Merge branch 'dev/adunn/visibility_test_perf' into 'main'

GPU Perf: Helps the compiler generate more optimal code for visibility tests by eliminating a bunch of unnecessary work. Since for visibility tests we only need a subset of the 'Surface' and 'SurfaceInteraction' data we dont need to compute/load all of it. Primarily, this improves the RTXDI and ReSTIR GI buckets.

See merge request lightspeedrtx/dxvk-remix-nv!964

src/dxvk/shaders/rtx/algorithm/geometry_resolver.slangh

@@ -960,7 +960,7 @@ void geometryResolverVertexOutputDebugView(
       uint iterations = 0;
       if (opaqueSurfaceMaterial.heightTextureIndex != BINDING_INDEX_INVALID && opaqueSurfaceMaterial.displaceIn > 0.0f)
       {
-        SurfaceInteraction tempSurfaceInteraction = surfaceInteractionCreate(surface, rayInteraction, ray);
+        SurfaceInteraction tempSurfaceInteraction = surfaceInteractionCreate<SurfaceGenerateTangents>(surface, rayInteraction, ray);
         const f16mat3 worldToTangent = f16mat3(tempSurfaceInteraction.interpolatedTangent, tempSurfaceInteraction.interpolatedBitangent, tempSurfaceInteraction.interpolatedNormal);
         const f16vec3 viewDirTangentSpace = normalize(mul(worldToTangent, rayInteraction.viewDirection));
         hitPos = pomCalculateTexcoord(opaqueSurfaceMaterial, tempSurfaceInteraction, viewDirTangentSpace, iterations);
@@ -980,7 +980,7 @@ void geometryResolverVertexOutputDebugView(
       float3 hitPos = float3(0.f);
       if (opaqueSurfaceMaterial.heightTextureIndex != BINDING_INDEX_INVALID && opaqueSurfaceMaterial.displaceIn > 0.0f)
       {
-        SurfaceInteraction tempSurfaceInteraction = surfaceInteractionCreate(surface, rayInteraction, ray);
+        SurfaceInteraction tempSurfaceInteraction = surfaceInteractionCreate<SurfaceGenerateTangents>(surface, rayInteraction, ray);
         const f16mat3 worldToTangent = f16mat3(tempSurfaceInteraction.interpolatedTangent, tempSurfaceInteraction.interpolatedBitangent, tempSurfaceInteraction.interpolatedNormal);
         const f16vec3 viewDirTangentSpace = normalize(mul(worldToTangent, rayInteraction.viewDirection));
         uint iterations = 0;

src/dxvk/shaders/rtx/algorithm/nee_cache_light.slangh

@@ -134,7 +134,7 @@ struct NEECacheUtils
     Ray ray = {};
     ray.origin = shadingPosition;
     ray.spreadAngle = spreadAngle;
-    surfaceInteraction = surfaceInteractionCreate(surface, rayInteracton, ray, false, footprintMode);
+    surfaceInteraction = surfaceInteractionCreate<SurfaceIgnoreTangents>(surface, rayInteracton, ray, false, footprintMode);
     surfaceInteraction.textureGradientX *= cb.neeCacheArgs.emissiveTextureSampleFootprintScale;
     surfaceInteraction.textureGradientY *= cb.neeCacheArgs.emissiveTextureSampleFootprintScale;
 
@@ -305,7 +305,7 @@ struct NEECacheUtils
 
     Ray dummyRay;
     dummyRay.direction = f16vec3(1.0);
-    SurfaceInteraction surfaceInteraction = surfaceInteractionCreate(surface, rayInteracton, dummyRay, false, kFootprintFromTextureCoordDiff);
+    SurfaceInteraction surfaceInteraction = surfaceInteractionCreate<SurfaceIgnoreTangents>(surface, rayInteracton, dummyRay, false, kFootprintFromTextureCoordDiff);
     triangleCenter = surfaceInteraction.position;
     triangleNormal = surfaceInteraction.triangleNormal;
     lightIntensity = vec3(0);

src/dxvk/shaders/rtx/algorithm/resolve.slangh

@@ -364,7 +364,7 @@ void resolveVertex<T: IBasePayloadState>(
 
   surface = surfaces[uint(rayInteraction.surfaceIndex)]; 
 
-  surfaceInteraction = surfaceInteractionCreate(surface, rayInteraction, ray);
+  surfaceInteraction = surfaceInteractionCreate<SurfaceGenerateTangents>(surface, rayInteraction, ray);
 
   // Process user clip planes and view distance
 
@@ -967,7 +967,7 @@ void resolveVertexUnordered<let RAY_FLAGS : uint8_t>(
     const Surface surface = surfaces[uint(rayInteraction.surfaceIndex)]; 
 
     if (isTriangle)
-      surfaceInteraction = surfaceInteractionCreate(surface, rayInteraction, ray);
+      surfaceInteraction = surfaceInteractionCreate<SurfaceGenerateTangents>(surface, rayInteraction, ray);
 
     // Test the hit surface against view distance and ignored surface clipping
 

src/dxvk/shaders/rtx/algorithm/visibility.slangh

@@ -38,7 +38,7 @@ static const uint visibilityModeOpaqueOnly = (1 << 5);                 // Only q
 // Visibility Helper Functions
 
 // Returns the attenuation factor for the surface, i.e. (1.0 - opacity)
-f16vec3 handleVisibilityVertex<let VisibilityMode : uint>(Ray ray, RayHitInfo rayHitInfo, bool usePreviousTLAS)
+f16vec3 handleVisibilityVertex<let VisibilityMode : uint>(Ray ray, RayHitInfo rayHitInfo, const bool usePreviousTLAS)
 {
   RayInteraction rayInteraction = rayInteractionCreate(ray, rayHitInfo);
 
@@ -57,13 +57,32 @@ f16vec3 handleVisibilityVertex<let VisibilityMode : uint>(Ray ray, RayHitInfo ra
       return 1.0;
   }
 
-  const Surface surface = surfaces[uint(rayInteraction.surfaceIndex)]; 
+  Surface surface = surfaces[uint(rayInteraction.surfaceIndex)]; 
 
-  if (surfaceIsDecal(surface))
-    // Decals do not cast shadows
-    return 1.0;
+  // Since during visibility testing, we dont need all the data, we can just explicitly
+  //   initialize these values and  help the compiler generate more optimal code paths.
+  const bool requiresNormalsAndTangents = VisibilityMode & (visibilityModeEnableSubsurfaceMaterials|visibilityModeEnableTranslucentMaterials);
+  {
+    surface.displaceIn = 0.f;
+    surface.decalSortOrder = 0;
+    surface.objectPickingValue = 0;
+    surface.hashPacked = 0;
+    surface.texcoordGenerationMode = TexGenMode::None;
+
+    if (!usePreviousTLAS)
+    {
+      surface.previousPositionBufferIndex = BINDING_INDEX_INVALID;
+      surface.prevObjectToWorld = mat4x3(1.f, 0.f, 0.f, 0.f,  0.f, 1.f, 0.f, 0.f,  0.f, 0.f, 1.f, 0.f);
+    }
+
+    if (!requiresNormalsAndTangents)
+    {
+      surface.normalBufferIndex = BINDING_INDEX_INVALID;
+      surface.normalObjectToWorld = getIdentityMatrix();
+    }
+  }
 
-  SurfaceInteraction surfaceInteraction = surfaceInteractionCreate(surface, rayInteraction, ray, usePreviousTLAS);
+  SurfaceInteraction surfaceInteraction = surfaceInteractionCreate<requiresNormalsAndTangents ? SurfaceGenerateTangents : SurfaceIgnoreTangents>(surface, rayInteraction, ray, usePreviousTLAS);
 
   if (isSurfaceClipped(surface, surfaceInteraction))
     // Ignore parts of surfaces that are under clip planes

src/dxvk/shaders/rtx/concept/surface/surface_interaction.slangh

@@ -35,6 +35,9 @@ static const uint kFootprintFromRayDirection = 1u;
 static const uint kFootprintFromTextureCoordDiff = 2u;
 static const uint kFootprintFromRayOriginClamped = 3u;
 
+static const bool SurfaceGenerateTangents = true;
+static const bool SurfaceIgnoreTangents = false;
+
 // Surface Interaction Helper Functions
 
 // Texture Gradient from Ray Cone Helper
@@ -170,7 +173,7 @@ void calcSpriteSheetAdjustment(
   textureGradientY *= uvSize;
 }
 
-SurfaceInteraction surfaceInteractionCreate(
+SurfaceInteraction surfaceInteractionCreate<let GenerateTangents : bool>(
   Surface surface, RayInteraction rayInteraction, Ray ray,
   bool usePreviousPositions = false, uint footprintMode = kFootprintFromRayDirection)
 {
@@ -440,18 +443,21 @@ SurfaceInteraction surfaceInteractionCreate(
 
     surfaceInteraction.textureCoordinates = interpolateHitAttribute(texcoords, bary);
 
-    // Note: World positions used here, the original PBRT usage uses model space vertex positions since it calculates the
-    // TBN  in local space, but since we generate the TBN frame in world space this is more applicable. Additionally
-    // it removes the need to keep the object space positions around which should ideally reduce register pressure.
-    genTangSpace(
+    if (GenerateTangents)
+    {
+      // Note: World positions used here, the original PBRT usage uses model space vertex positions since it calculates the
+      // TBN  in local space, but since we generate the TBN frame in world space this is more applicable. Additionally
+      // it removes the need to keep the object space positions around which should ideally reduce register pressure.
+      genTangSpace(
       texcoords, worldPositions, surfaceInteraction.interpolatedNormal,
       surfaceInteraction.interpolatedTangent, surfaceInteraction.interpolatedBitangent, surfaceInteraction.rawTangent, surfaceInteraction.rawBitangent);
 
-    // Derive triangle TBN based on interpolated TBN. This works in all cases tested,
-    // and should be much cheaper than calling genTangSpace again.
-    f16vec4 interpolatedToTriangle = quaternionCreateOrientation(surfaceInteraction.interpolatedNormal, surfaceInteraction.triangleNormal);
-    surfaceInteraction.triangleTangent = quaternionTransformVector(interpolatedToTriangle, surfaceInteraction.interpolatedTangent);
-    surfaceInteraction.triangleBitangent =  quaternionTransformVector(interpolatedToTriangle, surfaceInteraction.interpolatedBitangent);
+      // Derive triangle TBN based on interpolated TBN. This works in all cases tested,
+      // and should be much cheaper than calling genTangSpace again.
+      f16vec4 interpolatedToTriangle = quaternionCreateOrientation(surfaceInteraction.interpolatedNormal, surfaceInteraction.triangleNormal);
+      surfaceInteraction.triangleTangent = quaternionTransformVector(interpolatedToTriangle, surfaceInteraction.interpolatedTangent);
+      surfaceInteraction.triangleBitangent = quaternionTransformVector(interpolatedToTriangle, surfaceInteraction.interpolatedBitangent);
+    }
 
     // Convert texcoords from the 1st cascade to a corresponding cascade in a combined tiled cascades texture space
     if (surface.texcoordGenerationMode == uint(TexGenMode::CascadedViewPositions))