Sparse Matrix-Vector Product for the bmSparse Matrix Format in GPUs

Historically, the sparse matrix-vector product (SpMV) acceleration concentrated the central part of the research efforts devoted to sparse linear algebra kernels. Lately, other fundamental sparse operations have gained considerable attention. A remarkable example is the SpGEMM (sparse matrix-matrix product) operation, for which several routines and specific sparse formats have been proposed in the last few years. One of these storage formats, the bitmap-based bmSparse, shows promising performance on the SpGEMM kernel over massively parallel hardware platforms and storage advantages that could benefit other operations. In this work, we study and develop an SpMV kernel for the bmSparse format. The experimental results show that the implementation can reach up to a 4 \(\times \) speedup over cuSparse ’s CSR implementation, suggesting that this format is an interesting alternative for a potentially broad range of sparse matrix applications.