GoaLitiuM/FlaxEngine
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Optimize and refactor DDGI to use linear attention per-probe to dynamically control ray count and blend speed

Browse Source 
Probes will use lower ray count when behind the camera or not correlated with the view direction or far from geometry.
Probes nearby camera or with high instability in irradiance will maintain higher ray count.
Probes that use less rays will have slower blending to reduce artifacts.
Added probe instability, attention and stats debugging for devs.
This commit is contained in:
Wojtek Figat
2024-07-18 08:38:23 +02:00
parent ee02aa394a
commit aeff147b6d
3 changed files with 310 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

101
Source/Engine/Renderer/GI/DynamicDiffuseGlobalIllumination.cpp
View File

@@ -40,6 +40,19 @@
#define DDGI_PROBE_RESOLUTION_DISTANCE 14 // Resolution (in texels) for probe distance data (excluding 1px padding on each side)
#define DDGI_PROBE_UPDATE_BORDERS_GROUP_SIZE 8
#define DDGI_PROBE_CLASSIFY_GROUP_SIZE 32
#define DDGI_DEBUG_STATS 0 // Enables additional GPU-driven stats for probe/rays count
#define DDGI_DEBUG_INSTABILITY 0 // Enables additional probe irradiance instability debugging
#if DDGI_DEBUG_STATS
#include "Engine/Core/Collections/SamplesBuffer.h"
#define DDGI_DEBUG_STATS_FRAMES 60
struct StatsData
{
uint32 RaysCount;
uint32 ProbesCount;
};
#endif
GPU_CB_STRUCT(Data0 {
DynamicDiffuseGlobalIlluminationPass::ConstantsData DDGI;
@@ -52,11 +65,13 @@ GPU_CB_STRUCT(Data0 {
float ResetBlend;
float TemporalTime;
Int4 ProbeScrollClears[4];
Float3 ViewDir;
float Padding1;
});
GPU_CB_STRUCT(Data1 {
// TODO: use push constants on Vulkan or root signature data on DX12 to reduce overhead of changing single DWORD
Float2 Padding1;
Float2 Padding2;
uint32 CascadeIndex;
uint32 ProbeIndexOffset;
});
@@ -84,11 +99,21 @@ public:
int32 ProbesCountTotal = 0;
Int3 ProbeCounts = Int3::Zero;
GPUTexture* ProbesTrace = nullptr; // Probes ray tracing: (RGB: hit radiance, A: hit distance)
GPUTexture* ProbesData = nullptr; // Probes data: (RGB: world-space offset, A: state/data)
GPUTexture* ProbesData = nullptr; // Probes data: (RGB: probe-space offset, A: state/data)
GPUTexture* ProbesIrradiance = nullptr; // Probes irradiance (RGB: sRGB color)
GPUTexture* ProbesDistance = nullptr; // Probes distance (R: mean distance, G: mean distance^2)
GPUBuffer* ActiveProbes = nullptr; // List with indices of the active probes (built during probes classification to use indirect dispatches for probes updating), counter at 0
GPUBuffer* UpdateProbesInitArgs = nullptr; // Indirect dispatch buffer for active-only probes updating (trace+blend)
#if DDGI_DEBUG_STATS
GPUBuffer* StatsWrite = nullptr;
GPUBuffer* StatsRead = nullptr;
SamplesBuffer<uint32, DDGI_DEBUG_STATS_FRAMES> StatsProbes;
SamplesBuffer<uint32, DDGI_DEBUG_STATS_FRAMES> StatsRays;
uint32 StatsFrames = 0;
#endif
#if DDGI_DEBUG_INSTABILITY
GPUTexture* ProbesInstability = nullptr;
#endif
DynamicDiffuseGlobalIlluminationPass::BindingData Result;
FORCE_INLINE void Release()
@@ -99,6 +124,16 @@ public:
RenderTargetPool::Release(ProbesDistance);
SAFE_DELETE_GPU_RESOURCE(ActiveProbes);
SAFE_DELETE_GPU_RESOURCE(UpdateProbesInitArgs);
#if DDGI_DEBUG_STATS
SAFE_DELETE_GPU_RESOURCE(StatsWrite);
SAFE_DELETE_GPU_RESOURCE(StatsRead);
StatsProbes.Clear();
StatsRays.Clear();
StatsFrames = 0;
#endif
#if DDGI_DEBUG_INSTABILITY
RenderTargetPool::Release(ProbesInstability);
#endif
}
~DDGICustomBuffer()
@@ -373,12 +408,21 @@ bool DynamicDiffuseGlobalIlluminationPass::RenderInner(RenderContext& renderCont
INIT_TEXTURE(ProbesData, PixelFormat::R8G8B8A8_SNorm, probesCountTotalX, probesCountTotalY);
INIT_TEXTURE(ProbesIrradiance, PixelFormat::R11G11B10_Float, probesCountTotalX * (DDGI_PROBE_RESOLUTION_IRRADIANCE + 2), probesCountTotalY * (DDGI_PROBE_RESOLUTION_IRRADIANCE + 2));
INIT_TEXTURE(ProbesDistance, PixelFormat::R16G16_Float, probesCountTotalX * (DDGI_PROBE_RESOLUTION_DISTANCE + 2), probesCountTotalY * (DDGI_PROBE_RESOLUTION_DISTANCE + 2));
#if DDGI_DEBUG_INSTABILITY
INIT_TEXTURE(ProbesInstability, PixelFormat::R16_Float, probesCountTotalX * (DDGI_PROBE_RESOLUTION_IRRADIANCE + 2), probesCountTotalY * (DDGI_PROBE_RESOLUTION_IRRADIANCE + 2));
#endif
#undef INIT_TEXTURE
#define INIT_BUFFER(buffer, name) ddgiData.buffer = GPUDevice::Instance->CreateBuffer(TEXT(name)); if (!ddgiData.buffer || ddgiData.buffer->Init(desc2)) return true; memUsage += ddgiData.buffer->GetMemoryUsage();
GPUBufferDescription desc2 = GPUBufferDescription::Raw((probesCountCascade + 1) * sizeof(uint32), GPUBufferFlags::ShaderResource | GPUBufferFlags::UnorderedAccess);
INIT_BUFFER(ActiveProbes, "DDGI.ActiveProbes");
desc2 = GPUBufferDescription::Buffer(sizeof(GPUDispatchIndirectArgs) * Math::DivideAndRoundUp(probesCountCascade, DDGI_TRACE_RAYS_PROBES_COUNT_LIMIT), GPUBufferFlags::Argument | GPUBufferFlags::UnorderedAccess, PixelFormat::R32_UInt, nullptr, sizeof(uint32));
INIT_BUFFER(UpdateProbesInitArgs, "DDGI.UpdateProbesInitArgs");
#if DDGI_DEBUG_STATS
desc2 = GPUBufferDescription::Raw(sizeof(StatsData), GPUBufferFlags::UnorderedAccess);
INIT_BUFFER(StatsWrite, "DDGI.StatsWrite");
desc2 = desc2.ToStagingReadback();
INIT_BUFFER(StatsRead, "DDGI.StatsRead");
#endif
#undef INIT_BUFFER
LOG(Info, "Dynamic Diffuse Global Illumination probes: {0}, memory usage: {1} MB", probesCountTotal, memUsage / (1024 * 1024));
clear = true;
@@ -393,6 +437,9 @@ bool DynamicDiffuseGlobalIlluminationPass::RenderInner(RenderContext& renderCont
context->ClearUA(ddgiData.ProbesData, Float4::Zero);
context->ClearUA(ddgiData.ProbesIrradiance, Float4::Zero);
context->ClearUA(ddgiData.ProbesDistance, Float4::Zero);
#if DDGI_DEBUG_INSTABILITY
context->ClearUA(ddgiData.ProbesInstability, Float4::Zero);
#endif
}
ddgiData.LastFrameUsed = Engine::FrameCount;
@@ -486,6 +533,7 @@ bool DynamicDiffuseGlobalIlluminationPass::RenderInner(RenderContext& renderCont
data.ProbeScrollClears[cascadeIndex] = Int4(cascade.ProbeScrollClears, 0);
}
data.TemporalTime = renderContext.List->Setup.UseTemporalAAJitter ? RenderTools::ComputeTemporalTime() : 0.0f;
data.ViewDir = renderContext.View.Direction;
GBufferPass::SetInputs(renderContext.View, data.GBuffer);
context->UpdateCB(_cb0, &data);
context->BindCB(0, _cb0);
@@ -496,6 +544,10 @@ bool DynamicDiffuseGlobalIlluminationPass::RenderInner(RenderContext& renderCont
PROFILE_GPU_CPU_NAMED("Probes Update");
bool anyDirty = false;
uint32 threadGroupsX, threadGroupsY;
#if DDGI_DEBUG_STATS
uint32 zero[4] = {};
context->ClearUA(ddgiData.StatsWrite, zero);
#endif
for (int32 cascadeIndex = 0; cascadeIndex < cascadesCount; cascadeIndex++)
{
if (cascadeSkipUpdate[cascadeIndex])
@@ -556,6 +608,9 @@ bool DynamicDiffuseGlobalIlluminationPass::RenderInner(RenderContext& renderCont
context->BindSR(8, skybox);
context->BindSR(9, ddgiData.ActiveProbes->View());
context->BindUA(0, ddgiData.ProbesTrace->View());
#if DDGI_DEBUG_STATS
context->BindUA(1, ddgiData.StatsWrite->View());
#endif
context->DispatchIndirect(_csTraceRays[(int32)Graphics::GIQuality], ddgiData.UpdateProbesInitArgs, arg);
context->ResetUA();
context->ResetSR();
@@ -564,21 +619,55 @@ bool DynamicDiffuseGlobalIlluminationPass::RenderInner(RenderContext& renderCont
// Update probes irradiance and distance textures (one thread-group per probe)
{
PROFILE_GPU_CPU_NAMED("Update Probes");
// Distance
context->BindSR(0, ddgiData.Result.ProbesData);
context->BindSR(1, ddgiData.ProbesTrace->View());
context->BindSR(2, ddgiData.ActiveProbes->View());
context->BindUA(0, ddgiData.Result.ProbesIrradiance);
context->DispatchIndirect(_csUpdateProbesIrradiance, ddgiData.UpdateProbesInitArgs, arg);
context->BindUA(0, ddgiData.Result.ProbesDistance);
context->DispatchIndirect(_csUpdateProbesDistance, ddgiData.UpdateProbesInitArgs, arg);
context->ResetUA();
context->ResetSR();
// Irradiance
context->BindSR(1, ddgiData.ProbesTrace->View());
context->BindSR(2, ddgiData.ActiveProbes->View());
context->BindUA(0, ddgiData.Result.ProbesIrradiance);
context->BindUA(1, ddgiData.Result.ProbesData);
#if DDGI_DEBUG_INSTABILITY
context->BindUA(2, ddgiData.ProbesInstability->View());
#endif
context->DispatchIndirect(_csUpdateProbesIrradiance, ddgiData.UpdateProbesInitArgs, arg);
context->ResetUA();
context->ResetSR();
}
arg += sizeof(GPUDispatchIndirectArgs);
}
}
#if DDGI_DEBUG_STATS
// Update stats
{
StatsData stats;
if (void* mapped = ddgiData.StatsRead->Map(GPUResourceMapMode::Read))
{
Platform::MemoryCopy(&stats, mapped, sizeof(stats));
ddgiData.StatsRead->Unmap();
ddgiData.StatsProbes.Add(stats.ProbesCount);
ddgiData.StatsRays.Add(stats.RaysCount);
}
context->CopyBuffer(ddgiData.StatsRead, ddgiData.StatsWrite, sizeof(stats));
if (++ddgiData.StatsFrames >= DDGI_DEBUG_STATS_FRAMES)
{
ddgiData.StatsFrames = 0;
stats.ProbesCount = ddgiData.StatsProbes.Average();
stats.RaysCount = ddgiData.StatsRays.Average();
LOG(Info, "DDGI active probes: {}, traced rays: {} per frame, rays per probe: {}", stats.ProbesCount, stats.RaysCount, stats.ProbesCount > 0 ? stats.RaysCount / stats.ProbesCount : 0);
}
}
#endif
// Update probes border pixels
if (anyDirty)
{
@@ -718,7 +807,11 @@ bool DynamicDiffuseGlobalIlluminationPass::Render(RenderContext& renderContext,
{
// Pass DDGI data to the material
_debugMaterial->SetParameterValue(TEXT("ProbesData"), Variant(ddgiData.ProbesData));
#if DDGI_DEBUG_INSTABILITY
_debugMaterial->SetParameterValue(TEXT("ProbesIrradiance"), Variant(ddgiData.ProbesInstability));
#else
_debugMaterial->SetParameterValue(TEXT("ProbesIrradiance"), Variant(ddgiData.ProbesIrradiance));
#endif
_debugMaterial->SetParameterValue(TEXT("ProbesDistance"), Variant(ddgiData.ProbesDistance));
auto cb = _debugMaterial->GetShader()->GetCB(3);
if (cb)