Abstract:

Sparse matrixvector multiplication is often an important computational kernel in many scientific applications. This paper faces the ndiagonal sparse matrix, uses the CUDA program model and describes a new compress format of sparse matrix based on the DIA compress format (CDIA), and gives each thread finegrained task distribution. In order to fulfill the characteristics of the align access of memory in CUDA, we transpose the compress matrix and design a  finegrained algorithm and program and do  some optimization to the program. In the data experiment, our best implementation achieves up to 39.6Gflop/s in singleprecision and 19.6Gflop/s in doubleprecision, and enhances the performance by about 7.6% and 17.4% that of Nathan Bell’s and Michael Garland’s respectively.

Key words: GPU;CDIA;CUDA;sparse matrixvector multiplication