GPUProgrammingYanciZhangGameProgrammingIIOutlineGPUoverviewOpenGLshadinglanguageoverviewVertex/Geometry/FragmentshaderApplication:Per-pixelshadingGameProgrammingIIWhatisGPU?GPU:GraphicsProcessingUnitDevelopedrapidlyfrombeingprimitivedrawingdevicestobeingmajorcomputingresourcesExtremelypowerfulandflexibleprocessorTremendousmemorybandwidthandcomputationalpowerHighlevellanguageshaveemergedCapableofgeneral-purposecomputationbeyondgraphicsapplicationsGameProgrammingIIMotivationInmanyrespectsGPUismorepowerfulthanCPUComputationalpower:FLOPS(FloatingpointOperationsPerSecond)ParallelismBandwidthPerformancegrowthrateGameProgrammingIIFloatingPointCalculationFLOPS:AcommonbenchmarkmeasurementforratingthespeedofFPUCPUIntelCorei7980XE(quad-core):107.55GFLOPSGPUnVidiaGeForceGTX480:2.02TFLOPSModernGPUssupporthighprecision32-bitfloatingpointthroughoutthepipelineNosupportforadoubleprecisionformatGameProgrammingIIParallelismParallelism:allowssimultaneousoperationsatthesametimeCPUDonotadequatelyexploitparallelismDual-core,quad-coreGPUGeForceGTX480:512kernelsGameProgrammingIIBandwidthPeakperformanceofcomputersystemsisoftenfarinexcessofactualapplicationperformanceThebandwidthbetweenkeycomponentsultimatelydictatessystemperformanceCPU64bitsDDR3-2133dual-channel:17GB/sGPUGeForceGTX480:384bits,177.4GB/sGameProgrammingIIGettingFasterandFasterCPUAnnualgrowth~1.5x->decadegrowth~60xMoore’slawGPUAnnualgrowth~2.0x->decadegrowth>1000xFasterthanMoore’slawMulti-billiondollarvideogamemarketisapressurecookerthatdrivesinnovationGameProgrammingIIKeystoHigh-Perf.ComputingEfficientcomputationMaximizethehardwaredevotedtocomputationAllowparallelismTaskparallelismDataparallelismInstructionparallelismEnsureeachcomputationunitoperatesatmaximumefficiencyGameProgrammingIIKeystoHigh-Perf.ComputingEfficientcommunicationSimplyprovidinglargeam