SpV8: Pursuing Optimal Vectorization and Regular Computation Pattern in SpMV
Hosted in Virtual Platform
DescriptionSparse Matrix-Vector Multiplication is important in scientific and industry applications, and remains a well-known challenge due to the high irregularity. Most existing researches suffer from non-negligible penalty due to their complex computation patterns. In this paper, we propose SpV8, a novel approach that achieves maximal vectorization with regular computation pattern. We evaluate SpV8 on Intel CPU and compare with multiple state-of-art algorithms. The results show that SpV8 achieves up to 10x speedup (average 2.8x) against MKL SpMV routine, and up to 2.4x (average 1.4x) against the best existing approach. Moreover, SpMV features very low pre-processing overhead in all compared approaches.