Merge branch 'dev' into dev-tracy

.github/workflows/build.yaml

@@ -16,7 +16,7 @@ jobs:
     steps:
         - uses: actions/checkout@v3
           with:
-            submodules: 'true'
+            submodules: recursive
 
         - uses: ilammy/msvc-dev-cmd@v1.10.0
 
@@ -28,7 +28,7 @@ jobs:
           run: cmake -S . --preset=ALL --check-stamp-file "build\CMakeFiles\generate.stamp"
 
         - name: cmake build
-          run: cmake --build build --config Release
+          run: cmake --build build/ALL --config Release
 
         - name: create a tagged release and upload the archive
           uses: ncipollo/release-action@v1

BuildRelease.bat

@@ -7,9 +7,9 @@ if NOT "%1" == "" (
 
 echo Running preset %preset%
 
-cmake -S . --preset=%preset% --check-stamp-file "build\CMakeFiles\generate.stamp"
+cmake -S . --preset=%preset% --check-stamp-file "build\%preset%\CMakeFiles\generate.stamp"
 if %ERRORLEVEL% NEQ 0 exit 1
-cmake --build build --config Release
+cmake --build build/%preset% --config Release
 if %ERRORLEVEL% NEQ 0 exit 1
 
 pause

CMakePresets.json

@@ -12,20 +12,14 @@
 			"cacheVariables": {
 				"SKSE_SUPPORT_XBYAK": "ON"
 			},
-			"binaryDir": "${sourceDir}/build"
+			"binaryDir": "${sourceDir}/build/${presetName}"
 		},
 		{
 			"name": "vcpkg",
 			"hidden": true,
 			"cacheVariables": {
-				"CMAKE_TOOLCHAIN_FILE": {
-					"type": "STRING",
-					"value": "$env{VCPKG_ROOT}/scripts/buildsystems/vcpkg.cmake"
-				},
-				"VCPKG_OVERLAY_PORTS": {
-					"type": "STRING",
-					"value": "${sourceDir}/cmake/ports/"
-				  },
+				"CMAKE_TOOLCHAIN_FILE": "$env{VCPKG_ROOT}/scripts/buildsystems/vcpkg.cmake",
+				"VCPKG_OVERLAY_PORTS": "${sourceDir}/cmake/ports/",
 				"VCPKG_TARGET_TRIPLET": "x64-windows-static-md"
 			}
 		},
@@ -53,7 +47,8 @@
 			}
 		},
 		{
-			"name": "AE",
+			"name": "skyrim",
+			"hidden": true,
 			"cacheVariables": {
 				"BUILD_SKYRIM": true,
 				"ENABLE_SKYRIM_AE": "ON",
@@ -68,85 +63,64 @@
 				"msvc"
 			]
 		},
+		{
+			"name": "AE",
+			"cacheVariables": {
+				"ENABLE_SKYRIM_AE": "ON",
+				"ENABLE_SKYRIM_SE": "OFF",
+				"ENABLE_SKYRIM_VR": "OFF"
+			},
+			"inherits": "skyrim"
+		},
 		{
 			"name": "SE",
 			"cacheVariables": {
-				"BUILD_SKYRIM": true,
 				"ENABLE_SKYRIM_AE": "OFF",
 				"ENABLE_SKYRIM_SE": "ON",
 				"ENABLE_SKYRIM_VR": "OFF",
 				"BUILD_TESTS": "OFF"
 			},
-			"inherits": [
-				"common",
-				"vcpkg",
-				"win64",
-				"msvc"
-			]
+			"inherits": "skyrim"
 		},
 		{
 			"name": "VR",
 			"cacheVariables": {
-				"BUILD_SKYRIM": true,
 				"ENABLE_SKYRIM_AE": "OFF",
 				"ENABLE_SKYRIM_SE": "OFF",
 				"ENABLE_SKYRIM_VR": "ON",
 				"BUILD_TESTS": "OFF"
 			},
-			"inherits": [
-				"common",
-				"vcpkg",
-				"win64",
-				"msvc"
-			]
+			"inherits": "skyrim"
 		},
 		{
 			"name": "ALL",
 			"cacheVariables": {
-				"BUILD_SKYRIM": true,
 				"ENABLE_SKYRIM_AE": "ON",
 				"ENABLE_SKYRIM_SE": "ON",
 				"ENABLE_SKYRIM_VR": "ON",
 				"BUILD_TESTS": "OFF"
 			},
-			"inherits": [
-				"common",
-				"vcpkg",
-				"win64",
-				"msvc"
-			]
+			"inherits": "skyrim"
 		},
 		{
 			"name": "PRE-AE",
 			"cacheVariables": {
-				"BUILD_SKYRIM": true,
 				"ENABLE_SKYRIM_AE": "OFF",
 				"ENABLE_SKYRIM_SE": "ON",
 				"ENABLE_SKYRIM_VR": "ON",
 				"BUILD_TESTS": "OFF"
 			},
-			"inherits": [
-				"common",
-				"vcpkg",
-				"win64",
-				"msvc"
-			]
+			"inherits": "skyrim"
 		},
 		{
 			"name": "FLATRIM",
 			"cacheVariables": {
-				"BUILD_SKYRIM": true,
 				"ENABLE_SKYRIM_AE": "ON",
 				"ENABLE_SKYRIM_SE": "ON",
 				"ENABLE_SKYRIM_VR": "OFF",
 				"BUILD_TESTS": "OFF"
 			},
-			"inherits": [
-				"common",
-				"vcpkg",
-				"win64",
-				"msvc"
-			]
+			"inherits": "skyrim"
 		}
 	]
 }

features/Light Limit Fix/Shaders/LightLimitFix/ClusterCullingCS.hlsl

@@ -11,9 +11,9 @@ cbuffer PerFrame : register(b0)
 StructuredBuffer<ClusterAABB> clusters : register(t0);
 StructuredBuffer<StructuredLight> lights : register(t1);
 
-RWStructuredBuffer<uint> lightIndexCounter : register(u0);  //1
-RWStructuredBuffer<uint> lightIndexList : register(u1);     //MAX_CLUSTER_LIGHTS * 16^3
-RWStructuredBuffer<LightGrid> lightGrid : register(u2);     //16^3
+RWStructuredBuffer<uint> lightIndexCounter : register(u0);
+RWStructuredBuffer<uint> lightIndexList : register(u1);
+RWStructuredBuffer<LightGrid> lightGrid : register(u2);
 
 groupshared StructuredLight sharedLights[GROUP_SIZE];
 
@@ -25,22 +25,25 @@ bool LightIntersectsCluster(StructuredLight light, ClusterAABB cluster, int eyeI
 	return dot(dist, dist) <= (light.radius * light.radius);
 }
 
-[numthreads(16, 8, 8)] void main(uint3 groupId
-								 : SV_GroupID,
-								 uint3 dispatchThreadId
-								 : SV_DispatchThreadID,
-								 uint3 groupThreadId
-								 : SV_GroupThreadID,
-								 uint groupIndex
-								 : SV_GroupIndex) {
+[numthreads(NUMTHREAD_X, NUMTHREAD_Y, NUMTHREAD_Z)] void main(
+	uint3 groupId
+	: SV_GroupID, uint3 dispatchThreadId
+	: SV_DispatchThreadID, uint3 groupThreadId
+	: SV_GroupThreadID, uint groupIndex
+	: SV_GroupIndex) {
+	if (any(dispatchThreadId >= uint3(CLUSTER_BUILDING_DISPATCH_SIZE_X, CLUSTER_BUILDING_DISPATCH_SIZE_Y, CLUSTER_BUILDING_DISPATCH_SIZE_Z)))
+		return;
+
 	if (all(dispatchThreadId == 0)) {
 		lightIndexCounter[0] = 0;
 	}
 
 	uint visibleLightCount = 0;
 	uint visibleLightIndices[MAX_CLUSTER_LIGHTS];
 
-	uint clusterIndex = groupIndex + GROUP_SIZE * groupId.z;
+	uint clusterIndex = dispatchThreadId.x +
+	                    dispatchThreadId.y * CLUSTER_BUILDING_DISPATCH_SIZE_X +
+	                    dispatchThreadId.z * (CLUSTER_BUILDING_DISPATCH_SIZE_X * CLUSTER_BUILDING_DISPATCH_SIZE_Y);
 
 	ClusterAABB cluster = clusters[clusterIndex];
 
@@ -61,11 +64,13 @@ bool LightIntersectsCluster(StructuredLight light, ClusterAABB cluster, int eyeI
 		for (uint i = 0; i < batchSize; i++) {
 			StructuredLight light = lights[i];
 
-			if (visibleLightCount < MAX_CLUSTER_LIGHTS && (LightIntersectsCluster(light, cluster)
+			bool updateCluster = LightIntersectsCluster(light, cluster);
 #ifdef VR
-															  || LightIntersectsCluster(light, cluster, 1)
+			updateCluster = updateCluster || LightIntersectsCluster(light, cluster, 1);
 #endif  // VR
-																  )) {
+			updateCluster = updateCluster && (visibleLightCount < MAX_CLUSTER_LIGHTS);
+
+			if (updateCluster) {
 				visibleLightIndices[visibleLightCount] = lightOffset + i;
 				visibleLightCount++;
 			}
@@ -83,8 +88,11 @@ bool LightIntersectsCluster(StructuredLight light, ClusterAABB cluster, int eyeI
 		lightIndexList[offset + i] = visibleLightIndices[i];
 	}
 
-	lightGrid[clusterIndex].offset = offset;
-	lightGrid[clusterIndex].lightCount = visibleLightCount;
+	LightGrid output = {
+		offset, visibleLightCount, 0, 0
+	};
+
+	lightGrid[clusterIndex] = output;
 }
 
 //https://www.3dgep.com/forward-plus/#Grid_Frustums_Compute_Shader

features/Light Limit Fix/Shaders/LightLimitFix/Common.hlsli

@@ -1,11 +1,10 @@
 
-#define GROUP_SIZE (16 * 16 * 4)
+#define NUMTHREAD_X 16
+#define NUMTHREAD_Y 16
+#define NUMTHREAD_Z 4
+#define GROUP_SIZE (NUMTHREAD_X * NUMTHREAD_Y * NUMTHREAD_Z)
 #define MAX_CLUSTER_LIGHTS 128
 
-#define CLUSTER_BUILDING_DISPATCH_SIZE_X 16
-#define CLUSTER_BUILDING_DISPATCH_SIZE_Y 16
-#define CLUSTER_BUILDING_DISPATCH_SIZE_Z 16
-
 struct ClusterAABB
 {
 	float4 minPoint;

features/Light Limit Fix/Shaders/LightLimitFix/LightLimitFix.hlsli

@@ -31,15 +31,16 @@ StructuredBuffer<StrictLightData> strictLights : register(t53);
 
 bool GetClusterIndex(in float2 uv, in float z, out uint clusterIndex)
 {
+	const uint3 clusterSize = lightLimitFixSettings.ClusterSize.xyz;
+
 	if (z < strictLights[0].LightsNear || z > strictLights[0].LightsFar)
 		return false;
 
 	float clampedZ = clamp(z, strictLights[0].LightsNear, strictLights[0].LightsFar);
-	uint clusterZ = uint(max((log2(z) - log2(strictLights[0].LightsNear)) * 16.0 / log2(strictLights[0].LightsFar / strictLights[0].LightsNear), 0.0));
-	uint2 clusterDim = ceil(BufferDim / float2(16, 16));
-	uint3 cluster = uint3(uint2((uv * BufferDim) / clusterDim), clusterZ);
+	uint clusterZ = uint(max((log2(z) - log2(strictLights[0].LightsNear)) * clusterSize.z / log2(strictLights[0].LightsFar / strictLights[0].LightsNear), 0.0));
+	uint3 cluster = uint3(uint2(uv * clusterSize.xy), clusterZ);
 
-	clusterIndex = cluster.x + (16 * cluster.y) + (16 * 16 * cluster.z);
+	clusterIndex = cluster.x + (clusterSize.x * cluster.y) + (clusterSize.x * clusterSize.y * cluster.z);
 	return true;
 }
 

features/Skylighting/Shaders/Skylighting/Skylighting.hlsli

@@ -132,29 +132,6 @@ namespace Skylighting
 		return vl * step;
 	}
 
-	// http://torust.me/ZH3.pdf
-	// ZH hallucination that makes skylighting more directional
-	// skipped luminance because it's single channel
-	float hallucinateZH3(sh2 sh, float3 normal)
-	{
-		const static float factor = sqrt(5.0f / (16.0f * 3.1415926f));
-
-		float result = shFuncProductIntegral(sh, shEvaluateCosineLobe(normal)) / shPI;  // cosine lobe integral -> pi
-		if (all(abs(sh.yzw) < 1e-10))
-			return result;
-
-		float3 zonalAxis = normalize(sh.wyz);
-		float ratio = abs(dot(sh.wyz * float3(-1, -1, 0), zonalAxis)) / sh.x;
-		float zonalL2Coeff = sh.x * (0.08f * ratio + 0.6f * ratio * ratio);
-
-		float fZ = dot(zonalAxis, normal);
-		float zhDir = factor * (3.0f * fZ * fZ - 1.0f);
-
-		result += 0.25f * zonalL2Coeff * zhDir;
-
-		return saturate(result);
-	}
-
 	sh2 fauxSpecularLobeSH(float3 N, float3 V, float roughness)
 	{
 		// https://www.gdcvault.com/play/1026701/Fast-Denoising-With-Self-Stabilizing

package/Shaders/AmbientCompositeCS.hlsl

@@ -68,7 +68,7 @@ RWTexture2D<half3> DiffuseAmbientRW : register(u1);
 #	endif
 
 	sh2 skylighting = Skylighting::sample(skylightingSettings, SkylightingProbeArray, positionMS.xyz, normalWS);
-	half skylightingDiffuse = Skylighting::hallucinateZH3(skylighting, skylightingSettings.DirectionalDiffuse ? normalWS : float3(0, 0, 1));
+	half skylightingDiffuse = shFuncProductIntegral(skylighting, shEvaluateCosineLobe(skylightingSettings.DirectionalDiffuse ? normalWS : float3(0, 0, 1)));
 	skylightingDiffuse = Skylighting::mixDiffuse(skylightingSettings, skylightingDiffuse);
 
 	visibility = skylightingDiffuse;

package/Shaders/Common/SharedData.hlsli

@@ -112,6 +112,8 @@ struct LightLimitFixSettings
 	uint EnableLightsVisualisation;
 	uint LightsVisualisationMode;
 	uint pad0;
+
+	uint4 ClusterSize;
 };
 
 #	define SL_INCL_STRUCT

package/Shaders/Lighting.hlsl

@@ -327,7 +327,7 @@ VS_OUTPUT main(VS_INPUT input)
 	vsout.ScreenNormalTransform1.xyz = ScreenNormalTransform[1];
 	vsout.ScreenNormalTransform2.xyz = ScreenNormalTransform[2];
 #	else
-	float3x4 transMat = mul((float4x3)(ScreenProj[eyeIndex]), World[eyeIndex]);
+	float3x4 transMat = mul(ScreenProj[eyeIndex], World[eyeIndex]);
 
 #		if defined(MODELSPACENORMALS)
 	vsout.ScreenNormalTransform0.xyz = transMat[0].xyz;
@@ -2207,7 +2207,7 @@ PS_OUTPUT main(PS_INPUT input, bool frontFace
 	float3 reflectionDiffuseColor = diffuseColor + directionalAmbientColor;
 
 #	if defined(SKYLIGHTING)
-	float skylightingDiffuse = Skylighting::hallucinateZH3(skylightingSH, skylightingSettings.DirectionalDiffuse ? worldSpaceNormal : float3(0, 0, 1));
+	float skylightingDiffuse = shFuncProductIntegral(skylightingSH, shEvaluateCosineLobe(skylightingSettings.DirectionalDiffuse ? worldSpaceNormal : float3(0, 0, 1)));
 	skylightingDiffuse = Skylighting::mixDiffuse(skylightingSettings, skylightingDiffuse);
 	directionalAmbientColor = sRGB2Lin(directionalAmbientColor);
 	directionalAmbientColor *= skylightingDiffuse;

package/Shaders/RunGrass.hlsl

@@ -653,7 +653,7 @@ PS_OUTPUT main(PS_INPUT input, bool frontFace
 #						endif
 
 	sh2 skylightingSH = Skylighting::sample(skylightingSettings, SkylightingProbeArray, positionMSSkylight, normal);
-	float skylighting = Skylighting::hallucinateZH3(skylightingSH, skylightingSettings.DirectionalDiffuse ? normal : float3(0, 0, 1));
+	float skylighting = shFuncProductIntegral(skylightingSH, shEvaluateCosineLobe(skylightingSettings.DirectionalDiffuse ? normal : float3(0, 0, 1)));
 	skylighting = Skylighting::mixDiffuse(skylightingSettings, skylighting);
 
 	directionalAmbientColor = sRGB2Lin(directionalAmbientColor);

src/Buffer.h

@@ -83,12 +83,14 @@ class ConstantBuffer
 };
 
 template <typename T>
-D3D11_BUFFER_DESC StructuredBufferDesc(UINT a_count = 1)
+D3D11_BUFFER_DESC StructuredBufferDesc(UINT a_count = 1, bool cpu_access = true)
 {
 	D3D11_BUFFER_DESC desc{};
 	ZeroMemory(&desc, sizeof(desc));
-	desc.Usage = D3D11_USAGE_DYNAMIC;
+	desc.Usage = cpu_access ? D3D11_USAGE_DYNAMIC : D3D11_USAGE_DEFAULT;
 	desc.BindFlags = D3D11_BIND_SHADER_RESOURCE;
+	if (!cpu_access)
+		desc.BindFlags = desc.BindFlags | D3D11_BIND_UNORDERED_ACCESS;
 	desc.CPUAccessFlags = D3D11_CPU_ACCESS_WRITE;
 	desc.MiscFlags = D3D11_RESOURCE_MISC_BUFFER_STRUCTURED;
 	desc.StructureByteStride = sizeof(T);