Abstract
We propose a unified approach for implementing systolic computations on distributed memory multicomputers (DMMCs). Advantages of this approach include its applicability to a wide range of systolic algorithms, DMMCs and computer languages. It also incorporates the evaluation of expected execution times for different possible implementations of a given systolic computation and the generation of source code for a selected implementation. We present the details of the various stages of the unified approach including the generation of a systolic array, several partitioning schemes and related theorems, source code generation, execution time analysis and optimization based on performance models. The intended use of the approach is for building a compiler that would accept a systolic algorithm specification as input and would generate an efficient executable object program.