how on earth did RWMC even work before !?

Author: devsh

22.RaytracedAO/Renderer.cpp

@@ -1254,6 +1254,8 @@ void Renderer::initScreenSizedResources(
 		const bool baseIsKnown = cascadeLuminanceBase>std::numeric_limits<float>::min();
 		if (core::isnan<float>(cascadeLuminanceStart))
 			cascadeLuminanceStart = baseIsKnown ? (maxEmitterRadianceLuma*std::pow(cascadeLuminanceBase,-cascadeSegmentCount)):Emin;
+		// rationale, we don't have NEE and BRDF importance sampling samples with throughput <= 1.0
+		// However we have RIS, and that can complicate this assumption a bit
 		if (!baseIsKnown)
 			cascadeLuminanceBase = core::max(std::pow(maxEmitterRadianceLuma/cascadeLuminanceStart,1.f/cascadeSegmentCount),1.0625f);
 		std::cout << "Re-Weighting Monte Carlo = ENABLED [cascadeCount: "<<cascadeCount<<", start: "<<cascadeLuminanceStart<<", base: "<<cascadeLuminanceBase<<"]" << std::endl;

22.RaytracedAO/closestHit.comp

@@ -98,7 +98,7 @@ void main()
 				);
 
 				const vec3 origin = dPdBary*compactBary+lastVxPos;
-				generate_next_rays(
+				rayMask = generate_next_rays(
 					MAX_RAYS_GENERATED,material,frontfacing,vertex_depth,
 					scramble_start_state,sampleID,outPixelLocation,origin,
 					throughput,aovThroughputScale,contrib.albedo,contrib.worldspaceNormal

22.RaytracedAO/main.cpp

@@ -643,11 +643,19 @@ int main(int argc, char** argv)
 		mainSensorData.denoiserInfo.bloomIntensity = film.denoiserBloomIntensity;
 		mainSensorData.denoiserInfo.tonemapperArgs = std::string(film.denoiserTonemapperArgs);
 		mainSensorData.fileFormat = film.fileFormat;
+#ifndef UBER_TEST
+		mainSensorData.cascadeCount = 6;
+		mainSensorData.cascadeLuminanceBase = film.cascadeLuminanceBase;
+		mainSensorData.cascadeLuminanceStart = film.cascadeLuminanceStart;
+		mainSensorData.kappa = 1.f;
+		mainSensorData.Emin = 1.7f;
+#else
 		mainSensorData.cascadeCount = film.cascadeCount;
 		mainSensorData.cascadeLuminanceBase = film.cascadeLuminanceBase;
 		mainSensorData.cascadeLuminanceStart = film.cascadeLuminanceStart;
 		mainSensorData.kappa = mainSensorData.cascadeCount<2 ? 0.f:film.rfilter.kappa;
 		mainSensorData.Emin = film.rfilter.Emin;
+#endif
 		mainSensorData.envmapRegFactor = core::clamp(film.envmapRegularizationFactor, 0.0f, 0.8f);
 		mainSensorData.outputFilePath = std::filesystem::path(film.outputFilePath);
 		// handle missing output path

22.RaytracedAO/raygen.comp

@@ -152,7 +152,7 @@ void main()
 						else if (cascadeIx==higherCascade)
 							weight = splat.cascadeWeights[1];
 						nextSampleAccumulationCascade(
-							contrib.color*weight,coord,samplesPerPixelPerDispatch,
+							firstFrame,contrib.color*weight,coord,samplesPerPixelPerDispatch,
 							cascadeIx,pc.cummon.rcpFramesDispatched
 						);
 					}
@@ -162,7 +162,8 @@ void main()
 				// clear accumulations totally if beginning a new frame
 				if (firstFrame)
 				{
-					storeAccumulation(contrib.color,coord);
+					if (!isRWMCEnabled())
+						storeAccumulation(contrib.color,coord);
 					storeAlbedo(contrib.albedo,coord);
 					storeWorldspaceNormal(contrib.worldspaceNormal,coord);
 					storeMask(hideEnvmap&&(!hit) ? 1.f:0.f,coord);

22.RaytracedAO/raytraceCommon.glsl

@@ -104,16 +104,14 @@ void addAccumulation(in vec3 delta, in uvec3 coord)
 
 // TODO: use a R17G17B17_UNORM format matched to cascade range, then use 13 bits to store last spp count (max 8k spp renders)
 // This way we can avoid writing every cascade every path storage
-void nextSampleAccumulationCascade(in vec3 weightedDelta, uvec3 coord, in uint samplesPerPixelPerDispatch, in uint cascadeIndex, in float rcpN)
+void nextSampleAccumulationCascade(in bool firstFrame, in vec3 weightedDelta, uvec3 coord, in uint samplesPerPixelPerDispatch, in uint cascadeIndex, in float rcpN)
 {
 	// but leave first index in the array for the ray accumulation metadata, hence the +1
 	coord.z += (cascadeIndex+1u)*samplesPerPixelPerDispatch;
-	const vec3 prev = fetchAccumulation(coord);
+	const vec3 prev = firstFrame ? vec3(0.0):fetchAccumulation(coord);
 	const vec3 newVal = prev+(weightedDelta-prev)*rcpN;
-	// TODO: do a better check, compare actually encoded values for difference
-	const uvec3 diff = floatBitsToUint(newVal)^floatBitsToUint(prev);
-	if (bool((diff.x|diff.y|diff.z)&0x7ffffff0u))
-		storeAccumulation(newVal,coord);
+	// always store, cause we need to reset the value
+	storeAccumulation(newVal,coord);
 }
 void addAccumulationCascade(in vec3 weightedDelta, uvec3 coord, in uint samplesPerPixelPerDispatch, in uint cascadeIndex)
 {