Improve light AABB generation #2

com.unity.render-pipelines.core/ShaderLibrary/Macros.hlsl

@@ -46,6 +46,7 @@
 #define HALF_MIN_SQRT 0.0078125  // 2^-7 == sqrt(HALF_MIN), useful for ensuring HALF_MIN after x^2
 #define HALF_MAX 65504.0
 #define UINT_MAX 0xFFFFFFFFu
+#define INT_MAX  0x7FFFFFFF
 
 
 #ifdef SHADER_API_GLES

com.unity.render-pipelines.high-definition/CHANGELOG.md

@@ -133,6 +133,7 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 - On platforms that allow it skip the first mip of the depth pyramid and compute it alongside the depth buffer used for low res transparents.
 - When trying to install the local configuration package, if another one is already present the user is now asked whether they want to keep it or not.
 - Improved MSAA color resolve to fix issues when very bright and very dark samples are resolved together.
+- Improve performance of GPU light AABB generation
 
 ## [10.0.0] - 2019-06-10
 

com.unity.render-pipelines.high-definition/Runtime/Lighting/LightLoop/LightCullUtils.hlsl

@@ -3,31 +3,33 @@
 
 // Used to index into our SFiniteLightBound (g_data) and
 // LightVolumeData (_LightVolumeData) buffers.
-int GenerateLightCullDataIndex(int lightIndex, uint numVisibleLights, uint eyeIndex)
+uint GenerateLightCullDataIndex(uint lightIndex, uint numVisibleLights, uint eyeIndex)
 {
+    lightIndex = min(lightIndex, numVisibleLights - 1); // Stay within bounds
+
     // For monoscopic, there is just one set of light cull data structs.
     // In stereo, all of the left eye structs are first, followed by the right eye structs.
-    const int perEyeBaseIndex = (int)eyeIndex * (int)numVisibleLights;
+    const uint perEyeBaseIndex = eyeIndex * numVisibleLights;
     return (perEyeBaseIndex + lightIndex);
 }
 
 struct ScreenSpaceBoundsIndices
 {
-    int min;
-    int max;
+    uint min;
+    uint max;
 };
 
 // The returned values are used to index into our AABB screen space bounding box buffer
 // Usually named g_vBoundsBuffer.  The two values represent the min/max indices.
-ScreenSpaceBoundsIndices GenerateScreenSpaceBoundsIndices(int lightIndex, uint numVisibleLights, uint eyeIndex)
+ScreenSpaceBoundsIndices GenerateScreenSpaceBoundsIndices(uint lightIndex, uint numVisibleLights, uint eyeIndex)
 {
     // In the monoscopic mode, there is one set of bounds (min,max -> 2 * g_iNrVisibLights)
     // In stereo, there are two sets of bounds (leftMin, leftMax, rightMin, rightMax -> 4 * g_iNrVisibLights)
-    const int eyeRelativeBase = (int)eyeIndex * 2 * (int)numVisibleLights;
+    const uint eyeRelativeBase = eyeIndex * 2 * numVisibleLights;
 
     ScreenSpaceBoundsIndices indices;
     indices.min = eyeRelativeBase + lightIndex;
-    indices.max = eyeRelativeBase + lightIndex + (int)numVisibleLights;
+    indices.max = indices.min + numVisibleLights;
 
     return indices;
 }

com.unity.render-pipelines.high-definition/Runtime/Lighting/LightLoop/LightLoop.cs

@@ -130,8 +130,8 @@ struct SFiniteLightBound
         public Vector3 boxAxisY; // Scaled by the extents (half-size)
         public Vector3 boxAxisZ; // Scaled by the extents (half-size)
         public Vector3 center;   // Center of the bounds (box) in camera space
-        public Vector2 scaleXY;  // Scale applied to the top of the box to turn it into a truncated pyramid
-        public float radius;     // Circumscribed sphere for the bounds (box)
+        public float   scaleXY;  // Scale applied to the top of the box to turn it into a truncated pyramid (X = Y)
+        public float   radius;     // Circumscribed sphere for the bounds (box)
     };
 
     [GenerateHLSL]
@@ -696,7 +696,6 @@ enum ClusterDepthSource : int
             { "TileLightListGen_NoDepthRT_SrcBigTile", "TileLightListGen_DepthRT_SrcBigTile_Oblique", "TileLightListGen_DepthRT_MSAA_SrcBigTile_Oblique" }
         };
 
-        static int s_GenAABBKernel;
         static int s_GenListPerTileKernel;
         static int[,] s_ClusterKernels = new int[(int)ClusterPrepassSource.Count, (int)ClusterDepthSource.Count];
         static int[,] s_ClusterObliqueKernels = new int[(int)ClusterPrepassSource.Count, (int)ClusterDepthSource.Count];
@@ -879,8 +878,6 @@ void InitializeLightLoop(IBLFilterBSDF[] iBLFilterBSDFArray)
             m_MaxLightsOnScreen = m_MaxDirectionalLightsOnScreen + m_MaxPunctualLightsOnScreen + m_MaxAreaLightsOnScreen + m_MaxEnvLightsOnScreen;
             m_MaxPlanarReflectionOnScreen = lightLoopSettings.maxPlanarReflectionOnScreen;
 
-            s_GenAABBKernel = buildScreenAABBShader.FindKernel("ScreenBoundsAABB");
-
             // Cluster
             {
                 s_ClearVoxelAtomicKernel = clearClusterAtomicIndexShader.FindKernel("ClearAtomic");
@@ -1766,9 +1763,9 @@ void GetLightVolumeDataAndBound(LightCategory lightCategory, GPULightType gpuLig
                 fAltDx *= range; fAltDy *= range;
 
                 // Handle case of pyramid with this select (currently unused)
-                var altDist = Mathf.Sqrt(fAltDy * fAltDy + (true ? 1.0f : 2.0f) * fAltDx * fAltDx);
-                bound.radius = altDist > (0.5f * range) ? altDist : (0.5f * range);       // will always pick fAltDist
-                bound.scaleXY = squeeze ? new Vector2(0.01f, 0.01f) : new Vector2(1.0f, 1.0f);
+                var altDist   = Mathf.Sqrt(fAltDy * fAltDy + (true ? 1.0f : 2.0f) * fAltDx * fAltDx);
+                bound.radius  = altDist > (0.5f * range) ? altDist : (0.5f * range);       // will always pick fAltDist
+                bound.scaleXY = squeeze ? 0.01f : 1.0f;
 
                 lightVolumeData.lightAxisX = vx;
                 lightVolumeData.lightAxisY = vy;
@@ -1780,16 +1777,19 @@ void GetLightVolumeDataAndBound(LightCategory lightCategory, GPULightType gpuLig
             }
             else if (gpuLightType == GPULightType.Point)
             {
-                Vector3 vx = xAxisVS;
-                Vector3 vy = yAxisVS;
-                Vector3 vz = zAxisVS;
+                // Construct a view-space axis-aligned bounding cube around the bounding sphere.
+                // This allows us to utilize the same polygon clipping technique for all lights.
+                // Non-axis-aligned vectors may result in a larger screen-space AABB.
+                Vector3 vx = new Vector3(1, 0, 0);
+                Vector3 vy = new Vector3(0, 1, 0);
+                Vector3 vz = new Vector3(0, 0, 1);
 
                 bound.center = positionVS;
                 bound.boxAxisX = vx * range;
                 bound.boxAxisY = vy * range;
                 bound.boxAxisZ = vz * range;
-                bound.scaleXY.Set(1.0f, 1.0f);
-                bound.radius = range;
+                bound.scaleXY  = 1.0f;
+                bound.radius   = range;
 
                 // fill up ldata
                 lightVolumeData.lightAxisX = vx;
@@ -1810,7 +1810,7 @@ void GetLightVolumeDataAndBound(LightCategory lightCategory, GPULightType gpuLig
                 bound.boxAxisY = extents.y * yAxisVS;
                 bound.boxAxisZ = extents.z * zAxisVS;
                 bound.radius   = extents.magnitude;
-                bound.scaleXY.Set(1.0f, 1.0f);
+                bound.scaleXY  = 1.0f;
 
                 lightVolumeData.lightPos   = centerVS;
                 lightVolumeData.lightAxisX = xAxisVS;
@@ -1830,7 +1830,7 @@ void GetLightVolumeDataAndBound(LightCategory lightCategory, GPULightType gpuLig
                 bound.boxAxisY = extents.y * yAxisVS;
                 bound.boxAxisZ = extents.z * zAxisVS;
                 bound.radius   = extents.magnitude;
-                bound.scaleXY.Set(1.0f, 1.0f);
+                bound.scaleXY  = 1.0f;
 
                 lightVolumeData.lightPos   = centerVS;
                 lightVolumeData.lightAxisX = xAxisVS;
@@ -1850,7 +1850,7 @@ void GetLightVolumeDataAndBound(LightCategory lightCategory, GPULightType gpuLig
                 bound.boxAxisY = extents.y * yAxisVS;
                 bound.boxAxisZ = extents.z * zAxisVS;
                 bound.radius   = extents.magnitude;
-                bound.scaleXY.Set(1.0f, 1.0f);
+                bound.scaleXY  = 1.0f;
 
                 lightVolumeData.lightPos   = centerVS;
                 lightVolumeData.lightAxisX = xAxisVS;
@@ -2065,8 +2065,8 @@ void GetEnvLightVolumeDataAndBound(HDProbe probe, LightVolumeType lightVolumeTyp
                     bound.boxAxisX = influenceRightVS * influenceExtents.x;
                     bound.boxAxisY = influenceUpVS * influenceExtents.x;
                     bound.boxAxisZ = influenceForwardVS * influenceExtents.x;
-                    bound.scaleXY.Set(1.0f, 1.0f);
-                    bound.radius = influenceExtents.x;
+                    bound.scaleXY  = 1.0f;
+                    bound.radius   = influenceExtents.x;
                     break;
                 }
                 case LightVolumeType.Box:
@@ -2075,8 +2075,8 @@ void GetEnvLightVolumeDataAndBound(HDProbe probe, LightVolumeType lightVolumeTyp
                     bound.boxAxisX = influenceExtents.x * influenceRightVS;
                     bound.boxAxisY = influenceExtents.y * influenceUpVS;
                     bound.boxAxisZ = influenceExtents.z * influenceForwardVS;
-                    bound.scaleXY.Set(1.0f, 1.0f);
-                    bound.radius = influenceExtents.magnitude;
+                    bound.scaleXY  = 1.0f;
+                    bound.radius   = influenceExtents.magnitude;
 
                     // The culling system culls pixels that are further
                     //   than a threshold to the box influence extents.
@@ -2124,7 +2124,7 @@ void CreateBoxVolumeDataAndBound(OrientedBBox obb, LightCategory category, Light
             bound.boxAxisY = extentConservativeY * upVS;
             bound.boxAxisZ = extentConservativeZ * forwardVS;
             bound.radius   = extentConservativeMagnitude;
-            bound.scaleXY.Set(1.0f, 1.0f);
+            bound.scaleXY  = 1.0f;
 
             // The culling system culls pixels that are further
             //   than a threshold to the box influence extents.
@@ -3087,13 +3087,21 @@ static void GenerateLightsScreenSpaceAABBs(in BuildGPULightListParameters parame
         {
             if (parameters.totalLightCount != 0)
             {
-                // With XR single-pass, we have one set of light bounds per view to iterate over (bounds are in view space for each view)
-                cmd.SetComputeBufferParam(parameters.screenSpaceAABBShader, parameters.screenSpaceAABBKernel, HDShaderIDs.g_data, resources.convexBoundsBuffer);
-                cmd.SetComputeBufferParam(parameters.screenSpaceAABBShader, parameters.screenSpaceAABBKernel, HDShaderIDs.g_vBoundsBuffer, resources.AABBBoundsBuffer);
+                using (new ProfilingScope(cmd, ProfilingSampler.Get(HDProfileId.GenerateLightAABBs)))
+                {
+                    // With XR single-pass, we have one set of light bounds per view to iterate over (bounds are in view space for each view)
+                    cmd.SetComputeBufferParam(parameters.screenSpaceAABBShader, parameters.screenSpaceAABBKernel, HDShaderIDs.g_data, resources.convexBoundsBuffer);
+                    cmd.SetComputeBufferParam(parameters.screenSpaceAABBShader, parameters.screenSpaceAABBKernel, HDShaderIDs.g_vBoundsBuffer, resources.AABBBoundsBuffer);
+
+                    ConstantBuffer.Push(cmd, parameters.lightListCB, parameters.screenSpaceAABBShader, HDShaderIDs._ShaderVariablesLightList);
 
-                ConstantBuffer.Push(cmd, parameters.lightListCB, parameters.screenSpaceAABBShader, HDShaderIDs._ShaderVariablesLightList);
+                    const int threadsPerLight = 4;  // Shader: THREADS_PER_LIGHT (4)
+                    const int threadsPerGroup = 64; // Shader: THREADS_PER_GROUP (64)
 
-                cmd.DispatchCompute(parameters.screenSpaceAABBShader, parameters.screenSpaceAABBKernel, (parameters.totalLightCount + 7) / 8, parameters.viewCount, 1);
+                    int groupCount = HDUtils.DivRoundUp(parameters.totalLightCount * threadsPerLight, threadsPerGroup);
+
+                    cmd.DispatchCompute(parameters.screenSpaceAABBShader, parameters.screenSpaceAABBKernel, groupCount, parameters.viewCount, 1);
+                }
             }
         }
 
@@ -3408,12 +3416,7 @@ unsafe BuildGPULightListParameters PrepareBuildGPULightListParameters(  HDCamera
 
             // Screen space AABB
             parameters.screenSpaceAABBShader = buildScreenAABBShader;
-            parameters.screenSpaceAABBShader.shaderKeywords = null;
-            if (isProjectionOblique)
-            {
-                parameters.screenSpaceAABBShader.EnableKeyword("USE_OBLIQUE_MODE");
-            }
-            parameters.screenSpaceAABBKernel = s_GenAABBKernel;
+            parameters.screenSpaceAABBKernel = 0;
 
             // Big tile prepass
             parameters.runBigTilePrepass = hdCamera.frameSettings.IsEnabled(FrameSettingsField.BigTilePrepass);

com.unity.render-pipelines.high-definition/Runtime/Lighting/LightLoop/LightLoop.cs.hlsl

@@ -74,7 +74,7 @@ struct SFiniteLightBound
     float3 boxAxisY;
     float3 boxAxisZ;
     float3 center;
-    float2 scaleXY;
+    float scaleXY;
     float radius;
 };