1

Sparse Matrix-Vector Multiplication (SpMV) is an important computational kernel in scientific applications that tends to perform poorly on modern processors because of irregular memory accesses. GPU have evolved into a very attractive hardware platform for general purpose computations due to their high floating-point computation performance, which results in that GPGPU becomes the hot and popular topic in HPC. Therefore, we need to parallelize and optimize SpMV on GPGPU to get a better performance. In this paper, we studied the register-level blocking algorithm and the heuristic algorithm to optimize SpMV performance on GPGPU. Based on AMD Stream Computing, We propose an automatic performance tuning SpMV software package on GPGPU, with its name GOSpMV. Experimental results on AMD Radeon HD 3690 shows that, compared with the CSR format SpMV implementation on CPU, our GOSpMV can achieve an average speedup with 3.11 and the max speedup is 5.95