Abstract.We provide parallel matrix-vector multiply routines for 1D and 2D partitioned sparse square and rectangular matrices. We clearly give pseudocodes that perform necessary initializations for parallel execution. We show how to maximize overlapping between communication and computation through the proper usage of condensed storage by rows and condensed storage by columns formats of the sparse matrices. We give pseudocodes for multiplication routines which benefit from such overlaps.
Key words. parallel computing, sparse matrix-vector multiply, sparse matrix partitioning