Date of Award

1989

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

First Advisor

Jorge L. Aravena

Abstract

The goal of the research is the establishment of a formal methodology to develop computational structures more suitable for the changing nature of real-time signal processing and control applications. A major effort is devoted to the following question: Given a systolic array designed to execute a particular algorithm, what other algorithms can be executed on the same array? One approach for answering this question is based on a general model of array operations using graph-theoretic techniques. As a result, a systematic procedure is introduced that models array operations as a function of the compute cycle. As a consequence of the analysis, the dissertation develops the concept of fast algorithm realizations. This concept characterizes specific realizations that can be evaluated in a reduced number of cycles. It restricts the operations to remain in the same class but with reduced execution time. The concept takes advantage of the data dependencies of the algorithm at hand. This feature allows the modification of existing structures by reordering the input data. Applications of the principle allows optimum time band and triangular matrix product on arrays designed for dense matrices. A second approach for analyzing the families of algorithms implementable in an array, is based on the concept of array time constrained operation. The principle uses the number of compute cycle as an additional degree of freedom to expand the class of transformations generated by a single array. A mathematical approach, based on concepts from multilinear algebra, is introduced to model the recursive transformations implemented in linear arrays at each compute cycle. The proposed representation is general enough to encompass a large class of signal processing and control applications. A complete analytical model of the linear maps implementable by the array at each compute cycle is developed. The proposed methodology results in arrays that are more adaptable to the changing nature of operations. Lessons learned from analyzing existing arrays are used to design smart arrays for special algorithm realizations. Applications of the methodology include the design of flexible time structures and the ability to decompose a full size array into subarrays implementing smaller size problems.

Pages

189

Share

COinS