A new method of Sparse Matrix-Vector Multiplication on GPU

In this paper we present a new method called Hybrid Processing Method for Sparse Matrix Vector Multiplication (SpMV) on the system accelerated by Graphic Processing Units (GPU). The Hybrid Processing Method can select different computation model automatically by the characters of sparse matrix. Hence the self-adaptability of this method can overcome the deficiencies of the traditional algorithm. Our study is divided into two parts. Firstly, we propose a new sparse matrix storage format which named Compress Sparse Row with Non-zero Size (CSRNS). According to the new format, we designed Hybrid Processing Method for SpMV on GPU. This new method could automatically adopt different type of sparse matrix and effectively improve load balance. Secondly, we implement this new method by OpenCL Through experiment to prove this new method can availably enhance speedup ratio compare with traditional method.