Multiple back-end support for the armadillo linear algebra interface

Abstract

The Armadillo C++ library provides programmers with a high-level Matlab-like syntax for linear algebra. Its design aims at providing a good balance between speed and ease of use. It can be linked with different back-ends, i.e. different LAPACKcompliant libraries. In this work we present a novel run-time support of Armadillo, which gracefully extends mainstream implementation to enable dynamic back-end switching and multiple back-end support. The extension is specifically designed to not affect Armadillo class template prototypes, thus to be easily interoperable with future evolutions of the Armadillo library itself. The proposed software stack is then tested for functionality and performance against a kernel code extracted from an industrial application.

Accelerating Spectral Graph Analysis Through Wavefronts of Linear Algebra Operations

“…For the parallel coordination, we rely on FastFlow [9], [10], but several tools could have been exploited, like Intel TBB or OpenMP, that also provide interfaces to implement task-parallel programs [11]. For linear algebra functions, we rely on Dynamic-Armadillo [12], a customised version of Armadillo 1 (cf. IV-A1), but alternative approaches could rely on, for instance, ArrayFire 2 , which would equally meet the requirements for this work; in alternative, a low-level approach could be considered, as, for instance, direct interacting with Magma [13].…”

“…In this work, we exploit Dynamic-Armadillo [12], an extension that supports dynamic offloading of operations onto multiple back-ends. Specifically, the OpenBLAS [15] CPU back-end and the Magma [13] hybrid CPU/GPU back-end will be used.…”

Scaling Dense Linear Algebra on Multicore and Beyond: A Survey

“…Finally, libraries aimed to perform SVD on GPUs exist (e.g. Magma [13], CuSolver 5 and higher-level wrappers as Array-Fire 6 ) and their performance has been already assessed [15]. Although some good performance was observed, the capability to handle large problems is limited by the GPU memory, which is usually smaller than CPU memory; for this reason, this survey targets CPU-only libraries.…”