Graph Traverse Software Pipelining

Graph Traverse Software Pipelining

ID:37658003

大小:102.51 KB

页数:32页

时间:2019-05-27

Graph Traverse Software Pipelining_第1页
Graph Traverse Software Pipelining_第2页
Graph Traverse Software Pipelining_第3页
Graph Traverse Software Pipelining_第4页
Graph Traverse Software Pipelining_第5页
资源描述:

《Graph Traverse Software Pipelining》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库

1、GraphTraverseSoftwarePipeliningCristinaBarrado,EduardAyguadéandJesúsLabartaComputerArchitectureDepartment,UPC,Barcelona,SpainAcknowledgmentsThisworkhasbeensupportedbytheSpanishMinistryofEducationundercontractCICYT-TIC429/95andbytheCEPBA(EuropeanCenterforParallelismofBarcelona).0Graph

2、TraverseSoftwarePipeliningAbstractSoftwarepipeliningisbecomingwidelyusedasaloopexecutionmodelformicroprocessorssupportingahighinstructionlevelparallelism.Inthispaperwedescribeaheuristicmethodforsoftwarepipelining,namedGraphTraverseSoftwarePipelining(GTSP),thatdividestheschedulingprob

3、lemintwophases:schedulingonagraphandcodegeneration.ThefirstphaseencapsulatestheNP-completepartoftheproblem.Theschedulingisrepresentedonthedatadependencegraph(DDG)throughasetofschedulingrecurrencescomposedofschedulingedges,representingapartialorderingbetweenstatementsintheeachrecurren

4、ce.Aheuristicisgiventoincludesuchedgesinordertogenerateefficientschedules.Itimprovesprevioussoftwarepipeliningtechniquesachievingthemaximumthroughputfor98%ofabenchmarksetcomposedofmorethan500innermostloopsfromtheSpecFP95.Thesecondphaseisafastandautomaticalgorithmthattransformsthesche

5、dulinggeneratedontheDDGintoVLIW-likeinstructions.AmongtheadvantagesofGTSPweremarktheabstractiononthegraphthatbenefitsfromthelargedevelopedgraphtheory,theglobalknowledgeoftheheuristicduringtheschedulingphase,andthepossibilityofreachingmaximumthroughputscheduleswithoutunrollingtheDDG.1

6、.IntroductionCurrenttechnologytrendsareencouragingsystemdesignerstoofferhigh-levelsofpotentialparallelismwithinasinglechip.SuperscalarandVeryLongInstructionWordprocessors(VLIW)[Gwen95,Gwen94,Gwen96,...],basedonthereplicationofprocessorresources,issueseveralinstructionspercyclefromasi

7、nglethreadofexecution;thisparallelismisusallyreferredasinstruction-levelparallelism(ILP).Multiprocessorsonachip[ONHW96,...],basedonthereplicationofprocessors,offeracoarserthread-levelparallelism,whereeachprocessorhasitsownresourcesandbasicallysharesthehigherlevelsofthememoryhierarchy

8、.Inallcases,

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。