Automatic vectorization of communications for data-parallel programs

Abstract

Optimizing communication is a key issue in compiling dataparallel languages for distributed memory architectures. We examine here the cyclic distribution, and we derive symbolic expressions for communication sets under the only assumption that the initial parallel loop is defined by afline expressions of the indices. This technique relys on unimodular changes of basis. Analysis of the properties of communications leads to a tiling of the local memory addresses that provides maximal message vectorization.