MNN/source/backend/cpu/CPUSoftMaxInt8.hpp

40 lines
1.1 KiB
C++

//
// CPUSoftMaxInt8.hpp
// MNNCPU
//
// Created by MNN on 2023/4/22.
//
#ifndef CPUSoftMaxInt8_hpp
#define CPUSoftMaxInt8_hpp
#include "core/Execution.hpp"
#include <math.h>
namespace MNN {
class CPUSoftmaxInt8 : public Execution {
public:
CPUSoftmaxInt8(Backend *backend, int axis);
virtual ~CPUSoftmaxInt8() = default;
virtual ErrorCode onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) override;
virtual ErrorCode onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) override;
static Execution* create(const MNN::Op *op, Backend *backend);
void QuantizedSoftmax(const uint8_t *inputData, int outerSize, int targetAxis, int32_t inputBetaMultiplier,
int32_t inputBetaLeftShift, uint8_t *output_data, int threadNum);
private:
int32_t mInputMultiplier;
int mInputLeftShift;
int mDiffMin;
int mAxis;
int mInside;
int mOutside;
int mTargetAxis;
Tensor mStorage;
Tensor mTempOutput;
bool mNeedUnpackC4;
};
}
#endif /* CPUSoftMaxInt8_hpp */