软件流水中的循环展开优化  

Loop unrolling optimization for software pipelining

作　　者：李文龙[1] 刘利[1] 汤志忠[1] 

机构地区：[1]清华大学计算机科学与技术系,北京100084

出　　处：《北京航空航天大学学报》2004年第11期1111-1115,共5页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金资助项目 (60 173 0 10 )

摘　　要：在软件流水中应用循环展开可以实现分数值的启动间距 ,提高资源的利用率 ,同时基于展开的优化技术可以降低程序的资源需求和关键路径的长度 .提出了基于程序特性的展开因子算法UTBPC(UnrollingTimesBasedProgramCharacter istics) ,解决了循环展开的核心问题———展开因子的确定 ,同时提出了基于展开的软件数据预取优化技术 ,提高了软件数据预取的效率 .所有这些都在ORC(OpenRe searchCompiler)中实现 ,并对SPEC 2 0 0 0中的程序进行了测试 ,平均性能提高了 2 .6% .实验结果表明新提出的算法和基于展开的数据预取优化技术提高了编译器的整体性能 .Loop unrolling can make software pipelining achieve fractional initiation interval and improve resource utilization. Optimizations based on unrolling can reduce resource requirements and the heights of critical paths. An algorithm named UTBPC (unrolling times based program characteristics) for determining unrolling factors and unrolling based optimization for software data prefetching were proposed. These optimizations were implemented in ORC (open research compiler), and SPEC CPU2000 benchmarks were tested in Itanium processor. The average performance was improved by 2.6%. The results show that UTBPC algorithm and unrolling based optimization for software data prefetching can improve the overall performance of compilers.

关 键 词：计算机软件 软件流水 循环展开 展开因子 数据预取 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]