Merge pull request 'Implement frustum culling using job system' (#1) from frustum_culling_job into master

Reviewed-on: #1

Editor/Misaki.AoVolume.Editor.asmdef

@@ -2,7 +2,8 @@
     "name": "Misaki.AoVolume.Editor",
     "rootNamespace": "Misaki.AoVolume.Editor",
     "references": [
-        "GUID:33624ad552e4d7342a53d46280a5a916"
+        "GUID:33624ad552e4d7342a53d46280a5a916",
+        "GUID:d8b63aba1907145bea998dd612889d6b"
     ],
     "includePlatforms": [],
     "excludePlatforms": [],

Runtime/CustomPass/AoVolumePass.cs

@@ -2,7 +2,9 @@ using Misaki.AoVolume;
 using System.Buffers;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
+using Misaki.AoVolume.Shader;
 using Unity.Collections;
+using Unity.Jobs;
 using UnityEngine;
 using UnityEngine.Experimental.Rendering;
 using UnityEngine.Rendering;
@@ -11,6 +13,7 @@ using UnityEngine.Rendering.HighDefinition;
 internal class AoVolumePass : CustomPass
 {
     private NativeArray<OrientedBoundingBox> _volumeBounds;
+    private NativeList<VolumeData> _culledVolumeDatas;
     private ComputeBuffer _volumeBoundsBuffer;
     private ComputeBuffer _visibleIndicesBuffer;
     private ComputeBuffer _visibleVolumeCountBuffer;
@@ -42,6 +45,7 @@ internal class AoVolumePass : CustomPass
     protected override void Setup(ScriptableRenderContext renderContext, CommandBuffer cmd)
     {
         _volumeBounds = new NativeArray<OrientedBoundingBox>(maxVolumeOnScreen, Allocator.Persistent);
+        _culledVolumeDatas = new NativeList<VolumeData>(maxVolumeOnScreen, Allocator.Persistent);
         _volumeBoundsBuffer = new ComputeBuffer(maxVolumeOnScreen, Marshal.SizeOf<OrientedBoundingBox>());
         _visibleIndicesBuffer = new ComputeBuffer(maxVolumeOnScreen, Marshal.SizeOf<uint>(), ComputeBufferType.Raw);
         _visibleVolumeCountBuffer = new ComputeBuffer(1, Marshal.SizeOf<uint>(), ComputeBufferType.Counter);
@@ -58,13 +62,48 @@ internal class AoVolumePass : CustomPass
     {
         return Mathf.FloorToInt(Mathf.Log(Mathf.Min(hDCamera.actualWidth, hDCamera.actualHeight), 2)) - 1;
     }
+    private void FrustumCulling(CustomPassContext ctx, ref int volumeCount)
+    {
+        Frustum cameraFrustum = ctx.hdCamera.frustum;
+
+        GPUFrustum frustum = new();
+        frustum.normal0 = cameraFrustum.planes[0].normal;
+        frustum.dist0 = cameraFrustum.planes[0].distance;
+        frustum.normal1 = cameraFrustum.planes[1].normal;
+        frustum.dist1 = cameraFrustum.planes[1].distance;
+        frustum.normal2 = cameraFrustum.planes[2].normal;
+        frustum.dist2 = cameraFrustum.planes[2].distance;
+        frustum.normal3 = cameraFrustum.planes[3].normal;
+        frustum.dist3 = cameraFrustum.planes[3].distance;
+        frustum.normal4 = cameraFrustum.planes[4].normal;
+        frustum.dist4 = cameraFrustum.planes[4].distance;
+        frustum.normal5 = cameraFrustum.planes[5].normal;
+        frustum.dist5 = cameraFrustum.planes[5].distance;
+        frustum.corner0.xyz = cameraFrustum.corners[0];
+        frustum.corner1.xyz = cameraFrustum.corners[1]; 
+        frustum.corner2.xyz = cameraFrustum.corners[2];
+        frustum.corner3.xyz = cameraFrustum.corners[3];
+        frustum.corner4.xyz = cameraFrustum.corners[4];
+        frustum.corner5.xyz = cameraFrustum.corners[5];
+        frustum.corner6.xyz = cameraFrustum.corners[6];
+        frustum.corner7.xyz = cameraFrustum.corners[7];
+
+        _culledVolumeDatas.Clear();
+        FrustumCullingJob cullingJob = new()
+        {
+            GPUFrustum = frustum,
+            InputVolumeDatas = VolumeDatabase.Instance.VolumeDatas,
+            InputVolumeBounds = _volumeBounds,
+            OutputVolumeDataWriter = _culledVolumeDatas.AsParallelWriter()
+        };
+
+        JobHandle cullingJobHandle = cullingJob.Schedule(volumeCount, 64);
+        cullingJobHandle.Complete();
+        volumeCount = _culledVolumeDatas.Length;
+    }
 
     private void HierarchicalZCulling(CustomPassContext ctx, int volumeCount)
     {
-        for (var i = 0; i < volumeCount; i++)
-        {
-            _volumeBounds[i] = new OrientedBoundingBox(VolumeDatabase.Instance.VolumeDatas[i].worldMatrix);
-        }
         _volumeBoundsBuffer.SetData(_volumeBounds);
 
         ctx.cmd.SetComputeBufferParam(cullingShader, 0, "_VolumeBounds", _volumeBoundsBuffer);
@@ -81,7 +120,7 @@ internal class AoVolumePass : CustomPass
 
     private void RenderVisibleVolumes(CustomPassContext ctx)
     {
-        _volumeDataBuffer.SetData(VolumeDatabase.Instance.VolumeDatas);
+        _volumeDataBuffer.SetData(_culledVolumeDatas.AsArray());
 
         ctx.cmd.SetComputeBufferParam(renderingShader, 0, "_VolumeDatas", _volumeDataBuffer);
         ctx.cmd.SetComputeBufferParam(renderingShader, 0, "_VisibleVolumeCount", _visibleVolumeCountBuffer);
@@ -140,6 +179,7 @@ internal class AoVolumePass : CustomPass
             return;
         }
 
+        FrustumCulling(ctx, ref volumeCount);
         HierarchicalZCulling(ctx, volumeCount);
         RenderVisibleVolumes(ctx);
         //DebugVisibleVolumes();
@@ -149,6 +189,7 @@ internal class AoVolumePass : CustomPass
     protected override void Cleanup()
     {
         _volumeBounds.Dispose();
+        _culledVolumeDatas.Dispose();
         _volumeBoundsBuffer?.Dispose();
         _visibleIndicesBuffer?.Dispose();
         _visibleVolumeCountBuffer?.Dispose();

Runtime/Misaki.AoVolume.asmdef

@@ -5,7 +5,8 @@
         "GUID:457756d89b35d2941b3e7b37b4ece6f1",
         "GUID:df380645f10b7bc4b97d4f5eb6303d95",
         "GUID:e0cd26848372d4e5c891c569017e11f1",
-        "GUID:d8b63aba1907145bea998dd612889d6b"
+        "GUID:d8b63aba1907145bea998dd612889d6b",
+        "GUID:2665a8d13d1b3f18800f46e256720795"
     ],
     "includePlatforms": [],
     "excludePlatforms": [],

Runtime/Shader/FrustumCullingJob.cs

@@ -0,0 +1,119 @@
+using System.Runtime.CompilerServices;
+using Unity.Burst;
+using Unity.Collections;
+using Unity.Jobs;
+using Unity.Mathematics;
+using UnityEngine;
+
+namespace Misaki.AoVolume.Shader
+{
+    [BurstCompile]
+    internal struct FrustumCullingJob : IJobParallelFor
+    {
+        [ReadOnly] internal GPUFrustum GPUFrustum;
+        [ReadOnly] internal NativeArray<VolumeData> InputVolumeDatas;
+        internal NativeArray<OrientedBoundingBox> InputVolumeBounds;
+        internal NativeList<VolumeData>.ParallelWriter OutputVolumeDataWriter;
+
+        public void Execute(int index)
+        {
+            OrientedBoundingBox obb = new(InputVolumeDatas[index].worldMatrix);
+            if (FrustumObbIntersection(obb))
+            {
+                OutputVolumeDataWriter.AddNoResize(InputVolumeDatas[index]);
+                InputVolumeBounds[index] = obb;
+            }
+        }
+
+        private struct FrustumPlane
+        {
+            internal float3 Normal;
+            internal float Distance;
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static bool CheckOverlap(OrientedBoundingBox obb, float3 planeNormal, float planeDist)
+        {
+            float maxHalfDiagProj = obb.extent.x * math.abs(math.dot(obb.right, planeNormal)) +
+                                    obb.extent.y * math.abs(math.dot(obb.up, planeNormal)) +
+                                    obb.extent.z * math.abs(math.dot(GetForward(obb), planeNormal));
+            float centerProj = math.dot(obb.center, planeNormal) + planeDist;
+            return maxHalfDiagProj + centerProj >= 0;
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static Vector3 GetForward(OrientedBoundingBox obb)
+        {
+            return Vector3.Cross(obb.up, obb.right);
+        }
+
+        private bool FrustumObbIntersection(OrientedBoundingBox obb)
+        {
+            bool overlap = CheckOverlap(obb, GPUFrustum.normal0, GPUFrustum.dist0);
+            overlap &= CheckOverlap(obb, GPUFrustum.normal1, GPUFrustum.dist1);
+            overlap &= CheckOverlap(obb, GPUFrustum.normal2, GPUFrustum.dist2);
+            overlap &= CheckOverlap(obb, GPUFrustum.normal3, GPUFrustum.dist3);
+            overlap &= CheckOverlap(obb, GPUFrustum.normal4, GPUFrustum.dist4);
+            overlap &= CheckOverlap(obb, GPUFrustum.normal5, GPUFrustum.dist5);
+
+            var planes = new NativeArray<FrustumPlane>(3, Allocator.Temp);
+            planes[0] = new FrustumPlane
+            {
+                Normal = obb.right,
+                Distance = obb.extent.x
+            };
+            planes[1] = new FrustumPlane
+            {
+                Normal = obb.up,
+                Distance = obb.extent.y
+            };
+            planes[2] = new FrustumPlane
+            {
+                Normal = GetForward(obb),
+                Distance = obb.extent.z
+            };
+            for (int i = 0; i < 3; i++)
+            {
+                bool outsidePos = true;
+                bool outsideNeg = true;
+
+                float proj = math.dot(planes[i].Normal, GPUFrustum.corner0.xyz - obb.center);
+                outsidePos &= proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                proj = math.dot(planes[i].Normal, GPUFrustum.corner1.xyz - obb.center);
+                outsidePos &= proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                proj = math.dot(planes[i].Normal, GPUFrustum.corner2.xyz - obb.center);
+                outsidePos &= proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                proj = math.dot(planes[i].Normal, GPUFrustum.corner3.xyz - obb.center);
+                outsidePos &= proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                proj = math.dot(planes[i].Normal, GPUFrustum.corner4.xyz - obb.center);
+                outsideNeg &= -proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                proj = math.dot(planes[i].Normal, GPUFrustum.corner5.xyz - obb.center);
+                outsideNeg &= -proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                proj = math.dot(planes[i].Normal, GPUFrustum.corner6.xyz - obb.center);
+                outsideNeg &= -proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                proj = math.dot(planes[i].Normal, GPUFrustum.corner7.xyz - obb.center);
+                outsideNeg &= -proj > planes[i].Distance;
+                outsideNeg &= -proj > planes[i].Distance;
+
+                overlap &= !(outsidePos || outsideNeg);
+            }
+
+            planes.Dispose();
+            return overlap;
+        }
+    }
+}

Runtime/Shader/FrustumCullingJob.cs.meta

@@ -0,0 +1,3 @@
+fileFormatVersion: 2
+guid: 5403ee0538c94d9abc5aa44a761092d0
+timeCreated: 1740495091

Runtime/Shader/Includes/GeometryData.cs

@@ -1,4 +1,4 @@
-using UnityEngine;
+using Unity.Mathematics;
 using UnityEngine.Rendering;
 
 namespace Misaki.AoVolume
@@ -7,28 +7,28 @@ namespace Misaki.AoVolume
     internal struct GPUFrustum
     {
         // The data of the 6 planes of the frustum
-        public Vector3 normal0;
+        public float3 normal0;
         public float dist0;
-        public Vector3 normal1;
+        public float3 normal1;
         public float dist1;
-        public Vector3 normal2;
+        public float3 normal2;
         public float dist2;
-        public Vector3 normal3;
+        public float3 normal3;
         public float dist3;
-        public Vector3 normal4;
+        public float3 normal4;
         public float dist4;
-        public Vector3 normal5;
+        public float3 normal5;
         public float dist5;
 
         // The data of the 8 corners of the frustum
-        public Vector4 corner0;
-        public Vector4 corner1;
-        public Vector4 corner2;
-        public Vector4 corner3;
-        public Vector4 corner4;
-        public Vector4 corner5;
-        public Vector4 corner6;
-        public Vector4 corner7;
+        public float4 corner0;
+        public float4 corner1;
+        public float4 corner2;
+        public float4 corner3;
+        public float4 corner4;
+        public float4 corner5;
+        public float4 corner6;
+        public float4 corner7;
     }
 
     [GenerateHLSL(PackingRules.Exact, false)]
@@ -36,32 +36,32 @@ namespace Misaki.AoVolume
     {
         // 4 x float3 = 48 bytes.
         // TODO: pack the axes into 16-bit UNORM per channel, and consider a quaternionic representation.
-        public Vector3 center;
-        public Vector3 right;
-        public Vector3 up;
+        public float3 center;
+        public float3 right;
+        public float3 up;
 
-        public Vector3 extent;
+        public float3 extent;
 
         //public ushort quatX;
         //public ushort quatY;
         //public ushort quatZ;
         //public ushort quatW;
 
-        public readonly Vector3 Forward => Vector3.Cross(up, right);
+        public readonly float3 Forward => math.cross(up, right);
 
-        public OrientedBoundingBox(Matrix4x4 trs)
+        public OrientedBoundingBox(float4x4 trs)
         {
-            var vecX = (Vector3)trs.GetColumn(0);
-            var vecY = (Vector3)trs.GetColumn(1);
-            var vecZ = (Vector3)trs.GetColumn(2);
+            float3 vecX = trs.c0.xyz;
+            float3 vecY = trs.c1.xyz;
+            float3 vecZ = trs.c2.xyz;
 
-            center = trs.GetColumn(3);
-            right = vecX * (1.0f / vecX.magnitude);
-            up = vecY * (1.0f / vecY.magnitude);
+            center = trs.c3.xyz;
+            right = vecX * (1.0f / math.length(vecX));
+            up = vecY * (1.0f / math.length(vecY));
 
-            extent.x = 0.5f * vecX.magnitude;
-            extent.y = 0.5f * vecY.magnitude;
-            extent.z = 0.5f * vecZ.magnitude;
+            extent.x = 0.5f * math.length(vecX);
+            extent.y = 0.5f * math.length(vecY);
+            extent.z = 0.5f * math.length(vecZ);
         }
     }
 

Runtime/Shader/Includes/GeometryData.cs.hlsl

@@ -16,17 +16,17 @@ struct CullingData
 // PackingRules = Exact
 struct GPUFrustum
 {
-    float3 normal0;
+    float4 normal0;
     float dist0;
-    float3 normal1;
+    float4 normal1;
     float dist1;
-    float3 normal2;
+    float4 normal2;
     float dist2;
-    float3 normal3;
+    float4 normal3;
     float dist3;
-    float3 normal4;
+    float4 normal4;
     float dist4;
-    float3 normal5;
+    float4 normal5;
     float dist5;
     float4 corner0;
     float4 corner1;
@@ -42,10 +42,10 @@ struct GPUFrustum
 // PackingRules = Exact
 struct OrientedBoundingBox
 {
-    float3 center;
-    float3 right;
-    float3 up;
-    float3 extent;
+    float4 center;
+    float4 right;
+    float4 up;
+    float4 extent;
 };
 
 

Runtime/Shader/Includes/VolumeData.cs

@@ -1,4 +1,5 @@
 using System;
+using Unity.Mathematics;
 using UnityEngine;
 using UnityEngine.Rendering;
 
@@ -10,12 +11,12 @@ namespace Misaki.AoVolume
     {
         [HideInInspector]
         [NonSerialized]
-        public Matrix4x4 worldMatrix;
+        public float4x4 worldMatrix;
         [HideInInspector]
         [NonSerialized]
-        public Matrix4x4 inverseWorldMatrix;
+        public float4x4 inverseWorldMatrix;
 
-        public Vector3 size;
+        public float3 size;
 
         [Range(0.0f, 1.0f)]
         public float intensity;
@@ -26,7 +27,7 @@ namespace Misaki.AoVolume
 
         public static VolumeData Default => new()
         {
-            size = Vector3.one,
+            size = new float3(1f),
             intensity = 1.0f,
             falloff = 0.25f,
             normalFalloff = 0.0f