()->onReleaseBuffer(mWeight.get(), Backend::STATIC);
}
}
ErrorCode Convolution3D3x3::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();
if (mPadMode == PadMode_SAME) {
mPads.clear();
auto kernels = std::vector<int32_t>({mKernelDepth, 3, 3});
for (int i = 0; i < 3; ++i) {
int inputNeeded = output->length(i + 2) - 1 + kernels[i]; // stride = dialate = 1
mPads.push_back((inputNeeded - input->length(i + 2)) / 2);
}
}
auto CONVOLUTION_TILED_NUMBER = MNNGetConvolutionTileNumber();
mSourceBuffer.reset(Tensor::createDevice<float>({threadNumber, id, BLOCK_UNIT2, UP_DIV(ic, 4), CONVOLUTION_TILED_NUMBER, 4}));
mDestBuffer.reset(Tensor::createDevice<float>({threadNumber, od + 1, BLOCK_UNIT2, UP_DIV(oc, 4), CONVOLUTION_TILED_NUMBER, 4}));
mTempBuffer.reset(Tensor::createDevice<float>({threadNumber, BLOCK_UNIT2, 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 Convolution3D3x3::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
AUTOTIME;
auto input = inputs[0];
auto output = outputs[0];
auto CONVOLUTION_TILED_NUMBER = MNNGetConvolutionTileNumber();
const int inputWidth = input->length(4), inputHeight = input->length(3), inputDepth = input->length(2), ic_4 = UP_DIV(input->length(1), 4);
const int outputWidth = output->length(4), outputHeight = output->length(3), outputDepth = output->length(2), dc_4 = UP_DIV(output->length(1), 4);
const int padDepth = mPads[0], padHeight = mPads[1], padWidth = mPads[2], kernelDepth = mKernelDepth;
const int wUnit = UP_DIV(outputWidth, 2), hUnit = UP_DIV(outputHeight, 2), totalCount = wUnit * hUnit;
const int tileCount = UP_DIV(totalCount, CONVOLUTION_TILED_NUMBER);
auto postFunction = mPostFunction;
// MNN_PRINT("outputWidth=%d, outputHeight=%d\n", outputWidth, outputHeight);
const int threadNumber = ((CPUBackend*)backend())->threadNumber();
auto sourceTransformFunc = [=](int xIndex, int xC, const float* srcOrigin, float* _srcOrigin, float* dstBlock) {
const int dstStepD = BLOCK_UNIT2 * ic_4 * xC * 4;
// Source Transform
for (int xi = 0; xi < xC; ++xi) {
auto index = xIndex + xi;
auto dstUnit = _srcOrigin + 4 * xi;
int wIndex = index % wUnit;
int hIndex = index / wUnit;
int srcX = wIndex * 2 - padWidth;
int srcY = hIndex * 2 - padHeight;
int sy = ALIMAX(0, srcY) - srcY;
int ey = ALIMIN(srcY + 4, inputHeight) - srcY;
int sx = ALIMAX(0, srcX) - srcX;
int ex = ALIMIN(srcX + 4, inputWidth) - srcX;
auto srcStart = srcOrigin + (srcX + srcY * inputWidth) * 4;
memset(dstBlock, 0, SOURCE_BLOCK * sizeof(float));
for (int z = 0; z < ic_4; ++z) {
auto dstStart = dstUnit + z * 4 * xC;
auto src_z = srcStart + z * 4 * inputWidth * inputHeight * inputDepth;
for (int d = 0; d < inputDepth; ++d) {
if (ex > sx) {
// Extract One Block
for (int yy = sy; yy < ey; ++yy) {
auto dst_yy = dstBlock + yy * 16;
auto src_yy = src_z + (d * inputHeight + yy) * inputWidth * 4;
memcpy(dst_yy + 4 * sx, src_yy + sx * 4, 4 * (ex - sx) * sizeof(float));
}
}
// Transform
sourceTransform(dstBlock, dstStart + d * dstStepD, 4 * xC * ic_4);
}
}
}
};
auto destTransformFunc = [=](int xIndex, int xC, const float* srcOrigin, float* dstOrigin, float* dstBlock) {
for (int xi = 0; xi < xC; ++xi) {
auto index = xIndex + xi;
auto srcUnit = srcOrigin + 4 * xi;
int wIndex = index % wUnit;
int hIndex = index / wUnit;
int dstX = wIndex * 2;
int dstY = hIndex * 2;
auto dstStart = dstOrigin + 4 * (dstX + dstY * outputWidth);
for (int od = 0; od < outputDepth; ++od) {
auto _srcUnit = srcUnit + od * BLOCK_UNIT2 * dc_4 * xC * 4;
auto _dstStart = dstStart + od * outputHeight * outputWidth * 4;
for (int z = 0; z < dc_4; ++z) {
auto srcZ = _srcUnit + z * xC * 4;
auto dstZ = _dstStart + z * outputDepth * outputWidth * outputHeight * 4;
destTransform(srcZ, dstBlock, dc_4 * 4 * xC);
Vec4::save(dstZ, Vec4::load(dstBlock));
if (wIndex * 2 + 1 < outputWidth) {
Vec4::save(dstZ + 4, Vec4::load(dstBlock + 4));
}
if (hIndex * 2 + 1 < outputHeight) {
Vec4::save(dstZ + outputWidth * 4, Vec4::load(dstBlock + 8));
if (wIndex * 2 + 1 < outputWidth) {
Vec4::save(dstZ + outputWidth * 4 + 4, Vec4::load(dstBlock + 12));
}
}
}
}
}
};
auto gemmFunc = [=](int xC, int start, int end, const float* srcOrigin, const float* weight, float* dstOrigin) {
float* tempDst = dstOrigin + outputDepth * BLOCK_UNIT2 * 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 * BLOCK_UNIT2 + 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) * BLOCK_UNIT2 * ic_4 * xC * 4;
const float* _weight = weight + kd * BLOCK_UNIT2 * dc_4 * ic_4 * 16;
for (int i = start; i < end; ++i) {
if (xC == CONVOLUTION_TILED_NUMBER) {
MNNGemmFloatUnit_4(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(BLOCK_UNIT2, threadNumber);
gemmFunc(xC, tId * step, ALIMIN((tId + 1) * step, BLOCK_UNIT2), _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 dstBlock = mTempBuffer->host<float>() + tId * mTempBuffer->stride(0);
for (int tIndex = tileStart; tIndex < tileEnd; tIndex += tileStep) {
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, dstBlock);
if (threadNumber != tileStep) {
gemmConcurrencyFunc(xC, _srcOrigin, mWeight->host<float>(), _dstOrigin);
} else {
gemmFunc(xC, 0, BLOCK_UNIT2, _srcOrigin, mWeight->host<float>(), _dstOrigin);
}
destTransformFunc(xIndex, xC, _dstOrigin, dstOrigin, dstBlock);
}
};
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;
}
} // namespace MNN
Reference in New Issue View Git Blame Copy Permalink