
06571975
a
06571975
Stpiczy\'nski, Przemys{\l}aw
Semiautomatic acceleration of sparse matrixvector product using openacc.
Wyrzykowski, Roman (ed.) et al., Parallel processing and applied mathematics. 11th international conference, PPAM 2015, Krakow, Poland, September 69, 2015. Revised selected papers. Part II. Cham: Springer (ISBN 9783319321516/pbk; 9783319321523/ebook). Lecture Notes in Computer Science 9574, 143152 (2016).
2016
Cham: Springer
EN
sparse matrices
spmv
gpus
openacc
cusparse
doi:10.1007/9783319321523_14
Summary: The aim of this paper is to show that well known SPARSKIT SpMV routines for {\it EllpackItpack} and {\it Jagged Diagonal} formats can be easily and successfully adapted to a hybrid GPUaccelerated computer environment using OpenACC. We formulate general guidelines for simple steps that should be done to transform source codes with irregular data access into efficient OpenACC programs. We also advise how to improve the performance of such programs by tuning data structures to utilize hardware properties of GPUs. Numerical experiments show that our accelerated versions of SPARSKIT SpMV routines achieve the performance comparable with the performance of the corresponding CUSPARSE routines optimized by NVIDIA.