Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision > Struct Template Reference

RAII wrapper owning cuBLASLt descriptors and the selected heuristic algorithm. More...

Public Member Functions
	CublasLtMatMulPlan ()=default
	CublasLtMatMulPlan (const CublasLtMatMulPlan &)=delete
	CublasLtMatMulPlan (CublasLtMatMulPlan &&other) noexcept
	~CublasLtMatMulPlan ()
bool	isValid () const
CublasLtMatMulPlan &	operator= (const CublasLtMatMulPlan &)=delete
CublasLtMatMulPlan &	operator= (CublasLtMatMulPlan &&other) noexcept

Public Attributes
cublasLtMatmulAlgo_t	algorithm {}
bool	has_algorithm { false }
bool	has_bias_epilogue { false }
cublasLtMatrixLayout_t	layoutA { nullptr }
cublasLtMatrixLayout_t	layoutB { nullptr }
cublasLtMatrixLayout_t	layoutC { nullptr }
cublasLtMatmulDesc_t	matmul_desc { nullptr }
cublasLtMatmulPreference_t	preference { nullptr }

Detailed Description

template<typename TComputePrecision>
struct Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >

RAII wrapper owning cuBLASLt descriptors and the selected heuristic algorithm.

Owns:

• matmul_desc: operation descriptor (transpose flags, epilogue, bias pointer)
• layoutA, layoutB, layoutC: matrix memory layouts
• preference: algorithm preference used during heuristic search
• algorithm: selected algorithm (present when has_algorithm is true)

Non-copyable; move-only.

Constructor & Destructor Documentation

CublasLtMatMulPlan() [1/3]

template<typename TComputePrecision>

Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >::CublasLtMatMulPlan ( )

exportdefault

Here is the caller graph for this function:

~CublasLtMatMulPlan()

template<typename TComputePrecision>

Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >::~CublasLtMatMulPlan ( )

inlineexport

CublasLtMatMulPlan() [2/3]

template<typename TComputePrecision>

Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >::CublasLtMatMulPlan ( const CublasLtMatMulPlan< TComputePrecision > & )

exportdelete

Here is the call graph for this function:

CublasLtMatMulPlan() [3/3]

template<typename TComputePrecision>

Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >::CublasLtMatMulPlan ( CublasLtMatMulPlan< TComputePrecision > && other )

inlineexportnoexcept

Here is the call graph for this function:

Member Function Documentation

isValid()

template<typename TComputePrecision>

bool Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >::isValid ( ) const

inlineexport

Here is the caller graph for this function:

operator=() [1/2]

template<typename TComputePrecision>

CublasLtMatMulPlan & Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >::operator= ( const CublasLtMatMulPlan< TComputePrecision > & )

exportdelete

Here is the call graph for this function:

operator=() [2/2]

template<typename TComputePrecision>

CublasLtMatMulPlan & Mila::Dnn::Compute::Cuda::CublasLtMatMulPlan< TComputePrecision >::operator= ( CublasLtMatMulPlan< TComputePrecision > && other )

inlineexportnoexcept

Here is the call graph for this function:

---

The documentation for this struct was generated from the following file:

• /__w/Mila/Mila/Mila/Src/Dnn/Compute/Devices/Cuda/Operations/Common/CublasLtPlan.ixx