Performance portability of sparse matrix–vector multiplication implemented using OpenMP, OpenACC and SYCL

The aim of this paper is to study the performance portability of OpenMP, OpenACC and SYCL implementations of sparse matrix–vector product (SpMV) and its extended version in which the dot product of the input vector and the result is also calculated, for CSR and BSR storage formats, on Intel and AMD CPUs and NVIDIA GPU platforms. We compare it with the performance portability of much more sophisticated implementations provided by the vendors in their Intel oneAPI MKL and NVIDIA cuSPARSE libraries. Using the reformulated performance portability metric we show how it changes for various sparse matrices and which portable implementation and format achieve better performance portability. Numerical experiments show that the considered portable implementations for the CSR format usually achieve better performance than for the BSR format. On GPU, CSR OpenACC implementations for SpMV and SpMV-DOT tend to be the best. On CPU, CSR OpenMP implementation usually gives the best results for SpMV-DOT, while CSR OpenMP and BSR MKL achieve the best results for a similar number of matrices. • Portable implementations of SpMV and SpMV-DOT in OpenMP, OpenACC and SYCL. • The use of reformulated performance portability metric to compare efficiency of portable implementations on CPU and GPU platforms. • Comparison of the performance of portable implementations with their counterparts from Intel oneAPI MKL and NVIDIA cuSPARSE libraries. • In some cases, portable implementations perform better than those provided by vendors. • Analysis of the impact of matrix size and shape on the performance of portable implementations.