GPU Sparse Matrix Vector Multiplication Optimization Based on ELLB Storage Format

Abstract

ELLPACK(ELL) sparse matrix storage format has problems such as high storage consumption and low efficiency of sparse matrix vector multiplication(SpMV). To solve this problem, we propose a Graphic Processing Unit(GPU)-based efficient ELLPACK-Block(ELLB) sparse matrix storage format. Based on the original ELL storage format, this format adaptively divides the matrix into blocks according to the average number of non-zero elements in each row, and uses auxiliary matrices to improve the efficiency of SpMV solution. We use the ELLB storage format to solve the SpMV problem for different matrices. The experimental results show that compared with the Perfect Compressed Sparse Row(PCSR) format, the ELLB sparse matrix storage format saves 50 of the memory space, and the average efficiency of solving SpMV is increased by 7 times; compared with the Effective Compressed Sparse Row(ECSR) format, the memory space usage is increased by 25, but the solution of SpMV The efficiency is increased by an average of 7.65 times.