Automatic parallelization of loops with 2-dimension arrays for distributed memory supercomputers (matrix multiplication)

The parallel performance of loops with 2-dimension arrays, particularly matrix multiplication, is considered. The most effective algorithm of matrix multiplication for distributed memory supercomputers that requires data dependence analyses is designed. The method of automatic matrix distribution in a memory of each processor and processor communications are described. The program for matrix multiplication detection and replacing by parallel form is considered. There are some results of parallel program performance on supercomputer nCube 2S in this paper.