Practical parallel Fortran

The implementation of the Silicon Graphics automatic parallelizing Fortran compiler is described. It consists of three major parts: the Power Fortran accelerator (PFA), which discovers the parallelism, the Power Fortran compiler, which implements the parallelism, and the run-time library, which handles the synchronization of the parallel processes. All three parts are outlined, and a few benchmark results are presented. The Silicon Graphics parallel Fortran System provides fully automatic parallelization, allowing even naive users to take advantage of the Power Series multiprocessors. The PFA is a source-to-source preprocessor that runs as a compiler pass. Once a loop has been labeled as parallel, either by hand or by PFA, the compiler must generate code for it. The loop is spooled out and compiled as if it were a local, nested procedure in the Pascal sense. The library of routines deals with the parallel execution of the spooled loop subrouting.<
>