CUDA_4.0_Readiness_Tech_Brief

《CUDA_4.0_Readiness_Tech_Brief》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、CUDATOOLKIT4.0READINESSFORCUDAAPPLICATIONSMarch3,2011TechnicalBriefTECHNICALBRIEFINTRODUCTIONInNVIDIA®CUDATMToolkitversion4.0,amajoremphasishasbeenplacedonimprovingtheprogrammabilityofmulti-threadedandmulti-GPUapplicationsandonimprovingtheeaseofportingexistingcodetoCUDAC/C++.Thisdocumentdescri

2、besthekeyAPIchangesandimprovementsthathavebeenmadetowardthatend,particularlywheretheyhavethepotentialtoimpactexistingapplications.ThisdocumentalsohighlightsafewoftheimprovementsthathavebeenmadetothelibrariesbundledwiththeCUDAToolkit.Foracompletelistingofnewfeatures,pleaserefertotheCUDAToolkitR

3、eleaseNotes;forcompletedocumentationofCUDAAPIs,pleaserefertotheCUDAToolkitReferenceManual,theCUDACProgrammingGuide,andtheCUDAlibrarydocumentation.CUDATOOLKIT4.0READINESSFORCUDAAPPLICATIONS2TECHNICALBRIEFMULTI-GPUPROGRAMMINGInCUDAToolkit3.2andearlier,thereweretwobasicapproachesavailabletoexecut

4、eCUDAkernelsonmultipleGPUs(CUDA“devices”)concurrentlyfromasinglehostapplication:Useonehostthreadperdevice,sinceanygivenhostthreadcancallcudaSetDevice()atmostonetime.Usethepush/popcontextfunctionsprovidedbytheCUDADriverAPI.ForapplicationsthatdonotrequiretightcouplingofthevariousCUDAdeviceswit

5、hinasystem(e.g.,whenthedevicesareprocessingindependentdatasetswithlittleneedtocommunicateorsynchronizewitheachother),theseapproacheswereoftensufficient.However,forapplicationsthatneedatightercouplingoftheexecutionofworkacrossdevices,orwheretheuseofmultiplehostthreadsisotherwiseinconvenient,itw

6、ouldsometimesbebetterifasinglehostthreadcouldeasilylaunchworkontoanydevicesitneeded.CUDARuntimeAPITheCUDARuntimenowprovidesanativemeanstoaccomplishthis:ahostthreadcansimplycallcudaSetDevice()atanytime(ratherthanjustonceatthebeginningoftheprogram)tochangethecurrentlyactivedevice.Thishasthefollo

7、wingconsequences:Kernellauncheswillbeexecutedonthecurrentlyselecteddevice.Memoryallocationswillbemadeonthecurrentlyselecteddevice.Streamsandeventscreatedwillbeassociatedwiththecurrentlyselecteddevice.1Forexample:cudaSetDevice(0);//st