程序向量化中非规则访存问题研究  

Research on Irregular Memory Access Problem for Programs Vectorization

作　　者：徐金龙 赵荣彩 刘鹏 李晓亮 

机构地区：[1]数学工程与先进计算国家重点实验室 [2]解放军93010部队

出　　处：《计算机工程》2015年第12期86-90,共5页Computer Engineering

基　　金：国家"863"计划基金资助项目(2009AA01220);"核高基"重大专项(2009zx10036-001-001)

摘　　要：现有的程序向量化方法通常仅支持连续访存模式,而不支持非连续的内存访问。为实现更多程序的向量并行,提出一种向量化非规则访存处理方法。对访存特征进行检测分类,针对不同特征给出对应的向量化方案,同时设计收益分析方法,以保证向量化的有效性。实验结果表明,该方法可有效提高向量化能力,实现复杂访存形式的程序向量化。Current programs vectorization methods commonly support only continuous memory access forms,they do not support non-contiguous memory access.In order to implement more programs vector parallelisms,an effective method handing irregular memory access for vectorization is proposed.In this method,memory accesses are classified,and a feature detection method is provided.Different vectorization solutions are proposed corresponding to different memory access features.Finally,cost-benefit analysis method is provided to guarantee the effectiveness of vectorization.Experimental results indicate that this method improves the vectorization ability significantly,and the programs with complex memory access can be vectorized.

关 键 词：非连续访存 向量化 访存特征 数据重组 数组访存 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]