资源描述：
《Efficient parallelized particle filter design on CUDA》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、EFFICIENTPARALLELIZEDPARTICLEFILTERDESIGNONCUDAMin-AnChao,Chun-YuanChu,Chih-HaoChao,andAn-Yeu(Andy)WuGraduateInstituteofElectronicsEngineering,NationalTaiwanUniversityTaipeiCity10617,TaiwanABSTRACTThefirstparallelizationdesignofparticlefiltersonthetradi-tionalGPUispresentedin[
2、10].However,theissueofdataParticlefilteringiswidelyusedinnumerousnonlinearappli-localityismorecrucialinCUDAthantraditionalGPUsduecationswhichrequirereconfigurability,fastprototyping,andtotheexecutionmodelandmemoryhierarchy,whichisnotonlineparallelsignalprocessing.Theemergingco
3、mputingaddressedin[10].Therefore,furtheroptimizationforsuchplatform,CUDA,mayberegardedasthemostappealingplat-architectureisyetanunsolvedandchallengingproblem.formforsuchimplementation.However,therearenotyetInthispaper,weproposetwodesigntechniques,A)finite-literaturesexploring
4、howtoutilizeCUDAforparticlefilters.redrawimportance-maximizing(FRIM)prioreditingandThispareraimstoprovidetwodesigntechniques,A)finite-B)localizedresampling.Sincetheglobaloperationsofredrawimportance-maximizing(FRIM)prioreditingandB)particlefilters,whichdominatetheexecutiontime,
5、arepro-localizedresampling,forefficientimplementationofparticleportionaltothenumberofparticles(N),theFRIMpriorfiltersonCUDA,whichcanbeverifiedtoreduceglobaloper-editingaimstoreducethetotalparticlenumberwithaddi-ationsandprovidesignificantspeedup.Themodificationsontionalcalculatio
6、nateachlocalthread.Moreover,theFRIMalgorithmandarchitecturalmappingareevaluatedwithcon-prioreditingcanalsoimprovethequalityofdrawnparticles,ceptualandquantitativeanalysis.Fromtheclassicbearings-i.e.,increasetheeffectiveparticlenumber(Neff).Thisonlytrackingexperiments,theprop
7、oseddesignis5.73timesmakesresamplingavailabletobedonelocallywithtoler-fasterthanthedirectimplementationonGeForce9400m.ableperformancedegradation,whichsignificantlyreducesIndexTerms—Particlefilter,GPGPU,CUDAtime-consumingglobaloperations.Therestofthispaperisorganizedasfollows.S
8、ection2givesthebackgroundofparticlefilteringandthecharacteris-1.INTRODUCTION
