QR: lower-level API

Overview

Workspace queries, argument validation and the CTA ormqr kernel behind the QR entry points.

Functions

ortho_buffer_size() [1/2]

template<Backend B, typename T>

size_t batchlas::ortho_buffer_size	(	Queue &	ctx,
		const MatrixView< T, MatrixFormat::Dense > &	A,
		Transpose	transA,
		OrthoAlgorithm	algo = OrthoAlgorithm::Chol2 )

Required workspace, in bytes, for ortho(); arguments as for the call itself.

ortho_buffer_size() [2/2]

template<Backend B, typename T>

size_t batchlas::ortho_buffer_size	(	Queue &	ctx,
		const MatrixView< T, MatrixFormat::Dense > &	A,
		const MatrixView< T, MatrixFormat::Dense > &	M,
		Transpose	transA,
		Transpose	transM,
		OrthoAlgorithm	algo = OrthoAlgorithm::Chol2,
		size_t	iterations = 2 )

Required workspace, in bytes, for the external-basis ortho(); same arguments.

ormqx_cta()

template<Backend B, typename T>

Event batchlas::ormqx_cta	(	Queue &	ctx,
		const MatrixView< T, MatrixFormat::Dense > &	a_in,
		const VectorView< T > &	tau_in,
		const MatrixView< T, MatrixFormat::Dense > &	c_in,
		Uplo	factorization,
		Side	side,
		Transpose	trans,
		int32_t	k,
		const Span< std::byte > &	ws,
		size_t	cta_wg_size_multiplier = 1 )

CTA-optimized application of Q from a QR/QL factorization (ORMQx/UNMQx semantics) for very small matrices: applies the implicit Q of the Householder reflectors (A, TAU) from GEQRF (Upper) or GEQLF (Lower) to C.

Overwrites c_in with op(Q) * C (Side::Left) or C * op(Q) (Side::Right), using the first k reflectors.

geqrf_validate_params()

template<typename T>

void batchlas::geqrf_validate_params ( const MatrixView< T, MatrixFormat::Dense > & A )

inline

Validates the arguments of the positional geqrf() entry point.

Checks only non-negative extents. Rectangular A of either orientation is valid; the length of tau is checked by the option overloads.

Exceptions

batchlas::invalid_argument

on negative extents

Definition at line 57 of file geqrf.hh.

geqrf_buffer_size()

template<Backend B, typename T>

size_t batchlas::geqrf_buffer_size	(	Queue &	ctx,
		const MatrixView< T, MatrixFormat::Dense > &	A,
		Span< T >	tau )

Workspace, in bytes, that geqrf() needs for this shape on this queue.

orgqr_validate_params()

template<typename T>

void batchlas::orgqr_validate_params ( const MatrixView< T, MatrixFormat::Dense > & A )

inline

Validates the arguments of the positional orgqr() entry point.

Checks only non-negative extents. n <= m is not checked: an n > m view is served by the vendor kernel, the only one whose can_run admits it (src/ops/orgqr/orgqr.cc). The length of tau is checked by the option overloads.

Exceptions

batchlas::invalid_argument

on negative extents

Definition at line 55 of file orgqr.hh.

orgqr_buffer_size()

template<Backend B, typename T>

size_t batchlas::orgqr_buffer_size	(	Queue &	ctx,
		const MatrixView< T, MatrixFormat::Dense > &	A,
		Span< T >	tau )

Workspace, in bytes, that orgqr() needs for this shape on this queue.

ormqr_buffer_size()

template<Backend B, typename T>

size_t batchlas::ormqr_buffer_size	(	Queue &	ctx,
		const MatrixView< T, MatrixFormat::Dense > &	A,
		const MatrixView< T, MatrixFormat::Dense > &	C,
		Side	side,
		Transpose	trans,
		Span< T >	tau,
		int32_t	block_size_hint = 0 )

Workspace, in bytes, that ormqr() needs for these operands on this queue.

Makes the same kernel choice and block width as ormqr() from the same inputs, so the result is valid for a call with the same arguments and block_size_hint. evidence: docs/perf/qr.md#ormqr-one-route-resolution-for-the-call-and-its-size-query