Parallel Algorithm for Quasi-Band Matrix-Matrix Multiplication

Sparse matrices arise in many practical scenarios. As a result, support for efficient operations such as multiplication of sparse matrices (spmm) is considered to be an important research area. Often, sparse matrices also exhibit particular characteristics that can be used towards better parallel algorithmics. In this paper, we focus on quasi-band sparse matrices that have a large majority of the non-zeros along the diagonals. We design and implement an efficient algorithm for multiplying two such matrices on a many-core architecture such as a GPU. Our implementation outperforms the corresponding library implementation by a factor of 2x on average over a wide variety of quasi-band matrices from standard datasets. We analyze our performance over synthetic quasi-band matrices.