COSC: Combine Optimized Sparse Matrix-Vector Multiplication for CSR Format

Sparse matrix-vector multiplication (SpMV) is the most important computational kernel, since its widely application in the scientific and engineering fields. In this paper, a series optimization methods of SpMV for CSR format have been approached. We focus on the underlying, paralleling and working set, and implement the library named as Combine Optimized SpMV for CSR format (COSC) which is the first CSR SpMV library combining with above optimization strategies and incorporates the optimizations such as explicit software pipelining, SIMDization, index compression, array padding and etc. Experiments show that the SpMV operation with optimization of COSC has average of 2.11 speedup while comparing with the non-optimized one. The implementation of COSC can be settled in SpMV and improve efficiency of hot spot.