Symbolic synthesis of parallel processing systems

Abstract

The authors derive high-level parallel processing arrays for matrix computations using symbolic transformations. They propose a graphical language MGD (Mesh Graph Descriptor) as the basis for the transformations. The input to the synthesis system is the single-assignment form of matrix algorithms and the output is a structure of the synthesized parallel arrays. The synthesized arrays produced range from fully-parallel systolic arrays to limited-size parallel arrays. The approach is concise, verifiable, and easy to use. An example of LU decomposition illustrates the approach.<>