BLAS level 2

Overview

Batched matrix-vector operations.

Functions

gemv()

template<Backend B, typename T>

Event batchlas::gemv	(	Queue &	ctx,
		const MatrixView< T, MatrixFormat::Dense > &	A,
		const VectorView< T > &	X,
		const VectorView< T > &	Y,
		T	alpha,
		T	beta,
		Transpose	transA )

Batched general matrix-vector multiply.

For every batch item \(b\) computes

\[ y_b := \alpha \, \mathrm{op}(A_b) \, x_b + \beta \, y_b \]

where \(\mathrm{op}(A)\) is \(A\), \(A^T\) or \(A^H\) according to transA. For an m x n A, X has length n and Y length m under NoTrans, and the other way round under Trans / ConjTrans. The vectors' inc (element spacing) and stride (batch spacing) are honoured.

Also callable as gemv(ctx, A, x, y, GemvOptions<T>{...}), with owning Matrix / Vector arguments, and without B (taken from ctx.backend()).

Template Parameters

B	backend the call is compiled for; must match ctx's device
T	scalar type: float, double, std::complex<float> or std::complex<double>

Parameters

ctx	queue the work is enqueued on
A	batch of m x n matrices
X	batch of input vectors
Y	batch of output vectors; input scaled by beta, overwritten with the result
alpha	scale of the product
beta	scale of the input Y
transA	op() applied to A

Returns

event of the last enqueued kernel; Y is valid once it completes

Precondition

A, X and Y have the same batch size and conforming lengths.

Exceptions

batchlas::NoRouteError

in a build without the vendor BLAS for B when no native kernel can run the call (for example a heterogeneous A).

Note

No argument validation is done up front: a shape the native kernels refuse goes to the vendor library, which reports it.

See also

GemvOptions, docs/cpp-api.md