// Copyright 1998-2017 Epic Games, Inc. All Rights Reserved.

/*=============================================================================
	DistanceFieldShadowing.usf
=============================================================================*/

#include "Common.usf"
#include "DeferredShadingCommon.usf"
#include "DistanceFieldLightingShared.usf"
#include "DistanceFieldShadowingShared.usf"

uint ObjectBoundingGeometryIndexCount;

float4 FetchObjectDataFloat4(uint SourceIndex)
{
	return float4(ObjectData[4 * SourceIndex + 0], ObjectData[4 * SourceIndex + 1], ObjectData[4 * SourceIndex + 2], ObjectData[4 * SourceIndex + 3]);
}

void CopyCulledObjectData(uint DestIndex, uint SourceIndex)
{
	RWCulledObjectBounds[DestIndex] = float4(ObjectBounds[4 * SourceIndex + 0], ObjectBounds[4 * SourceIndex + 1], ObjectBounds[4 * SourceIndex + 2], ObjectBounds[4 * SourceIndex + 3]);

	UNROLL
	for (uint VectorIndex = 0; VectorIndex < CULLED_OBJECT_DATA_STRIDE; VectorIndex++)
	{
		// Note: only copying the first CULLED_OBJECT_DATA_STRIDE of the original object data
		RWCulledObjectData[DestIndex * CULLED_OBJECT_DATA_STRIDE + VectorIndex] = FetchObjectDataFloat4(SourceIndex * OBJECT_DATA_STRIDE + VectorIndex);
	}

	float3 LocalVolumeBoundsMin;
	float3 LocalVolumeBoundsMax;
	LoadGlobalObjectLocalVolumeBoundsMinMax(SourceIndex, LocalVolumeBoundsMin, LocalVolumeBoundsMax);

	float3 LocalBoundsVertices[8];
	LocalBoundsVertices[0] = float3(LocalVolumeBoundsMin.x, LocalVolumeBoundsMin.y, LocalVolumeBoundsMin.z);
	LocalBoundsVertices[1] = float3(LocalVolumeBoundsMax.x, LocalVolumeBoundsMin.y, LocalVolumeBoundsMin.z);
	LocalBoundsVertices[2] = float3(LocalVolumeBoundsMin.x, LocalVolumeBoundsMax.y, LocalVolumeBoundsMin.z);
	LocalBoundsVertices[3] = float3(LocalVolumeBoundsMax.x, LocalVolumeBoundsMax.y, LocalVolumeBoundsMin.z);
	LocalBoundsVertices[4] = float3(LocalVolumeBoundsMin.x, LocalVolumeBoundsMin.y, LocalVolumeBoundsMax.z);
	LocalBoundsVertices[5] = float3(LocalVolumeBoundsMax.x, LocalVolumeBoundsMin.y, LocalVolumeBoundsMax.z);
	LocalBoundsVertices[6] = float3(LocalVolumeBoundsMin.x, LocalVolumeBoundsMax.y, LocalVolumeBoundsMax.z);
	LocalBoundsVertices[7] = float3(LocalVolumeBoundsMax.x, LocalVolumeBoundsMax.y, LocalVolumeBoundsMax.z);

	float3 MinViewSpacePosition = float3(2000000, 2000000, 2000000);
	float3 MaxViewSpacePosition = float3(-2000000, -2000000, -2000000);

	float4 M0 = FetchObjectDataFloat4(SourceIndex * OBJECT_DATA_STRIDE + 11);
	float4 M1 = FetchObjectDataFloat4(SourceIndex * OBJECT_DATA_STRIDE + 12);
	float4 M2 = FetchObjectDataFloat4(SourceIndex * OBJECT_DATA_STRIDE + 13);
	float4 M3 = FetchObjectDataFloat4(SourceIndex * OBJECT_DATA_STRIDE + 14);

	float4x4 LocalToWorld = float4x4(M0, M1, M2, M3);

	float3 ViewSpaceBoundsVertices[8];

	for (uint i = 0; i < 8; i++)
	{
		float3 WorldBoundsPosition = mul(float4(LocalBoundsVertices[i], 1), LocalToWorld).xyz;
		float3 ViewSpacePosition = mul(float4(WorldBoundsPosition, 1), WorldToShadow).xyz;
		MinViewSpacePosition = min(MinViewSpacePosition, ViewSpacePosition);
		MaxViewSpacePosition = max(MaxViewSpacePosition, ViewSpacePosition);
		ViewSpaceBoundsVertices[i] = ViewSpacePosition;
	}

	float3 ObjectXAxis = (ViewSpaceBoundsVertices[1] - ViewSpaceBoundsVertices[0]) / 2.0f;
	float3 ObjectYAxis = (ViewSpaceBoundsVertices[2] - ViewSpaceBoundsVertices[0]) / 2.0f;
	float3 ObjectZAxis = (ViewSpaceBoundsVertices[4] - ViewSpaceBoundsVertices[0]) / 2.0f;

	RWCulledObjectBoxBounds[DestIndex * CULLED_OBJECT_BOX_BOUNDS_STRIDE + 0] = float4(MinViewSpacePosition, 0);
	RWCulledObjectBoxBounds[DestIndex * CULLED_OBJECT_BOX_BOUNDS_STRIDE + 1] = float4(MaxViewSpacePosition, 0);
	RWCulledObjectBoxBounds[DestIndex * CULLED_OBJECT_BOX_BOUNDS_STRIDE + 2] = float4(ObjectXAxis / max(dot(ObjectXAxis, ObjectXAxis), .0001f), 0);
	RWCulledObjectBoxBounds[DestIndex * CULLED_OBJECT_BOX_BOUNDS_STRIDE + 3] = float4(ObjectYAxis / max(dot(ObjectYAxis, ObjectYAxis), .0001f), 0);
	RWCulledObjectBoxBounds[DestIndex * CULLED_OBJECT_BOX_BOUNDS_STRIDE + 4] = float4(ObjectZAxis / max(dot(ObjectZAxis, ObjectZAxis), .0001f), 0);
}

float4 ShadowConvexHull[12];
float4 ShadowBoundingSphere;
uint NumShadowHullPlanes;

bool ShadowConvexHullIntersectSphere(float3 SphereOrigin, float SphereRadius)
{
	float3 TranslatedSphereOrigin = SphereOrigin + ShadowBoundingSphere.xyz;

	for (uint PlaneIndex = 0; PlaneIndex < NumShadowHullPlanes; PlaneIndex++)
	{
		float4 PlaneData = ShadowConvexHull[PlaneIndex];
		float PlaneDistance = dot(PlaneData.xyz, TranslatedSphereOrigin) - PlaneData.w;

		if (PlaneDistance > SphereRadius)
		{
			return false;
		}
	}
	
	return true;
}

[numthreads(UPDATEOBJECTS_THREADGROUP_SIZE, 1, 1)]
void CullObjectsForShadowCS(
	uint3 GroupId : SV_GroupID,
	uint3 DispatchThreadId : SV_DispatchThreadID,
    uint3 GroupThreadId : SV_GroupThreadID) 
{
	uint ObjectIndex = DispatchThreadId.x;

#define USE_FRUSTUM_CULLING 1
#if USE_FRUSTUM_CULLING

	if (DispatchThreadId.x == 0)
	{
		// RWObjectIndirectArguments is zeroed by a clear before this shader, only need to set things that are non-zero (and are not read by this shader as that would be a race condition)
		// IndexCount, NumInstances, StartIndex, BaseVertexIndex, FirstInstance
		RWObjectIndirectArguments[0] = ObjectBoundingGeometryIndexCount;
	}

	GroupMemoryBarrierWithGroupSync();

	uint SourceIndex = ObjectIndex;

	if (ObjectIndex < NumSceneObjects)
	{
		float4 ObjectBoundingSphere = float4(ObjectBounds[4 * SourceIndex + 0], ObjectBounds[4 * SourceIndex + 1], ObjectBounds[4 * SourceIndex + 2], ObjectBounds[4 * SourceIndex + 3]);
	
		if (ShadowBoundingSphere.w == 0 && ShadowConvexHullIntersectSphere(ObjectBoundingSphere.xyz, ObjectBoundingSphere.w)
			|| ShadowBoundingSphere.w > 0 && dot(ShadowBoundingSphere.xyz - ObjectBoundingSphere.xyz, ShadowBoundingSphere.xyz - ObjectBoundingSphere.xyz) < Square(ShadowBoundingSphere.w + ObjectBoundingSphere.w))
		{
			uint DestIndex;
			InterlockedAdd(RWObjectIndirectArguments[1], 1U, DestIndex);
			CopyCulledObjectData(DestIndex, SourceIndex);
		}
	}

#else

	if (DispatchThreadId.x == 0)
	{
		// IndexCount, NumInstances, StartIndex, BaseVertexIndex, FirstInstance
		RWObjectIndirectArguments[0] = ObjectBoundingGeometryIndexCount;
		RWObjectIndirectArguments[1] = NumSceneObjects;
	}

	uint SourceIndex = ObjectIndex;
	uint DestIndex = ObjectIndex;

	if (ObjectIndex < NumSceneObjects)
	{
		CopyCulledObjectData(DestIndex, SourceIndex);
	}
#endif
}

RWBuffer<uint> RWShadowTileHeadDataUnpacked;
RWBuffer<uint> RWShadowTileArrayData;

[numthreads(THREADGROUP_SIZEX, THREADGROUP_SIZEY, 1)]
void ClearTilesCS(
	uint3 GroupId : SV_GroupID,
	uint3 DispatchThreadId : SV_DispatchThreadID,
    uint3 GroupThreadId : SV_GroupThreadID) 
{
    uint TileIndex = DispatchThreadId.y * ShadowTileListGroupSize.x + DispatchThreadId.x;
	
	RWShadowTileHeadDataUnpacked[TileIndex * 2 + 0] = TileIndex;
	RWShadowTileHeadDataUnpacked[TileIndex * 2 + 1] = 0;
}

struct FShadowObjectCullVertexOutput
{
	nointerpolation float4 PositionAndRadius : TEXCOORD0;
	nointerpolation uint ObjectIndex : TEXCOORD1;
};

float ConservativeRadiusScale;
float MinRadius;

/** Used when culling objects into screenspace tile lists */
void ShadowObjectCullVS(
	float4 InPosition : ATTRIBUTE0,
	uint ObjectIndex : SV_InstanceID,
	out FShadowObjectCullVertexOutput Output,
	out float4 OutPosition : SV_POSITION
	)
{
	float4 ObjectPositionAndRadius = LoadObjectPositionAndRadius(ObjectIndex);
	// ConservativeRadiusScale pushes the sphere vertices out so the triangles between them lie completely outside the sphere
	// MinRadius is for conservative rasterization
	float EffectiveRadius = (ObjectPositionAndRadius.w + MinRadius) * ConservativeRadiusScale;
	float3 WorldPosition = InPosition.xyz * EffectiveRadius + ObjectPositionAndRadius.xyz;
	OutPosition = mul(float4(WorldPosition, 1), WorldToShadow);

	// Clamp the vertex to the near plane if it is in front of the near plane
	if (OutPosition.z < 0)
	{
		OutPosition.z = 0.000001f;
		OutPosition.w = 1.0f;
	}

	Output.PositionAndRadius = ObjectPositionAndRadius;
	Output.ObjectIndex = ObjectIndex;
} 

/** Intersects a single object with the tile and adds to the intersection list if needed. */
void ShadowObjectCullPS(
	FShadowObjectCullVertexOutput Input, 
	in float4 SVPos : SV_POSITION,
	out float4 OutColor : SV_Target0)
{
	OutColor = 0;
	
	uint2 TilePosition = (uint2)SVPos.xy;
	uint TileIndex = TilePosition.y * ShadowTileListGroupSize.x + TilePosition.x;

#define OBJECT_OBB_INTERSECTION 1
#if OBJECT_OBB_INTERSECTION
	
	float3 ShadowTileMin;
	float3 ShadowTileMax;

	float2 TilePositionForCulling = float2(TilePosition.x, ShadowTileListGroupSize.y - TilePosition.y);
	//@todo - why is this expand needed
	float TileExpand = 1;
	ShadowTileMin.xy = (TilePositionForCulling - TileExpand) / (float2)ShadowTileListGroupSize * 2 - 1;
	ShadowTileMax.xy = (TilePositionForCulling + 1) / (float2)ShadowTileListGroupSize * 2 - 1;
	// Extrude toward light to avoid culling objects between the light and the shadow frustum
	ShadowTileMin.z = -1000;
	ShadowTileMax.z = 1;

	float3 ObjectViewSpaceMin;
	float3 ObjectViewSpaceMax;
	LoadObjectViewSpaceBox(Input.ObjectIndex, ObjectViewSpaceMin, ObjectViewSpaceMax);

	BRANCH
	// Separating axis test on the AABB
	// Note: don't clip by near plane, objects closer to the light can still cast into the frustum
	if (all(ObjectViewSpaceMax.xy > ShadowTileMin.xy) && all(ObjectViewSpaceMin < ShadowTileMax))
	{
		float3 ObjectCenter = .5f * (ObjectViewSpaceMin + ObjectViewSpaceMax);
		float3 MinProjections = 500000;
		float3 MaxProjections = -500000;

		{
			float3 Corners[8];
			Corners[0] = float3(ShadowTileMin.x, ShadowTileMin.y, ShadowTileMin.z);
			Corners[1] = float3(ShadowTileMax.x, ShadowTileMin.y, ShadowTileMin.z);
			Corners[2] = float3(ShadowTileMin.x, ShadowTileMax.y, ShadowTileMin.z);
			Corners[3] = float3(ShadowTileMax.x, ShadowTileMax.y, ShadowTileMin.z);
			Corners[4] = float3(ShadowTileMin.x, ShadowTileMin.y, ShadowTileMax.z);
			Corners[5] = float3(ShadowTileMax.x, ShadowTileMin.y, ShadowTileMax.z);
			Corners[6] = float3(ShadowTileMin.x, ShadowTileMax.y, ShadowTileMax.z);
			Corners[7] = float3(ShadowTileMax.x, ShadowTileMax.y, ShadowTileMax.z);

			float3 ObjectAxisX;
			float3 ObjectAxisY; 
			float3 ObjectAxisZ;
			LoadObjectAxes(Input.ObjectIndex, ObjectAxisX, ObjectAxisY, ObjectAxisZ);

			UNROLL
			for (int i = 0; i < 8; i++)
			{
				float3 CenterToVertex = Corners[i] - ObjectCenter;
				float3 Projections = float3(dot(CenterToVertex, ObjectAxisX), dot(CenterToVertex, ObjectAxisY), dot(CenterToVertex, ObjectAxisZ));
				MinProjections = min(MinProjections, Projections);
				MaxProjections = max(MaxProjections, Projections);
			}
		}

		BRANCH
		// Separating axis test on the OBB
		if (all(MinProjections < 1) && all(MaxProjections > -1))
		{
			uint ArrayIndex;
			InterlockedAdd(RWShadowTileHeadDataUnpacked[TileIndex * 2 + 1], 1U, ArrayIndex);

			if (ArrayIndex < ShadowMaxObjectsPerTile)
			{
				uint DataIndex = (ArrayIndex * (uint)(ShadowTileListGroupSize.x * ShadowTileListGroupSize.y + .5f) + TileIndex);
				RWShadowTileArrayData[DataIndex * SHADOW_TILE_ARRAY_DATA_STRIDE] = Input.ObjectIndex;
			}
		}
	}
	
#else
	{
		uint ArrayIndex;
		InterlockedAdd(RWShadowTileHeadDataUnpacked[TileIndex * 2 + 1], 1U, ArrayIndex);

		if (ArrayIndex < ShadowMaxObjectsPerTile)
		{
			uint DataIndex = (ArrayIndex * (uint)(ShadowTileListGroupSize.x * ShadowTileListGroupSize.y + .5f) + TileIndex);
			RWShadowTileArrayData[DataIndex * SHADOW_TILE_ARRAY_DATA_STRIDE] = Input.ObjectIndex;
		}
	}
#endif
}

RWTexture2D<float2> RWShadowFactors;
float2 NumGroups;

/** From point being shaded toward light, for directional lights. */
float3 LightDirection;
float4 LightPositionAndInvRadius;
float LightSourceRadius;
float RayStartOffsetDepthScale;
float3 TanLightAngleAndNormalThreshold;
int4 ScissorRectMinAndSize;

/** Min and Max depth for this tile. */
groupshared uint IntegerTileMinZ;
groupshared uint IntegerTileMaxZ;

/** Inner Min and Max depth for this tile. */
groupshared uint IntegerTileMinZ2;
groupshared uint IntegerTileMaxZ2;

/** Number of objects affecting the tile, after culling. */
groupshared uint TileNumObjects0;
groupshared uint TileNumObjects1;

void CullObjectsToTileWithGather(
	float SceneDepth, 
	uint ThreadIndex, 
	uint2 GroupId, 
	float TraceDistance, 
	float MinDepth, 
	out uint NumIntersectingObjects, 
	out bool bTileShouldComputeShadowing,
	out uint GroupIndex)
{
	// Initialize per-tile variables
    if (ThreadIndex == 0) 
	{
        IntegerTileMinZ = 0x7F7FFFFF;     
        IntegerTileMaxZ = 0;
		IntegerTileMinZ2 = 0x7F7FFFFF;  
		IntegerTileMaxZ2 = 0;
		TileNumObjects0 = 0;
		TileNumObjects1 = 0;
    }

    GroupMemoryBarrierWithGroupSync();
    
	if (SceneDepth > MinDepth)
	{
		// Use shared memory atomics to build the depth bounds for this tile
		// Each thread is assigned to a pixel at this point
		//@todo - move depth range computation to a central point where it can be reused by all the frame's tiled deferred passes!
		InterlockedMin(IntegerTileMinZ, asuint(SceneDepth));
		InterlockedMax(IntegerTileMaxZ, asuint(SceneDepth));
	}

    GroupMemoryBarrierWithGroupSync();

    float MinTileZ = asfloat(IntegerTileMinZ);
    float MaxTileZ = asfloat(IntegerTileMaxZ);

	float HalfZ = .5f * (MinTileZ + MaxTileZ);

	// Compute a second min and max Z, clipped by HalfZ, so that we get two depth bounds per tile
	// This results in more conservative tile depth bounds and fewer intersections
	if (SceneDepth >= HalfZ && SceneDepth > MinDepth)
	{
		InterlockedMin(IntegerTileMinZ2, asuint(SceneDepth));
	}

	if (SceneDepth <= HalfZ && SceneDepth > MinDepth)
	{
		InterlockedMax(IntegerTileMaxZ2, asuint(SceneDepth));
	}

	GroupMemoryBarrierWithGroupSync();
	
	float MinTileZ2 = asfloat(IntegerTileMinZ2);
	float MaxTileZ2 = asfloat(IntegerTileMaxZ2);

	float3 ViewTileMin;
	float3 ViewTileMax;

	float3 ViewTileMin2;
	float3 ViewTileMax2;

	float ExpandRadius = 0;

	// Note: this code is assuming a centered projection, aka no translation present in ViewToClip
	// Stereo rendering uses an off center projection
	float2 TanViewFOV = GetTanHalfFieldOfView();
	// tan(FOV) = HalfUnitPlaneWidth / 1, so TanViewFOV * 2 is the size of the whole unit view plane
	// We are operating on a subset of that defined by ScissorRectMinAndSize
	float2 TileSize = TanViewFOV * 2 * ScissorRectMinAndSize.zw / ((float2)View.ViewSizeAndInvSize.xy * NumGroups);
	float2 UnitPlaneMin = -TanViewFOV + TanViewFOV * 2 * (ScissorRectMinAndSize.xy - View.ViewRectMin.xy) * View.ViewSizeAndInvSize.zw;
	float2 UnitPlaneTileMin = (GroupId.xy * TileSize + UnitPlaneMin) * float2(1, -1);
	float2 UnitPlaneTileMax = ((GroupId.xy + 1) * TileSize + UnitPlaneMin) * float2(1, -1);

	ViewTileMin.xy = min(MinTileZ * UnitPlaneTileMin, MaxTileZ2 * UnitPlaneTileMin) - ExpandRadius;
	ViewTileMax.xy = max(MinTileZ * UnitPlaneTileMax, MaxTileZ2 * UnitPlaneTileMax) + ExpandRadius;
	ViewTileMin.z = MinTileZ - ExpandRadius;
	ViewTileMax.z = MaxTileZ2 + ExpandRadius;
	ViewTileMin2.xy = min(MinTileZ2 * UnitPlaneTileMin, MaxTileZ * UnitPlaneTileMin) - ExpandRadius;
	ViewTileMax2.xy = max(MinTileZ2 * UnitPlaneTileMax, MaxTileZ * UnitPlaneTileMax) + ExpandRadius;
	ViewTileMin2.z = MinTileZ2 - ExpandRadius;
	ViewTileMax2.z = MaxTileZ + ExpandRadius;

	float3 ViewGroup0Center = (ViewTileMax + ViewTileMin) / 2;
	float3 WorldGroup0Center = mul(float4(ViewGroup0Center, 1), View.ViewToTranslatedWorld).xyz - View.PreViewTranslation;
	float Group0BoundingRadius = length(ViewGroup0Center - ViewTileMax);

	float3 ViewGroup1Center = (ViewTileMax2 + ViewTileMin2) / 2;
	float3 WorldGroup1Center = mul(float4(ViewGroup1Center, 1), View.ViewToTranslatedWorld).xyz - View.PreViewTranslation;
	float Group1BoundingRadius = length(ViewGroup1Center - ViewTileMax2);

#if POINT_LIGHT
	float3 LightVector0 = LightPositionAndInvRadius.xyz - WorldGroup0Center;
	float LightVector0Length = length(LightVector0);
	float3 LightVector1 = LightPositionAndInvRadius.xyz - WorldGroup1Center;
	float LightVector1Length = length(LightVector1);
	float3 LightDirection0 = LightVector0 / LightVector0Length;
	float3 LightDirection1 = LightVector1 / LightVector1Length;;
	float RayLength0 = LightVector0Length;
	float RayLength1 = LightVector1Length;

	// Don't operate on tiles completely outside of the light's influence
	bTileShouldComputeShadowing = LightVector0Length < 1.0f / LightPositionAndInvRadius.w + Group0BoundingRadius
		|| LightVector1Length < 1.0f / LightPositionAndInvRadius.w + Group1BoundingRadius;

#else
	float3 LightDirection0 = LightDirection;
	float3 LightDirection1 = LightDirection;
	float RayLength0 = TraceDistance;
	float RayLength1 = TraceDistance;

	bTileShouldComputeShadowing = SceneDepth > MinDepth;

#endif

	BRANCH
	if (bTileShouldComputeShadowing)
	{
		uint NumCulledObjects = GetCulledNumObjects();

		// Compute per-tile lists of affecting objects through bounds culling
		// Each thread now operates on a sample instead of a pixel
		LOOP
		for (uint ObjectIndex = ThreadIndex; ObjectIndex < NumCulledObjects; ObjectIndex += THREADGROUP_TOTALSIZE)
		{
			float4 SphereCenterAndRadius = LoadObjectPositionAndRadius(ObjectIndex);

			BRANCH
			if (RaySegmentHitSphere(WorldGroup0Center, LightDirection0, RayLength0, SphereCenterAndRadius.xyz, SphereCenterAndRadius.w + Group0BoundingRadius))
			{
				uint ListIndex;
				InterlockedAdd(TileNumObjects0, 1U, ListIndex);
				// Don't overwrite on overflow
				ListIndex = min(ListIndex, (uint)(MAX_INTERSECTING_OBJECTS - 1));
				IntersectingObjectIndices[MAX_INTERSECTING_OBJECTS * 0 + ListIndex] = ObjectIndex; 
			}

			BRANCH
			if (RaySegmentHitSphere(WorldGroup1Center, LightDirection1, RayLength1, SphereCenterAndRadius.xyz, SphereCenterAndRadius.w + Group1BoundingRadius))
			{
				uint ListIndex;
				InterlockedAdd(TileNumObjects1, 1U, ListIndex);
				// Don't write out of bounds on overflow
				ListIndex = min(ListIndex, (uint)(MAX_INTERSECTING_OBJECTS - 1));
				IntersectingObjectIndices[MAX_INTERSECTING_OBJECTS * 1 + ListIndex] = ObjectIndex; 
			}
		}
	}

	GroupMemoryBarrierWithGroupSync();

	GroupIndex = SceneDepth > MaxTileZ2 ? 1 : 0;
	NumIntersectingObjects = min(GroupIndex == 0 ? TileNumObjects0 : TileNumObjects1, (uint)MAX_INTERSECTING_OBJECTS);
}

float MinDepth;
float MaxDepth;
uint DownsampleFactor;

[numthreads(THREADGROUP_SIZEX, THREADGROUP_SIZEY, 1)]
void DistanceFieldShadowingCS(
	uint3 GroupId : SV_GroupID,
	uint3 DispatchThreadId : SV_DispatchThreadID,
    uint3 GroupThreadId : SV_GroupThreadID) 
{
	uint ThreadIndex = GroupThreadId.y * THREADGROUP_SIZEX + GroupThreadId.x;

	float2 ScreenUV = float2((DispatchThreadId.xy * DownsampleFactor + ScissorRectMinAndSize.xy + .5f) * View.BufferSizeAndInvSize.zw);
	float2 ScreenPosition = (ScreenUV.xy - View.ScreenPositionScaleBias.wz) / View.ScreenPositionScaleBias.xy;

	float SceneDepth = CalcSceneDepth(ScreenUV);
	float4 HomogeneousWorldPosition = mul(float4(ScreenPosition * SceneDepth, SceneDepth, 1), View.ScreenToWorld);
	float3 OpaqueWorldPosition = HomogeneousWorldPosition.xyz / HomogeneousWorldPosition.w;

	// Distance for directional lights to trace
	float TraceDistance = TanLightAngleAndNormalThreshold.z;

	uint NumIntersectingObjects = GetCulledNumObjects();
	uint CulledDataParameter = 0;
	bool bTileShouldComputeShadowing = SceneDepth > MinDepth && SceneDepth < MaxDepth;

#define USE_CULLING 1
#if USE_CULLING

#if SCATTER_TILE_CULLING

	if (bTileShouldComputeShadowing)
	{
		GetShadowTileCulledData(OpaqueWorldPosition, CulledDataParameter, NumIntersectingObjects);
	}

#else

	//@todo - support MinDepth for tile culling
	float MinDepthForTileCulling = 0;
	CullObjectsToTileWithGather(SceneDepth, ThreadIndex, GroupId.xy, TraceDistance, MinDepthForTileCulling, NumIntersectingObjects, bTileShouldComputeShadowing, CulledDataParameter);

#endif
#endif // USE_CULLING

	float Result = 0;

#define COMPUTE_SHADOWING 1
#if COMPUTE_SHADOWING
	BRANCH
	if (bTileShouldComputeShadowing)
	{
		{
			// World space offset along the start of the ray to avoid incorrect self-shadowing
			float RayStartOffset = 2 + RayStartOffsetDepthScale * SceneDepth;
			// Keeps result from going all the way sharp
			float MinSphereRadius = .4f;
			// Maintain reasonable culling bounds
			float MaxSphereRadius = 100;

			#if POINT_LIGHT

				float3 LightVector = LightPositionAndInvRadius.xyz - OpaqueWorldPosition;
				float LightVectorLength = length(LightVector);
				float3 WorldRayStart = OpaqueWorldPosition + LightVector / LightVectorLength * RayStartOffset;
				float3 WorldRayEnd = LightPositionAndInvRadius.xyz;
				float MaxRayTime = LightVectorLength;
				float MaxAngle = tan(10 * PI / 180.0f);
				// Comparing tangents instead of angles, but tangent is always increasing in this range
				float TanLightAngle = min(LightSourceRadius / LightVectorLength, MaxAngle);

			#else

				float3 WorldRayStart = OpaqueWorldPosition + LightDirection * RayStartOffset;
				float3 WorldRayEnd = OpaqueWorldPosition + LightDirection * TraceDistance;
				float MaxRayTime = TraceDistance;
				float TanLightAngle = TanLightAngleAndNormalThreshold.x;

			#endif

			#if SCATTER_TILE_CULLING
				bool bUseScatterTileCulling = true;
			#else
				bool bUseScatterTileCulling = false;
			#endif
		
			#if USE_CULLING
				bool bUseCulling = true;
			#else
				bool bUseCulling = false;
			#endif

			float SubsurfaceDensity = 0;
			bool bUseSubsurfaceTransmission = false;

#if !FORWARD_SHADING
			FGBufferData GBufferData = GetGBufferData(ScreenUV);

			BRANCH
			if (IsSubsurfaceModel(GBufferData.ShadingModelID))
			{
				// Note: this has to match shadowmap logic
				// Derive density from a heuristic using opacity, tweaked for useful falloff ranges and to give a linear depth falloff with opacity
				SubsurfaceDensity = -.05f * log(1 - min(GBufferData.CustomData.a, .999f));
				bUseSubsurfaceTransmission = true;
			}
#endif

			Result = ShadowRayTraceThroughCulledObjects(
				WorldRayStart, 
				WorldRayEnd, 
				MaxRayTime, 
				TanLightAngle, 
				MinSphereRadius, 
				MaxSphereRadius, 
				SubsurfaceDensity,
				CulledDataParameter, 
				NumIntersectingObjects,
				bUseCulling,
				bUseScatterTileCulling,
				bUseSubsurfaceTransmission);
		}
	}

#else
	//Result = bTileShouldComputeShadowing;
	Result = NumIntersectingObjects / 5000.0f;
#endif

	RWShadowFactors[DispatchThreadId.xy] = float2(Result, SceneDepth);
}

Texture2D ShadowFactorsTexture;
SamplerState ShadowFactorsSampler;

float FadePlaneOffset;
float InvFadePlaneLength;
float NearFadePlaneOffset;
float InvNearFadePlaneLength;

void DistanceFieldShadowingUpsamplePS(
	in float4 UVAndScreenPos : TEXCOORD0,
	in float4 SVPos : SV_POSITION,
	out float4 OutColor : SV_Target0)
{
	// Distance field shadowing was computed at 0,0 regardless of viewrect min
	float2 DistanceFieldUVs = UVAndScreenPos.xy - ScissorRectMinAndSize.xy * View.BufferSizeAndInvSize.zw;
	float SceneDepth = CalcSceneDepth(UVAndScreenPos.xy);

#define BILATERAL_UPSAMPLE 1
#if BILATERAL_UPSAMPLE && UPSAMPLE_REQUIRED
	float2 LowResBufferSize = floor(View.BufferSizeAndInvSize.xy / DOWNSAMPLE_FACTOR);
	float2 LowResTexelSize = 1.0f / LowResBufferSize;
	float2 Corner00UV = floor(DistanceFieldUVs * LowResBufferSize - .5f) / LowResBufferSize + .5f * LowResTexelSize;
	float2 BilinearWeights = (DistanceFieldUVs - Corner00UV) * LowResBufferSize;

	float2 TextureValues00 = Texture2DSampleLevel(ShadowFactorsTexture, ShadowFactorsSampler, Corner00UV, 0).xy;
	float2 TextureValues10 = Texture2DSampleLevel(ShadowFactorsTexture, ShadowFactorsSampler, Corner00UV + float2(LowResTexelSize.x, 0), 0).xy;
	float2 TextureValues01 = Texture2DSampleLevel(ShadowFactorsTexture, ShadowFactorsSampler, Corner00UV + float2(0, LowResTexelSize.y), 0).xy;
	float2 TextureValues11 = Texture2DSampleLevel(ShadowFactorsTexture, ShadowFactorsSampler, Corner00UV + LowResTexelSize, 0).xy;

	float4 CornerWeights = float4(
		(1 - BilinearWeights.y) * (1 - BilinearWeights.x), 
		(1 - BilinearWeights.y) * BilinearWeights.x,
		BilinearWeights.y * (1 - BilinearWeights.x),
		BilinearWeights.y * BilinearWeights.x);

	float Epsilon = .0001f;

	float4 CornerDepths = abs(float4(TextureValues00.y, TextureValues10.y, TextureValues01.y, TextureValues11.y));
	float4 DepthWeights = 1.0f / (abs(CornerDepths - SceneDepth.xxxx) + Epsilon);
	float4 FinalWeights = CornerWeights * DepthWeights;

	float InterpolatedResult = 
		(FinalWeights.x * TextureValues00.x 
		+ FinalWeights.y * TextureValues10.x
		+ FinalWeights.z * TextureValues01.x 
		+ FinalWeights.w * TextureValues11.x)
		/ dot(FinalWeights, 1);

	float Output = InterpolatedResult;

#else
	float Output = Texture2DSampleLevel(ShadowFactorsTexture, ShadowFactorsSampler, DistanceFieldUVs, 0).x;
#endif

	float FarBlendFactor = 1.0f - saturate((SceneDepth - FadePlaneOffset) * InvFadePlaneLength);
	Output = lerp(1, Output, FarBlendFactor);

	float NearBlendFactor = saturate((SceneDepth - NearFadePlaneOffset) * InvNearFadePlaneLength);
	Output = lerp(1, Output, NearBlendFactor);

	OutColor = EncodeLightAttenuation(half4(Output, Output, Output, Output));
}