Automatically Pipelining Loop Iterations on Distributed-Memory Machines

Offner, Carl D.

HP Development Company

关键词: dependence analysis;    compiler optimizations;    High Performance Fortran;    HPF;    parallelism;    distributed memory;   

RP-ID  :  HPL-2005-176

学科分类：计算机科学（综合）

美国|英语

来源: HP Labs

【 摘 要 】

This paper really has two parts: 1. We start with an expository discussion of LU decomposition, including a motivation as well as a clean derivation of the standard algorithm. 2. Then we show how a clever HPF compiler could infer--simply from the HPF data mapping directives--that the outer loop can be pipelined. This inference is made possible by an extension of the ordinary notion of dependence to include what are here called spatial dependences. 31 Pages