一种关于DO－loop并行划分的新观点与新方法  

A NEW THEORY AND METHOD OFPARALLEL PARTITIONING OF DO-LOOPS

作　　者：刘键[1] 谢卫[1] 朱晓梅[1] 谷秋艳 

机构地区：[1]华中理工大学计算机科学与工程系

出　　处：《计算机学报》1996年第7期520-529,共10页Chinese Journal of Computers

摘　　要：本文提出了一个分配相关新概念以及与此相应的基于迭代空间等价分类的ＤＯ－ｌｏｏｐ并行划分的新观点与新方法．这种方法的主要特点是：①是关于以ＤＯ－ｌｏｏｐ并行划分的一个一般的统一的方法，能解决所有ＤＣ－ｌｏｏｐ的并行划分问题．②能准确地挖掘出程序中所有ＤＯ－ｌｏｏｐ的并行性，并且同时自动完成数据划分与计算划分．③最适用于ＭＩＭＤ与ＳＰＭＤ的大粒度并行划分．④可以和任务级并行划分技术、向量化技术、流水线技术结合起来，从而全面挖掘各主要抽象级并行性．In this paper,a new concept.allocation dependence, as well as the corresponding new theory and methods based on equivalence classification of iterationspace, are proposed. The theory has many advantages, such as:It is a generaland unified theory for parallel partitioning of DO-loops, and can solve all the problems in parallel partitioning of DO-loops; It can accurately extract the parallelism of all the DO-loops in the program, and automatically complete the data decomposition and the computation decomposition simutaneously; It can solvelarge grain parallel partitioning for MIMD and SPMD systems; Associated withparallel partitioning techniques in task level, vectorization techniques, and pipelinetechniques, it can extract the parallelism in all main abstract levels.

关 键 词：分配相关 链接点 DO-loop 并行划分 计算机 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]