MNN/source/backend/cpu/compute/ConvInt8TiledExecutor.cpp

1380 lines
62 KiB
C++

// ConvInt8TiledExecutor.cpp
// MNN
//
// Created by MNN on 2019/5/17.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "ConvInt8TiledExecutor.hpp"
#include "ConvolutionTiledExecutor.hpp"
#include "core/Macro.h"
#include "core/BufferAllocator.hpp"
#include <math.h>
#include "backend/cpu/CPUBackend.hpp"
#include "core/Concurrency.h"
#include "core/TensorUtils.hpp"
#define QUANT_INFO_BYTES 4
namespace MNN {
ConvInt8TiledExecutor::ConvInt8TiledExecutor(Backend* backend, const Op* op): CPUConvolution(op->main_as_Convolution2D()->common(), backend) {}
ConvInt8TiledExecutor::ConvInt8TiledExecutor(Backend* backend, const Op* op, std::shared_ptr<ResourceInt8> res): CPUConvolution(op->main_as_Convolution2D()->common(), backend), mResourceInt8(res) {
if (!res->mDynamicQuant) {
mMutableResource.reset(new MutableResourceInt8(res, backend));
mValid = mMutableResource->mValid;
}
}
ConvInt8TiledExecutor::~ConvInt8TiledExecutor() {
// Do nothing
}
bool ConvInt8TiledExecutor::onClone(Backend* bn, const Op* op, Execution** dst) {
return false;
}
ErrorCode ConvInt8TiledExecutor::onResize(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
if (nullptr != mMutableResource) {
mMutableResource->updateInputOutputScale(TensorUtils::getQuantInfo(inputs[0]), TensorUtils::getQuantInfo(outputs[0]));
}
CPUConvolution::onResize(inputs, outputs);
ConvolutionTiledExecutor::setIm2ColParameter(mIm2ColParamter, mCommon, inputs[0], outputs[0], mPadX, mPadY, static_cast<CPUBackend*>(backend())->functions(), static_cast<CPUBackend*>(backend())->int8Functions());
return NO_ERROR;
}
void ConvInt8TiledExecutor::reorderWeight(uint8_t* dst, const uint8_t* src, int32_t* info, int32_t initval) {
// weight shape = {blockNum, UP_DIV(oc, UNIT), UP_DIV(ic, SRC_UNIT) * kernelCount / blockNum, UNIT, SRC_UNIT};
MNN_ASSERT(dst != nullptr && src != nullptr);
int blockNum = info[0];
int oc = info[1];
int ic = info[2];
int kernelCount = info[3];
int UNIT = info[4];
int SRC_UNIT = info[5];
int blockL = UP_DIV(ic / blockNum, SRC_UNIT) * kernelCount;
int stride0 = ROUND_UP(oc, UNIT) * SRC_UNIT * blockL; // weight->stride(0)
int stride1 = blockL * SRC_UNIT * UNIT; // weight->stride(1)
int stride2 = UNIT * SRC_UNIT; // weight->stride(2)
int weightlen = stride0 * blockNum;
memset(dst, initval, weightlen);
for (int k = 0; k < kernelCount; ++k) {
auto icDivU = UP_DIV(ic, SRC_UNIT);
const auto srcK = src + k;
for (int y = 0; y < ic; ++y) {
const int yOutSide = y / SRC_UNIT;
const int yInSide = y % SRC_UNIT;
int blockId = (yOutSide + k * icDivU) / blockL;
int blockInsideId = (yOutSide + k * icDivU) % blockL;
auto dstY = dst + blockId * stride0 + blockInsideId * stride2 + yInSide;
const auto srcY = srcK + y * kernelCount;
for (int x = 0; x < oc; ++x) {
const int xOutSide = x / UNIT;
const int xInSide = x % UNIT;
const int dstIndex = xOutSide * stride1 + xInSide * SRC_UNIT;
const int srcIndex = x * kernelCount * ic;
dstY[dstIndex] = srcY[srcIndex];
}
}
}
}
void ConvInt8TiledExecutor::packWeightAndQuantInfo(int8_t* dstbuffer, const int8_t* weight, const int8_t* quantInfo, int32_t* info, int infoBytes) {
int blockNum = info[0];
int ocDiv = info[1];
int blockL = info[2];
int UNIT = info[3];
int SRC_UNIT = info[4];
auto ocUp4 = info[5];
auto src0 = weight; // int8 weight: [blocknum, oc/hp, ic/lp*(kx*ky)/blocknum, hp, lp]
auto src1 = quantInfo; // dequant scale
auto src2 = src1 + infoBytes * ocUp4 * blockNum; // dequant bias
int stride0 = info[1] * info[2] * info[3] * info[4];
int stride1 = info[2] * info[3] * info[4];
for (int bl = 0; bl < blockNum; ++bl) {
auto blockPtr = dstbuffer + bl * (stride0 + ocUp4 * 2 * infoBytes);
for (int hU = 0; hU < ocDiv; ++hU) {
int scaleCount = ALIMIN(ocUp4 - hU * UNIT, UNIT);
auto hUPtr = blockPtr + hU * (stride1 + 2 * UNIT * infoBytes);
memcpy(hUPtr, src0 + bl * stride0 + hU * stride1, stride1);
memcpy(hUPtr + stride1, src1 + (bl * ocUp4 + hU * UNIT) * infoBytes, scaleCount * infoBytes);
memcpy(hUPtr + stride1 + scaleCount * infoBytes, src2 + (bl * ocUp4 + hU * UNIT) * infoBytes, scaleCount * infoBytes);
}
}
}
static void GetResourceInt8(std::shared_ptr<CPUConvolution::ResourceInt8> resource, std::shared_ptr<ConvolutionCommon::Int8Common> quantCommon, const Convolution2D* conv2d, Backend* backend, AutoStorage<int8_t>& reorderedQuantInfo) {
// common parameters
int outputCount = conv2d->common()->outputCount();
auto core = static_cast<CPUBackend*>(backend)->functions();
int inputChannel = conv2d->common()->inputCount();
int kernelSize = conv2d->common()->kernelX() * conv2d->common()->kernelY();
int LSize = inputChannel * kernelSize;
int ocUp4 = ROUND_UP(outputCount, core->pack);
bool useCachedMmap = backend->getRuntime()->hint().useCachedMmap > 1;
int dequantCnt = quantCommon->alpha.size();
if (quantCommon->asymmetric) {
dequantCnt /= 2;
}
int blockNum = dequantCnt / outputCount;
int scaleSize = blockNum * ocUp4; // pack size.
int blockSize = LSize / blockNum;
int originOffset = 0;
resource->mActBits = 8;
if (quantCommon->canUseInt4) {
originOffset = -8;
resource->mActBits = 4;
}
resource->mBlockNum = blockNum;
// alloc memory
resource->mOriginBias.reset(Tensor::createDevice<int32_t>({ocUp4})); // float
auto success = backend->onAcquireBuffer(resource->mOriginBias.get(), Backend::STATIC);
resource->mWeightKernelSum.reset(Tensor::createDevice<uint8_t>({QUANT_INFO_BYTES * ocUp4}));
success = backend->onAcquireBuffer(resource->mWeightKernelSum.get(), Backend::STATIC);
if (!success) {
MNN_ERROR("Alloc memory error\n");
return;
}
if (useCachedMmap) {
return;
}
reorderedQuantInfo.reset(2 * scaleSize * QUANT_INFO_BYTES);
if (reorderedQuantInfo.get() == nullptr) {
MNN_ERROR("Memory not enough\n");
return;
}
// Save bias
::memset(resource->mOriginBias->host<float>(), 0, ocUp4 * sizeof(float));
if (conv2d->bias()) {
::memcpy(resource->mOriginBias->host<float>(), conv2d->bias()->data(), outputCount * sizeof(float));
} else {
::memset(resource->mOriginBias->host<float>(), 0, ocUp4 * sizeof(float));
}
// Save weight quant alpha and zero: wf=alpha*wi+zero
auto alphaPtr = reinterpret_cast<float*>(reorderedQuantInfo.get());
auto biasPtr = reinterpret_cast<float*>(reinterpret_cast<uint8_t*>(alphaPtr) + scaleSize * QUANT_INFO_BYTES);
if (outputCount % core->pack != 0) {
::memset(alphaPtr, 0, scaleSize * QUANT_INFO_BYTES);
::memset(biasPtr, 0, scaleSize * QUANT_INFO_BYTES);
}
auto quanInfoPtr = quantCommon->alpha.get();
int h = quantCommon->alpha.size();
if (quantCommon->asymmetric) {
for (int i = 0; i < blockNum; ++i) {
auto dstAlpha = alphaPtr + i * ocUp4;
auto dstBias = biasPtr + i * ocUp4;
for (int j = 0; j < outputCount; ++j) {
int scaleIndex = j * blockNum + i;
dstAlpha[j] = quanInfoPtr[2 * scaleIndex + 1];
dstBias[j] = quanInfoPtr[2 * scaleIndex] + (float)originOffset * dstAlpha[j];
}
}
} else {
for (int i = 0; i < blockNum; ++i) {
auto dstAlpha = alphaPtr + i * ocUp4;
auto dstBias = biasPtr + i * ocUp4;
for (int j = 0; j < outputCount; ++j) {
int scaleIndex = j * blockNum + i;
dstAlpha[j] = quanInfoPtr[scaleIndex];
dstBias[j] = (float)originOffset * dstAlpha[j];
}
}
}
// Save float weight kernel sum
auto weightKernelSum = resource->mWeightKernelSum->host<float>();
auto realWeightData = quantCommon->weight.get();
::memset(weightKernelSum, 0, resource->mWeightKernelSum->size());
for (int j = 0; j < outputCount; ++j) {
float sum = 0.f;
for (int k = 0; k < blockNum; ++k) {
int scaleIndex = k + j * blockNum;
float scale = 0;
float bias = 0;
if (quantCommon->asymmetric) {
scale = quanInfoPtr[2 * scaleIndex + 1];
bias = quanInfoPtr[2 * scaleIndex];
} else {
scale = quanInfoPtr[scaleIndex];
bias = 0;
}
int tmp = 0;
if (quantCommon->canUseInt4) {
for (int i = 0; i < blockSize; ++i) {
int l_index = k * blockSize + i;
int w_idx = (j * blockNum * blockSize + l_index);
int w_offset = w_idx / 2;
int w_mask = w_idx % 2;
uint8_t s = realWeightData[w_offset];
int val = w_idx % 2 ? s & 0x0f : s >> 4;
tmp += (val - 8);
}
} else {
for (int i = 0; i < blockSize; ++i) {
int l_index = k * blockSize + i;
tmp += (int)realWeightData[j * blockNum * blockSize + l_index];
}
}
sum += (tmp * scale + blockSize * bias);
}
weightKernelSum[j] = sum;
}
}
DenseConvInt8TiledExecutor::DenseConvInt8TiledExecutor(Backend* backend, const Op* op, std::shared_ptr<ConvolutionCommon::Int8Common> quanCommon) : ConvInt8TiledExecutor(backend, op) {
auto convOp = op->main_as_Convolution2D();
auto core = static_cast<CPUBackend*>(backend)->int8Functions();
auto gcore = static_cast<CPUBackend*>(backend)->functions();
mResourceInt8.reset(new CPUConvolution::ResourceInt8);
mResourceInt8->mDynamicQuant = true;
mResourceInt8->mWeightAsymmetricQuant = quanCommon->asymmetric;
AutoStorage<int8_t> reorderedQuantInfo;
GetResourceInt8(mResourceInt8, quanCommon, convOp, backend, reorderedQuantInfo);
int blockNum = mResourceInt8->mBlockNum;
// dynamic quant
int UNIT, SRC_UNIT, DST_XUNIT;
core->MNNGetGemmUnit(&UNIT, &SRC_UNIT, &DST_XUNIT);
int pack = gcore->pack;
auto weightLength = quanCommon->weight.size();
int kernelCount = mCommon->kernelX() * mCommon->kernelY();
int oc = convOp->common()->outputCount();
int ic = convOp->common()->inputCount();
bool directReadInt4weight = (kernelCount == 1 && ROUND_UP(oc, UNIT) == oc && ROUND_UP(ic, SRC_UNIT) == ic);
bool useCachedMmap = backend->getRuntime()->hint().useCachedMmap > 1;
#ifdef MNN_KLEIDIAI_ENABLED
if(quanCommon->canUseInt4) {
bool bFP16 = gcore->bytes == 2 ? true : false;
bool bAsym = quanCommon->asymmetric;
size_t blkSize = mBlockNum == 1 ? 0 : ic / mBlockNum;
KleidiAI::AccelType accelType = KleidiAI::getQIntAccelType(4, bAsym, blkSize);
KleidiAI& kai = KleidiAI::getInstance();
if(!kai.isLoaded(accelType)) {
kai.setLoaded(accelType);
kai.printInfo(accelType);
}
if(kai.canAccelerate(accelType)) {
mAccelType = accelType;
int n = oc;
int k = ic;
int packedWeightSize = kai.getRhsPackedSize(mAccelType, n, k, blkSize);
//Alloc packed weight tensor.
mResourceInt8->mWeightInt8.reset(Tensor::createDevice<uint8_t>({packedWeightSize}));
bool success = backend->onAcquireBuffer(mResourceInt8->mWeightInt8.get(), Backend::STATIC);
if (!success) {
MNN_ERROR("Out of static memory!\n");
return;
}
size_t paraNum = blockNum * ROUND_UP(oc, pack);
float *scalePtr = mResourceInt8->mOriginScale->host<float>();
float *zeroPtr = mResourceInt8->mOriginScale->host<float>() + paraNum;
float *biasPtr = mResourceInt8->mOriginBias->host<float>();
//Reload some parameters to fit ukernels' layout.
auto quanInfoPtr = quanCommon->alpha.get();
if(bAsym) {
for(int i = 0; i < paraNum; i++) {
zeroPtr[i] = quanInfoPtr[i * 2];
scalePtr[i] = quanInfoPtr[i * 2 + 1];
}
} else {
if(blkSize != 0) {
memcpy(scalePtr, (uint8_t*)quanInfoPtr, paraNum * sizeof(float));
}
}
//Run rhs pack.
auto weightPackedData = mResourceInt8->mWeightInt8->host<uint8_t>();
kai.runRhsPack(mAccelType, 1, n, k, blkSize, 0/*unused*/,
(uint8_t*)quanCommon->weight.get(),
(const void*)scalePtr, (const void*)zeroPtr, (const void*)biasPtr,
weightPackedData, directReadInt4weight);
return;
}
}
#endif
int lU = UP_DIV(ic / blockNum, SRC_UNIT) * kernelCount;
std::vector<int> shape = {blockNum, UP_DIV(oc, UNIT), lU, UNIT, SRC_UNIT};
if (quanCommon->canUseInt4) {
shape[4] = SRC_UNIT / 2;
}
auto quantlen = 2 * mResourceInt8->mBlockNum * ROUND_UP(oc, pack) * QUANT_INFO_BYTES;
auto weightlen = shape[0] * shape[1] * shape[2] * shape[3] * shape[4];
mResourceInt8->mWeightInt8.reset(Tensor::createDevice<uint8_t>({weightlen + quantlen}));
auto res = backend->onAcquireBuffer(mResourceInt8->mWeightInt8.get(), Backend::STATIC);
if (!res) {
MNN_ERROR("weight acquire buffer error\n");
return;
}
if (useCachedMmap) {
return;
}
AutoStorage<int8_t> weightReordered(weightlen);
if (weightReordered.get() == nullptr) {
MNN_ERROR("Memory not enough\n");
return;
}
/* 1. reorder weight */
if (quanCommon->canUseInt4 && directReadInt4weight) {
// int4 weight reorder
mResourceInt8->mWeightAsymmetricQuant = true;
int hU = UP_DIV(oc, UNIT);
int lU = UP_DIV(ic, SRC_UNIT);
int hP = UNIT;
int lP = SRC_UNIT;
auto srcPtr = (uint8_t*)quanCommon->weight.get();
auto dstPtr = (uint8_t*)weightReordered.get();
::memset(dstPtr, 0, weightlen);
// Pack two int4-weight to one int8-weight.
int cnt = lP * hP / 4;
int L = lU * lP;
int blockL = lU / blockNum;
int stride0 = (lP * hP) * hU * blockL;
int stride1 = (lP * hP) * blockL;
for (int i = 0; i < hU; ++i) {
for (int j = 0; j < lU; ++j) {
int blockId = j / blockL;
int blockkInsideId = j % blockL;
for (int k = 0; k < cnt; ++k) {
int dstIndx0 = (blockId * stride0 + i * stride1 + blockkInsideId * lP * hP) / 2 + (2 * k);
int hpId0 = (2 * k + 1) / lP;
int lpId0 = (2 * k) % lP;
int hpId1 = (2 * (k + cnt) + 1) / lP;
int lpId1 = (2 * (k + cnt)) % lP;
int srcIndx0 = ((i * hP + hpId0) * L + (j * lP + lpId0)) / 2;
int srcIndx1 = ((i * hP + hpId1) * L + (j * lP + lpId1)) / 2;
int s0 = (srcPtr[srcIndx0] >> 4);
int s1 = (srcPtr[srcIndx0] & 15);
int s2 = (srcPtr[srcIndx1] >> 4);
int s3 = (srcPtr[srcIndx1] & 15);
int d0 = s0 * 16 + s2;
int d1 = s1 * 16 + s3;
dstPtr[dstIndx0] = d0;
dstPtr[dstIndx0 + 1] = d1;
}
}
}
} else {
// std::shared_ptr<Tensor> srcWeight;
int blocksize = ic * kernelCount / blockNum;
int originOffset = 0;
int32_t info[6] = {blockNum, oc, ic, kernelCount, UNIT, SRC_UNIT};
if (quanCommon->canUseInt4) {
originOffset = -8;
mResourceInt8->mWeightAsymmetricQuant = true;
auto srcPtr = reinterpret_cast<uint8_t*>(quanCommon->weight.get());
std::vector<int8_t> tmpWeight(weightLength * 2, originOffset);
for (int j = 0; j < oc; ++j) {
for (int k = 0; k < blockNum; ++k) {
for (int i = 0; i < blocksize; ++i) {
int index = j * blockNum * blocksize + k * blocksize + i;
uint8_t w_ = srcPtr[index / 2];
int truew = index % 2 ? (w_ & 0x0f) : (w_ >> 4);
tmpWeight[index] = truew - 8;
}
}
}
AutoStorage<uint8_t> packedInt8weight(weightlen * 2);
if (packedInt8weight.get() == nullptr) {
MNN_ERROR("Weight reorder memory not enough!\n");
return;
}
reorderWeight(packedInt8weight.get(), (uint8_t*)tmpWeight.data(), info, originOffset);
// pack two int4 to int8
int leng = weightlen * 2;
auto srcint4Ptr = (int8_t*)packedInt8weight.get();
auto dstint4Ptr = (uint8_t*)weightReordered.get();
int permuteUnit = UNIT * SRC_UNIT;
int halfPermuteStride = static_cast<int32_t>(permuteUnit / 2);
for (int i = 0; i < leng / permuteUnit; ++i) {
auto src0 = srcint4Ptr + i * permuteUnit;
auto dst0 = dstint4Ptr + i * halfPermuteStride;
for (int j = 0; j < halfPermuteStride; ++j) {
int s0 = src0[j];
int s1 = src0[j + halfPermuteStride];
int d = (s0 + 8) * 16 + (s1 + 8);
dst0[j] = d;
}
}
} else {
reorderWeight((uint8_t*)weightReordered.get(), (uint8_t*)quanCommon->weight.get(), info, 0);
}
}
/* 2. put weight and quantInfo together */
int32_t params[6] = {shape[0], shape[1], shape[2], shape[3], shape[4], quantlen / (2 * QUANT_INFO_BYTES * blockNum)};
ConvInt8TiledExecutor::packWeightAndQuantInfo(mResourceInt8->mWeightInt8->host<int8_t>(), (int8_t*)weightReordered.get(), reorderedQuantInfo.get(), params, QUANT_INFO_BYTES);
// Relu/Relu6 post parameters
auto postPtr = getPostParameters();
mResourceInt8->mReluThreshold.resize(2);
mResourceInt8->mReluThreshold[0] = postPtr[2];
mResourceInt8->mReluThreshold[1] = postPtr[3];
if (gcore->bytes == 2) {
gcore->MNNFp32ToLowp(mResourceInt8->mReluThreshold.data(), reinterpret_cast<int16_t*>(mResourceInt8->mReluThreshold.data()), 2);
}
}
static void _computeAlphaScaleOfflineQuant(Backend* backend, const Convolution2D* conv2d, std::shared_ptr<CPUConvolution::ResourceInt8> resourceInt8) {
/* Used to compute weight quant scale and bias and weightKernelSum of type float. */
bool quanBuffer = (conv2d->quanParameter() != nullptr && conv2d->quanParameter()->buffer() != nullptr);
MNN_ASSERT(quanBuffer || resourceInt8);
auto core = static_cast<CPUBackend*>(backend)->functions();
// common parameters
int outputCount = conv2d->common()->outputCount();
int LSize = conv2d->common()->inputCount() * conv2d->common()->kernelX() * conv2d->common()->kernelY();
int ocUp4 = ROUND_UP(outputCount, core->pack);
int8_t* weightOrigin;
// Save weight quant scale and bias: wf=scale*wi+bias
std::shared_ptr<Tensor> scaleBias(Tensor::createDevice<uint8_t>({2 * ocUp4 * core->bytes}));
auto success = backend->onAcquireBuffer(scaleBias.get(), Backend::STATIC);
if (!success) {
MNN_ERROR("Alloc dequant scaleBias memory error\n");
return;
}
auto alphaPtr = scaleBias->host<float>();
auto biasPtr = reinterpret_cast<float*>(reinterpret_cast<uint8_t*>(alphaPtr) + ocUp4 * core->bytes);
::memset(alphaPtr, 0, 2 * ocUp4 * core->bytes);
// Load quant scale and bias
weightOrigin = resourceInt8->mWeightInt8->host<int8_t>();
auto wZero = resourceInt8->mWeightQuantZero->host<int32_t>(); // has packed to outputUp4
auto wScale = resourceInt8->mOriginScale->host<float>();
int h = ocUp4;
MNN_ASSERT(4 == core->bytes);
for (int i=0; i< h; ++i) {
alphaPtr[i] = wScale[i];
biasPtr[i] = (-1.f) * wZero[i] * wScale[i];
}
resourceInt8->mOriginScale = scaleBias;
// Compute float weightKernelSum
resourceInt8->mWeightKernelSum.reset(Tensor::createDevice<uint8_t>({ocUp4 * 4}));
success = backend->onAcquireBuffer(resourceInt8->mWeightKernelSum.get(), Backend::STATIC);
if (!success) {
MNN_ERROR("Alloc dequant mWeightKernelSum memory error\n");
return;
}
auto weightKernelSum = resourceInt8->mWeightKernelSum->host<float>();
for (int i = 0; i < outputCount; ++i) {
int sum = 0;
for (int j = 0; j < LSize; ++j) {
sum = sum + static_cast<int>(weightOrigin[j + i * LSize]);
}
auto scale = alphaPtr[i];
auto bias = biasPtr[i];
weightKernelSum[i] = static_cast<float>(sum) * scale + LSize * bias;
}
}
DenseConvInt8TiledExecutor::DenseConvInt8TiledExecutor(Backend* backend, const Op* op, std::shared_ptr<ResourceInt8> res) : ConvInt8TiledExecutor(backend, op, res) {
// offline quant
auto convOp = op->main_as_Convolution2D();
auto core = static_cast<CPUBackend*>(backend)->int8Functions();
auto gcore = static_cast<CPUBackend*>(backend)->functions();
int pack = gcore->pack;
int ic = convOp->common()->inputCount();
int oc = convOp->common()->outputCount();
int kernelCount = convOp->common()->kernelX() * convOp->common()->kernelY();
int UNIT, SRC_UNIT, DST_XUNIT;
core->MNNGetGemmUnit(&UNIT, &SRC_UNIT, &DST_XUNIT);
int lU = UP_DIV(ic, SRC_UNIT) * kernelCount;
std::vector<int> shape = {1, UP_DIV(oc, UNIT), lU, UNIT, SRC_UNIT};
int weightlen = shape[0] * shape[1] * shape[2] * shape[3] * shape[4];
AutoStorage<uint8_t> weightReordered(weightlen);
if (!weightReordered.get()) {
MNN_ERROR("Memory not enough for quant model weight reorder\n");
return;
}
int32_t info[6] = {1, oc, ic, kernelCount, UNIT, SRC_UNIT};
ConvInt8TiledExecutor::reorderWeight((uint8_t*)weightReordered.get(), mResourceInt8->mWeightInt8->host<uint8_t>(), info, 0);
_computeAlphaScaleOfflineQuant(backend, convOp, mResourceInt8);
auto quantlen = mResourceInt8->mOriginScale->size();
mResourceInt8->mWeightInt8.reset(Tensor::createDevice<uint8_t>({weightlen + quantlen}));
auto allocSuc = backend->onAcquireBuffer(mResourceInt8->mWeightInt8.get(), Backend::STATIC);
if (!allocSuc) {
MNN_ERROR("Buffer alloc error!\n");
return;
}
int32_t params[6] = {shape[0], shape[1], shape[2], shape[3], shape[4], quantlen/ (2 * QUANT_INFO_BYTES * 1)};
ConvInt8TiledExecutor::packWeightAndQuantInfo(mResourceInt8->mWeightInt8->host<int8_t>(), (int8_t*)weightReordered.get(), mResourceInt8->mOriginScale->host<int8_t>(), params, QUANT_INFO_BYTES);
mGemmKernel = core->Int8GemmKernel;
#ifdef MNN_USE_SSE
int actBits = convOp->symmetricQuan()->nbits();
if (actBits <= 7) {
mGemmKernel = core->Int8GemmKernelFast;
}
#else
if(convOp->symmetricQuan()->method() == QuantizeAlgo_OVERFLOW_AWARE){
mGemmKernel = core->Int8GemmKernelFast;
}
#endif
}
DenseConvInt8TiledExecutor::DenseConvInt8TiledExecutor(Backend* backend, const Op* op, const DenseConvInt8TiledExecutor& exe)
: ConvInt8TiledExecutor(backend, op, exe.mResourceInt8), mGemmKernel(exe.mGemmKernel) {
}
DenseConvInt8TiledExecutor::~DenseConvInt8TiledExecutor() {
// Do nothing
}
bool DenseConvInt8TiledExecutor::onClone(Backend* bn, const Op* op, Execution** dst) {
if (nullptr == dst) {
return true;
}
auto exe = new DenseConvInt8TiledExecutor(bn, op, *this);
if (!exe->valid()) {
return false;
}
#ifdef MNN_KLEIDIAI_ENABLED
exe->mAccelType = this->mAccelType;
#endif
*dst = exe;
return true;
}
void DenseConvInt8TiledExecutor::getPackParameter(int* Unit, int* srcUnit, int* DestUnit, const CoreInt8Functions* core) {
core->MNNGetGemmUnit(Unit, srcUnit, DestUnit);
}
ErrorCode DenseConvInt8TiledExecutor::onResize(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
// default option
mUseBatchQuan = false;
mQuantFirst = true;
auto option = static_cast<CPUBackend*>(backend())->getRuntime()->hint().dynamicQuantOption;
int batch = inputs[0]->batch();
int inC = inputs[0]->channel();
auto output = outputs[0];
int inputPlane = batch * inputs[0]->width() * inputs[0]->height();
auto planeSize = output->width() * output->height() * output->batch();
auto core = static_cast<CPUBackend*>(backend())->int8Functions();
auto gcore =static_cast<CPUBackend*>(backend())->functions();
int UNIT, SRC_UNIT, DST_XUNIT;
core->MNNGetGemmUnit(&UNIT, &SRC_UNIT, &DST_XUNIT);
int kernelCount = mCommon->kernelY() * mCommon->kernelX();
bool fastway = (kernelCount == 1) && (output->width() == inputs[0]->width()) && (output->height() == inputs[0]->height()) && (mCommon->strideX() * mCommon->strideY()) == 1;
if (inputPlane > 1) {
mUseBatchQuan = true;
}
if (!fastway) { // general conv
mQuantFirst = false;
if (planeSize > 1) {
mUseBatchQuan = true;
}
if (option == 1) { // lowest level.
mQuantFirst = true;
mUseBatchQuan = false;
}
}
float weightBytes = mResourceInt8->mActBits == 4 ? 0.5 : 1;
mBlockNum = mResourceInt8->mBlockNum;
#ifdef MNN_KLEIDIAI_ENABLED
KleidiAI& kai = KleidiAI::getInstance();
if(mResourceInt8->mDynamicQuant && mResourceInt8->mActBits == 4 && kai.canAccelerate(mAccelType)) {
MNN_ASSERT(kai.isLoaded(mAccelType));
const size_t m = inputs[0]->batch(); //lhs vector number.
const size_t n = outputs[0]->channel(); //rhs vector number.
const size_t k = inputs[0]->channel(); //vector size.
const size_t blkSize = mBlockNum == 1 ? 0 : k / mBlockNum;
int packedSize = kai.getLhsQuantedPackedSize(mAccelType, m, k, blkSize);
int elementSize = kai.isHalf() ? sizeof(__fp16) : sizeof(float);
if(m > 1 && !kai.isLinear()) {
int srcSize = m * k * elementSize;
int dstSize = m * n * elementSize;
int extraSize = srcSize > dstSize ? srcSize : dstSize;
packedSize += extraSize;
}
//Split mTempIm2ColBuffer as two parts for linear/tile transfer:
//Part0: Lhs_packed.
//Part1: Lhs/Dst before transfer.
mTempIm2ColBuffer.reset(Tensor::createDevice<int8_t>({packedSize}));
bool success = backend()->onAcquireBuffer(mTempIm2ColBuffer.get(), Backend::DYNAMIC);
if (!success) {
MNN_ERROR("Out of dynamic memory!\n");
return OUT_OF_MEMORY;
}
backend()->onReleaseBuffer(mTempIm2ColBuffer.get(), Backend::DYNAMIC);
return NO_ERROR;
}
#endif
bool useExtraScale = true;
if (mResourceInt8->mDynamicQuant == false) {
mMutableResource->updateInputOutputScale(TensorUtils::getQuantInfo(inputs[0]), TensorUtils::getQuantInfo(outputs[0]));
if (mMutableResource->mResource->mUseConvQuan) {
useExtraScale = false;
}
CPUConvolution::onResize(inputs, outputs);
ConvolutionTiledExecutor::setIm2ColParameter(mIm2ColParamter, mCommon, inputs[0], outputs[0], mPadX, mPadY, gcore, core);
mBlockNum = 1;
mQuantFirst = true;
mUseBatchQuan = false;
} else { // Dynamic Quant kernels
CPUConvolution::onResize(inputs, outputs);
// Gemm Kernel
mGemmKernel = core->Int8GemmKernel;
if (mResourceInt8->mActBits == 4) {
mGemmKernel = core->Int8GemmKernel_W4;
}
mQuantFunc = core->MNNFloat2Int8;
if (gcore->bytes == 2 && gcore->pack == 8) {
mGemmKernel = core->MNNGemmInt8AddBiasScale_Unit_FP16;
if (mResourceInt8->mActBits == 4) {
mGemmKernel = core->MNNGemmInt8AddBiasScale_w4_Unit_FP16;
}
mQuantFunc = core->DynamicQuanInput_ARM82;
mQuantAndReorderFunc = core->DynamicQuanInputAndReorder_ARM82;
}
// A axisSum kernel
mSumByAxisLFunc = gcore->MNNSumByAxisLForMatmul_A;
ConvolutionTiledExecutor::setIm2ColParameter(mIm2ColParamter, mCommon, inputs[0], outputs[0], mPadX, mPadY, gcore, core);
int ocUp4 = ROUND_UP(outputs[0]->channel(), gcore->pack);
}
// input scale buffer
const int threads = static_cast<CPUBackend*>(backend())->threadNumber();
// Im2col info
int im2colBytes = 1;
const int L2Size = 2048;
int tileLimitByC = UP_DIV(L2Size, mIm2ColParamter.kernelCountUnit * SRC_UNIT);
if (mQuantFirst == false) {
im2colBytes = gcore->bytes;
tileLimitByC = 1;
}
int ic = inputs[0]->channel();
int tileLimit = 0;
int outC = output->channel();
int outC4 = UP_DIV(outC, gcore->pack);
auto icDiv4KernelCount = mIm2ColParamter.kernelCountUnit;
mSplitByOc = true;
// flop and io
float flop = gcore->bytes * planeSize * (ROUND_UP(output->channel(), gcore->pack) * icDiv4KernelCount * SRC_UNIT / 1024.0 / 1024.0 / 1024.0);
float ios = (((CPUBackend*)backend())->getTensorSize(outputs[0], true) + ((CPUBackend*)backend())->getTensorSize(inputs[0], true) + ((CPUBackend*)backend())->getTensorSize(mResourceInt8->mWeightInt8.get()) * weightBytes) / (1024.0 * 1024.0 * 1024.0);
if (threads < planeSize) { // Thread split by output nhw.
tileLimit = ALIMIN(tileLimitByC, UP_DIV(planeSize, threads));
mIm2ColCount = UP_DIV(tileLimit, DST_XUNIT);
auto DynamicDestUnit = DST_XUNIT * mIm2ColCount;
mTileCount = UP_DIV(planeSize, DynamicDestUnit);
if (mTileCount > threads) {
mSplitByOc = false;
}
}
if (mSplitByOc) {
tileLimit = ALIMIN(tileLimitByC, planeSize);
mIm2ColCount = UP_DIV(tileLimit, DST_XUNIT);
auto DynamicDestUnit = DST_XUNIT * mIm2ColCount;
mTileCount = UP_DIV(planeSize, DynamicDestUnit);
auto ocPerThread = UP_DIV(outC4, threads);
auto threadNeed = UP_DIV(outC4, ocPerThread);
int totalWork = outC4;
int part = 1;
if (UNIT > gcore->pack) { // AVX512:UNIT=64,pack=16
MNN_ASSERT(UNIT % gcore->pack == 0);
int ocDivUnit = UP_DIV(outC4 * gcore->pack, UNIT);
ocPerThread = UP_DIV(ocDivUnit, threads);
threadNeed = UP_DIV(ocDivUnit, ocPerThread);
totalWork = ocDivUnit;
part = UNIT / gcore->pack;
}
mThreadNums = ALIMIN(threads, threadNeed);
mDivides.resize(threads+1);
mDivides[0] = 0;
static_cast<CPUBackend *>(backend())->computeDivideSizes(totalWork, mDivides.data() + 1, flop / ios);
for (int i = 0; i < mDivides.size(); ++i) {
mDivides[i] *= part;
}
}
if (!mSplitByOc) {
mThreadNums = ALIMIN(threads, mTileCount);
mDivides.resize(threads+1);
mDivides[0] = 0;
static_cast<CPUBackend *>(backend())->computeDivideSizes(mTileCount, mDivides.data() + 1, flop / ios);
}
int ocUp4 = ROUND_UP(outC, gcore->pack);
int k = mThreadNums;
int workPT = DST_XUNIT * mIm2ColCount;
if (mSplitByOc) {
k = 1; // Use one thread to finish im2col.
workPT = mTileCount * DST_XUNIT * mIm2ColCount;
}
auto bufferAlloc = static_cast<CPUBackend*>(backend())->getBufferAllocator();
auto blitInfoSize = ConvolutionTiledExecutor::computeBlitInfoSize(workPT, mIm2ColParamter.ow, mIm2ColParamter.kernelX * mIm2ColParamter.kernelY, k);
mBlitInfoStride = blitInfoSize.second;
mBlitInfo = bufferAlloc->alloc(blitInfoSize.first);
int im2colBuffSize = DST_XUNIT * mIm2ColCount * icDiv4KernelCount * SRC_UNIT;
if (!mSplitByOc) {
mTempIm2ColBuffer.reset(Tensor::createDevice<int8_t>({threads, im2colBuffSize * im2colBytes}));
mTempSrcSum.resize(threads * mBlockNum * DST_XUNIT * mIm2ColCount * 4); // Use 4 bytes to save kernel sum.
} else {
mTempIm2ColBuffer.reset(Tensor::createDevice<int8_t>({mTileCount, im2colBuffSize * im2colBytes}));
mTempSrcSum.resize(mTileCount * mBlockNum * DST_XUNIT * mIm2ColCount * 4); // Use 4 bytes to save kernel sum.
}
auto success = backend()->onAcquireBuffer(mTempIm2ColBuffer.get(), Backend::DYNAMIC);
if (useExtraScale) {
// dequant_scale: area * sizeof(float)
// quant_scale: area * sizeof(float)
// absmax: [threads, area*core->bytes)]
int size = DST_XUNIT * mIm2ColCount * QUANT_INFO_BYTES;
if (mUseBatchQuan) {
if (mQuantFirst == false) {
size = 2 * mIm2ColCount * DST_XUNIT * QUANT_INFO_BYTES + mIm2ColCount * DST_XUNIT * gcore->bytes;
} else {
size = 2 * inputPlane * QUANT_INFO_BYTES + inputPlane * gcore->bytes;
}
}
mBatchQuantInfo.reset(Tensor::createDevice<int8_t>({threads, size}));
success &= backend()->onAcquireBuffer(mBatchQuantInfo.get(), Backend::DYNAMIC);
}
if (!success || mBlitInfo.invalid()) {
return OUT_OF_MEMORY;
}
if (false == mResourceInt8->mDynamicQuant) {
bufferAlloc->free(mBlitInfo);
backend()->onReleaseBuffer(mTempIm2ColBuffer.get(), Backend::DYNAMIC);
if (useExtraScale) {
backend()->onReleaseBuffer(mBatchQuantInfo.get(), Backend::DYNAMIC);
}
return NO_ERROR;
}
// set im2col tensor info
if (mQuantFirst) {
mQuantInput.reset((Tensor::createDevice<int8_t>({batch, mIm2ColParamter.ih, mIm2ColParamter.iw, ROUND_UP(inC, gcore->pack)})));
} else if (!mSplitByOc){
mQuantInput.reset((Tensor::createDevice<int8_t>({threads, im2colBuffSize * 1})));
// mIm2ColParamter.ic = inC;
} else {
mQuantInput.reset((Tensor::createDevice<int8_t>({mTileCount, im2colBuffSize * 1})));
}
success &= backend()->onAcquireBuffer(mQuantInput.get(), Backend::DYNAMIC);
// set dynamic quant buffer
// set compute buffer
if (!mUseBatchQuan) {
mTempMaxMinValueBuffer.reset(Tensor::createDevice<uint8_t>({threads, 2 * gcore->bytes}));
if (mQuantFirst) {
mDynamicBias.reset(Tensor::createDevice<uint8_t>({ocUp4 * 4}));
} else {
mDynamicBias.reset(Tensor::createDevice<uint8_t>({threads, ocUp4 * 4}));
}
success &= backend()->onAcquireBuffer(mDynamicBias.get(), Backend::DYNAMIC);
success &= backend()->onAcquireBuffer(mTempMaxMinValueBuffer.get(), Backend::DYNAMIC);
}
mAccumBuffer.reset(Tensor::createDevice<int32_t>({threads, DST_XUNIT * ocUp4}));
success &= backend()->onAcquireBuffer(mAccumBuffer.get(), Backend::DYNAMIC);
if (!success) {
return OUT_OF_MEMORY;
}
bufferAlloc->free(mBlitInfo);
backend()->onReleaseBuffer(mTempIm2ColBuffer.get(), Backend::DYNAMIC);
backend()->onReleaseBuffer(mBatchQuantInfo.get(), Backend::DYNAMIC);
backend()->onReleaseBuffer(mQuantInput.get(), Backend::DYNAMIC);
if (mUseBatchQuan == false) {
backend()->onReleaseBuffer(mDynamicBias.get(), Backend::DYNAMIC);
backend()->onReleaseBuffer(mTempMaxMinValueBuffer.get(), Backend::DYNAMIC);
}
backend()->onReleaseBuffer(mAccumBuffer.get(), Backend::DYNAMIC);
return NO_ERROR;
}
ErrorCode DenseConvInt8TiledExecutor::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
const auto input = inputs[0];
auto output = outputs[0];
auto core = static_cast<CPUBackend*>(backend())->int8Functions();
auto gcore = static_cast<CPUBackend*>(backend())->functions();
#ifdef MNN_KLEIDIAI_ENABLED
KleidiAI& kai = KleidiAI::getInstance();
if(mResourceInt8->mDynamicQuant && mResourceInt8->mActBits == 4 && kai.canAccelerate(mAccelType)) {
MNN_ASSERT(kai.isLoaded(mAccelType));
const size_t m = input->batch(); //lhs vector number.
const size_t n = output->channel(); //rhs vector number.
const size_t k = input->channel(); //vector size.
const size_t blkSize = mBlockNum == 1 ? 0 : k / mBlockNum;
bool bHalf = kai.isHalf();
size_t elementSize = bHalf ? sizeof(__fp16) : sizeof(float);
size_t lhsPackedSize = kai.getLhsQuantedPackedSize(mAccelType, m, k, blkSize);
auto lhs = input->host<uint8_t>();
auto lhsPacked = mTempIm2ColBuffer->host<int8_t>();
auto rhsPacked = mResourceInt8->mWeightInt8->host<uint8_t>();
auto dst = output->host<uint8_t>();
uint8_t *linearLhs, *linearDst;
if(m > 1 && !kai.isLinear()) {
linearLhs = (uint8_t *)lhsPacked + lhsPackedSize;
linearDst = linearLhs;
} else {
linearLhs = lhs;
linearDst = dst;
}
int threadNum = static_cast<CPUBackend*>(backend())->threadNumber();
int threadNeed, vecPerThread;
//Dynamic quant pack lhs.
if(m == 1) {
kai.runLhsQuantPack(mAccelType, 1, k, blkSize, 1, linearLhs, lhsPacked);
} else {
if(!kai.isLinear()) {
if(bHalf) {
KleidiAIUtil::transferNC4HW4ToNCHW((__fp16 *)lhs, (__fp16 *)linearLhs, m, k);
} else {
KleidiAIUtil::transferNC4HW4ToNCHW((float *)lhs, (float *)linearLhs, m, k);
}
}
vecPerThread = kai.getVecNumPerThread(m, threadNum, kai.getMr(mAccelType, m));
threadNeed = m % vecPerThread == 0 ? m / vecPerThread : (m / vecPerThread + 1);
size_t srcStride = vecPerThread * k * elementSize;
auto BatchDynamicQuant = [=, &kai](int tId) {
auto threadSrc = linearLhs + tId * srcStride;
auto threadDst = lhsPacked + kai.getLhsQuantedPackedOffset(mAccelType, m, tId * vecPerThread, k, blkSize);
int vecNum = (tId == threadNeed - 1) ? (m - vecPerThread * tId) : vecPerThread; //Last threadN may less than vecPerThread.
kai.runLhsQuantPack(mAccelType, vecNum, k, blkSize, kai.getMr(mAccelType, m), threadSrc, threadDst);
};
MNN_CONCURRENCY_BEGIN(tId, threadNeed) {
BatchDynamicQuant((int)tId);
}
MNN_CONCURRENCY_END();
}
//Run matmul.
if(kai.bSupportSme2()) {
//SME prefer running on single thread to obtain better performance/power consumption ratio.
threadNum = 1;
}
vecPerThread = kai.getVecNumPerThread(n, threadNum, kai.getNStep(mAccelType));
threadNeed = n % vecPerThread == 0 ? n / vecPerThread : (n / vecPerThread + 1);
auto ThreadFunction = [=, &kai](int tId) {
auto threadRhsPacked = rhsPacked + kai.getRhsPackedOffset(mAccelType, tId * vecPerThread, k, blkSize);
auto threadDst = linearDst + kai.getDstOffset(0, tId * vecPerThread, n, elementSize);
int vecNum = (tId == threadNeed - 1) ? (n - vecPerThread * tId) : vecPerThread; //Last threadN may less than vecPerThread.
kai.runMatmul(mAccelType, m, vecNum, k, blkSize, lhsPacked, threadRhsPacked, threadDst, n * elementSize, elementSize, FLT_MAX, -FLT_MAX);
};
MNN_CONCURRENCY_BEGIN(tId, threadNeed) {
ThreadFunction((int)tId);
}
MNN_CONCURRENCY_END();
if(m > 1 && !kai.isLinear()) {
if(bHalf) {
KleidiAIUtil::transferNCHWToNC4HW4((__fp16 *)linearDst, (__fp16 *)dst, m, n);
} else {
KleidiAIUtil::transferNCHWToNC4HW4((float *)linearDst, (float *)dst, m, n);
}
}
return NO_ERROR;
}
#endif
int UNIT__, SRC_UNIT, DST_XUNIT;
core->MNNGetGemmUnit(&UNIT__, &SRC_UNIT, &DST_XUNIT);
auto blitProc = core->MNNPackC4Int8ForMatMul_A;
const int plane = output->batch() * mIm2ColParamter.oh * mIm2ColParamter.ow;
const int batch = input->batch();
const int PackUnit = gcore->pack;
const int dstZStep = plane * PackUnit;
const int ocDiv4 = UP_DIV(output->channel(), PackUnit);
const int ocUp4 = ROUND_UP(output->channel(), PackUnit);
const auto kernelCountUnitDouble = mIm2ColParamter.kernelCountUnit;
const auto col_buffer_unit_size = kernelCountUnitDouble * DST_XUNIT * SRC_UNIT * sizeof(int8_t);
const auto col_buffer_size = col_buffer_unit_size * mIm2ColCount;
const int dstBytes = static_cast<CPUBackend*>(backend())->getBytes(backend(), output);
const int blockL = kernelCountUnitDouble / mBlockNum; // source depthQuad for each block.
float weightBytes = 1.f;
int weight_step_Y = weightBytes * (UNIT__ * SRC_UNIT);
int src_step_Y = DST_XUNIT * SRC_UNIT;
int inputPlane = batch * input->width() * input->height();
auto im2colPtr = mTempIm2ColBuffer->host<int8_t>();
if (SRC_UNIT > PackUnit) {
memset(im2colPtr, 0, mTempIm2ColBuffer->size());
}
const auto weightDataPtr = mResourceInt8->mWeightInt8->host<int8_t>();
auto srcKernelSumPtr = mTempSrcSum.data();
auto outputDataPtr = output->host<int8_t>();
uint8_t* biasPtr = nullptr;
int32_t inputZeroPoint = 0;
if (nullptr != mMutableResource.get()) {
biasPtr = mMutableResource->mBiasFloat->host<uint8_t>();
inputZeroPoint = mMutableResource->mInputZeroPoint;
if (mBatchQuantInfo.get()) {
float scalein = TensorUtils::getQuantInfo(inputs[0])[0];
float scaleou = TensorUtils::getQuantInfo(outputs[0])[0];
for (int i = 0; i < DST_XUNIT * mIm2ColCount; ++i) {
mBatchQuantInfo->host<float>()[i] = scalein / scaleou;
}
}
}
auto SingleDynamicQuant = [&] (uint8_t* floatPtr, int32_t& inputZero, uint8_t* inputDequantScale, uint8_t* matmulBiasPtr, int inputsize, int threads, uint8_t* maxMinValPtr, int8_t* int8ptr) {
float quantscale = 0.f;
float dequantscale = 0.f;
float zeropoint = 0;
/* Count max and min value to compute input scale and zeropoint */
auto findMaxMinValueFunction = [&]() {
auto perThreadWorkCount = inputsize;
auto minValPtrTid = reinterpret_cast<float*>(maxMinValPtr);
auto maxValPtrTid = reinterpret_cast<float*>(maxMinValPtr + gcore->bytes);
auto inputDataPtrTid = reinterpret_cast<float*>(floatPtr);
gcore->MNNCountMaxMinValue(inputDataPtrTid, minValPtrTid, maxValPtrTid, inputsize);
};
findMaxMinValueFunction();
float maxVal = 0;
float minVal = 0;
if (gcore->bytes == 4) {
maxVal = (reinterpret_cast<float*>(maxMinValPtr))[1];
minVal = (reinterpret_cast<float*>(maxMinValPtr))[0];
}
if (gcore->bytes == 2) {
std::vector<float> _mVal(2);
gcore->MNNLowpToFp32(reinterpret_cast<int16_t*>(maxMinValPtr), _mVal.data(), 2);
maxVal = _mVal[1];
minVal = _mVal[0];
}
/* Dynamic quant */
if (mIm2ColParamter.padX > 0 || mIm2ColParamter.padY > 0) { // Ensure "0.0f" included in range.
if (minVal > 0.f) {
minVal = 0.f;
} else if (maxVal < 0.f){
maxVal = 0.f;
} else {
//
}
}
float range = maxVal - minVal;
if (fabs(range) < 1e-7) {
zeropoint = (-1 * maxVal)-128;
quantscale = 1.0f;
dequantscale = 1.0f;
} else {
quantscale = 255.0f / range;
dequantscale = range / 255.0f;
zeropoint = roundf(-minVal * 255.f / range) - 128.0f;
}
auto sizeDiv = UP_DIV(inputsize, PackUnit);
mQuantFunc((float*)floatPtr , int8ptr, sizeDiv, &quantscale, -128, 127, &zeropoint, 0);
/* bias float */
int offset = 0;
auto scale_ = (float*)inputDequantScale;
auto unitsize = mBatchQuantInfo->length(1) / QUANT_INFO_BYTES;
for (int i = 0; i < unitsize; ++i) {
scale_[i] = dequantscale;
}
auto biasfp32 = mResourceInt8->mOriginBias->host<float>();
float zerofp32 = (zeropoint + offset) * dequantscale;
gcore->MNNDynamicUpdateConvBiasScale((float*)matmulBiasPtr, biasfp32, mResourceInt8->mWeightKernelSum->host<float>(), &zerofp32, UP_DIV(output->channel(), 4));
// Move step for A and B for each block computing
inputZero = zeropoint;
biasPtr = matmulBiasPtr;
};
auto BatchDynamicQuant = [&](uint8_t* floatPtr, int32_t& inputZero, uint8_t* inputDequantScale, int LU, int EP, int LP, int32_t availableThreads, int8_t* dstInt8) {
// Allocate input max/sum/dequant/quant buffer
auto quantPtr = inputDequantScale + EP * QUANT_INFO_BYTES;
auto maxPtr = inputDequantScale + 2 * EP * QUANT_INFO_BYTES;
// compute sum and absmax
int divlu = UP_DIV(LU, availableThreads);
MNN_CONCURRENCY_BEGIN (tId, availableThreads) {
auto batchMax = reinterpret_cast<float*>(maxPtr + tId * EP * gcore->bytes);
auto ptr_ = reinterpret_cast<float*>(floatPtr + tId * divlu * gcore->bytes * EP * LP);
gcore->MNNAbsMax((float*)floatPtr, batchMax, LU, EP, LP);
} MNN_CONCURRENCY_END();
// Compute quant scale
gcore->MNNQuantScale((float*)maxPtr, (float*)quantPtr, (float*)inputDequantScale, availableThreads, EP);
// quant
auto scale_ptr = reinterpret_cast<float*>(quantPtr);
gcore->MNNDynamicQuant((float*)floatPtr, dstInt8, scale_ptr, LU, EP, LP);
inputZero = 0;
};
ssize_t oneScale = mUseBatchQuan ? 0 : 1;
if (mUseBatchQuan) {
biasPtr = mResourceInt8->mOriginBias->host<uint8_t>();
}
int8_t* inputDataPtr = nullptr; // Matmul input.
auto im2colSrc = input->host<uint8_t>(); // if not quant first, im2colSrc is original float input data.
auto inputsize = UP_DIV(input->channel(), PackUnit) * PackUnit * batch * input->height() * input->width();
if (mQuantFirst) { // quant first, then im2col
if (mUseBatchQuan) {
int icDiv4 = UP_DIV(input->channel(), PackUnit);
int availableT = (inputPlane > 500 && icDiv4 > mThreadNums) ? mThreadNums : 1;
BatchDynamicQuant(input->host<uint8_t>(), inputZeroPoint, mBatchQuantInfo->host<uint8_t>(), icDiv4, inputPlane, PackUnit, mThreadNums, mQuantInput->host<int8_t>());
inputDataPtr = mQuantInput->host<int8_t>();
} else if (mResourceInt8->mDynamicQuant) {
SingleDynamicQuant(input->host<uint8_t>(), inputZeroPoint, mBatchQuantInfo->host<uint8_t>(), mDynamicBias->host<uint8_t>(), inputsize, 1, mTempMaxMinValueBuffer->host<uint8_t>(), mQuantInput->host<int8_t>());
inputDataPtr = mQuantInput->host<int8_t>();
} else {
// offline quant.
inputDataPtr = input->host<int8_t>();
}
im2colSrc = (uint8_t*)inputDataPtr;
}
if (mResourceInt8->mActBits == 4) {
weightBytes = 0.5;
weight_step_Y *= 0.5;
}
int blockunit = ocUp4 * 2 * QUANT_INFO_BYTES + blockL * weight_step_Y * UP_DIV(output->channel(), UNIT__);
auto inputchannel = input->channel();
SumByAxisParams sumParams;
sumParams.oneScale = oneScale;
sumParams.SRC_UNIT = SRC_UNIT;
sumParams.blockNum = mBlockNum;
sumParams.DST_XUNIT = DST_XUNIT;
sumParams.col_buffer_unit_size = col_buffer_unit_size;
sumParams.kernelCountUnitDouble = kernelCountUnitDouble;
sumParams.valid = inputchannel % SRC_UNIT;
sumParams.kernelxy = mIm2ColParamter.kernelX * mIm2ColParamter.kernelY;
sumParams.LU = UP_DIV(inputchannel, SRC_UNIT);
int im2colBytes = mQuantFirst == true ? 1 : gcore->bytes;
auto tileSplitFunction = [&](int tId, int eStartIndex, int eEndIndex, int estep) {
auto ocDivThread = ocDiv4;
float* reluPtr = mResourceInt8->mReluThreshold.data();
QuanPostTreatParameters quanParam;
quanParam.blockNum = mBlockNum;
float* accumbuff = nullptr;
uint8_t* extraScale = nullptr;
uint8_t* ptrExtraScale = nullptr;
if (mBatchQuantInfo.get() && mQuantFirst) {
extraScale = mBatchQuantInfo->host<uint8_t>();
ptrExtraScale = extraScale;
}
if (mBlockNum > 1) {
accumbuff = reinterpret_cast<float*>(mAccumBuffer->host<int8_t>() + tId * mAccumBuffer->stride(0) * sizeof(int32_t));
}
#ifdef MNN_USE_SSE
if (mResourceInt8->mDynamicQuant) {
quanParam.extraBias = mResourceInt8->mWeightKernelSum->host<float>();
}
#endif
if (dstBytes != 1) {
quanParam.useInt8 = 0;
quanParam.fp32minmax = reluPtr;
} else {
quanParam.maxValue = mMutableResource->mClampMax;
if (mResourceInt8->mRelu) {
quanParam.minValue = mMutableResource->mOutputZeroPoint;
} else {
quanParam.minValue = mMutableResource->mClampMin;
}
}
auto biasFloatTid = reinterpret_cast<float*>(biasPtr);
auto weightPtrTid = weightDataPtr;
if (mBlockNum == 1) {
quanParam.biasFloat = biasFloatTid;
}
// auto im2colPtr = mTempIm2ColBuffer->host<int8_t>();
auto colAddr = im2colPtr + tId * mTempIm2ColBuffer->stride(0);
auto srcPtr = (int8_t const **)(mBlitInfo.ptr() + tId * mBlitInfoStride.first);
auto el = (int32_t *)(srcPtr + mBlitInfoStride.second);
auto xKernelSumPtrTid = reinterpret_cast<float*>(srcKernelSumPtr + tId * mBlockNum * DST_XUNIT * mIm2ColCount * 4);
int32_t info[5];
info[1] = mIm2ColParamter.iw * mIm2ColParamter.ih * batch;
info[2] = static_cast<int32_t>(col_buffer_unit_size);
info[3] = mIm2ColParamter.strideX;
for (int tIndex = eStartIndex; tIndex < eEndIndex; tIndex += estep) {
const int xIndexStart = tIndex * DST_XUNIT * mIm2ColCount;
auto outputInTilePtr = outputDataPtr + xIndexStart * PackUnit * dstBytes;
int realDstCount = ALIMIN(plane - xIndexStart, DST_XUNIT * mIm2ColCount);
ptrExtraScale = mUseBatchQuan ? (extraScale + xIndexStart * QUANT_INFO_BYTES) : extraScale;
// im2col
auto res = ConvolutionTiledExecutor::turnIm2ColToBlitInfo((const float**)srcPtr, el, xIndexStart, realDstCount, mIm2ColParamter, (uint8_t*)im2colSrc, im2colBytes);
int number = res.first;
bool needZero = res.second;
if (needZero && mQuantFirst) {
#ifdef MNN_USE_SSE
::memset(colAddr, inputZeroPoint + 128, col_buffer_size);
#else
::memset(colAddr, inputZeroPoint, col_buffer_size);
#endif
} else if (needZero) {
::memset(colAddr, 0, mTempIm2ColBuffer->stride(0));
}
info[0] = number;
info[4] = realDstCount;
int8_t* colAddrTemp = colAddr;
if (mQuantFirst && number > 0) {
blitProc(colAddr, srcPtr, info, el);
colAddrTemp = colAddr;
} else if (number > 0) {
if (SRC_UNIT > PackUnit && !needZero) {
memset(colAddr, 0, mTempIm2ColBuffer->stride(0));
}
info[2] = realDstCount;
gcore->MNNGeneralIm2Col((float*)colAddr, (float const**)srcPtr, info, el, SRC_UNIT, PackUnit); // colAddr: [lu, realDstCount, lp]
}
if (!mQuantFirst) {
auto ptrInputscale = mBatchQuantInfo->host<uint8_t>() + tId * mBatchQuantInfo->stride(0);
if (mUseBatchQuan) {
BatchDynamicQuant((uint8_t*)colAddr, inputZeroPoint, ptrInputscale, kernelCountUnitDouble, realDstCount, SRC_UNIT, 1, mQuantInput->host<int8_t>() + tId * mQuantInput->stride(0));
} else {
biasFloatTid = reinterpret_cast<float*>(mDynamicBias->host<uint8_t>() + tId * mDynamicBias->stride(0));
auto maxMinPtr = mTempMaxMinValueBuffer->host<uint8_t>() + tId * mTempMaxMinValueBuffer->stride(0);
SingleDynamicQuant((uint8_t*)colAddr, inputZeroPoint, ptrInputscale, (uint8_t*)biasFloatTid, kernelCountUnitDouble*realDstCount*SRC_UNIT, 1, maxMinPtr, mQuantInput->host<int8_t>() + tId * mQuantInput->stride(0));
if (mBlockNum == 1) {
quanParam.biasFloat = biasFloatTid;
}
}
extraScale = ptrInputscale;
ptrExtraScale = ptrInputscale;
colAddrTemp = mQuantInput->host<int8_t>() + tId * mQuantInput->stride(0);
}
if (mResourceInt8->mWeightAsymmetricQuant) {
MNN_ASSERT(mBatchQuantInfo.get() && mBatchQuantInfo->host<float>());
mSumByAxisLFunc(xKernelSumPtrTid, colAddrTemp, (float*)ptrExtraScale, realDstCount, sumParams);
}
auto ptrX = xKernelSumPtrTid;
do {
int step = ALIMIN(DST_XUNIT, realDstCount);
quanParam.extraScale = (float*)ptrExtraScale;
if (mBlockNum > 1) {
memset(accumbuff, 0, ocUp4 * 4 * DST_XUNIT);
quanParam.accumBuffer = accumbuff;
}
int8_t* saveResult = nullptr;
for (int k = 0; k < mBlockNum; ++k) {
quanParam.biasFloat = nullptr;
quanParam.fp32minmax = nullptr;
if (k == 0) {
quanParam.biasFloat = biasFloatTid;
}
if (k == mBlockNum - 1) {
quanParam.fp32minmax = reluPtr;
saveResult = outputInTilePtr;
}
quanParam.srcKernelSum = ptrX + k * step;
mGemmKernel(saveResult, colAddrTemp + k * blockL * step * SRC_UNIT, weightPtrTid + k * blockunit, blockL, dstZStep * dstBytes, ocDivThread, &quanParam, step);
}
ptrX += (step * mBlockNum);
realDstCount-=step;
outputInTilePtr += DST_XUNIT * PackUnit * dstBytes;
colAddrTemp += col_buffer_unit_size;
ptrExtraScale = mUseBatchQuan ? (ptrExtraScale + step * QUANT_INFO_BYTES) : extraScale;
} while(realDstCount > 0);
}
};
auto ocSplitFunction = [&](int threads) { // Thread split by OC
auto colAddr = mTempIm2ColBuffer->host<int8_t>();
auto srcPtr = (int8_t const **)(mBlitInfo.ptr());
auto el = (int32_t *)(srcPtr + mBlitInfoStride.second);
auto xKernelSumPtrTid = reinterpret_cast<float*>(srcKernelSumPtr);
int32_t info[5];
info[1] = mIm2ColParamter.iw * mIm2ColParamter.ih * batch;
info[2] = static_cast<int32_t>(col_buffer_unit_size);
info[3] = mIm2ColParamter.strideX;
float* reluPtr = mResourceInt8->mReluThreshold.data();
int8_t* matmulInput;
if (mQuantFirst) { // im2col
auto res = ConvolutionTiledExecutor::turnIm2ColToBlitInfo((const float**)srcPtr, el, 0, plane, mIm2ColParamter, (uint8_t*)im2colSrc, im2colBytes);
int number = res.first;
bool needZero = res.second;
if (needZero) {
#ifdef MNN_USE_SSE
::memset(colAddr, inputZeroPoint + 128, mTempIm2ColBuffer->size());
#else
::memset(colAddr, inputZeroPoint, mTempIm2ColBuffer->size());
#endif
}
info[0] = number;
info[4] = plane;
if (number > 0) {
blitProc(colAddr, srcPtr, info, el);
}
matmulInput = colAddr;
}
if (false == mQuantFirst) {
int realDstCount = plane;
int start = 0;
auto im2colDst = colAddr;
auto ptrInputscale = mBatchQuantInfo->host<uint8_t>();
auto int8Ptr = mQuantInput->host<int8_t>();
int sizePacked = 0;
while (realDstCount > 0) {
int work = std::min(realDstCount, DST_XUNIT);
sizePacked += (work * SRC_UNIT * kernelCountUnitDouble);
auto res = ConvolutionTiledExecutor::turnIm2ColToBlitInfo((const float**)srcPtr, el, start, work, mIm2ColParamter, (uint8_t*)im2colSrc, im2colBytes);
int number = res.first;
bool needZero = res.second;
if (needZero) {
::memset(im2colDst, 0, col_buffer_unit_size * gcore->bytes);
}
info[0] = number;
info[2] = work;
if (number > 0) { // im2col
gcore->MNNGeneralIm2Col((float*)im2colDst, (float const**)srcPtr, info, el, SRC_UNIT, PackUnit); // colAddr: [lu, realDstCount, lp]
}
// batch quant
if (mUseBatchQuan) {
BatchDynamicQuant((uint8_t*)im2colDst, inputZeroPoint, ptrInputscale, kernelCountUnitDouble, work, SRC_UNIT, 1, int8Ptr);
ptrInputscale += (work * sizeof(int32_t));
int8Ptr += col_buffer_unit_size;
}
realDstCount -= work;
start += work;
im2colDst += (col_buffer_unit_size * gcore->bytes);
}
if (!mUseBatchQuan) {
SingleDynamicQuant((uint8_t*)colAddr, inputZeroPoint, ptrInputscale, mDynamicBias->host<uint8_t>(), sizePacked, 1, mTempMaxMinValueBuffer->host<uint8_t>(), mQuantInput->host<int8_t>());
}
matmulInput = mQuantInput->host<int8_t>();
}
if (mResourceInt8->mWeightAsymmetricQuant) {
MNN_ASSERT(mBatchQuantInfo.get() && mBatchQuantInfo->host<float>());
mSumByAxisLFunc(xKernelSumPtrTid, matmulInput, mBatchQuantInfo->host<float>(), plane, sumParams);
}
MNN_CONCURRENCY_BEGIN(tId, threads) {
int ocIndex = PackUnit * mDivides[tId];
auto ocDivThread = ALIMIN(mDivides[tId + 1] - mDivides[tId], ocDiv4 - mDivides[tId]);
if (ocIndex < ocUp4) {
auto colAddrTemp = matmulInput;
QuanPostTreatParameters quanParam;
quanParam.blockNum = mBlockNum;
uint8_t* extraScale = nullptr; // input scale for batch dynamic quant.
uint8_t* ptrExtraScale = nullptr;
float* accumbuff = nullptr;
if (mBatchQuantInfo.get()) {
extraScale = mBatchQuantInfo->host<uint8_t>();
ptrExtraScale = extraScale;
}
if (mBlockNum > 1) {
accumbuff = reinterpret_cast<float*>(mAccumBuffer->host<int8_t>() + tId * mAccumBuffer->stride(0) * sizeof(int32_t));
}
#ifdef MNN_USE_SSE
if (mResourceInt8->mDynamicQuant) {
quanParam.extraBias = mResourceInt8->mWeightKernelSum->host<float>() + ocIndex;
}
#endif
if (dstBytes != 1) {
quanParam.useInt8 = 0;
quanParam.fp32minmax = reluPtr;
} else {
quanParam.maxValue = mMutableResource->mClampMax;
if (mResourceInt8->mRelu) {
quanParam.minValue = mMutableResource->mOutputZeroPoint;
} else {
quanParam.minValue = mMutableResource->mClampMin;
}
}
auto outputInTilePtr = outputDataPtr + ocIndex * plane * dstBytes;
const auto biasFloatTid = reinterpret_cast<float*>(biasPtr + ocIndex * 4);
const auto weightPtrTid = weightDataPtr + static_cast<int32_t>(ocIndex * blockL * SRC_UNIT * weightBytes + ocIndex * 2 * QUANT_INFO_BYTES);
int realDstCount = plane;
auto ptrX = xKernelSumPtrTid;
do {
int step = ALIMIN(DST_XUNIT, realDstCount);
quanParam.extraScale = (float*)ptrExtraScale;
if (mBlockNum > 1) {
memset(accumbuff, 0, ocUp4 * 4 * DST_XUNIT);
quanParam.accumBuffer = accumbuff;
}
int8_t* saveResult = nullptr;
for (int k = 0; k < mBlockNum; ++k) {
quanParam.biasFloat = nullptr;
quanParam.fp32minmax = nullptr;
if (k == 0) {
quanParam.biasFloat = (float*)biasFloatTid;
}
if (k == mBlockNum - 1) {
quanParam.fp32minmax = reluPtr;
saveResult = outputInTilePtr;
}
quanParam.srcKernelSum = ptrX + k * step;
mGemmKernel(saveResult, colAddrTemp + k * blockL * step * SRC_UNIT, weightPtrTid + k * blockunit, blockL, dstZStep * dstBytes, ocDivThread, &quanParam, step);
}
ptrX += (step * mBlockNum);
realDstCount-=step;
outputInTilePtr += DST_XUNIT * PackUnit * dstBytes;
colAddrTemp += col_buffer_unit_size;
ptrExtraScale = mUseBatchQuan ? (ptrExtraScale + step * QUANT_INFO_BYTES) : extraScale;
} while(realDstCount > 0);
}
}
MNN_CONCURRENCY_END();
};
const int threads = static_cast<CPUBackend*>(backend())->threadNumber();
if (!mSplitByOc) {
MNN_CONCURRENCY_BEGIN(tId, threads) {
if (mDivides[tId + 1] - mDivides[tId] > 0) {
tileSplitFunction((int)tId, mDivides[tId], mDivides[tId + 1], 1);
}
}
MNN_CONCURRENCY_END();
} else {
ocSplitFunction(threads);
}
return NO_ERROR;
}
} // namespace MNN