Optimizing sparse matrix-vector multiplication on CUDA

In recent years, GPUs have attracted the attention of many application developers as powerful massively parallel system. CUDA as a general purpose parallel computing architecture make GPUs an appealing choice to solve many complex computational problems in a more efficient way. In this paper, we discuss implementing optimizing spare matrix-vector multiplication on NVIDIA GPUs using CUDA programming model. We outline three optimizations include: (1) optimized CSR storage format, (2) optimized threads mapping, and (3) avoiding divergence judgment. We experimentally evaluate our optimizations on GeForce 9600 GTX, connect to Windows xp 64-bit system. In comparison with NVIDIA´s SpMV library and NVIDIA´s CUDDPA library, the results show that optimizing sparse matrix-vector multiplication on CUDA achieves better performance than other SpMV implementations.