wehub-resource-sync
79 lines
3.1 KiB
C++
79 lines
3.1 KiB
C++
//
|
|
// ConvolutionCommon.hpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2020/03/02.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#ifndef ConvolutionCommon_hpp
|
|
#define ConvolutionCommon_hpp
|
|
#include "AutoStorage.h"
|
|
#include "Execution.hpp"
|
|
#include "MNN_generated.h"
|
|
namespace MNN {
|
|
class MNN_PUBLIC ConvolutionCommon : public Execution {
|
|
public:
|
|
struct Int8Common {
|
|
AutoStorage<int8_t> weight;
|
|
// alpha (fp32) and alphaHalf (fp16 raw bits) are alternative views of the same scale/zero
|
|
// data. alphaIsFp16 records the canonical (disk-side) form. Use getAlphaFloat() /
|
|
// getAlphaHalf() to access; they lazily fill the alternate view on first request.
|
|
AutoStorage<float> alpha;
|
|
AutoStorage<int16_t> alphaHalf;
|
|
AutoStorage<float> weightFloat;
|
|
const IDSTQuan* quan;
|
|
bool asymmetric;
|
|
bool alphaIsFp16 = false;
|
|
std::vector<int8_t> weightMap;
|
|
bool canUseInt4 = false;
|
|
bool canUseInt3 = false;
|
|
bool canUseInt2 = false;
|
|
Backend* backend = nullptr;
|
|
int originBits = 8;
|
|
int alphaSize;
|
|
// Lazy accessors. Element count is alphaSize.
|
|
const float* getAlphaFloat();
|
|
const int16_t* getAlphaHalf();
|
|
};
|
|
static std::shared_ptr<Int8Common> load(const Op* op, Backend* backend = nullptr, bool forceFloat = false, bool forceInt8 = false, void* weightPtr = nullptr);
|
|
// if can not get quant bits, return 0
|
|
static int getQuantBitFromExternalFile(const Op* op);
|
|
static void getConvParameters(std::shared_ptr<ConvolutionCommon::Int8Common> *quanCommon, Backend* backend, const MNN::Op *op, const float** originWeight, int* originWeightSize);
|
|
static bool getConvInt8Parameters(const MNN::Op* op, std::shared_ptr<Int8Common>& quanCommon, Backend* backend,
|
|
const int8_t*& weight, int& weightSize, float* scale, int32_t* bias, int ocUpHp);
|
|
|
|
// Return padX, padY
|
|
static std::pair<int, int> convolutionPad(const Tensor* input, const Tensor* output,
|
|
const Convolution2DCommon* common);
|
|
// Return padLeft, padTop, padRight, padBottom
|
|
static std::tuple<int, int, int, int> convolutionPadFull(const Tensor* input, const Tensor* output,
|
|
const Convolution2DCommon* common);
|
|
static std::pair<int, int> convolutionTransposePad(const Tensor* input, const Tensor* output,
|
|
const Convolution2DCommon* common);
|
|
struct Im2ColParameter {
|
|
int32_t padX;
|
|
int32_t padY;
|
|
int32_t dilateX;
|
|
int32_t dilateY;
|
|
int32_t strideX;
|
|
int32_t strideY;
|
|
int32_t kernelX;
|
|
int32_t kernelY;
|
|
int32_t icDiv4;
|
|
int32_t kernelCountUnit;
|
|
int32_t iw;
|
|
int32_t ih;
|
|
int32_t ow;
|
|
int32_t oh;
|
|
int32_t srcZStep;
|
|
int32_t srcYStep;
|
|
int32_t packCUnit;
|
|
int32_t destICStride;
|
|
int32_t ic; // ic packed by LP, used by im2col dst data.
|
|
int32_t icup4; // ic packed by LP, used by im2col src data.
|
|
};
|
|
};
|
|
} // namespace MNN
|
|
#endif
|