HPF COMPILER FOR PARALLEL COMPUTERS - IMPLEMENTATION AND PERFORMANCE EVALUATION ON CENJU-3

This paper presents compilation techniques for High Performance Fortra n (HPF) used in authors' prototype HPF compiler for parallel computers . HPF is the first standard of data parallel language to allow program mers to express data parallelism explicitly with a set of compiler dir ectives. The techniques include computation mapping, communication gen eration, and run-time support. Extension of HPF with special directive s for specifying computation mapping is also proposed. Preliminary exp erimental results of some benchmark programs on NEC Cenju-3 demonstrat e the efficiency of authors' implementation.