A technique for mapping sparse matrix computations into regular processor arrays

Abstract

A technique for mapping irregular sparse matrix computations into regular parallel networks is proposed. It is based on regularization of the original irregular graph of an algorithm. For this aim, we use a mapping of an original index space corresponding to dense matrices into a new one, which corresponds to a chosen sparse-matrix storage scheme. This regularization is followed by space-time mappings, which transform the algorithm graph into resulting networks. The proposed approach is illustrated by the example of mapping matrix-vector multiplications.