On the synthesis of parallel programs from tensor product formulas for block recursive algorithms

Abstract

This paper presents a methodology for synthesizing parallel programs for block recursive algorithms such as fast Fourier transforms and Strassen's matrix multiplication algorithm. A block recursive algorithm is expressed as a tensor product formula which consists of matrix sums, matrix products, direct sums, tensor products, componentwise matrix operations, and stride permutations. These mathematical operations can be mapped to high-level programming language constructs such as iteration, sequential composition, parallel composition, vector operations, and index computation. Translation of a tensor product formula consisting of these primitives into a parallel program involves determination of the proper indexing schemes for the arrays. This paper gives an algorithm to determine the indexing scheme and the code required for the index computation. Various parallel programs can be synthesized by manipulating tensor product formulas to exploit different computational structures. In this paper, we discuss some issues involved in formula manipulation for a particular target machine, the Cray Y-MP.

This work was supported in part by DARPA, order number 7898, monitored by NIST under grant number 60NANB1D1151, DARPA, order number 7899, monitored by NIST under grant number 60NANB1D1150, and Ohio State University Seed Grant, No. 221337.