ComputationalGraphMatMul.h

#pragma once
 
#if _MSC_VER >= 1900 && defined(_M_X64)
 
#include "ComputationalGraph.h"
#include "Tensor.h"
 
namespace FLImaging
{
    namespace AI
    {
        template<typename T>
        class CCuda_ComputationalGraphMatMul_Cublas;
 
        template <typename T>
        class FL_EXPORT CComputationalGraphMatMul : public CComputationalGraph<T>
        {
        private:
            CComputationalGraphMatMul();
 
        protected:
            CComputationalGraphMatMul(const CComputationalGraphMatMul<T>& cg);
 
        public:
            CComputationalGraphMatMul(const CComputationalBase<T>& cbOperand1, const CComputationalBase<T>& cbOperand2);
            CComputationalGraphMatMul(const CComputationalBase<T>& cbOperand1, const CTensor<T>& tsrOperand2);
            CComputationalGraphMatMul(const CTensor<T>& tsrOperand1, const CComputationalBase<T>& cbOperand2);
            CComputationalGraphMatMul(const CTensor<T>& tsrOperand1, const CTensor<T>& tsrOperand2);
            virtual ~CComputationalGraphMatMul();
 
            virtual CTensor<T>& Forward() override;
            virtual CTensor<T>* Backward() override;
            virtual CComputationalBase<T>* Clone() const override;
            virtual const CResult PrintNodeParamInfo() const override;
 
            virtual const CResult GetBinaryData(Base::CFLData& fldBinary, bool bSuperClass = false, int32_t i32Version = -1, bool bDumpMode = false) const override;
            virtual const CResult GetBinaryData(Base::CFLData* pFldBinary, bool bSuperClass = false, int32_t i32Version = -1, bool bDumpMode = false) const override;
 
            virtual const CResult SetBinaryData(const Base::CFLData& fldBinary, int64_t* pI64Offset = nullptr) override;
            virtual const CResult SetBinaryData(const Base::CFLData* pFldBinary, int64_t* pI64Offset = nullptr) override;
 
            virtual const std::vector<int64_t>& GetEstimatedShape(bool bRecursive = true) const override;
 
            virtual int64_t GetRequiredDedicatedMemory(bool bTraining = false, bool bRecursively = true, int64_t i64BatchSize = 1) const override;
            virtual int64_t GetRequiredTemporaryMemory(bool bTraining = false, bool bRecursively = true, int64_t i64BatchSize = 1, int64_t i64MemoryIndex = 0) const override;
 
 
            DeclareGetClassType();
            SupportToDuplicateObjectWithoutCreateNewObject(CComputationalGraphMatMul, *this);
 
        protected:
            CTensor<T> m_tsrOprTranspose0;
            CTensor<T> m_tsrOprTranspose1;
 
        protected:
            #ifdef CUBLAS_MODE
            CCuda_ComputationalGraphMatMul_Cublas<T>* m_pCublas;
            #endif
 
        public:
            DeclareGetSignletonObject(CComputationalGraphMatMul);
        };
 
        #define CCGFMatMul(...) (*(new CComputationalGraphMatMul<float>(__VA_ARGS__)))
        #define CCGDMatMul(...) (*(new CComputationalGraphMatMul<double>(__VA_ARGS__)))
 
        #define CCGTMatMul(T, ...) (*(new CComputationalGraphMatMul<T>(__VA_ARGS__)))
    }
}
 
#endif