Interactivek-DTreeGPURaytracing
DanielReiterHorn
JeremySugerman
MikeHouston
PatHanrahan
StanfordUniversity
Abstract
Overthepastfewyears,thepowerfulcomputationratesandhighmemorybandwidthofGPUshaveattractede ortstorunraytracingonGPUs.OurworkextendsFoleyetal.’sGPUk-dtreeresearch.Weporttheirkd-restartalgorithmfrommulti-pass,usingCPUloadbalancing,tosinglepass,usingcurrentGPUs’branchingandloopingabilities.Weintroducethreeoptimizations:apacketizedformulation,atechniqueforrestartingpartiallydownthetreeinsteadofattheroot,andasmall, xed-sizestackthatischeckedbeforeresortingtorestart.Ouroptimizedimplementationachieves15-18millionprimaryrayspersecondand16-27millionshadowrayspersecondonourtestscenes.
OursystemalsotakesadvantageofGPUs’strengthsatrasterizationandshadingtoo eramodewhererasterizationreplaceseyeraysceneintersection,andprimaryhitsandlo-calshadingareproducedwithstandardDirect3Dcode.For1024x1024renderingsofoursceneswithshadowsandPhongshading,weachieve12-18framespersecond.Finally,wein-vestigatethee ciencyofourimplementationrelativetothecomputationalresourcesofourGPUsandalsocompareitagainstconventionalCPUsandtheCellprocessor,whichbothhavebeenshowntoraytracewell.
CRCategories:I.3.1[ComputerGraphics]:HardwareArchitecture—GraphicsprocessorsI.2.2[ComputerGraph-ics]:Raytracing—Systems
Keywords:ProgrammableGraphicsHardware,DataParallelComputing,StreamComputing,GPUComputing,Brook
1Introduction
Raytracingisaworkloadwithheavycomputationalrequire-ments,butabundantparallelism.Asaresult,researchershaveexploredmanynovelcustomandemergingarchitec-turesinpursuitofinteractivity[Purcelletal.2002;Woopetal.2005;Benthinetal.2006].OurworkextendstheworkofFoleyetal.[2005]withalgorithmicimprovementsandmodi cationsforthenewcapabilitiesofcurrentGPUs.Intheirwork,Foleyetal.adoptedk-dtreeaccelerationstructuresforGPUraytracingbyintroducingtwostack-freetraversalalgorithms,butidentifytheirperformanceasun-competitivewithoptimizedCPUimplementations.Ourim-plementationimprovestheirsbyreplacingtheirmulti-passapproachwithasinglepassapproachrelyingontherecentadditionofloopingandbranchingfunctionalitytoGPUs.
{danielrh,
yoel,mhouston,hanrahan}@graphics.stanford.edu
Weaddthreemajorenhancementstotheirkd-restartalgo-rithm:packetization,push-down,andshort-stack.Packe-tizationcombinesCPUraypacketideas[Waldetal.2001]withrestart,andtakesadvantageoftheGPU’shighere -ciencyrunningfour-wideSIMDcode.push-downlocalizesraystothesub-treetheyoverlapinsteadofalwaysrestartingfromthetreeroot.short-stackaddsasmall, xed-sizestacktomaintainthelastNpushes,andfallsbacktorestartingonunder ow.Thenetimpactismorethananorderofmag-nitudeperformanceimprovementoverFoleyetal.andatraversalalgorithmthatinourscenesvisitsfewerthan3%morenodesascomparedwiththeconventionalstack-basedapproach.
OursystemalsoseekstoexploittheGPU’snaturalstrengthsbygeneratingprimaryhitsviarasterizationandcomputinglocalshadingusingstandardprogrammablefrag-mentshading,andtracesraysforsecondarye ectssuchasshadowsandspecularbounces.Forcomparisonwithpriorworkandotherarchitectures,wewillalsopresentresultsfromcastingprimaryrayswithoursystem.OursystemisimplementedwithacombinationofBrook[Bucketal.2004]forthecomputationalkernelsandDirect3D9[Mi-crosoft2003]forrasterization,shading,anddisplay.WeuseATI’sCTM[ATI2006b]toolkittoworkarounddrivercom-pilerbugsandgatherstatistics,butallofourGPUshadercodeisstandardpixel-shader3.0.OnanATIX1900XTX[ATI2006a],our1024x1024sceneswithshadowsandPhongshadingrenderat12-18framespersecond.
Finally,weanalyzetheperformanceofourimple-mentationandcompareitwithworkonx86CPUsandtheSony/Toshiba/IBMCellBroadbandEngineProcessorTM(Cell)[Benthinetal.2006].Weidentifytheprimarybottlenecksandstructuraline cienciesofourenvi-ronmentaswellashowarchitecturaldi erenceschangesomeofthealgorithmicoptions.
2RelatedWork
Therehasbeenasigni cantinterestinstudyingraytrac-ingonparallelarchitectures.Benthinetal.[2006]designedaboundingvolumehierarchy(BVH)raytracerontheCellarchitectureusingasoftware-managedcachetosaveband-widthandsoftwarethreadstohidethelatencyonacachemiss.Theirraytracerperformsnearlyaswellpercoreandperclockasacommodityx86,resultingin60millionpri-maryrayspersecondand21millionprimaryplusshadowrayspersecondontheconferencescene.Thisgivesthemclock-for-clocka6.5xperformanceadvantageoverasinglecorex86.Sugermanetal.[2006]implementedak-dtreeraytracerontheCellwithasoftwaremanagedcache,butwithoutfrustumcullingorsoftwarethreads.Thisversionachieves17.4millionprimaryrayspersecondontheRobotstestscene,thesameperformanceastheGPUraytracerwepresenthere.
Therealsohasbeenresearchoncustomraytracingar-chitectures.ArecentexampleistheSaarCORprocessor[Schmittleretal.2002].Thischipfeaturesdedicatedk-dtree
Abstract GPU-based Rende... 暂无评价 7页 免费 GIST An Interactive, GPU... 28页 免费 Interactive k-d tree GPU... 8页 10财富值 Distributed Video ...
KD-tree比赛题目CUDA_IT/计算机_专业资料。属于积分2011 年 CUDA 校园程序设计...Interactive k-D Tree GPU Raytracing, I3D 2007 http://en.wikipedia.org/...
Interactive k-d tree GPU... 8页 10财富值 Fast penumbra calculatio... 暂无评价 8页 免费 Cost prediction in ray t... 10页 免费 distributed ray ...
tree implementations Limitations and problems The UNIVERSITY of NORTH CAROLINA ...2007. Interactive k-D tree GPU raytracing. In Proc. I3D’07 (to appear...
Interactive k-d tree G... 8页 4下载券 Interactive RayTracing... 14页 免费 RAYTRACING AS A TOO... 7页 免费 Raytracing and the cam... 暂无...
GH05 KD-Tree Acceleration Structures for a GPU Raytracer Tim Foley, Jeremy Sugerman Stanford University Motivation ? Accelerated raytracing – On commodity ...
GH05 KD-Tree Acceleration Structures for a GPU Raytracer Tim Foley, Jeremy Sugerman Stanford University Motivation ? Accelerated raytracing – On commodity ...
热门文档