作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2008, Vol. 34 ›› Issue (11): 77-79. doi: 10.3969/j.issn.1000-3428.2008.11.028

• 软件技术与数据库 • 上一篇    下一篇

分布内存系统中流水并行代码的自动生成

龚雪容1,陆林生2,赵荣彩1   

  1. (1. 解放军信息工程大学计算机科学与技术系,郑州 450002;2. 江南计算技术研究所,无锡 214083)
  • 收稿日期:1900-01-01 修回日期:1900-01-01 出版日期:2008-06-05 发布日期:2008-06-05

Automatic Generation of Pipeline Parallel Code in Distributed Memory System

GONG Xue-rong1, LU Lin-sheng2, ZHAO Rong-cai1   

  1. (1. Department of Computer Science and Technology, PLA Information Engineering University, Zhengzhou 450002; 2. Jiangnan Institute of Computing Technology, Wuxi 2l4083)
  • Received:1900-01-01 Revised:1900-01-01 Online:2008-06-05 Published:2008-06-05

摘要: 并行循环分为DOALL和DOACROSS。DOACROSS循环携带数据依赖,在并行执行时需要通信支持,对于可以精确分析依赖关系的DOACROSS循环可通过流水并行方式提高性能。该文针对流水并行代码的自动生成进行讨论,包括数据依赖关系图和流水关系图的建立、流水并行判别准则和流水代码的自动生成等。实验证明流水并行后能获得较好的加速比。

关键词: 流水并行, 数据依赖关系图, 流水关系图, 流水通信

Abstract: Parallel loops are divided into two kinds——DOALL and DOACROSS. Loops with data dependencies are often referred as DOACROSS loops. If the dependencies of DOACROSS loop can be precisely determined by compiler, pipeline parallel code for them can be created to improve the performance. This paper discusses the algorithms of creating the data dependence relation graph and pipeline relation graph, the discrimination rules of the pipeline parallel, and how to create the pipeline parallel code automatically. Experimental results show that the speedup ratio is satisfied with pipeline parallel.

Key words: pipeline parallel, data dependence relation graph, pipeline relation graph, pipeline communication

中图分类号: