high performance computing on graphics processing units: hgpu.org

GPU is the suitable equipment for accelerating computing-intensive applications in order to get the higher throughput for High Performance Computing (HPC). Sparse Matrix-Vector Multiplication (SpMV) is the core algorithm of HPC, so the SpMV’s throughput on GPU may affect the throughput on HPC platform. In the paper, we focus on the latency of reduction routine in SpMV included in CUSP, such as accessing shared memory and bank conflicting while multiple threads simultaneously accessing the same bank. We provide shuffle method to reduce the partial results instead of reducing in the shared memory in order to improve the throughput of SpMV on Kepler GPU. Experiments show that shuffle method can improve the throughput up to 9% of the original routine of SpMV in CUSP on average.