Support for irregular computation in high performance Fortran

Abstract

High Performance Fortran is an extension of the Fortran 95 language designed for data-parallel computing. It was designed to address the need for a vendor-independent, high-level programming language for parallel machines with distributed memory. The array language of Fortran 95 naturally allows parallel programs where the principal parallel idiom is evaluation of a scalar function for every element of a Fortran array.

Algorithms based on irregular meshes, graphs, sets, ensembles of particles (in N-body models of molecules, galaxies, etc.) and other irregular structures, while they may be data-parallel, do not naturally fit the array parallel style. HPF version 2.0, due to be released in December 1996, addresses some of these issues. It does so without enlarging the underlying Fortran language. I will discuss the proposed language extensions, and some alternatives that were not included, for irregular data mapping, user-specified computation mapping, reductions in parallel loops, and multiple process (or task) parallelism.