Ray Casting Architectures for Volume

MERL–AMITSUBISHIELECTRICRESEARCHLABORATORY,HanspeterPfistery,DeborahSilverz,ToddA.CookxTR-99-17April1999AbstractReal-timevisualizationoflargevolumedatasetsdemandshighperformancecompu-tation,pushingthestorage,processing,anddatacommunicationrequirementstothelimitsofcurrenttechnology.Generalpurposeparallelprocessorshavebeenusedtovi-sualizemoderatesizedatasetsatinteractiveframerates;however,thecostandsizeofthesesupercomputersinhibitsthewidespreaduseforreal-timevisualization.Thispapersurveysseveralspecialpurposearchitecturesthatseektorendervolumesatinteractiverates.Thesespecializedvisualizationacceleratorshavecost,performance,andsizead-vantagesoverparallelprocessors.Allarchitecturesimplementraycastingusingparallelandpipelinedhardware.Weintroduceanewmetricthatnormalizesperformancetocomparethesearchitectures.ThearchitecturesincludedinthissurveyareVOGUE,VIRIM,ArrayBasedRayCasting,EM-Cube,andVIZARDII.Wealsodiscussfutureapplicationsofspecialpurposeaccelerators.ToappearinIEEETransactionsonVisualizationandComputerGraphicsThisworkmaynotbecopiedorreproducedinwholeorinpartforanycommercialpurpose.Permissiontocopyinwholeorinpartwithoutpaymentoffeeisgrantedfornonprofiteducationalandresearchpurposesprovidedthatallsuchwholeorpartialcopiesincludethefollowing:anoticethatsuchcopyingisbypermissionofMitsubishiElectricInformationTechnologyCenterAmerica;anacknowledgmentoftheauthorsandindividualcontributionstothework;andallapplicableportionsofthecopyrightnotice.Copying,reproduction,orrepublishingforanyotherpurposeshallrequirealicensewithpaymentoffeetoMitsubishiElectricInformationTechnologyCenterAmerica.Allrightsreserved.CopyrightcMitsubishiElectricInformationTechnologyCenterAmerica,1999201Broadway,Cambridge,Massachusetts02139HarveyRayisaPh.D.studentatRutgersStateUniversity,Email:[e-mail]fisteriswithMitsubishiElectricResearch,Email:pfi[e-mail]ociateprofessoratRutgersStateUniversity,Email:[e-mail]nddevelopmentengineeratImprovSystemInc.,Email:[e-mail]:–1.Firstprinting,TR-99-17,April199911IntroductionVolumevisualizationisanimportanttooltoviewandanalyzelargeamountsofdatafromvariousscientificdisciplines.Ithasnumerousapplicationsinareassuchasbiomedicine,geophysics,computationalfluiddynamics,finiteelementmodels,andcomputationalchemistry.Numericalsimulationsandsamplingdevicessuchasmagneticresonanceimaging(MRI),computedtomography(CT),satelliteimaging,andsonararecommonsourcesoflarge3Ddatasets.Thesedatasetsaregenerallyanywherefrom1283to10243andmaybenon-symmetric(i.e.,10241024512).Volumerenderinginvolvestheprojectionofavolumedatasetontoa2Dimageplane.FromFigure1weseethatavolumedatasetisorganizedasa3Darrayofvolumeelements,orvoxels1.Voxelsrepresentvariousphysicalcharacteristics,suchasdensity,temperature,velocity,andpressure.Othermeasurements,suchasareaandvolume,canbeextractedfromthevolumedatasets.Volumedatamaycontainmorethanahundredmillionvoxelvaluesrequiringalargeamountofstorage.InFigure1,thevoxelsareuniforminsizeandregularlyspacedonarectilineargrid.Othertypesofvolumedatacanbeclassifiedintocurvilineargrids,whichcanbethoughtofasresultingfromawarpingofaregulargrid,andunstructuredgrids,whichconsistofarbitraryshapedcells.ThispaperpresentsasurveyofrecentcustomvolumerenderingFigure1:Volumedataset.architecturesthatseektoachieveinteractivevolumerenderingforrectilineardatasets.Asurveyofothermethodsusedtoachieverealtimevolumerenderingispresentedin[1].Themotivationforcustomvolumerenderersisdiscussedinthenextsection.Severalothercustomarchitecturesexist[2,3,4,5,6,7,8,9]butwerenotpresentedbecausetheyareeitherrelatedtothearchitecturespresentedhereorarenotconsideredtoberecent.Section3presentsthreeparallelvolumerenderingalgorithmsthatareimplementedbythearchitecturesinthispaper.MajorcomponentsofavolumerenderingsystemarediscussedinSection4.FivespecializedvolumerenderingarchitecturesaresurveyedinSection5.AnewmetricisintroducedinSection6tocompareeacharchitecture.AcomparisonofthesurveyedarchitecturesispresentedinSection7andadiscussionispresentedinsection8.FuturetrendsforspecializedrenderingarchitecturesarepresentedinSection9.2NeedforCustomVisualizationArchitecturesAreal-timevolumerenderingsystemisimportantforthefollowingreasons[10]:1)tovisualizerapidlychanging4D(spatial-temporal)datasets,2)forreal-timeexplorationof3Ddatasets(e.g.,virtualreality),3)forinteractivemanipulationofvisualizationparameters(e.g.,classification),and4)interactivevolumegraphics[11].Asthesamplingratesofdevicesbecomefaster,itwillbepossibletogenerateseveral3Ddatasetsatinteractiverates;real-timevolumerenderingisrequiredtovisualizethesedynamicallychanging1Note,thetermvoxelhasbeenusedtorefertopointsamplesandcubicvolumeelements.Thepaperssurveyedhereusebothdefinitionsforillustrationpurposes.Therefore,figuresinthispaperwilluseapointsamplerepresentationoraunitvolumerepresentationofavoxelasnecessary.MERL-TR-99-17April19992datasets