程序地循环优化方法1.0.ppt

《程序地循环优化方法1.0.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、提升循环级并行IntelCopyright©2002IntelCorporationAgendaIntroductionWhoCares?DefinitionLoopDependenceandRemovalDependencyIdentificationLabSummaryIntroductionLoopsmustmeetcertaincriteria…IterationIndependenceMemoryDisambiguationHighLoopCountEtc…WhoCares实现真正的并行:OpenMPAutoParallelization…显式的指令级并行ILP

2、(InstructionLevelParallelism)StreamingSIMD(MMX,SSE,SSE2,…)SoftwarePipeliningonIntel®Itanium™ProcessorRemoveDependenciesfortheOut-of-OrderCoreMoreInstructionsruninparallelonIntelItanium-Processor自动编译器并行HighLevelOptimizationsDefinitioninta[MAX];for(J=0;J

3、:IterationYofaloopisindependentofwhenorwhetheriterationXhappens图例OpenMP:TrueParallelismSIMD:VectorizationSWP:SoftwarePipeliningOOO:Out-of-OrderCoreILP:InstructionLevelParallelismGreen:BenefitsfromconceptYellow:SomeBenefitsfromConceptRed:NoBenefitfromConceptAgendaDefinitionWhoCares?LoopDep

4、endenceandRemovalDataDependenciesRemovingDependenciesDataAmbiguityandtheCompilerDependencyRemovalLabSummaryFlowDependencyReadAfterWriteCross-IterationFlowDependence:Variableswrittenthenreadindifferentiterationsfor(J=1;J

5、adCross-IterationAnti-Dependence:Variableswrittenthenreadindifferentiterationsfor(J=1;J

6、[1]=B[1];A[2]=C[1];A[2]=B[1];A[3]=C[1];IntraIterationDependencyDependencywithinaniterationHurtsILPMaybeautomaticallyremovedbycompilerK=1;for(J=1;J

7、ParallelismNotalldependenciescanberemovedfor(J=1;J