Skip to content

Commit

Permalink
New radiance volume
Browse files Browse the repository at this point in the history
  • Loading branch information
@tippesi
tippesi committed Jan 2, 2025
1 parent 0c5a770 commit 2c033ef
Show file tree
Hide file tree
Showing 24 changed files with 860 additions and 138 deletions.
436 changes: 433 additions & 3 deletions data/shader/ddgi/ddgi.hsh
View file

Large diffs are not rendered by default.

2 changes: 1 addition & 1 deletion data/shader/ddgi/probeDebug.fsh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,6 @@ void main() {

materialIdxFS = GetProbeState(instanceID) == PROBE_STATE_ACTIVE ?
pushConstants.probeActiveMaterialIdx : pushConstants.probeInactiveMaterialIdx;
//materialIdxFS = probeAge > 32 ? pushConstants.probeOffsetMaterialIdx : materialIdxFS;
materialIdxFS = probeAge < 32 ? pushConstants.probeOffsetMaterialIdx : materialIdxFS;

}
27 changes: 20 additions & 7 deletions data/shader/ddgi/probeState.csh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,11 +47,12 @@ void main() {
GetProbeHistoryInfo(ivec3(gl_WorkGroupID.xyz), cascadeIndex, historyProbeCoord, reset);

uint historyBaseIdx = Flatten3D(ivec3(historyProbeCoord.xzy), ivec3(gl_NumWorkGroups.xzy));
//historyBaseIdx = baseIdx;

if (gl_LocalInvocationID.x == 0u) {
backFaceHits = 0u;
inCellHits = 0u;
probeState = GetProbeState(historyBaseIdx);
probeState = floatBitsToUint(historyProbeStates[historyBaseIdx].x);
temporalCellHits = (probeState == PROBE_STATE_NEW || reset) ? 0.01 : historyProbeStates[historyBaseIdx].y;
temporalBackFaceRatio = (probeState == PROBE_STATE_NEW || reset) ? 0.25 : historyProbeStates[historyBaseIdx].z;
}
Expand All @@ -68,11 +69,11 @@ void main() {
for(uint i = gl_LocalInvocationIndex; i < probeRayCount; i += workGroupOffset) {
RayHit hit = UnpackRayHit(hits[rayBaseIdx + i]);

bool backface = hit.hitDistance <= 0.0;
bool backface = float(hit.radiance.a) <= 0.0;
if (backface) {
atomicAdd(backFaceHits, uint(1));
}
if (hit.hitDistance < extendedSize && !backface) {
if (float(hit.radiance.a) < extendedSize && !backface) {
atomicAdd(inCellHits, uint(1));
}
}
Expand All @@ -83,24 +84,36 @@ void main() {
uint historyProbeState = probeState;
probeState = PROBE_STATE_INACTIVE;

temporalCellHits = mix(float(inCellHits), temporalCellHits, ddgiData.hysteresis);
float hysteresis = ddgiData.hysteresis;
if (inCellHits == 0) {
hysteresis = 0.5;
}

// Use temporally stable information to decide probe state
temporalCellHits = mix(float(inCellHits), temporalCellHits, ddgiData.hysteresis);
// Use temporally stable information to decide probe state (remeber hits are in )
if (temporalCellHits > 0.01) {
probeState = PROBE_STATE_ACTIVE;
}

float backFaceRatio = float(backFaceHits) / probeRayCount;
temporalBackFaceRatio = mix(float(backFaceRatio), temporalBackFaceRatio, ddgiData.hysteresis);

if (temporalBackFaceRatio > 0.25) {
if (backFaceRatio < 0.05) {
temporalBackFaceRatio = backFaceRatio;
}

if (backFaceRatio > 0.5) {
temporalBackFaceRatio = backFaceRatio;
}

if (temporalBackFaceRatio > 0.15) {
probeState = PROBE_STATE_INACTIVE;
}

probeStates[baseIdx].x = uintBitsToFloat(probeState);
probeStates[baseIdx].y = temporalCellHits;
probeStates[baseIdx].z = temporalBackFaceRatio;
probeStates[baseIdx].w = (historyProbeState == PROBE_STATE_NEW || reset) ? 1.0 : min(256.0, historyProbeStates[historyBaseIdx].w + 1.0);
probeStates[baseIdx].w = (historyProbeState == PROBE_STATE_NEW || reset || probeState == PROBE_STATE_ACTIVE && historyProbeState == PROBE_STATE_INACTIVE) ? 0.0 : min(256.0, historyProbeStates[historyBaseIdx].w + 1.0);
}

}
155 changes: 111 additions & 44 deletions data/shader/ddgi/probeUpdate.csh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,9 +16,16 @@
#include <../common/flatten.hsh>
#include <../common/utility.hsh>
#include <../common/barrier.hsh>
#include <../common/types.hsh>

#ifdef IRRADIANCE
#if defined(IRRADIANCE)
layout (local_size_x = 6, local_size_y = 6) in;
#elif defined(RADIANCE)
#ifdef LOWER_RES_RADIANCE
layout (local_size_x = 14, local_size_y = 14) in;
#else
layout (local_size_x = 30, local_size_y = 30) in;
#endif
#else
#ifdef LOWER_RES_MOMENTS
layout (local_size_x = 6, local_size_y = 6) in;
Expand All @@ -27,8 +34,10 @@ layout (local_size_x = 14, local_size_y = 14) in;
#endif
#endif

#ifdef IRRADIANCE
#if defined(IRRADIANCE)
layout (set = 3, binding = 0, rgb10_a2) writeonly uniform image2DArray irradiance;
#elif defined(RADIANCE)
layout (set = 3, binding = 0, rgb10_a2) writeonly uniform image2DArray radiance;
#else
layout (set = 3, binding = 0, rg16f) writeonly uniform image2DArray moment;
#endif
Expand All @@ -45,50 +54,64 @@ layout(std430, set = 3, binding = 3) buffer HistoryProbeOffsets {
vec4 historyProbeOffsets[];
};

const uint sharedSize = 32;
const uint sharedSize = 128;

#ifdef IRRADIANCE
#if defined(IRRADIANCE) || defined(RADIANCE)
struct RayData {
// This seems to be a better layout than two vec3's
// My guess is that vec3's get expanded to vec4's
#ifndef AE_HALF_FLOAT
vec4 direction;
vec4 radiance;
#else
AeF16x4 direction;
AeF16x4 radiance;
#endif
};
#else
struct RayData {
vec3 direction;
float dist;
AeF16x4 direction;
};
#endif
shared RayData rayData[sharedSize];
vec4 GetHistoryFromHigherCascade(ivec2 pixel, int cascadeIndex) {
vec4 GetHistoryFromHigherCascade(ivec2 pixel, int cascadeIndex) {
ivec3 workGroup = ivec3(gl_WorkGroupID.xyz);
workGroup.y %= ddgiData.volumeProbeCount.y;
vec3 probePosition = vec3(gl_WorkGroupID.xyz) * ddgiData.cascades[cascadeIndex].cellSize.xyz + ddgiData.cascades[cascadeIndex].volumeMin.xyz;
vec3 probePosition = vec3(workGroup) * ddgiData.cascades[cascadeIndex].cellSize.xyz + ddgiData.cascades[cascadeIndex].volumeMin.xyz;
// Note: Cascades are orderd from biggest (index 0) to smallest (index cascadeCount - 1)
cascadeIndex = max(cascadeIndex - 1, 0);
vec3 localPosition = probePosition - ddgiData.cascades[cascadeIndex].volumeMin.xyz;
ivec3 baseCell = ivec3(localPosition / ddgiData.cascades[cascadeIndex].cellSize.xyz);
ivec3 cascadeProbeOffset = ivec3(0, cascadeIndex * ddgiData.volumeProbeCount.y, 0);
ivec3 baseCell = ivec3(localPosition / ddgiData.cascades[cascadeIndex].cellSize.xyz) + cascadeProbeOffset;
bool reset;
ivec3 historyProbeCoord;
GetProbeHistoryInfo(baseCell, cascadeIndex, historyProbeCoord, reset);
#ifdef IRRADIANCE
#if defined(IRRADIANCE)
ivec2 res = ivec2(ddgiData.volumeIrradianceRes);
#elif defined(RADIANCE)
ivec2 res = ivec2(ddgiData.volumeRadianceRes);
#else
ivec2 res = ivec2(ddgiData.volumeMomentsRes);
#endif
ivec2 historyResOffset = (res + ivec2(2)) * ivec2(historyProbeCoord.xz) + ivec2(1);
ivec3 historyVolumeCoord = ivec3(historyResOffset + pixel, int(historyProbeCoord.y));
#ifdef IRRADIANCE
#if defined(IRRADIANCE)
vec3 lastResult = texelFetch(irradianceVolume, historyVolumeCoord, 0).rgb;
return vec4(lastResult, 0.0);
#elif defined(RADIANCE)
vec3 lastResult = texelFetch(radianceVolume, historyVolumeCoord, 0).rgb;
return vec4(lastResult, 0.0);
#else
vec2 lastResult = texelFetch(momentsVolume, historyVolumeCoord, 0).rg;
return vec4(lastResult, 0.0, 0.0);
Expand All @@ -108,32 +131,36 @@ void main() {
uint historyBaseIdx = Flatten3D(ivec3(historyProbeCoord.xzy), ivec3(gl_NumWorkGroups.xzy));
uint probeState = floatBitsToUint(historyProbeStates[historyBaseIdx].x);
float probeAge = historyProbeStates[historyBaseIdx].w;
float probeAge = probeStates[baseIdx].w;
vec4 probeOffset = reset ? vec4(0.0, 0.0, 0.0, 1.0) : historyProbeOffsets[historyBaseIdx];
uint rayBaseIdx = baseIdx * ddgiData.rayCount;
uint probeRayCount = GetProbeRayCount(probeState);
uint groupSize = gl_WorkGroupSize.x * gl_WorkGroupSize.y;
#ifdef IRRADIANCE
#if defined(IRRADIANCE)
ivec2 res = ivec2(ddgiData.volumeIrradianceRes);
#elif defined(RADIANCE)
ivec2 res = ivec2(ddgiData.volumeRadianceRes);
#else
ivec2 res = ivec2(ddgiData.volumeMomentsRes);
#endif
ivec2 pix = ivec2(gl_LocalInvocationID);
vec2 coord = (vec2(pix) + vec2(0.5)) / vec2(res);
vec3 N = OctahedronToUnitVector(coord);
AeF16x3 N = AeF16x3(OctahedronToUnitVector(coord));
ivec2 resOffset = (res + ivec2(2)) * ivec2(gl_WorkGroupID.xz) + ivec2(1);
ivec3 volumeCoord = ivec3(resOffset + pix, int(gl_WorkGroupID.y));
float cellLength = length(ddgiData.cascades[cascadeIndex].cellSize.xyz);
float maxDepth = cellLength * 0.75;
AeF16 cellLength = AeF16(length(ddgiData.cascades[cascadeIndex].cellSize.xyz));
AeF16 maxDepth = cellLength * AeF16(0.75);
vec4 result = vec4(0.0);
AeF16 depthSharpness = AeF16(ddgiData.depthSharpness);
AeF16x4 result = AeF16x4(0.0);
vec3 newProbeOffset = probeOffset.xyz;
for (uint i = 0; i < probeRayCount; i += sharedSize) {
// We might not have a multiple of shared size in terms of rays
Expand All @@ -142,26 +169,39 @@ void main() {
// Load rays cooperatively
for (uint j = gl_LocalInvocationIndex; j < loadRayCount; j += groupSize) {
RayHit hit = UnpackRayHit(hits[rayBaseIdx + i + j]);
rayData[j].direction.rgb = hit.direction;
#ifdef IRRADIANCE
rayData[j].direction.rgb = hit.direction.xyz;
#if defined(IRRADIANCE) || defined(RADIANCE)
rayData[j].radiance.rgb = hit.radiance.rgb;
rayData[j].direction.a = hit.hitDistance;
rayData[j].direction.a = hit.radiance.a;
#else
rayData[j].dist = hit.hitDistance;
#ifndef AE_HALF_FLOAT
float dist = hit.radiance.a;
dist = dist < 0.0 ? dist * 0.2 : dist;
float hitDistance = min(maxDepth, dist);
#else
AeF16 dist = hit.radiance.a;
dist = dist < AeF16(0.0) ? dist * AeF16(0.2) : dist;
AeF16 hitDistance = min(maxDepth, dist);
#endif
rayData[j].direction.w = hitDistance;
#endif
}
barrier();
float maxWeight = 0.0;
// Iterate over all rays in the shared memory
for (uint j = 0; j < loadRayCount; j++) {
#ifdef IRRADIANCE
float weight = max(0.0, dot(N, rayData[j].direction.rgb));
#if defined(IRRADIANCE)
AeF16 weight = max(0.0hf, dot(N, rayData[j].direction.rgb));
if (weight >= 0.00001) {
vec3 radiance = vec3(rayData[j].radiance);
result += vec4(radiance, 1.0) * weight;
if (weight >= 0.00001hf) {
AeF16x3 radiance = AeF16x3(rayData[j].radiance);
result += AeF16x4(radiance, 1.0hf) * weight;
}
const float probeOffsetDistance = max(ddgiData.cascades[cascadeIndex].cellSize.w * 0.05, 0.5);
Expand All @@ -183,16 +223,22 @@ void main() {
//newProbeOffset -= rayData[j].direction.xyz * (probeOffsetDistance - dist) * 0.001 * probeOffset.w / probeOffsetDistance;
}
}
#elif defined(RADIANCE)
AeF16 weight = max(0.0hf, dot(N, rayData[j].direction.rgb));
weight = pow(weight, 256.0hf);
AeF16x3 radiance = rayData[j].radiance.xyz;
if (weight > 0.15)
result += AeF16x4(radiance, 1.0hf) * weight;
#else
float dist = rayData[j].dist;
dist = dist < 0.0 ? dist * 0.2 : dist;
AeF16 weight = max(0.0hf, dot(N, rayData[j].direction.xyz));
float hitDistance = min(maxDepth, dist);
float weight = max(0.0, dot(N, rayData[j].direction));
AeF16 hitDistance = rayData[j].direction.w;
weight = pow(weight, ddgiData.depthSharpness);
if (weight >= 0.00000001) {
result += vec4(hitDistance, sqr(hitDistance), 0.0, 1.0) * weight;
weight = pow(weight, depthSharpness);
if (weight >= 0.00000001hf) {
result += AeF16x4(hitDistance, sqr(hitDistance), 0.0hf, 1.0hf) * weight;
}
#endif
}
Expand All @@ -205,24 +251,25 @@ void main() {
ivec3 historyVolumeCoord = ivec3(historyResOffset + pix, int(historyProbeCoord.y));
// Use a dynamic hysteris based on probe age to accumulate more efficiently at the beginning of a probes life
float hysteresis = min(ddgiData.hysteresis, probeAge / (probeAge + 1.0));
//hysteresis = ddgiData.hysteresis;
float hysteresis = clamp(probeAge / (probeAge + 1.0), 0.9, ddgiData.hysteresis);
hysteresis = ddgiData.hysteresis;
#ifdef IRRADIANCE
#if defined(IRRADIANCE)
vec3 lastResult = texelFetch(irradianceVolume, historyVolumeCoord, 0).rgb;
if (probeState == PROBE_STATE_NEW || reset) {
//lastResult = GetHistoryFromHigherCascade(pix, cascadeIndex).rgb;
lastResult = GetHistoryFromHigherCascade(pix, cascadeIndex).rgb;
}
vec3 resultOut = lastResult;
if (result.w > 0.0) {
result.xyz /= result.w;
result.xyz = pow(result.xyz, vec3(1.0 / ddgiData.volumeGamma));
vec4 fullResult = vec4(result);
fullResult.xyz /= fullResult.w;
fullResult.xyz = pow(fullResult.xyz, vec3(1.0 / ddgiData.volumeGamma));
if (probeState == PROBE_STATE_NEW || reset) {
resultOut = mix(result.xyz, lastResult, hysteresis);
resultOut = mix(fullResult.xyz, lastResult, hysteresis);
}
else {
resultOut = mix(result.xyz, lastResult, hysteresis);
resultOut = mix(fullResult.xyz, lastResult, hysteresis);
}
}
imageStore(irradiance, volumeCoord, vec4(resultOut, 0.0));
Expand All @@ -233,18 +280,38 @@ void main() {
probeOffset.w = max(0.0, reset ? 1.0 : probeOffset.w - 0.01);
probeOffsets[baseIdx] = ddgiData.optimizeProbes > 0 ? probeOffset : vec4(0.0, 0.0, 0.0, 1.0);
}
#elif defined(RADIANCE)
vec3 lastResult = texelFetch(radianceVolume, historyVolumeCoord, 0).rgb;
if (probeState == PROBE_STATE_NEW || reset) {
lastResult = GetHistoryFromHigherCascade(pix, cascadeIndex).rgb;
}
vec3 resultOut = lastResult;
if (result.w > 0.0) {
vec4 fullResult = vec4(result);
fullResult.xyz /= fullResult.w;
fullResult.xyz = pow(fullResult.xyz, vec3(1.0 / ddgiData.volumeGamma));
if (probeState == PROBE_STATE_NEW || reset) {
resultOut = mix(fullResult.xyz, lastResult, hysteresis);
}
else {
resultOut = mix(fullResult.xyz, lastResult, hysteresis);
}
}
imageStore(radiance, volumeCoord, vec4(resultOut, 0.0));
#else
vec2 lastResult = texelFetch(momentsVolume, historyVolumeCoord, 0).rg;
if (probeState == PROBE_STATE_NEW || reset) {
//lastResult = GetHistoryFromHigherCascade(pix, cascadeIndex).rg;
lastResult = GetHistoryFromHigherCascade(pix, cascadeIndex).rg;
}
vec2 resultOut = lastResult;
if (result.w > 0.0) {
vec4 fullResult = vec4(result);
if (probeState == PROBE_STATE_NEW || reset) {
resultOut = result.xy / result.w;
resultOut = fullResult.xy / fullResult.w;
}
else {
resultOut = mix(result.xy / result.w, lastResult, hysteresis);
resultOut = mix(fullResult.xy / fullResult.w, lastResult, hysteresis);
}
}
Expand Down
Loading

0 comments on commit 2c033ef

Please sign in to comment.