mirror of https://github.com/alibaba/MNN.git
827 lines
31 KiB
C++
827 lines
31 KiB
C++
//
|
|
// ConvolutionCommon.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2020/03/02.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "ConvolutionCommon.hpp"
|
|
#include <math.h>
|
|
#include "backend/cpu/compute/CommonOptFunction.h"
|
|
#include "backend/cpu/CPUBackend.hpp"
|
|
#include "half.hpp"
|
|
#include "core/OpCommonUtils.hpp"
|
|
|
|
namespace MNN {
|
|
|
|
namespace IDSTDecoder {
|
|
|
|
static inline void *MNNMemoryAllocAlignZeroAlign(size_t size) {
|
|
return MNNMemoryCallocAlign(size, MNN_MEMORY_ALIGN_DEFAULT);
|
|
}
|
|
|
|
static int ReadBlobDim(BaseLoader* myfile, unsigned int* shape, int shapeBufCnt, bool useInt32) {
|
|
uint8_t uSize = 0;
|
|
myfile->read((char*)&uSize, 1);
|
|
if (uSize > 4) {
|
|
printf("Read shape error!\n");
|
|
return 0;
|
|
}
|
|
int copyLength = uSize;
|
|
if (copyLength > shapeBufCnt) {
|
|
copyLength = shapeBufCnt;
|
|
}
|
|
if (useInt32) {
|
|
myfile->read((char*)shape, sizeof(unsigned int) * copyLength);
|
|
} else {
|
|
uint16_t shape_i16[32] = {0};
|
|
myfile->read((char*)shape_i16, sizeof(uint16_t) * copyLength);
|
|
for (int i = 0; i < copyLength; ++i) {
|
|
shape[i] = shape_i16[i];
|
|
}
|
|
}
|
|
return copyLength;
|
|
}
|
|
|
|
static double _log2(double x) {
|
|
return log(x) / log(2);
|
|
}
|
|
|
|
static uint32_t atLestBitsCnt(uint32_t n) {
|
|
for (uint32_t i = 0; i < 32; i++) {
|
|
int32_t t = n << i;
|
|
if (t < 0)
|
|
return 32 - i - (((t << 1) == 0) ? 1 : 0);
|
|
}
|
|
return 0;
|
|
}
|
|
|
|
static void SplitBufToArray(uint8_t *buf, size_t bufLen, uint8_t *arr, size_t arrLen, size_t iNeedBits) {
|
|
unsigned char cMask = (1 << (iNeedBits)) - 1;
|
|
unsigned char *tmp = (unsigned char *)buf;
|
|
int iOffset = 0;
|
|
for (unsigned int i = 0; i < arrLen; i++) {
|
|
unsigned char idx = 0;
|
|
long uShift = 8 - iNeedBits - iOffset % 8;
|
|
if (uShift < 0) {
|
|
idx = (tmp[iOffset / 8] << (0 - uShift)) & cMask;
|
|
idx |= (tmp[(iOffset / 8) + 1] >> (8 + uShift)) & cMask;
|
|
} else {
|
|
idx = (tmp[iOffset / 8] >> uShift) & cMask;
|
|
}
|
|
iOffset += iNeedBits;
|
|
if (iOffset % 8 == 0) {
|
|
tmp += iOffset / 8;
|
|
iOffset = 0;
|
|
}
|
|
arr[i] = idx;
|
|
}
|
|
}
|
|
|
|
// fixme!!! not efficiency
|
|
typedef struct _SIMPLE_SET {
|
|
int8_t *UniSet;
|
|
uint32_t UniSetSize;
|
|
uint32_t CurUniCnt;
|
|
} SIMPLE_SET, *PSIMPLE_SET;
|
|
|
|
static PSIMPLE_SET CreateSimpleSet(uint32_t maxSize) {
|
|
PSIMPLE_SET set = (PSIMPLE_SET)calloc(1, sizeof(SIMPLE_SET));
|
|
if (set == nullptr)
|
|
return nullptr;
|
|
set->UniSet = (int8_t *)calloc(maxSize, sizeof(int8_t));
|
|
set->UniSetSize = maxSize;
|
|
set->CurUniCnt = 0;
|
|
return set;
|
|
}
|
|
|
|
static void SimpleRank(int8_t *data, uint32_t cnt, int up) {
|
|
if (up) {
|
|
for (uint32_t i = 0; i < cnt; i++) {
|
|
for (uint32_t j = i + 1; j < cnt; j++) {
|
|
if (data[i] > data[j]) {
|
|
int8_t tmp = data[i];
|
|
data[i] = data[j];
|
|
data[j] = tmp;
|
|
}
|
|
}
|
|
}
|
|
} else {
|
|
for (uint32_t i = 0; i < cnt; i++) {
|
|
for (uint32_t j = i + 1; j < cnt; j++) {
|
|
if (data[i] < data[j]) {
|
|
int8_t tmp = data[i];
|
|
data[i] = data[j];
|
|
data[j] = tmp;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
static void InsertSimpleSet(PSIMPLE_SET set, int8_t value) {
|
|
if (set->CurUniCnt >= set->UniSetSize)
|
|
return;
|
|
for (uint32_t i = 0; i < set->CurUniCnt; i++) {
|
|
if (set->UniSet[i] == value)
|
|
return;
|
|
}
|
|
set->UniSet[set->CurUniCnt++] = value;
|
|
// SimpleRank(set->UniSet, set->CurUniCnt, 1);
|
|
}
|
|
|
|
static void DestorySimpleSet(PSIMPLE_SET set) {
|
|
if (set->UniSet != nullptr)
|
|
free(set->UniSet);
|
|
free(set);
|
|
}
|
|
|
|
typedef struct _SIMPLE_MAP {
|
|
int8_t *CharCharMap;
|
|
uint32_t CharMapSize;
|
|
uint32_t CurMapCnt;
|
|
} SIMPLE_MAP, *PSIMPLE_MAP;
|
|
|
|
static PSIMPLE_MAP CreateSimpleMap(uint32_t MaxCnt) {
|
|
PSIMPLE_MAP map = (PSIMPLE_MAP)calloc(1, sizeof(SIMPLE_MAP));
|
|
if (map == nullptr)
|
|
return nullptr;
|
|
map->CharMapSize = MaxCnt * sizeof(int8_t);
|
|
map->CurMapCnt = 0;
|
|
map->CharCharMap = (int8_t *)calloc(1, MaxCnt * 2);
|
|
return map;
|
|
}
|
|
|
|
static void DestroySimpleMap(PSIMPLE_MAP map) {
|
|
if (map->CharCharMap)
|
|
free(map->CharCharMap);
|
|
free(map);
|
|
}
|
|
|
|
static void InsertMap(PSIMPLE_MAP map, int8_t k, int8_t v) {
|
|
for (uint32_t i = 0; i < map->CurMapCnt; i++) {
|
|
if (map->CharCharMap[i * 2] == k) {
|
|
map->CharCharMap[i * 2 + 1] = v;
|
|
return;
|
|
}
|
|
}
|
|
if (map->CurMapCnt >= map->CharMapSize)
|
|
return;
|
|
map->CharCharMap[map->CurMapCnt * 2] = k;
|
|
map->CharCharMap[map->CurMapCnt * 2 + 1] = v;
|
|
map->CurMapCnt++;
|
|
}
|
|
|
|
static int8_t FindInMap(PSIMPLE_MAP map, int8_t k, int *found) {
|
|
for (uint32_t i = 0; i < map->CurMapCnt; i++) {
|
|
if (map->CharCharMap[i * 2] == k) {
|
|
if (found != nullptr)
|
|
*found = 1;
|
|
return map->CharCharMap[i * 2 + 1];
|
|
}
|
|
}
|
|
if (found != nullptr)
|
|
*found = 0;
|
|
return 0;
|
|
}
|
|
|
|
static bool isLinearSample(const std::vector<int8_t>& sample, int bit) {
|
|
const int offset = 1 << (bit - 1);
|
|
const int size = 1 << bit;
|
|
if (sample.size() != size) {
|
|
return false;
|
|
}
|
|
for (int i = 0; i < sample.size(); i++) {
|
|
if (static_cast<int>(sample[i]) != i - offset) {
|
|
return false;
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
|
|
static void ReadQuanInfo(BaseLoader* s, size_t* len, ConvolutionCommon::Int8Common* result, bool shapeInt32) {
|
|
*len = 1;
|
|
// blob shape
|
|
unsigned int shape[32] = {0};
|
|
uint32_t shapeDim = (uint32_t)ReadBlobDim(s, shape, 32, shapeInt32);
|
|
if (shapeDim == 0 || shapeDim > 32)
|
|
return;
|
|
for (uint32_t i = 0; i < shapeDim; i++)
|
|
*len *= shape[i];
|
|
|
|
// sample
|
|
uint32_t sampleCnt = 0;
|
|
s->read((char*)&sampleCnt, 1);
|
|
if (sampleCnt == 0) {
|
|
sampleCnt = 256;
|
|
}
|
|
result->weightMap.resize(sampleCnt);
|
|
auto samples = result->weightMap.data();
|
|
if (samples == nullptr)
|
|
return;
|
|
s->read((char*)samples, sampleCnt);
|
|
SimpleRank(samples, sampleCnt, 1);
|
|
uint32_t idxBitsCnt = atLestBitsCnt(sampleCnt);
|
|
result->canUseInt4 = idxBitsCnt == 4;
|
|
}
|
|
|
|
static int8_t *ReadQuanData_c(BaseLoader* s, size_t* len, ConvolutionCommon::Int8Common* result, bool shapeInt32, bool forceQuant) {
|
|
int8_t *blob = nullptr;
|
|
uint8_t *idxBuf = nullptr;
|
|
size_t dataCnt = 1;
|
|
|
|
do {
|
|
// blob shape
|
|
unsigned int shape[32] = {0};
|
|
uint32_t shapeDim = (uint32_t)ReadBlobDim(s, shape, 32, shapeInt32);
|
|
if (shapeDim == 0 || shapeDim > 32)
|
|
break;
|
|
for (uint32_t i = 0; i < shapeDim; i++)
|
|
dataCnt *= shape[i];
|
|
|
|
// sample
|
|
uint32_t sampleCnt = 0;
|
|
s->read((char*)&sampleCnt, 1);
|
|
if (sampleCnt == 0) {
|
|
sampleCnt = 256;
|
|
}
|
|
result->weightMap.resize(sampleCnt);
|
|
auto samples = result->weightMap.data();
|
|
if (samples == nullptr)
|
|
break;
|
|
s->read((char*)samples, sampleCnt);
|
|
SimpleRank(samples, sampleCnt, 1);
|
|
uint32_t idxBitsCnt = atLestBitsCnt(sampleCnt);
|
|
idxBitsCnt = idxBitsCnt < 1 ? 1 : idxBitsCnt;
|
|
// index
|
|
size_t idxBufSize = ceil(idxBitsCnt * dataCnt * 0.125);
|
|
idxBuf = (uint8_t *)MNNMemoryAllocAlignZeroAlign(idxBufSize);
|
|
if (nullptr == idxBuf) {
|
|
MNN_ERROR("Not enought memory\n");
|
|
break;
|
|
}
|
|
s->read((char*)idxBuf, idxBufSize);
|
|
bool linear = isLinearSample(result->weightMap, idxBitsCnt);
|
|
if (linear) {
|
|
result->originBits = idxBitsCnt;
|
|
}
|
|
if (linear && (idxBitsCnt == 4 || idxBitsCnt == 8)) {
|
|
if (!forceQuant && idxBitsCnt == 4) {
|
|
// back to float, 4bit to 8bit
|
|
*len = dataCnt;
|
|
blob = (int8_t *)MNNMemoryAllocAlignZeroAlign((size_t)UP_DIV(dataCnt, 2) * 2);
|
|
for (int i = 0; i < idxBufSize; i++) {
|
|
int val = idxBuf[i];
|
|
int x1 = val / 16;
|
|
int x2 = val % 16;
|
|
blob[2 * i] = x1 - 8;
|
|
blob[2 * i + 1] = x2 - 8;
|
|
}
|
|
} else {
|
|
// keep quant
|
|
blob = (int8_t*)idxBuf;
|
|
idxBuf = nullptr;
|
|
if (idxBitsCnt == 4) {
|
|
result->canUseInt4 = true;
|
|
} else {
|
|
for (int i = 0; i < idxBufSize; i++) {
|
|
blob[i] = (int)blob[i] - 128;
|
|
}
|
|
}
|
|
*len = idxBufSize;
|
|
}
|
|
} else {
|
|
blob = (int8_t *)MNNMemoryAllocAlignZeroAlign((size_t)UP_DIV(dataCnt, 2) * 2);
|
|
if (nullptr == blob) {
|
|
break;
|
|
}
|
|
bool success = true;
|
|
int offset = (1 << (idxBitsCnt-1));
|
|
do {
|
|
if (linear) {
|
|
SplitBufToArray(idxBuf, (uint32_t)idxBufSize, (uint8_t*)blob, (uint32_t)dataCnt, (uint32_t)idxBitsCnt);
|
|
auto src = (uint8_t*)blob;
|
|
auto dst = blob;
|
|
for (int i=0; i<dataCnt; ++i) {
|
|
dst[i] = (int)src[i] - offset;
|
|
}
|
|
break;
|
|
}
|
|
// split index value into bytes
|
|
uint8_t* idxBytes = (uint8_t *)MNNMemoryAllocAlignZeroAlign(dataCnt * sizeof(uint8_t));
|
|
if (idxBitsCnt == 0 || nullptr == idxBytes) {
|
|
success = false;
|
|
break;
|
|
}
|
|
SplitBufToArray(idxBuf, (uint32_t)idxBufSize, idxBytes, (uint32_t)dataCnt, (uint32_t)idxBitsCnt);
|
|
int i = 0;
|
|
for (; i < dataCnt; i++) {
|
|
if (idxBytes[i] >= sampleCnt) {
|
|
MNN_PRINT("iNeedBits is %u\nRead quan weights error with idx:%d\n", idxBitsCnt, (int)idxBytes[i]);
|
|
success = false;
|
|
break;
|
|
}
|
|
blob[i] = samples[idxBytes[i]];
|
|
}
|
|
MNNMemoryFreeAlign(idxBytes);
|
|
} while (false);
|
|
|
|
if (!success) {
|
|
MNNMemoryFreeAlign(blob);
|
|
blob = nullptr;
|
|
break;
|
|
}
|
|
if (len) {
|
|
*len = blob ? dataCnt : 0;
|
|
}
|
|
if (result->originBits <= 4 && forceQuant) {
|
|
// Reduce blob to 4 bit
|
|
result->canUseInt4 = true;
|
|
auto sizeDiv2 = UP_DIV(dataCnt, 2);
|
|
auto newBlob = (int8_t *)MNNMemoryAllocAlign((size_t)sizeDiv2, MNN_MEMORY_ALIGN_DEFAULT);
|
|
for (int i=0; i<sizeDiv2; ++i) {
|
|
auto s0 = blob[2*i+0] + 8;
|
|
auto s1 = blob[2*i+1] + 8;
|
|
newBlob[i] = (s0 << 4) + s1;
|
|
}
|
|
MNNMemoryFreeAlign(blob);
|
|
blob = newBlob;
|
|
}
|
|
}
|
|
} while (0);
|
|
|
|
if (idxBuf != nullptr)
|
|
MNNMemoryFreeAlign(idxBuf);
|
|
|
|
return blob;
|
|
}
|
|
|
|
static int8_t *ReadSparseQuanData_c(BaseLoader* myfile, size_t* len, const float* alpha_ptr, size_t alpha_size, ConvolutionCommon::Int8Common* result, bool useInt32) { // MNN_ERROR("sparse:%d\n", 1);
|
|
unsigned int shape[32];
|
|
uint32_t ucMapSize = 0;
|
|
PSIMPLE_SET setWeight = CreateSimpleSet(256);
|
|
if (setWeight == nullptr) {
|
|
return nullptr;
|
|
}
|
|
std::shared_ptr<unsigned int> __autoReleaseSetWeight(nullptr, [setWeight](void *) { DestorySimpleSet(setWeight); });
|
|
unsigned int nnz;
|
|
unsigned char iIdxNeedBits;
|
|
int8_t *blob = nullptr;
|
|
// 1. weights blob shape(unsigned int32)
|
|
int ShapeDim = ReadBlobDim(myfile, shape, 32, useInt32);
|
|
size_t Size = sizeof(int8_t);
|
|
for (int i = 0; i < ShapeDim; i++)
|
|
Size *= shape[i];
|
|
blob = (int8_t *)MNNMemoryAllocAlignZeroAlign((size_t)Size);
|
|
if (blob == nullptr)
|
|
return nullptr;
|
|
// 2. nnz
|
|
myfile->read((char *)&nnz, 4);
|
|
// 3. max_step use # bits () (unsigned char)
|
|
myfile->read((char *)&iIdxNeedBits, 1);
|
|
// read idx array
|
|
// 4. buf for steps ceil(nnz*step need bits/8)
|
|
AutoStorage<unsigned char> arrIdxBuffer(nnz);
|
|
unsigned char *arrIdx = arrIdxBuffer.get();
|
|
if (nullptr == arrIdx) {
|
|
return nullptr;
|
|
}
|
|
{
|
|
size_t bufLen = (size_t)(ceil(0.125 * iIdxNeedBits * nnz));
|
|
char *buf = (char *)MNNMemoryAllocAlignZeroAlign(bufLen * sizeof(char));
|
|
if (nullptr == buf) {
|
|
return nullptr;
|
|
}
|
|
myfile->read((char *)buf, bufLen);
|
|
SplitBufToArray((uint8_t *)buf, (uint32_t)bufLen, (uint8_t *)arrIdx, (uint32_t)nnz, (uint32_t)iIdxNeedBits);
|
|
MNNMemoryFreeAlign(buf);
|
|
}
|
|
// 5. Avalable values Count(unsigned char)
|
|
myfile->read((char *)&ucMapSize, 1);
|
|
if (0 == ucMapSize) {
|
|
ucMapSize = 256;
|
|
}
|
|
result->weightMap.resize(ucMapSize);
|
|
// 6. valueset(signed char * valueset_size)
|
|
for (int i = 0; i < ucMapSize; i++) {
|
|
int8_t tmp;
|
|
myfile->read((char *)&tmp, 1);
|
|
InsertSimpleSet(setWeight, tmp);
|
|
result->weightMap[i] = tmp;
|
|
}
|
|
SimpleRank(setWeight->UniSet, setWeight->CurUniCnt, 1);
|
|
// map<unsigned char, signed char> mapWeight;
|
|
PSIMPLE_MAP mapWeight = CreateSimpleMap(256);
|
|
if (mapWeight == nullptr) {
|
|
return nullptr;
|
|
}
|
|
std::shared_ptr<unsigned int> __autoReleaseMapWeight(nullptr, [mapWeight](void *) { DestroySimpleMap(mapWeight); });
|
|
|
|
for (int i = 0; i < setWeight->CurUniCnt; i++) {
|
|
InsertMap(mapWeight, i, setWeight->UniSet[i]);
|
|
}
|
|
// unsigned char iIdx = 0;
|
|
// 7. none zero weights indexes(nnz*ceil(log2(Avalable_values_Count))/8)
|
|
AutoStorage<unsigned char> arrWeightIdxBuffer(nnz);
|
|
unsigned char *arrWeightIdx = arrWeightIdxBuffer.get();
|
|
if (nullptr == arrWeightIdx) {
|
|
return nullptr;
|
|
}
|
|
int iDataNeedBits = (int)ceil(_log2(ucMapSize));
|
|
iDataNeedBits = iDataNeedBits < 1 ? 1 : iDataNeedBits;
|
|
{
|
|
size_t bufLen = (size_t)(ceil(0.125 * iDataNeedBits * nnz));
|
|
char *buf = (char *)MNNMemoryAllocAlignZeroAlign(bufLen * sizeof(char));
|
|
if (nullptr == buf) {
|
|
return nullptr;
|
|
}
|
|
myfile->read((char *)buf, bufLen);
|
|
SplitBufToArray((uint8_t *)buf, (uint32_t)bufLen, (uint8_t *)arrWeightIdx, (uint32_t)nnz,
|
|
(uint32_t)iDataNeedBits);
|
|
MNNMemoryFreeAlign(buf);
|
|
}
|
|
// set blob data with idx and weight idx
|
|
{
|
|
if (alpha_size == 2 * shape[0]) {
|
|
const int min_value = -(1 << (iDataNeedBits - 1));
|
|
auto alphaPtr = alpha_ptr;
|
|
int area = Size / shape[0];
|
|
for (int i = 0; i < shape[0]; i++) {
|
|
float min = alphaPtr[2*i];
|
|
float scale = alphaPtr[2*i+1];
|
|
int zeroQuant = min_value;
|
|
if (scale > 1e-6) {
|
|
zeroQuant = round((0.0f - min) / scale) + min_value;
|
|
}
|
|
memset(blob+area*i, zeroQuant, area * sizeof(signed char));
|
|
}
|
|
} else {
|
|
memset(blob, 0, Size * sizeof(signed char)); //backward compability with previous symmetric weight quant
|
|
}
|
|
int iPreIdx = 0;
|
|
for (int i = 0; i < nnz; i++) {
|
|
iPreIdx += arrIdx[i];
|
|
int found = 0;
|
|
int8_t value = FindInMap(mapWeight, arrWeightIdx[i], &found);
|
|
if (!found) {
|
|
MNN_ERROR("Read quan weights error with idx:%d\n", arrWeightIdx[i]);
|
|
MNNMemoryFreeAlign(blob);
|
|
return nullptr;
|
|
}
|
|
blob[iPreIdx] = value;
|
|
}
|
|
}
|
|
*len = Size;
|
|
return blob;
|
|
}
|
|
|
|
|
|
} // namespace IDSTDecoder
|
|
|
|
std::shared_ptr<ConvolutionCommon::Int8Common> ConvolutionCommon::load(const Op* op, Backend* backend, bool forceFloat, bool forceInt8) {
|
|
auto conv = op->main_as_Convolution2D();
|
|
auto quan = conv->quanParameter();
|
|
std::shared_ptr<ConvolutionCommon::Int8Common> result(new Int8Common);
|
|
result->quan = quan;
|
|
size_t buffer_size = 0, alpha_size = 0;
|
|
const int8_t* buffer_ptr = nullptr;
|
|
const float* alpha_ptr = nullptr;
|
|
std::unique_ptr<int8_t[]> external_buffer;
|
|
size_t weightLength = 0;
|
|
int8_t *buffer = nullptr;
|
|
bool useCachedMmap = false;
|
|
if (backend && backend->getRuntime()) {
|
|
useCachedMmap = backend->getRuntime()->hint().useCachedMmap > 1;
|
|
}
|
|
if (USE_EXTERNAL_DATA(conv) && op->externalPath() && quan->type() == 8) {
|
|
std::unique_ptr<FileLoader> external(new FileLoader(op->externalPath()->c_str()));
|
|
auto param = op->main_as_Convolution2D();
|
|
external->offset(param->external()->data()[0]);
|
|
result->weightFloat.reset(param->external()->data()[1] / sizeof(float));
|
|
external->read((char*)(result->weightFloat.get()), param->external()->data()[1]);
|
|
return result;
|
|
}
|
|
if (USE_EXTERNAL_DATA(conv) && op->externalPath() && quan->buffer() == nullptr) {
|
|
auto external_info = conv->external()->data();
|
|
buffer_size = external_info[1];
|
|
alpha_size = external_info[2] / sizeof(float);
|
|
std::unique_ptr<FileLoader> external_file(new FileLoader(op->externalPath()->c_str()));
|
|
external_file->offset(external_info[0]);
|
|
if (useCachedMmap) {
|
|
if (alpha_size) {
|
|
result->alpha.reset(alpha_size);
|
|
IDSTDecoder::ReadQuanInfo(external_file.get(), &weightLength, result.get(), quan->shapeInt32());
|
|
}
|
|
} else {
|
|
// external data
|
|
if (0 != buffer_size) {
|
|
if (1 == quan->type() && !forceFloat) {
|
|
buffer = IDSTDecoder::ReadQuanData_c(external_file.get(), &weightLength, result.get(), quan->shapeInt32(), forceInt8);
|
|
} else {
|
|
external_buffer.reset(new int8_t[buffer_size]);
|
|
buffer_ptr = external_buffer.get();
|
|
external_file->read((char*)buffer_ptr, buffer_size);
|
|
}
|
|
}
|
|
if (0 != alpha_size) {
|
|
result->alpha.reset(alpha_size);
|
|
if (nullptr == result->alpha.get()) {
|
|
MNN_PRINT("Alloc memory error for extract idst int8\n");
|
|
return nullptr;
|
|
}
|
|
alpha_ptr = result->alpha.get();
|
|
external_file->read((char*)alpha_ptr, alpha_size * sizeof(float));
|
|
}
|
|
}
|
|
} else {
|
|
if (quan->buffer()) {
|
|
buffer_size = quan->buffer()->size();
|
|
buffer_ptr = quan->buffer()->data();
|
|
}
|
|
if (quan->alpha()) {
|
|
alpha_size = quan->alpha()->size();
|
|
alpha_ptr = quan->alpha()->data();
|
|
result->alpha.reset(alpha_size);
|
|
if (nullptr == result->alpha.get()) {
|
|
MNN_PRINT("Alloc memory error for extract idst int8\n");
|
|
return nullptr;
|
|
}
|
|
::memcpy(result->alpha.get(), alpha_ptr, alpha_size * sizeof(float));
|
|
}
|
|
}
|
|
if (quan->index() != nullptr) {
|
|
if (forceFloat) {
|
|
// Expand sparse to dense
|
|
result->weightFloat.reset(quan->weightSize());
|
|
if (nullptr == result->weightFloat.get()) {
|
|
return nullptr;
|
|
}
|
|
::memset(result->weightFloat.get(), 0, quan->weightSize() * sizeof(float));
|
|
auto index = quan->index()->data();
|
|
auto indexSize = quan->index()->size();
|
|
if (nullptr == alpha_ptr || alpha_size != indexSize) {
|
|
MNN_ERROR("The model is error, don't has alpha but has index\n");
|
|
return nullptr;
|
|
}
|
|
for (uint32_t i=0; i<indexSize; ++i) {
|
|
result->weightFloat.get()[index[i]] = alpha_ptr[i];
|
|
}
|
|
} // Otherwise needn't treat, just return result with quan info
|
|
return result;
|
|
}
|
|
|
|
std::unique_ptr<MemoryLoader> originBuffer(new MemoryLoader((unsigned char*)buffer_ptr));
|
|
if (1 == quan->type() && weightLength == 0) {
|
|
buffer = IDSTDecoder::ReadQuanData_c(originBuffer.get(), &weightLength, result.get(), quan->shapeInt32(), forceInt8);
|
|
}
|
|
if (2 == quan->type()) {
|
|
buffer = IDSTDecoder::ReadSparseQuanData_c(originBuffer.get(), &weightLength, alpha_ptr, alpha_size, result.get(), quan->shapeInt32());
|
|
}
|
|
// read fp16 data
|
|
if (3 == quan->type()) {
|
|
if (useCachedMmap) {
|
|
weightLength = buffer_size / sizeof(half_float::half);
|
|
result->weightFloat.reset(weightLength);
|
|
return result;
|
|
}
|
|
weightLength = buffer_size / sizeof(half_float::half);
|
|
std::vector<int8_t> tempHalfWeight(buffer_size);
|
|
::memcpy(tempHalfWeight.data(), buffer_ptr, buffer_size);
|
|
auto halfWeight = reinterpret_cast<half_float::half *>(tempHalfWeight.data());
|
|
result->weightFloat.reset(weightLength);
|
|
if (nullptr == result->weightFloat.get()) {
|
|
MNN_PRINT("Alloc memory error for extract fp16 back to float\n");
|
|
return nullptr;
|
|
}
|
|
std::transform(halfWeight, halfWeight + weightLength, result->weightFloat.get(),
|
|
[](half_float::half h) { return float(h); });
|
|
return result;
|
|
}
|
|
|
|
// weight int8 only
|
|
if (4 == quan->type()) {
|
|
weightLength = buffer_size;
|
|
result->weight.reset(weightLength);
|
|
::memcpy(result->weight.get(), buffer_ptr, weightLength);
|
|
}
|
|
|
|
bool oldType4 = (quan->type() == 4 && quan->aMin() == 0 && std::abs(quan->quantScale()) < 1e-6);
|
|
if (quan->readType() != 0 || oldType4) {
|
|
result->asymmetric = true;
|
|
} else {
|
|
result->asymmetric = false;
|
|
}
|
|
if (!useCachedMmap) {
|
|
if (result->weight.get() == nullptr) {
|
|
if (nullptr == buffer) {
|
|
MNN_PRINT("Alloc memory error for extract idst int8\n");
|
|
return nullptr;
|
|
}
|
|
result->weight.set(buffer, weightLength);
|
|
}
|
|
int outputCount = 0;
|
|
if (result->asymmetric) {
|
|
outputCount = result->alpha.size() / 2;
|
|
// clampMin is minVal in asymmetric quant, clampMin = -(2^(bit))
|
|
// and old version clampMin is -128
|
|
float clampMin = quan->aMin() == 0 ? -128 : quan->aMin();
|
|
if (clampMin < 0) {
|
|
for (int o = 0; o < outputCount; ++o) {
|
|
result->alpha.get()[2 * o] = result->alpha.get()[2 * o] - clampMin * result->alpha.get()[2 * o + 1];
|
|
}
|
|
}
|
|
} else {
|
|
outputCount = result->alpha.size(); // backward compability with previous symmetric quantization
|
|
}
|
|
if (!quan->has_scaleInt()) {
|
|
float extraFactor = quan->quantScale();
|
|
// for old type 4 models, their quan->quantScale is 0. which will introduce a bug here
|
|
if (oldType4) {
|
|
extraFactor = 1.0f;
|
|
} else if (extraFactor != 1.0f) {
|
|
for (int o=0; o<result->alpha.size(); ++o) {
|
|
result->alpha.get()[o] *= extraFactor;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
if (forceInt8) {
|
|
return result;
|
|
}
|
|
if (!quan->has_scaleInt() || forceFloat) {
|
|
// Back to float
|
|
result->weightFloat.reset(weightLength);
|
|
if (nullptr == result->weightFloat.get()) {
|
|
MNN_PRINT("Alloc memory error for extract idst int8/ Back to float\n");
|
|
return nullptr;
|
|
}
|
|
int outputCount = 0;
|
|
if (result->asymmetric) {
|
|
outputCount = result->alpha.size() / 2;
|
|
} else {
|
|
outputCount = result->alpha.size();
|
|
}
|
|
int partWeightSize = weightLength / outputCount;
|
|
for (int o = 0; o < outputCount; ++o) {
|
|
float min = 0.0f;
|
|
float alpha = 0.0f;
|
|
if (result->asymmetric) {
|
|
min = result->alpha.get()[2*o];
|
|
alpha = result->alpha.get()[2*o+1];
|
|
} else {
|
|
alpha = result->alpha.get()[o];
|
|
}
|
|
auto dstW = result->weightFloat.get() + o * partWeightSize;
|
|
auto srcW = result->weight.get() + o * partWeightSize;
|
|
for (int v=0; v < partWeightSize; ++v) {
|
|
dstW[v] = (float)srcW[v] * alpha + min;
|
|
}
|
|
}
|
|
result->weight.release();
|
|
result->alpha.release();
|
|
}
|
|
return result;
|
|
}
|
|
|
|
void ConvolutionCommon::getConvParameters(std::shared_ptr<Int8Common> *quanCommon, Backend* backend, const MNN::Op *op, const float** originWeight, int* originWeightSize) {
|
|
auto conv2d = op->main_as_Convolution2D();
|
|
*originWeight = nullptr;
|
|
*originWeightSize = 0;
|
|
if (nullptr != conv2d->quanParameter()) {
|
|
bool forceFloat = conv2d->quanParameter()->index() != nullptr;
|
|
*quanCommon = load(op, backend, forceFloat);
|
|
*originWeight = (*quanCommon)->weightFloat.get();
|
|
*originWeightSize = (*quanCommon)->weightFloat.size();
|
|
}
|
|
if (*originWeight == nullptr) {
|
|
*originWeight = conv2d->weight()->data();
|
|
*originWeightSize = conv2d->weight()->size();
|
|
}
|
|
}
|
|
|
|
bool ConvolutionCommon::getConvInt8Parameters(const MNN::Op* op, std::shared_ptr<Int8Common>& quanCommon, Backend* backend,
|
|
const int8_t*& weight, int& weightSize, float*& scale, int32_t*& bias, int32_t*& weightQuantZeroPoint) {
|
|
// Compability for old quant model
|
|
auto conv2d = op->main_as_Convolution2D();
|
|
int outputCount = conv2d->common()->outputCount();
|
|
weightSize = 0;
|
|
// fix xcode UndefinedBehaviorSanitizer
|
|
if (conv2d->symmetricQuan() && conv2d->symmetricQuan()->weight() != nullptr) {
|
|
weight = conv2d->symmetricQuan()->weight()->data();
|
|
weightSize = conv2d->symmetricQuan()->weight()->size();
|
|
}
|
|
if (conv2d->quanParameter() && (conv2d->quanParameter()->buffer() || conv2d->external())) { // int8 weight
|
|
quanCommon = ConvolutionCommon::load(op, backend, false, true);
|
|
MNN_ASSERT(quanCommon != nullptr);
|
|
weight = quanCommon->weight.get();
|
|
weightSize = quanCommon->weight.size();
|
|
}
|
|
if (weight == nullptr) {
|
|
MNN_ERROR("ConvolutionCommon::getConvInt8Parameters: No weight data!");
|
|
return false;
|
|
}
|
|
bool weightAsy = false;
|
|
if (quanCommon && quanCommon->asymmetric) {
|
|
weightAsy = true;
|
|
}
|
|
|
|
if (conv2d->symmetricQuan() && conv2d->symmetricQuan()->bias() && conv2d->symmetricQuan()->scale()) {
|
|
// Compability for old model
|
|
MNN_ASSERT(conv2d->symmetricQuan()->bias()->size() == outputCount && conv2d->symmetricQuan()->scale()->size() == outputCount);
|
|
::memcpy(bias, conv2d->symmetricQuan()->bias()->data(), outputCount * sizeof(int32_t));
|
|
::memcpy(scale, conv2d->symmetricQuan()->scale()->data(), outputCount * sizeof(float));
|
|
return true;
|
|
}
|
|
if (conv2d->bias()) {
|
|
::memcpy(bias, conv2d->bias()->data(), outputCount * sizeof(float));
|
|
}
|
|
if (conv2d->quanParameter() && conv2d->quanParameter()->alpha()) {
|
|
auto alphaAndBeta = conv2d->quanParameter()->alpha()->data();
|
|
int quantCount = conv2d->quanParameter()->alpha()->size();
|
|
if (false == weightAsy) { // symmetric quant
|
|
::memcpy(scale, conv2d->quanParameter()->alpha()->data(), quantCount * sizeof(float));
|
|
} else if (true == weightAsy) { // asymmetric
|
|
// int ocx2 = 2 * outputCount;
|
|
int scaleSize = quantCount / 2;
|
|
float clampMin = conv2d->quanParameter()->aMin() == 0 ? -128 : conv2d->quanParameter()->aMin();
|
|
for (int i = 0; i < scaleSize; ++i) {
|
|
weightQuantZeroPoint[i] = static_cast<int32_t>(roundf((-1) * alphaAndBeta[2 * i] / alphaAndBeta[2 * i + 1]) + clampMin);
|
|
scale[i] = alphaAndBeta[2 * i + 1];
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
MNN_ERROR("ConvolutionCommon::getConvInt8Parameters: No bias & scale data!");
|
|
return false;
|
|
}
|
|
|
|
std::pair<int, int> ConvolutionCommon::convolutionPad(const Tensor *input, const Tensor *output,
|
|
const Convolution2DCommon *mCommon) {
|
|
if (mCommon->padMode() == PadMode_SAME) {
|
|
int kernelWidthSize = (mCommon->kernelX() - 1) * mCommon->dilateX() + 1;
|
|
int kernelHeightSize = (mCommon->kernelY() - 1) * mCommon->dilateY() + 1;
|
|
|
|
int padNeededWidth = (output->width() - 1) * mCommon->strideX() + kernelWidthSize - input->width();
|
|
int padNeededHeight = (output->height() - 1) * mCommon->strideY() + kernelHeightSize - input->height();
|
|
auto mPadX = padNeededWidth / 2;
|
|
auto mPadY = padNeededHeight / 2;
|
|
return std::make_pair(mPadX, mPadY);
|
|
}
|
|
auto mPadX = mCommon->padX();
|
|
auto mPadY = mCommon->padY();
|
|
if (nullptr != mCommon->pads() && mCommon->pads()->size() >= 2) {
|
|
mPadX = mCommon->pads()->data()[1];
|
|
mPadY = mCommon->pads()->data()[0];
|
|
}
|
|
return std::make_pair(mPadX, mPadY);
|
|
}
|
|
|
|
std::tuple<int, int, int, int> ConvolutionCommon::convolutionPadFull(const Tensor* input, const Tensor* output,
|
|
const Convolution2DCommon* common) {
|
|
auto pad = convolutionPad(input, output, common);
|
|
int iw = input->width();
|
|
int ih = input->height();
|
|
int ow = output->width();
|
|
int oh = output->height();
|
|
|
|
int right = (ow - 1) * common->strideX() + (common->kernelX() - 1) * common->dilateX() - pad.first;
|
|
int padRight = 0;
|
|
if (right >= iw) {
|
|
padRight = right - iw + 1;
|
|
}
|
|
int bottom = (oh - 1) * common->strideY() + (common->kernelY() - 1) * common->dilateY() - pad.second;
|
|
int padBottom = 0;
|
|
if (bottom >= ih) {
|
|
padBottom = bottom - ih + 1;
|
|
}
|
|
return std::make_tuple(pad.first, pad.second, padRight, padBottom);
|
|
}
|
|
|
|
std::pair<int, int> ConvolutionCommon::convolutionTransposePad(const Tensor *input, const Tensor *output,
|
|
const Convolution2DCommon *mCommon) {
|
|
if (mCommon->padMode() == PadMode_SAME) {
|
|
const int outputWidth = output->width();
|
|
const int outputHeight = output->height();
|
|
|
|
const int outputWidthPadded = (input->width() - 1) * mCommon->strideX() + mCommon->kernelX();
|
|
const int outputHeightPadded = (input->height() - 1) * mCommon->strideY() + mCommon->kernelY();
|
|
|
|
const int padNeededWidth = outputWidthPadded - outputWidth;
|
|
const int padNeededHeight = outputHeightPadded - outputHeight;
|
|
|
|
auto mPadX = padNeededWidth / 2;
|
|
auto mPadY = padNeededHeight / 2;
|
|
return std::make_pair(mPadX, mPadY);
|
|
}
|
|
auto mPadX = mCommon->padX();
|
|
auto mPadY = mCommon->padY();
|
|
if (nullptr != mCommon->pads() && mCommon->pads()->size() >= 2) {
|
|
mPadY = mCommon->pads()->data()[0];
|
|
mPadX = mCommon->pads()->data()[1];
|
|
}
|
|
return std::make_pair(mPadX, mPadY);
|
|
}
|
|
|
|
} // namespace MNN
|