Internal

Functions
template<Target target = Target::HostTask, typename scalar_t >
void	slate::internal::geqrf (Matrix< scalar_t > &&A, Matrix< scalar_t > &&T, std::vector< scalar_t * > dwork_array, size_t work_size, int64_t ib, int max_panel_threads, int priority)
	QR factorization of a column of tiles.
template<typename scalar_t >
void	slate::internal::geqrf (internal::TargetType< Target::HostTask >, Matrix< scalar_t > &A, Matrix< scalar_t > &T, std::vector< scalar_t * > dwork_array, size_t work_size, int64_t ib, int max_panel_threads, int priority)
	QR factorization of a column of tiles, HostTask implementation.
template<typename scalar_t >
void	slate::internal::geqrf (internal::TargetType< Target::HostNest >, Matrix< scalar_t > &A, Matrix< scalar_t > &T, std::vector< scalar_t * > dwork_array, size_t work_size, int64_t ib, int max_panel_threads, int priority)
	QR factorization of a column of tiles, HostNest implementation.
template<typename scalar_t >
void	slate::internal::geqrf (internal::TargetType< Target::HostBatch >, Matrix< scalar_t > &A, Matrix< scalar_t > &T, std::vector< scalar_t * > dwork_array, size_t work_size, int64_t ib, int max_panel_threads, int priority)
	QR factorization of a column of tiles, HostBatch implementation.
template<typename scalar_t >
void	slate::internal::geqrf (internal::TargetType< Target::Devices >, Matrix< scalar_t > &A, Matrix< scalar_t > &T, std::vector< scalar_t * > dwork_array, size_t work_size, int64_t ib, int max_panel_threads, int priority)
	QR factorization of a column of tiles, device implementation.
template<Target target = Target::HostTask, typename scalar_t >
void	slate::internal::ttmqr (Side side, Op op, Matrix< scalar_t > &&A, Matrix< scalar_t > &&T, Matrix< scalar_t > &&C, int tag)
	Distributed multiply matrix by Q from QR triangle-triangle factorization of column of tiles.
template<typename scalar_t >
void	slate::internal::ttmqr (internal::TargetType< Target::HostTask >, Side side, Op op, Matrix< scalar_t > &A, Matrix< scalar_t > &T, Matrix< scalar_t > &C, int tag)
	Distributed multiply matrix by Q from QR triangle-triangle factorization of column of tiles, host implementation.
template<Target target = Target::HostTask, typename scalar_t >
void	slate::internal::ttqrt (Matrix< scalar_t > &&A, Matrix< scalar_t > &&T)
	Distributed QR triangle-triangle factorization of column of tiles.
template<typename scalar_t >
void	slate::internal::ttqrt (internal::TargetType< Target::HostTask >, Matrix< scalar_t > &A, Matrix< scalar_t > &T)
	Distributed QR triangle-triangle factorization, host implementation.
template<Target target = Target::HostTask, typename scalar_t >
void	slate::internal::unmqr (Side side, Op op, Matrix< scalar_t > &&V, Matrix< scalar_t > &&T, Matrix< scalar_t > &&C, Matrix< scalar_t > &&W, int priority, int64_t queue_index)
	Multiply matrix by Q from local QR factorization.
template<Target target, typename scalar_t >
void	slate::internal::unmqr (internal::TargetType< target >, Side side, Op op, Matrix< scalar_t > V, Matrix< scalar_t > &T, Matrix< scalar_t > &C, Matrix< scalar_t > &W, int priority, int64_t queue_index)
	Multiply matrix by Q from local QR factorization.

Detailed Description

Function Documentation

geqrf() [1/3]

template<typename scalar_t >

void slate::internal::geqrf	(	internal::TargetType< Target::HostBatch >	,
		Matrix< scalar_t > &	A,
		Matrix< scalar_t > &	T,
		std::vector< scalar_t * >	dwork_array,
		size_t	work_size,
		int64_t	ib,
		int	max_panel_threads,
		int	priority
	)

QR factorization of a column of tiles, HostBatch implementation.

Forwarding to HostTask as there is no implementation currently.

geqrf() [2/3]

template<typename scalar_t >

void slate::internal::geqrf	(	internal::TargetType< Target::HostNest >	,
		Matrix< scalar_t > &	A,
		Matrix< scalar_t > &	T,
		std::vector< scalar_t * >	dwork_array,
		size_t	work_size,
		int64_t	ib,
		int	max_panel_threads,
		int	priority
	)

QR factorization of a column of tiles, HostNest implementation.

Forwarding to HostTask as there is no implementation currently.

geqrf() [3/3]

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

void slate::internal::geqrf	(	Matrix< scalar_t > &&	A,
		Matrix< scalar_t > &&	T,
		std::vector< scalar_t * >	dwork_array,
		size_t	work_size,
		int64_t	ib,
		int	max_panel_threads,
		int	priority
	)

QR factorization of a column of tiles.

Dispatches to target implementations.

ttmqr()

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

void slate::internal::ttmqr	(	Side	side,
		Op	op,
		Matrix< scalar_t > &&	A,
		Matrix< scalar_t > &&	T,
		Matrix< scalar_t > &&	C,
		int	tag
	)

Distributed multiply matrix by Q from QR triangle-triangle factorization of column of tiles.

Dispatches to target implementations. todo: This assumes A and T have already been communicated as needed. However, it necessarily handles communication for C. Tag is used in geqrf to differentiate communication for look-ahead panel from rest of trailing matrix.

ttqrt() [1/2]

template<typename scalar_t >

void slate::internal::ttqrt	(	internal::TargetType< Target::HostTask >	,
		Matrix< scalar_t > &	A,
		Matrix< scalar_t > &	T
	)

Distributed QR triangle-triangle factorization, host implementation.

Assumes panel tiles reside on host.

ttqrt() [2/2]

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

void slate::internal::ttqrt	(	Matrix< scalar_t > &&	A,
		Matrix< scalar_t > &&	T
	)

Distributed QR triangle-triangle factorization of column of tiles.

Each rank has one triangular tile, the result of local geqrf panel. Dispatches to target implementations.

unmqr()

template<Target target, typename scalar_t >

void slate::internal::unmqr	(	internal::TargetType< target >	,
		Side	side,
		Op	op,
		Matrix< scalar_t >	V,
		Matrix< scalar_t > &	T,
		Matrix< scalar_t > &	C,
		Matrix< scalar_t > &	W,
		int	priority,
		int64_t	queue_index
	)

Multiply matrix by Q from local QR factorization.

C = op(Q) C for side = left, or C = C op(Q) for side = right. Assumes V and T are each a single block-column. Assumes W and C have the same dimensions and distribution. This corresponds to larfb( ..., direct=Forward, storev=Columnwise, ... ). This does not include applying the distributed triangle-triangle reductions.