GPU-based Arbitrarily Sparse Matrix Vector Multiplication

In this paper, we propose three novel approaches to performing vector multiplication on arbitrarily sparse matrices. We developed the proposed methods specifically for implementation on the Graphics Processing Unit with an emphasis on increasing speed for very large sparse matrices. The novel method, Boxed-TJDS, BoxedCRS and Hybrid Boxed-TJDS, were designed using the preexisting sparse matrix storage formats CRS and TJDS. We perform an empirical analysis of the the novel approaches against an SSE implementation of the mathematical operation using matrices of varying sparsity and size. Our results show that the three novel methods represent a range of possibilities for sparse matrix vector multiplication. While the Hybrid and BoxedTJDS methods perform moderately slower than and comparable to the SSE implementation, respectively, the Boxed-CRS approaches performs considerably faster than SSE as the matrix size increases in non-zero values. ∗{colbert,gaelen,eurip}@cs.ucf.edu