Contributions to High Performance Fortran Compilation

This study presents our contributions to High Performance Fortran HPF compilation HPF is a data parallel language based on Fortran Directives are used to specify parallelism and data mapping onto distributed memory parallel architectures We aim at translating a global addressing implicit communication HPF pro gram into a message passing parallel model with local addressing and explicit communication The generated code must be both correct and especially e cient Our approach is based on a formalization of compilation technical problems into a mathematical framework and on using standard algorithms for generating optimized code This document is divided into four parts After the introduction which presents scienti c computing and its challenges we describe in the rst part the HPF language which solves some of the encountered problems The weak squaring of HPF with the state of the art in compilation at the time of its design is outlined Compilation techniques developped for HPF are also detailed as the related work In the second part we discuss HPF design issues We rst analyze HPF language design issues show de ciencies and suggest new solutions A simple compilation technique to avoid a language extension aiming at explicitly declaring overlap areas is described Finally we present a possible extension to enable several mappings to be associated to subroutine arguments In the third part we present the compilation techniques First we describe a linear algebra framework for compiling computation and communication This framework is applied and extended to possibly parallel I O Compilation of HPF remappings is then addressed including optimizations to avoid useless remappings and an optimal technique to manage the actual communication taking advantage of broadcast and load balancing opportunities linked to data replication Perfor mance comparisons against the Digital HPF compiler are also presented In the last part we present our prototype HPF compiler around lines of code developed within the PIPS automatic parallelizer environment The im plementation including its weaknesses and extensions is detailed In particular our prototype handles all static and dynamic HPF mapping speci cations as well as parallelism directives The runtime is based on the public domain PVM commu nication library Experiments with codes generated by our compiler on a network of workstations are also described Finally we conclude