Software

FZJ-2024-04705

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png tfQMRgpu v0.9 - a header-only GPU-accelerated iterative linear solver for multiple right-hand sides

Baumeister, P. (Corresponding author)FZJ* ; Nassyr, S.FZJ* ; Li, Y.

2023

This record in other databases:

Please use a persistent id in citations: doi:10.5281/ZENODO.8333498

Abstract: tfQMRgpu is a C++ header-only library written in CUDA. Version 0.9 aims to offer the transpose-free Quasi Minimal Residual method for the iterative solving of linear problems with multiple right-hand sides using graphical processors (GPUs).The library is shipped with the default use case of complex block-sparse operators with a set of available block sizes to be selected before compilation. C-interfaces expose the APIs for both, single precision (float) and double precision, to be called from other programming languages such as C, Fortran, Julia or Python.The block-sparse matrix solver version can be linked as shared object tfQMRgpu.so.The header-only feature allows to provide a custom C++/CUDA action operators to tfQMRgpu::solve<action_t> and compiles with the host code.

---

Contributing Institute(s):

1. Jülich Supercomputing Center (JSC)

Research Program(s):

1. 5111 - Domain-Specific Simulation & Data Life Cycle Labs (SDLs) and Research Groups (POF4-511) (POF4-511)

Appears in the scientific report 2024

---