- 1、本文档共38页,可阅读全部内容。
- 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
- 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
OptimizingParallelReductioninCUDAMarkHarris
NVIDIADeveloperTechnology
2ParallelReductionCommonandimportantdataparallelprimitiveEasytoimplementinCUDAHardertogetitrightServesasagreatoptimizationexampleWe’llwalkstepbystepthrough7differentversionsDemonstratesseveralimportantoptimizationstrategies
3ParallelReductionTree-basedapproachusedwithineachthreadblockNeedtobeabletousemultiplethreadblocksToprocessverylargearraysTokeepallmultiprocessorsontheGPUbusyEachthreadblockreducesaportionofthearrayButhowdowecommunicatepartialresultsbetweenthreadblocks?475911142531704163
4Problem:GlobalSynchronizationIfwecouldsynchronizeacrossallthreadblocks,couldeasilyreduceverylargearrays,right?GlobalsyncaftereachblockproducesitsresultOnceallblocksreachsync,continuerecursivelyButCUDAhasnoglobalsynchronization.Why?ExpensivetobuildinhardwareforGPUswithhighprocessorcountWouldforceprogrammertorunfewerblocks(nomorethan#multiprocessors*#residentblocks/multiprocessor)toavoiddeadlock,whichmayreduceoverallefficiencySolution:poseintomultiplekernelsKernellaunchservesasaglobalsynchronizationpointKernellaunchhasnegligibleHWoverhead,lowSWoverhead
5Solution:KernelpositionAvoidglobalsyncbyposingcomputationintomultiplekernelinvocationsInthecaseofreductions,codeforalllevelsisthesameRecursivekernelinvocation475911142531704163475911142531704163475911142531704163475911142531704163475911142531704163475911142531704163475911142531704163475911142531704163475911142531704163Level0:
8blocksLevel1:
1block
6WhatisOurOptimizationGoal?WeshouldstrivetoreachGPUpeakperformanceChoosetherightmetric:GFLOP/s:forcompute-boundkernelsBandwidth:formemory-boundkernelsReductionshaveverylowarithmeticintensity1flopperelementloaded(bandwidth-optimal)ThereforeweshouldstriveforpeakbandwidthWilluseG80GPU
您可能关注的文档
- 课件参考文案gaa30785cad sbd.pdf
- 主起落架伸缩系统.pdf
- 文稿章生成函数reffgenerating functions04.pdf
- 级埃尔沃德插图nimitz-class aircraft carriers174.pdf
- 案例评论教程.pdf
- 用户手册激光测振仪.pdf
- 六届团支书联席会复习宝典级线代课件佚名.pptx
- 自动空调g k l mpathfinder维修手册.pdf
- 广州万达电影乐园初步概念设计乘车计划场景ngzhou presentation.pptx
- 新媒体直播平台实施.pdf
- 西宁市湟源县2023-2024学年七年级下学期期末数学易错题汇总卷(含答案).doc
- 昌吉州吉木萨尔县2023-2024学年七年级下学期期末数学易错题整理卷(含答案).doc
- 遵义道真2023-2024学年七年级下学期期末数学易错题汇总卷(含答案).doc
- 牡丹江牡丹江2023-2024学年七年级下学期期末数学复习卷(含答案).doc
- 上饶市弋阳县2023-2024学年七年级下学期期末数学易错题汇总卷(含答案).doc
- 德州市夏津县2023-2024学年七年级下学期期末数学易错题汇总卷(含答案).doc
- 襄樊市樊城区2023-2024学年七年级下学期期末数学易错题汇总卷(含答案).doc
- 孝感市孝昌县2023-2024学年七年级下学期期末数学押题卷(含答案).doc
- 昆明市寻甸回族彝族自治县2023-2024学年七年级下学期期末数学压轴卷(含答案).doc
- 东营市垦利县2023-2024学年七年级下学期期末数学复习卷(含答案).doc
文档评论(0)