Solving matrix problems with no size restriction on a systolic array processor

Abstract

In this paper we propose several data structures partitioning and transformation schemes, in order to get an efficient execution of various matrix algorithms without any size resriction. The following matrix operations are considered:

-Matrix-Matrix multiplication -Solving triangular matrix equations -L-U decomposition -Inverses of triangular and dense matrices

All these algorithms are to be executed on a problem-size independent sprial systolic array processor. The array topology is fixed, and a simple feedback and control are needed.

For all the algorithms that have been considered, the PE's utilization tends to the maximum possible value.