The SPAD test:循环级投机并行化方法  

The SPAD test:a Technique for Speculative Exploiting Loop-level Parallelism

作　　者：周静[1] 曾国荪[1] 

机构地区：[1]同济大学计算机科学及技术系

出　　处：《小型微型计算机系统》2007年第11期1932-1936,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(60173026)资助;上海科委重大项目(03DZ15029)资助;上海高校网格技术E-研究院(200301-1)资助

摘　　要：传统的并行编译技术能够在编译期间进行相关性分析,有效地并行化循环程序,但是对于程序运行时潜在的并行性却无能为力.因此,并行编译技术必须使用实时依赖分析技术,尽可能挖掘循环级并行性.本文提出仿射依赖关系,消除了循环迭代依赖;基于投机并行思想,提出了SPAD方法.实例分析表明,SPAD是有效的.与LRPD和SPNT方法相比较,SPAD做了重要的改进,因此是更通用的投机并行化方案.Traditional parallelizing compiler techniques can analysis the dependences and effectively parallel loops at compiletime, but cannot exploit the potential parallelism of loops at run-time. Therefore, parallelizing compilers must rely on the run-time dependence analysis techniques to detect parallelism as possible. We propose the affine data dependence which eliminates all cross-iteration data dependences. Meanwhile, a novel test, called the speculative parallelization based affine dependence （SPAD） test, is developed based upon the speculative parallelization. Some case studies show that the SPAD-test is effective and efficient. Compared to the LRPD test and SPNT test ,the SPAD-test makes significant improvement ,and is therefore a more general scheme of speculative parallelization test.

关 键 词：并行编译 投机并行 循环级并行 依赖分析 

分 类 号：TP338[自动化与计算机技术—计算机系统结构]