trmm: Triangular matrix multiply

Functions
template<Target target = Target::HostTask, typename scalar_t >
void	slate::internal::trmm (Side side, scalar_t alpha, TriangularMatrix< scalar_t > &&A, Matrix< scalar_t > &&B, int priority, int64_t queue_index)
	Triangular matrix multiply.
template<typename scalar_t >
void	slate::internal::trmm (internal::TargetType< Target::HostTask >, Side side, scalar_t alpha, TriangularMatrix< scalar_t > &A, Matrix< scalar_t > &B, int priority, int64_t queue_index)
	Triangular matrix multiply.
template<typename scalar_t >
void	slate::internal::trmm (internal::TargetType< Target::HostNest >, Side side, scalar_t alpha, TriangularMatrix< scalar_t > &A, Matrix< scalar_t > &B, int priority, int64_t queue_index)
	Triangular matrix multiply.
template<typename scalar_t >
void	slate::internal::trmm (internal::TargetType< Target::HostBatch >, Side side, scalar_t alpha, TriangularMatrix< scalar_t > &A, Matrix< scalar_t > &B, int priority, int64_t queue_index)
	Triangular matrix multiply.
template<typename scalar_t >
void	slate::internal::trmm (internal::TargetType< Target::Devices >, Side side, scalar_t alpha, TriangularMatrix< scalar_t > &A, Matrix< scalar_t > &B, int priority, int64_t queue_index)
	Triangular matrix multiply.
template<Target target = Target::HostTask, typename scalar_t >
void	slate::work::trmm (Side side, scalar_t alpha, TriangularMatrix< scalar_t > A, Matrix< scalar_t > B, uint8_t *bcast, uint8_t *gemm, int64_t lookahead)
	Triangular matrix multiply.

Detailed Description

Function Documentation

trmm() [1/6]

template<typename scalar_t >

void slate::internal::trmm	(	internal::TargetType< Target::Devices >	,
		Side	side,
		scalar_t	alpha,
		TriangularMatrix< scalar_t > &	A,
		Matrix< scalar_t > &	B,
		int	priority,
		int64_t	queue_index
	)

Triangular matrix multiply.

GPU device batched cuBLAS implementation.

trmm() [2/6]

template<typename scalar_t >

void slate::internal::trmm	(	internal::TargetType< Target::HostBatch >	,
		Side	side,
		scalar_t	alpha,
		TriangularMatrix< scalar_t > &	A,
		Matrix< scalar_t > &	B,
		int	priority,
		int64_t	queue_index
	)

Triangular matrix multiply.

Host batched implementation.

trmm() [3/6]

template<typename scalar_t >

void slate::internal::trmm	(	internal::TargetType< Target::HostNest >	,
		Side	side,
		scalar_t	alpha,
		TriangularMatrix< scalar_t > &	A,
		Matrix< scalar_t > &	B,
		int	priority,
		int64_t	queue_index
	)

Triangular matrix multiply.

Host nested OpenMP implementation.

trmm() [4/6]

template<typename scalar_t >

void slate::internal::trmm	(	internal::TargetType< Target::HostTask >	,
		Side	side,
		scalar_t	alpha,
		TriangularMatrix< scalar_t > &	A,
		Matrix< scalar_t > &	B,
		int	priority,
		int64_t	queue_index
	)

Triangular matrix multiply.

Host OpenMP task implementation.

trmm() [5/6]

template<Target target = Target::HostTask, typename scalar_t >

void slate::internal::trmm	(	Side	side,
		scalar_t	alpha,
		TriangularMatrix< scalar_t > &&	A,
		Matrix< scalar_t > &&	B,
		int	priority,
		int64_t	queue_index
	)

Triangular matrix multiply.

Dispatches to target implementations.

trmm() [6/6]

template<Target target = Target::HostTask, typename scalar_t >

void slate::work::trmm	(	Side	side,
		scalar_t	alpha,
		TriangularMatrix< scalar_t >	A,
		Matrix< scalar_t >	B,
		uint8_t *	bcast,
		uint8_t *	gemm,
		int64_t	lookahead
	)

Triangular matrix multiply.

Note A and B are passed by value, so we can transpose if needed (for side = right) without affecting caller.

Template Parameters

target	One of HostTask, HostNest, HostBatch, Devices.
scalar_t	One of float, double, std::complex<float>, std::complex<double>.

Parameters

[in]	side	Whether A appears on the left or on the right of B: Side::Left: \(B = \alpha A B\) Side::Right: \(B = \alpha B A\)
[in]	alpha	The scalar alpha.
[in]	A	If side = left, the m-by-m triangular matrix A; if side = right, the n-by-n triangular matrix A.
[in,out]	B	On entry, the m-by-n matrix B. On exit, overwritten by the result \(\alpha A B\) or \(\alpha B A\).
[in]	bcast	A raw pointer to a dummy vector data.. The dummy vector is used for OpenMP dependencies tracking, not based on the actual data. Entries in the dummy vector represent each column of matrix \(A\) and each row of matrix \(B\). The size of bcast should be number of block columns of matrix \(A\).
[in]	gemm	A raw pointer to a dummy vector data. The dummy vector is used for OpenMP dependencies tracking, not based on the actual data. Entries in the dummy vector represent each column of matrix \(A\) and each row of matrix \(B\). The size of gemm should be number of block columns of matrix \(A\).
[in]	lookahead	Number of blocks to overlap communication and computation. lookahead >= 0. Default 1.