void RendererGPUBenchmark(FSynthBenchmarkResults& InOut, const FSceneView& View, uint32 WorkScale, bool bDebugOut) { check(IsInRenderingThread()); // two RT to ping pong so we force the GPU to flush it's pipeline TRefCountPtr<IPooledRenderTarget> RTItems[3]; { FPooledRenderTargetDesc Desc(FPooledRenderTargetDesc::Create2DDesc(FIntPoint(GBenchmarkResolution, GBenchmarkResolution), PF_B8G8R8A8, TexCreate_None, TexCreate_RenderTargetable | TexCreate_ShaderResource, false)); GRenderTargetPool.FindFreeElement(Desc, RTItems[0], TEXT("Benchmark0")); GRenderTargetPool.FindFreeElement(Desc, RTItems[1], TEXT("Benchmark1")); Desc.Extent = FIntPoint(1, 1); Desc.Flags = TexCreate_CPUReadback; // needs TexCreate_ResolveTargetable? Desc.TargetableFlags = TexCreate_None; GRenderTargetPool.FindFreeElement(Desc, RTItems[2], TEXT("BenchmarkReadback")); } // set the state RHISetBlendState(TStaticBlendState<>::GetRHI()); RHISetRasterizerState(TStaticRasterizerState<>::GetRHI()); RHISetDepthStencilState(TStaticDepthStencilState<false,CF_Always>::GetRHI()); { // larger number means more accuracy but slower, some slower GPUs might timeout with a number to large const uint32 IterationCount = 70; const uint32 MethodCount = ARRAY_COUNT(InOut.GPUStats); // 0 / 1 uint32 DestRTIndex = 0; const uint32 TimerSampleCount = IterationCount * MethodCount + 1; static FRenderQueryRHIRef TimerQueries[TimerSampleCount]; static uint32 PassCount[IterationCount]; for(uint32 i = 0; i < TimerSampleCount; ++i) { TimerQueries[i] = GTimerQueryPool.AllocateQuery(); } if(!TimerQueries[0]) { UE_LOG(LogSynthBenchmark, Warning, TEXT("GPU driver does not support timer queries.")); } // TimingValues are in Seconds per GPixel FTimingSeries TimingSeries[MethodCount]; for(uint32 MethodIterator = 0; MethodIterator < MethodCount; ++MethodIterator) { TimingSeries[MethodIterator].Init(IterationCount); } check(MethodCount == 5); InOut.GPUStats[0] = FSynthBenchmarkStat(TEXT("ALUHeavyNoise"), 1.0f / 4.601f, TEXT("s/GigaPix")); InOut.GPUStats[1] = FSynthBenchmarkStat(TEXT("TexHeavy"), 1.0f / 7.447f, TEXT("s/GigaPix")); InOut.GPUStats[2] = FSynthBenchmarkStat(TEXT("DepTexHeavy"), 1.0f / 3.847f, TEXT("s/GigaPix")); InOut.GPUStats[3] = FSynthBenchmarkStat(TEXT("FillOnly"), 1.0f / 25.463f, TEXT("s/GigaPix")); InOut.GPUStats[4] = FSynthBenchmarkStat(TEXT("Bandwidth"), 1.0f / 1.072f, TEXT("s/GigaPix")); // e.g. on NV670: Method3 (mostly fill rate )-> 26GP/s (seems realistic) // reference: http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_processing_units theoretical: 29.3G/s RHIEndRenderQuery(TimerQueries[0]); // multiple iterations to see how trust able the values are for(uint32 Iteration = 0; Iteration < IterationCount; ++Iteration) { for(uint32 MethodIterator = 0; MethodIterator < MethodCount; ++MethodIterator) { // alternate between forward and backward (should give the same number) // uint32 MethodId = (Iteration % 2) ? MethodIterator : (MethodCount - 1 - MethodIterator); uint32 MethodId = MethodIterator; uint32 QueryIndex = 1 + Iteration * MethodCount + MethodId; // 0 / 1 const uint32 SrcRTIndex = 1 - DestRTIndex; GRenderTargetPool.VisualizeTexture.SetCheckPoint(RTItems[DestRTIndex]); RHISetRenderTarget(RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, FTextureRHIRef()); // decide how much work we do in this pass PassCount[Iteration] = (Iteration / 10 + 1) * WorkScale; RunBenchmarkShader(View, MethodId, RTItems[SrcRTIndex], PassCount[Iteration]); RHICopyToResolveTarget(RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, RTItems[DestRTIndex]->GetRenderTargetItem().ShaderResourceTexture, false, FResolveParams()); /*if(bGPUCPUSync) { // more consistent timing but strangely much faster to the level that is unrealistic FResolveParams Param; Param.Rect = FResolveRect(0, 0, 1, 1); RHICopyToResolveTarget( RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, RTItems[2]->GetRenderTargetItem().ShaderResourceTexture, false, Param); void* Data = 0; int Width = 0; int Height = 0; RHIMapStagingSurface(RTItems[2]->GetRenderTargetItem().ShaderResourceTexture, Data, Width, Height); RHIUnmapStagingSurface(RTItems[2]->GetRenderTargetItem().ShaderResourceTexture); }*/ RHIEndRenderQuery(TimerQueries[QueryIndex]); // ping pong DestRTIndex = 1 - DestRTIndex; } } { uint64 OldAbsTime = 0; RHIGetRenderQueryResult(TimerQueries[0], OldAbsTime, true); GTimerQueryPool.ReleaseQuery(TimerQueries[0]); #if !UE_BUILD_SHIPPING FBenchmarkGraph BenchmarkGraph(IterationCount, IterationCount, *(FPaths::ScreenShotDir() + TEXT("GPUSynthBenchmarkGraph.bmp"))); #endif for(uint32 Iteration = 0; Iteration < IterationCount; ++Iteration) { uint32 Results[MethodCount]; for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId) { uint32 QueryIndex = 1 + Iteration * MethodCount + MethodId; uint64 AbsTime; RHIGetRenderQueryResult(TimerQueries[QueryIndex], AbsTime, true); GTimerQueryPool.ReleaseQuery(TimerQueries[QueryIndex]); Results[MethodId] = AbsTime - OldAbsTime; OldAbsTime = AbsTime; } double SamplesInGPix = PassCount[Iteration] * GBenchmarkResolution * GBenchmarkResolution / 1000000000.0; for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId) { double TimeInSec = Results[MethodId] / 1000000.0; double TimingValue = TimeInSec / SamplesInGPix; // TimingValue in Seconds per GPixel TimingSeries[MethodId].SetEntry(Iteration, (float)TimingValue); } #if !UE_BUILD_SHIPPING { // This is for debugging and we don't want to change the output but we still use "InOut". // That shouldn't hurt, as we override the values after that anyway. for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId) { InOut.GPUStats[MethodId].SetMeasuredTime(TimingSeries[MethodId].GetEntry(Iteration)); } float LocalGPUIndex = InOut.ComputeGPUPerfIndex(); // * 0.01 to get it in 0..1 range // * 0.5f to have 100 is the middle BenchmarkGraph.DrawBar(Iteration, LocalGPUIndex * 0.01f * 0.5f); } #endif } for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId) { float Confidence = 0.0f; float TimingValue = TimingSeries[MethodId].ComputeValue(Confidence); if(Confidence > 0) { InOut.GPUStats[MethodId].SetMeasuredTime(TimingValue, Confidence); } UE_LOG(LogSynthBenchmark, Display, TEXT(" ... %.3f GigaPix/s, Confidence=%.0f%% '%s'"), 1.0f / InOut.GPUStats[MethodId].GetMeasuredTime(), Confidence, InOut.GPUStats[MethodId].GetDesc()); } UE_LOG(LogSynthBenchmark, Display, TEXT("")); #if !UE_BUILD_SHIPPING if(bDebugOut) { BenchmarkGraph.Save(); } #endif } } }
void RendererGPUBenchmark(FRHICommandListImmediate& RHICmdList, FSynthBenchmarkResults& InOut, const FSceneView& View, float WorkScale, bool bDebugOut) { check(IsInRenderingThread()); // two RT to ping pong so we force the GPU to flush it's pipeline TRefCountPtr<IPooledRenderTarget> RTItems[3]; { FPooledRenderTargetDesc Desc(FPooledRenderTargetDesc::Create2DDesc(FIntPoint(GBenchmarkResolution, GBenchmarkResolution), PF_B8G8R8A8, FClearValueBinding::None, TexCreate_None, TexCreate_RenderTargetable | TexCreate_ShaderResource, false)); GRenderTargetPool.FindFreeElement(Desc, RTItems[0], TEXT("Benchmark0")); GRenderTargetPool.FindFreeElement(Desc, RTItems[1], TEXT("Benchmark1")); Desc.Extent = FIntPoint(1, 1); Desc.Flags = TexCreate_CPUReadback; // needs TexCreate_ResolveTargetable? Desc.TargetableFlags = TexCreate_None; GRenderTargetPool.FindFreeElement(Desc, RTItems[2], TEXT("BenchmarkReadback")); } // set the state RHICmdList.SetBlendState(TStaticBlendState<>::GetRHI()); RHICmdList.SetRasterizerState(TStaticRasterizerState<>::GetRHI()); RHICmdList.SetDepthStencilState(TStaticDepthStencilState<false,CF_Always>::GetRHI()); { // larger number means more accuracy but slower, some slower GPUs might timeout with a number to large const uint32 IterationCount = 70; const uint32 MethodCount = ARRAY_COUNT(InOut.GPUStats); // 0 / 1 uint32 DestRTIndex = 0; const uint32 TimerSampleCount = IterationCount * MethodCount + 1; static FRenderQueryRHIRef TimerQueries[TimerSampleCount]; static float LocalWorkScale[IterationCount]; for(uint32 i = 0; i < TimerSampleCount; ++i) { TimerQueries[i] = GTimerQueryPool.AllocateQuery(); } const bool bSupportsTimerQueries = (TimerQueries[0] != NULL); if(!bSupportsTimerQueries) { UE_LOG(LogSynthBenchmark, Warning, TEXT("GPU driver does not support timer queries.")); // Temporary workaround for GL_TIMESTAMP being unavailable and GL_TIME_ELAPSED workaround breaking drivers #if PLATFORM_MAC GLint RendererID = 0; float PerfScale = 1.0f; [[NSOpenGLContext currentContext] getValues:&RendererID forParameter:NSOpenGLCPCurrentRendererID]; { switch((RendererID & kCGLRendererIDMatchingMask)) { case kCGLRendererATIRadeonX4000ID: // AMD 7xx0 & Dx00 series - should be pretty beefy PerfScale = 1.2f; break; case kCGLRendererATIRadeonX3000ID: // AMD 5xx0, 6xx0 series - mostly OK case kCGLRendererGeForceID: // Nvidia 6x0 & 7x0 series - mostly OK PerfScale = 2.0f; break; case kCGLRendererIntelHD5000ID: // Intel HD 5000, Iris, Iris Pro - not dreadful PerfScale = 4.2f; break; case kCGLRendererIntelHD4000ID: // Intel HD 4000 - quite slow PerfScale = 7.5f; break; case kCGLRendererATIRadeonX2000ID: // ATi 4xx0, 3xx0, 2xx0 - almost all very slow and drivers are now very buggy case kCGLRendererGeForce8xxxID: // Nvidia 3x0, 2x0, 1x0, 9xx0, 8xx0 - almost all very slow case kCGLRendererIntelHDID: // Intel HD 3000 - very, very slow and very buggy driver default: PerfScale = 10.0f; break; } } InOut.GPUStats[0] = FSynthBenchmarkStat(TEXT("ALUHeavyNoise"), 1.0f / 4.601f, TEXT("s/GigaPix")); InOut.GPUStats[1] = FSynthBenchmarkStat(TEXT("TexHeavy"), 1.0f / 7.447f, TEXT("s/GigaPix")); InOut.GPUStats[2] = FSynthBenchmarkStat(TEXT("DepTexHeavy"), 1.0f / 3.847f, TEXT("s/GigaPix")); InOut.GPUStats[3] = FSynthBenchmarkStat(TEXT("FillOnly"), 1.0f / 25.463f, TEXT("s/GigaPix")); InOut.GPUStats[4] = FSynthBenchmarkStat(TEXT("Bandwidth"), 1.0f / 1.072f, TEXT("s/GigaPix")); InOut.GPUStats[0].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 4.601f)) ); InOut.GPUStats[1].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 7.447f)) ); InOut.GPUStats[2].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 3.847f)) ); InOut.GPUStats[3].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 25.463f)) ); InOut.GPUStats[4].SetMeasuredTime( FTimeSample(PerfScale, PerfScale * (1.0f / 1.072f)) ); #endif return; } // TimingValues are in Seconds FTimingSeries TimingSeries[MethodCount]; // in 1/1000000 Seconds uint64 TotalTimes[MethodCount]; for(uint32 MethodIterator = 0; MethodIterator < MethodCount; ++MethodIterator) { TotalTimes[MethodIterator] = 0; TimingSeries[MethodIterator].Init(IterationCount); } check(MethodCount == 5); InOut.GPUStats[0] = FSynthBenchmarkStat(TEXT("ALUHeavyNoise"), 1.0f / 4.601f, TEXT("s/GigaPix")); InOut.GPUStats[1] = FSynthBenchmarkStat(TEXT("TexHeavy"), 1.0f / 7.447f, TEXT("s/GigaPix")); InOut.GPUStats[2] = FSynthBenchmarkStat(TEXT("DepTexHeavy"), 1.0f / 3.847f, TEXT("s/GigaPix")); InOut.GPUStats[3] = FSynthBenchmarkStat(TEXT("FillOnly"), 1.0f / 25.463f, TEXT("s/GigaPix")); InOut.GPUStats[4] = FSynthBenchmarkStat(TEXT("Bandwidth"), 1.0f / 1.072f, TEXT("s/GigaPix")); // e.g. on NV670: Method3 (mostly fill rate )-> 26GP/s (seems realistic) // reference: http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_processing_units theoretical: 29.3G/s RHICmdList.EndRenderQuery(TimerQueries[0]); // multiple iterations to see how trust able the values are for(uint32 Iteration = 0; Iteration < IterationCount; ++Iteration) { for(uint32 MethodIterator = 0; MethodIterator < MethodCount; ++MethodIterator) { // alternate between forward and backward (should give the same number) // uint32 MethodId = (Iteration % 2) ? MethodIterator : (MethodCount - 1 - MethodIterator); uint32 MethodId = MethodIterator; uint32 QueryIndex = 1 + Iteration * MethodCount + MethodId; // 0 / 1 const uint32 SrcRTIndex = 1 - DestRTIndex; GRenderTargetPool.VisualizeTexture.SetCheckPoint(RHICmdList, RTItems[DestRTIndex]); SetRenderTarget(RHICmdList, RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, FTextureRHIRef()); // decide how much work we do in this pass LocalWorkScale[Iteration] = (Iteration / 10.f + 1.f) * WorkScale; RunBenchmarkShader(RHICmdList, View, MethodId, RTItems[SrcRTIndex], LocalWorkScale[Iteration]); RHICmdList.CopyToResolveTarget(RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, RTItems[DestRTIndex]->GetRenderTargetItem().ShaderResourceTexture, false, FResolveParams()); /*if(bGPUCPUSync) { // more consistent timing but strangely much faster to the level that is unrealistic FResolveParams Param; Param.Rect = FResolveRect(0, 0, 1, 1); RHICmdList.CopyToResolveTarget( RTItems[DestRTIndex]->GetRenderTargetItem().TargetableTexture, RTItems[2]->GetRenderTargetItem().ShaderResourceTexture, false, Param); void* Data = 0; int Width = 0; int Height = 0; RHIMapStagingSurface(RTItems[2]->GetRenderTargetItem().ShaderResourceTexture, Data, Width, Height); RHIUnmapStagingSurface(RTItems[2]->GetRenderTargetItem().ShaderResourceTexture); }*/ RHICmdList.EndRenderQuery(TimerQueries[QueryIndex]); // ping pong DestRTIndex = 1 - DestRTIndex; } } { uint64 OldAbsTime = 0; // flushes the RHI thread to make sure all RHICmdList.EndRenderQuery() commands got executed. RHICmdList.ImmediateFlush(EImmediateFlushType::FlushRHIThread); RHICmdList.GetRenderQueryResult(TimerQueries[0], OldAbsTime, true); GTimerQueryPool.ReleaseQuery(RHICmdList, TimerQueries[0]); for(uint32 Iteration = 0; Iteration < IterationCount; ++Iteration) { uint32 Results[MethodCount]; for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId) { uint32 QueryIndex = 1 + Iteration * MethodCount + MethodId; uint64 AbsTime; RHICmdList.GetRenderQueryResult(TimerQueries[QueryIndex], AbsTime, true); GTimerQueryPool.ReleaseQuery(RHICmdList, TimerQueries[QueryIndex]); uint64 RelTime = AbsTime - OldAbsTime; TotalTimes[MethodId] += RelTime; Results[MethodId] = RelTime; OldAbsTime = AbsTime; } for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId) { float TimeInSec = Results[MethodId] / 1000000.0f; // to normalize from seconds to seconds per GPixel float SamplesInGPix = LocalWorkScale[Iteration] * GBenchmarkResolution * GBenchmarkResolution / 1000000000.0f; // TimingValue in Seconds per GPixel TimingSeries[MethodId].SetEntry(Iteration, TimeInSec / SamplesInGPix); } } if(bSupportsTimerQueries) { for(uint32 MethodId = 0; MethodId < MethodCount; ++MethodId) { float Confidence = 0.0f; // in seconds per GPixel float NormalizedTime = TimingSeries[MethodId].ComputeValue(Confidence); if(Confidence > 0) { FTimeSample TimeSample(TotalTimes[MethodId] / 1000000.0f, NormalizedTime); InOut.GPUStats[MethodId].SetMeasuredTime(TimeSample, Confidence); } } } } }