Compiling Data Parallel Languages for Shared Virtual Memory Systems

This deliverable gives a detailed language specification of a data-parallel programming language for shared virtual memory systems. In addition to data parallelism, SVM-Fortran supports task parallelism. The language provides features for locality optimization that are extensions of the HPF template concept. Based on the system-provided global address space in SVM systems, a less restrictive concept can be implemented, i.e. it is not the compilers task to analyze array references. Unlike templates in HPF, SVM-Fortran templates are used for work distribution, e.g. scheduling parallel loops. The advantage of this design is that the computation of the schedule based on the distribution is much cheaper than computing the work distribution based on a reference to a distributed array as it has to be done in HPF. This is especially true for irregular distributions. This deliverable also describes the compile time and run time techniques applied in the implementation of SVM-Fortran on the Intel Paragon