Optimizing sparse matrix-vector multiplication in NEC SX-Aurora vector engine

Abstract

Sparse Matrix-Vector multiplication (SpMV) is an essential piece of code used in many High Performance Computing (HPC) applications. As previous literature shows, achieving efficient vectorization and performance in modern multi-core systems is nothing straightforward. It is important then to revisit the current stateof-the-art matrix formats and optimizations to be able to deliver deliver high performance in long vector architectures. In this tech-report, we describe how to develop an efficient implementation that achieves high throughput in the NEC Vector Engine: a 256 element-long vector architecture. Combining several pre-processing and kernel optimizations we obtain an average 12% improvement over a base SELLC-s implementation on a heterogeneous set of 24 matrices.