RWMC is broken, Persistent workgroups are broken (light and camera doesn't update)

devshgraphicsprogramming · devshgraphicsprogramming · commit 5d3b3b27a9b4 · 2026-03-31T04:25:58.000+02:00
diff --git a/31_HLSLPathTracer/app_resources/hlsl/compute_render_scene_impl.hlsl b/31_HLSLPathTracer/app_resources/hlsl/compute_render_scene_impl.hlsl
@@ -87,7 +87,7 @@ void tracePixel(int32_t2 coords)
 	}
 
 	scene_type scene;
-	scene.updateLight(renderPushConstants.generalPurposeLightMatrix);
+	scene.updateLight(renderPushConstants.lightMatrix());
 
 	typename variant_types::raygen_type rayGen;
 	rayGen.pixOffsetParam = pixOffsetParam;
diff --git a/31_HLSLPathTracer/app_resources/hlsl/render_common.hlsl b/31_HLSLPathTracer/app_resources/hlsl/render_common.hlsl
@@ -12,15 +12,28 @@ NBL_CONSTEXPR uint32_t MaxSamplesLog2 = MAX_SAMPLES_LOG2;
 
 struct RenderPushConstants
 {
-    // TODO: cut down the MVP into a compact raygen matrix to get this whole struct to less than 112 bytes!
-	float32_t4x4 invMVP;
-	float32_t3x4 generalPurposeLightMatrix;
+    float32_t3x4 lightMatrix()
+    {
+        float32_t4x3 retval;
+        retval[0] = lightX;
+        retval[1] = lightY;
+        retval[2] = cross(lightX,lightY)*lightZscale;
+        retval[3] = lightPos;
+        return hlsl::transpose(retval);
+    }
+
     uint64_t pSampleSequence;
+	float32_t4x4 invMVP;
+    float32_t3 lightX;
+    float32_t3 lightY;
+    float32_t lightZscale;
+    float32_t3 lightPos;
     // TODO: compact a bit and refactor
     uint32_t sampleCount : MAX_SAMPLES_LOG2;
     uint32_t depth : MAX_DEPTH_LOG2;
     uint32_t sequenceSampleCountLog2 : 5;
     uint32_t unused : 13;
+    uint32_t unused1;
 };
 #undef MAX_SAMPLES_LOG2
 #undef MAX_DEPTH_LOG2
diff --git a/31_HLSLPathTracer/main.cpp b/31_HLSLPathTracer/main.cpp
@@ -1218,12 +1218,21 @@ class HLSLComputePathtracer final : public SimpleWindowedApplication, public Bui
 				const float32_t4x4 modelViewProjectionMatrix = nbl::hlsl::math::linalg::promoted_mul(viewProjectionMatrix, modelMatrix);
 				const float32_t4x4 invMVP = hlsl::inverse(modelViewProjectionMatrix);
 
+				pc.pSampleSequence = m_sequenceBuffer->getDeviceAddress();
 				pc.invMVP = invMVP;
-				pc.generalPurposeLightMatrix = hlsl::float32_t3x4(transpose(m_lightModelMatrix));
+				{
+					const auto matT = hlsl::float32_t4x3(m_lightModelMatrix);
+					pc.lightX = matT[0];
+					pc.lightY = matT[1];
+					// Z had a length and a direction, can point colinear or opposite cross product
+					const auto recon = hlsl::cross(matT[0], matT[1]);
+					pc.lightZscale = hlsl::sign(hlsl::dot(recon,matT[2]))*hlsl::length(matT[2])/hlsl::length(recon);
+					pc.lightPos = matT[3];
+					assert(pc.lightMatrix()==hlsl::transpose(matT));
+				}
 				pc.sampleCount = guiControlled.spp;
 				pc.depth = guiControlled.depth;
 				pc.sequenceSampleCountLog2 = m_sequenceSamplesLog2;
-				pc.pSampleSequence = m_sequenceBuffer->getDeviceAddress();
 				if (guiControlled.useRWMC)
 				{
 					rwmcPushConstants.renderPushConstants = pc;

Original file line number	Diff line number	Diff line change
`@@ -87,7 +87,7 @@ void tracePixel(int32_t2 coords)`
`87`	`87`	`}`
`88`	`88`
`89`	`89`	`scene_type scene;`
`90`		`- scene.updateLight(renderPushConstants.generalPurposeLightMatrix);`
	`90`	`+ scene.updateLight(renderPushConstants.lightMatrix());`
`91`	`91`
`92`	`92`	`typename variant_types::raygen_type rayGen;`
`93`	`93`	`rayGen.pixOffsetParam = pixOffsetParam;`