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

Optimize inactive probes update in DDGI with faster Jump Flood instead of Flood Fill

Browse Source
This commit is contained in:
Wojtek Figat
2026-01-06 00:01:17 +01:00
parent 0975ac2a25
commit 32e725392b
3 changed files with 46 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
Source/Engine/Renderer/GI/DynamicDiffuseGlobalIllumination.cpp
View File

@@ -42,7 +42,7 @@
#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_PROBE_EMPTY_AREA_DENSITY 10 // Spacing (in probe grid) between fallback probes placed into empty areas to provide valid GI for nearby dynamic objects or transparency
#define DDGI_PROBE_EMPTY_AREA_DENSITY 8 // Spacing (in probe grid) between fallback probes placed into empty areas to provide valid GI for nearby dynamic objects or transparency
#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
@@ -76,7 +76,8 @@ GPU_CB_STRUCT(Data0 {
GPU_CB_STRUCT(Data1 {
// TODO: use push constants on Vulkan or root signature data on DX12 to reduce overhead of changing single DWORD
Float2 Padding2;
float Padding2;
int32 StepSize;
uint32 CascadeIndex;
uint32 ProbeIndexOffset;
});
@@ -594,10 +595,17 @@ bool DynamicDiffuseGlobalIlluminationPass::RenderInner(RenderContext& renderCont
// For inactive probes, search nearby ones to find the closest valid for quick fallback when sampling irradiance
{
PROFILE_GPU_CPU_NAMED("Update Inactive Probes");
// TODO: this could run within GPUComputePass during Trace Rays or Update Probes to overlap compute works
context->BindUA(0, ddgiData.Result.ProbesData);
int32 iterations = Math::Min(probesCounts.MaxValue() - 1, DDGI_PROBE_EMPTY_AREA_DENSITY);
for (int32 i = 0; i < iterations; i++)
Data1 data;
data.CascadeIndex = cascadeIndex;
int32 iterations = Math::CeilToInt(Math::Log2((float)Math::Min(probesCounts.MaxValue(), DDGI_PROBE_EMPTY_AREA_DENSITY) + 1.0f));
for (int32 i = iterations - 1; i >= 0; i--)
{
data.StepSize = Math::FloorToInt(Math::Pow(2, (float)i) + 0.5f); // Jump Flood step size
context->UpdateCB(_cb1, &data);
context->Dispatch(_csUpdateInactiveProbes, threadGroupsX, 1, 1);
}
context->ResetUA();
}

5
Source/Shaders/GI/DDGI.hlsl
View File

@@ -26,8 +26,9 @@
#endif
#define DDGI_SRGB_BLENDING 1 // Enables blending in sRGB color space, otherwise irradiance blending is done in linear space
#define DDGI_DEFAULT_BIAS 0.2f // Default value for DDGI sampling bias
#define DDGI_FALLBACK_COORDS_ENCODE(coord) ((float3)coord / 128.0f)
#define DDGI_FALLBACK_COORDS_DECODE(data) (uint3)(data.xyz * 128.0f)
#define DDGI_FALLBACK_COORDS_ENCODE(coord) ((float3)(coord + 1) / 128.0f)
#define DDGI_FALLBACK_COORDS_DECODE(data) (uint3)(data.xyz * 128.0f - 1)
#define DDGI_FALLBACK_COORDS_VALID(data) (length(data.xyz) > 0)
//#define DDGI_DEBUG_CASCADE 0 // Forces a specific cascade to be only in use (for debugging)
// DDGI data for a constant buffer

70
Source/Shaders/GI/DDGI.shader
View File

@@ -27,7 +27,7 @@
#define DDGI_PROBE_CLASSIFY_GROUP_SIZE 32
#define DDGI_PROBE_RELOCATE_ITERATIVE 1 // If true, probes relocation algorithm tries to move them in additive way, otherwise all nearby locations are checked to find the best position
#define DDGI_PROBE_RELOCATE_FIND_BEST 1 // If true, probes relocation algorithm tries to move to the best matching location within nearby area
#define DDGI_PROBE_EMPTY_AREA_DENSITY 10 // Spacing (in probe grid) between fallback probes placed into empty areas to provide valid GI for nearby dynamic objects or transparency
#define DDGI_PROBE_EMPTY_AREA_DENSITY 8 // Spacing (in probe grid) between fallback probes placed into empty areas to provide valid GI for nearby dynamic objects or transparency
#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
@@ -49,7 +49,8 @@ uint FrameIndexMod8;
META_CB_END
META_CB_BEGIN(1, Data1)
float2 Padding2;
float Padding2;
int StepSize;
uint CascadeIndex;
uint ProbeIndexOffset;
META_CB_END
@@ -364,38 +365,35 @@ void CS_UpdateProbesInitArgs()
#ifdef _CS_UpdateInactiveProbes
globallycoherent RWTexture2D<snorm float4> RWProbesData : register(u0);
RWTexture2D<snorm float4> RWProbesData : register(u0);
void CheckNearbyProbe(inout uint3 fallbackCoords, inout uint probeState, uint3 probeCoords, int3 probeCoordsEnd, int3 offset)
void CheckNearbyProbe(inout uint3 fallbackCoords, inout uint probeState, inout float minDistance, uint3 probeCoords, int3 probeCoordsEnd, int3 offset)
{
uint3 nearbyCoords = (uint3)clamp(((int3)probeCoords + offset), int3(0, 0, 0), probeCoordsEnd);
uint nearbyIndex = GetDDGIScrollingProbeIndex(DDGI, CascadeIndex, nearbyCoords);
float4 nearbyData = RWProbesData[GetDDGIProbeTexelCoords(DDGI, CascadeIndex, nearbyIndex)];
uint nearbyState = DecodeDDGIProbeState(nearbyData);
uint3 nearbyFallbackCoords = DDGI_FALLBACK_COORDS_DECODE(nearbyData);
if (nearbyState != DDGI_PROBE_STATE_INACTIVE)
float nearbyDist = distance((float3)nearbyCoords, (float3)probeCoords);
if (DecodeDDGIProbeState(nearbyData) != DDGI_PROBE_STATE_INACTIVE && nearbyDist < minDistance)
{
// Use nearby probe
fallbackCoords = nearbyCoords;
probeState = nearbyState;
probeState = DDGI_PROBE_STATE_ACTIVE;
minDistance = nearbyDist;
return;
}
// TODO: optimize distance check with squared dst comparision
else if (distance((float3)nearbyFallbackCoords, (float3)probeCoords) < distance((float3)fallbackCoords, (float3)probeCoords))
nearbyCoords = DDGI_FALLBACK_COORDS_DECODE(nearbyData);
nearbyDist = distance((float3)nearbyCoords, (float3)probeCoords);
if (DDGI_FALLBACK_COORDS_VALID(nearbyData) && nearbyDist < minDistance)
{
// Check if fallback probe is actually active (not some leftover memory)
nearbyIndex = GetDDGIScrollingProbeIndex(DDGI, CascadeIndex, nearbyFallbackCoords);
nearbyData = RWProbesData[GetDDGIProbeTexelCoords(DDGI, CascadeIndex, nearbyIndex)];
nearbyState = DecodeDDGIProbeState(nearbyData);
if (nearbyState != DDGI_PROBE_STATE_INACTIVE)
{
// Use fallback of the nearby probe
fallbackCoords = nearbyFallbackCoords;
probeState = DDGI_PROBE_STATE_ACTIVE;
}
// Use fallback probe
fallbackCoords = nearbyCoords;
probeState = DDGI_PROBE_STATE_ACTIVE;
minDistance = nearbyDist;
}
}
// Compute shader to store closest valid probe coords inside inactive probes data for quick fallback lookup when sampling irradiance.
// Uses Jump Flood algorithm.
META_CS(true, FEATURE_LEVEL_SM5)
[numthreads(DDGI_PROBE_CLASSIFY_GROUP_SIZE, 1, 1)]
void CS_UpdateInactiveProbes(uint3 DispatchThreadId : SV_DispatchThreadID)
@@ -409,32 +407,26 @@ void CS_UpdateInactiveProbes(uint3 DispatchThreadId : SV_DispatchThreadID)
int2 probeDataCoords = GetDDGIProbeTexelCoords(DDGI, CascadeIndex, probeIndex);
float4 probeData = RWProbesData[probeDataCoords];
uint probeState = DecodeDDGIProbeState(probeData);
BRANCH
if (probeState == DDGI_PROBE_STATE_INACTIVE)
{
// Find the closest active probe (flood fill)
// Find the closest active probe (Jump Flood)
int3 probeCoordsEnd = (int3)DDGI.ProbesCounts - int3(1, 1, 1);
// Corners
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(1, 1, 1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(-1, 1, 1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(1, -1, 1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(-1, -1, 1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(1, 1, -1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(-1, 1, -1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(1, -1, -1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(-1, -1, -1));
// Sides
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(1, 0, 0));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(0, 1, 0));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(0, 0, 1));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(-1, 0, 0));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(0, -1, 0));
CheckNearbyProbe(fallbackCoords, probeState, probeCoords, probeCoordsEnd, int3(0, 0, -1));
float minDistance = 1e27f;
UNROLL for (int z = -1; z <= 1; z++)
UNROLL for (int y = -1; y <= 1; y++)
UNROLL for (int x = -1; x <= 1; x++)
{
int3 offset = int3(x, y, z) * StepSize;
CheckNearbyProbe(fallbackCoords, probeState, minDistance, probeCoords, probeCoordsEnd, offset);
}
}
// Ensure all threads (within dispatch) got proepr data before writing back to the same memory
DeviceMemoryBarrierWithGroupSync();
// Ensure all threads (within dispatch) got proper data before writing back to the same memory
AllMemoryBarrierWithGroupSync();
// Write modified probe data back (remain inactive)
BRANCH
if (probeState != DDGI_PROBE_STATE_INACTIVE && DispatchThreadId.x < ProbesCount && fallbackCoords.x != 1000)
{
RWProbesData[probeDataCoords] = EncodeDDGIProbeData(DDGI_FALLBACK_COORDS_ENCODE(fallbackCoords), DDGI_PROBE_STATE_INACTIVE, 0.0f);