MNN/source/backend/opencl/execution/buffer/SelfAttentionBufExecution.hpp

77 lines
3.0 KiB
C++

//
// FmhaV2Execution.hpp
// MNN
//
// Created by MNN on 2024/06/03.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifdef MNN_SUPPORT_TRANSFORMER_FUSE
#ifndef SelfAttentionBufExecution_hpp
#define SelfAttentionBufExecution_hpp
#include "backend/opencl/execution/image/CommonExecution.hpp"
namespace MNN {
namespace OpenCL {
class SelfAttentionBufImpl {
public:
SelfAttentionBufImpl(const MNN::Op *op, Backend *backend);
~SelfAttentionBufImpl() = default;
ErrorCode onResize(Backend *backend, const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs);
ErrorCode onExecute(Backend *backend, const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs);
private:
int getLocalSize(int size, int maxGroupSize);
float mScale;
int mQseqSplitNum = 1;
std::shared_ptr<Tensor> mTempQ, mTempK, mTempQK, mTempSoftMax, mTempV, mTempQKV;
std::shared_ptr<Tensor> mTempTrans;
int mNumHead = 0, mHeadDim = 0;
std::vector<std::shared_ptr<KernelWrap>> mKernel_split;
std::vector<std::shared_ptr<KernelWrap>> mKernel_qk;
std::vector<std::shared_ptr<KernelWrap>> mKernel_softmax;
std::vector<std::shared_ptr<KernelWrap>> mKernel_qkv;
std::vector<std::shared_ptr<KernelWrap>> mKernel_clip;
std::vector<std::shared_ptr<KernelWrap>> mKernel_trans;
std::vector<std::vector<uint32_t>> mGlobalWorkSizeSplit;
std::vector<std::vector<uint32_t>> mLocalWorkSizeSplit;
std::vector<std::vector<uint32_t>> mGlobalWorkSizeClip;
std::vector<std::vector<uint32_t>> mLocalWorkSizeClip;
std::vector<std::vector<uint32_t>> mGlobalWorkSizeQk;
std::vector<std::vector<uint32_t>> mLocalWorkSizeQk;
std::vector<std::vector<uint32_t>> mGlobalWorkSizeSoftMax;
std::vector<std::vector<uint32_t>> mLocalWorkSizeSoftMax;
std::vector<std::vector<uint32_t>> mGlobalWorkSizeQkv;
std::vector<std::vector<uint32_t>> mLocalWorkSizeQkv;
std::vector<std::vector<uint32_t>> mGlobalWorkSizeTrans;
std::vector<std::vector<uint32_t>> mLocalWorkSizeTrans;
uint32_t mMaxWorkGroupSize;
OpenCLBackend *mOpenCLBackend;
size_t mQkGlobal_size[3];
int mSoftmaxShape[4];
int mByte = 4;
cl_recording_qcom mRecording{NULL};
};
class SelfAttentionBufExecution : public CommonExecution {
public:
SelfAttentionBufExecution(const MNN::Op *op, Backend *backend);
SelfAttentionBufExecution(std::shared_ptr<SelfAttentionBufImpl> impl, const MNN::Op *op, Backend *backend);
virtual ~SelfAttentionBufExecution() = 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;
virtual bool onClone(Backend* bn, const Op* op, Execution** dst) override;
private:
std::shared_ptr<SelfAttentionBufImpl> mImpl;
};
} // namespace OpenCL
} // namespace MNN
#endif /* SelfAttentionBufExecution_hpp */
#endif/* MNN_SUPPORT_TRANSFORMER_FUSE */