Parallelization in an HPF language processor : Toward reality in scientific simulations: NEC's 21st Century Odyssey

We have developed an HPF (high Performance Fortran) language processor for SX-4 series, SX-5 series and Cenju-3/4 systems. HPF is a de facto standard data-parallel language mainly aimed at distributed memory multiprocessor systems. HPF allows users to develop parallelized programs by only specifying how to map data onto processors. The HPF compiler partitions computation among processors based on specified mapping information, and generates necessary data transfers. Therefore, both how to map computation onto processors and how to achieve high-speed data transfer are important for the efficient implementation of HPF compilers. This paper describes automatic parallelization and data transfer technology in NEC's HPF language processor. This paper also discusses the utilization of shared memory parallelization and vectorization on SX-4 and SX-5 series.