Date of Award

1991

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Doris L. Carver

Abstract

To exploit parallelism in Fortran code, this dissertation consists of a study of the following three issues: (1) recurrence resolution in Do-loops for vector processing, (2) dependence cycle statement ordering in Do-loops for parallel processing, and (3) sub-routine parallelization. For recurrence resolution, the major findings include: (1) the node splitting algorithm cannot be used directly to break an essential antidependence link, of which the source variable that results in antidependence is itself the sink variable of another true dependence so a correction method is proposed, (2) a sink variable renaming technique is capable of breaking an antidependence and/or output-dependence link, (3) for recurrences formed by only true dependences, a dynamic dependence concept and the derived technique are powerful, and (4) by integrating related techniques, an algorithm for resolving a general multistatement recurrence is developed. The performance of a parallel loop is determined by the level of parallelism and the time delay due to interprocessor communication and synchronization. For a dependence cycle of a single parallel loop executed in a general synchronization mode, the parallelism exposed varies with the alignment of statements. Statements are reordered on the basis of execution-time of the loop as estimated at compile-time. An improved timing formula and a derived statement ordering algorithm are proposed. Further extension of this algorithm to multiple perfectly nested Do-loops with simple global dependence cycle is also presented. The subroutine is a potential source for parallel processing. Several problems must be solved for subroutine parallelization: (1) the precedence of parallel executions of subroutines, (2) identification of the optimum execution mode for each subroutine and (3) the restructuring of a serial program. A five-step approach to parallelize called subroutines for a calling subroutine is proposed: (1) computation of control dependence, (2) approximation of the global effects of subroutines, (3) analysis of data dependence, (4) identification of execution mode, and (5) restructuring of calling and called subroutines. Application of these five steps in a recursive manner to different levels of calling subroutines in a program addresses the parallelization of subroutines.

Pages

147

Share

COinS