面向循环的混合向量化方法研究  

Hybrid Vectorization Method Research for Loops

作　　者：徐金龙[1,2] 赵荣彩[1,2] 丁锐[1,2] 

机构地区：[1]解放军信息工程大学,郑州450002 [2]数学工程与先进计算国家重点实验室,郑州450001

出　　处：《小型微型计算机系统》2014年第12期2764-2769,共6页Journal of Chinese Computer Systems

基　　金：国家"八六三"高技术研究发展计划项目(2009AA01220)资助;国家"核高基"重大专项项目(2009zx10036-001-001)资助

摘　　要：目前主流的两种向量化算法基于循环(loop-based)的向量化算法和超字并行(SLP)算法分别适用于不同特征的循环,前者适用于发掘迭代间的并行,后者擅于发掘迭代内的并行.向量化算法的选择将直接影响最终的向量化效果.为了增强向量化能力,提高向量化效率,根据目标循环的特征确定了向量化任务分配方案,并设计了一种混合的向量化框架来实现这种任务分配方案,旨在将两种向量化方法结合,充分发挥各自优势.同时,为了明确目标循环并行性特征,本文给出了迭代间和迭代内的向量并行性检测方法.实验结果及分析表明,本文提供的方法明显提高了向量化的识别能力,能够实现更合理的向量化任务分配.There are two mainstream vectorization approaches：loop-based vectorization focus on exploiting inter-iteration parallelism, SLP focus on exploiting inner-iteration parallelism. Different vectorization algorithm bring us different vectorization result. To improve the capacity and the efficiency of the vectorization tool, this paper bring out an vectorizition task selection scheme, which will attach some loop to appropriate vectorization approach. And design a hybrid vectorizition frame to achieve the scheme. The two vectorization approaches above are combined in the frame, and their advantages can be better used. Besides,to get the parallel feature of loop, we proposed a method for detecting parallelism in both inter-iteration and inner-iteration. The experimental result show that our way obvi- ously improves the ability to identify the parallelism. And can achieve a more reasonable assignment of vectorization task.

关 键 词：基于循环 超字并行 迭代间 迭代内 依赖 

分 类 号：TP314[自动化与计算机技术—计算机软件与理论]