TY  - EJOUR
AU  - Baumeister, Paul F
AU  - Nassyr, Stepan
TI  - tfQMRgpu: A GPU-accelerated linear solver with block-sparse complex result matrix
JO  - The journal of supercomputing
SN  - 0920-8542
CY  - Dordrecht [u.a.]
PB  - Springer Science + Business Media B.V
M1  - FZJ-2023-05135
PY  - 2023
AB  - Linear solvers are a central component of many applications in physics and engineering. In this work we present a software package for simultaneously solving with multiple right-hand sides using the vast compute performance and memory bandwidth of graphical processors. Using the transpose-free quasi minimal residual method iterative linear solving does not require the implementation of an adjoint operator. This C++/CUDA software packet has two ways of being employed. The precompiled version of this library offers linear solving for single and double precision block-sparse complex matrices with interfaces to various programming languages, in particular C, Fortran, Python and Julia. Furthermore, the core algorithm is available for custom implementations of any linear operator as a C++ header-only library. We showcase a matrix-free approach of a custom operator for a finite-difference stencil application solving the three-dimensional Helmholtz equation and compare the performance of the matrix-free approach against the block-sparse matrix version, both on NVIDIA hardware.
LB  - PUB:(DE-HGF)25
DO  - DOI:10.21203/rs.3.rs-3574519/v1
UR  - https://juser.fz-juelich.de/record/1019082
ER  -