// Copyright Epic Games, Inc. All Rights Reserved.

/*============================================================================= 

=============================================================================*/

#include "Common.ush"
#include "DeferredShadingCommon.ush"
#include "BRDF.ush"
#include "SHCommon.ush"
#include "ReflectionEnvironmentShared.ush"
#include "MonteCarlo.ush"
#include "SkyLightingShared.ush"

struct FCopyToCubeFaceVSOutput
{
	float2 UV : TEXCOORD0;
	float3 ScreenVector : TEXCOORD1;
	float4 Position : SV_POSITION;
};

void CopyToCubeFaceVS(
	in float2 InPosition : ATTRIBUTE0,
	in float2 InUV       : ATTRIBUTE1,
	out FCopyToCubeFaceVSOutput Out
	)
{	
	DrawRectangle(float4(InPosition.xy, 0, 1), InUV, Out.Position, Out.UV);
	Out.ScreenVector = mul(float4(Out.Position.xy, 1, 0), View.ScreenToTranslatedWorld).xyz;
}

int CubeFace;

Texture2D SceneColorTexture;
SamplerState SceneColorSampler;

#if !SHADING_PATH_DEFERRED
Texture2D SceneDepthTexture;
#endif
SamplerState SceneDepthSampler;

/** 
 * X = 0 if capturing sky light, 1 if capturing reflection capture with MaxDistance fade, 2 otherwise, 
 * Y = Sky distance threshold, 
 * Z = whether a skylight's lower hemisphere should be replaced with LowerHemisphereColor.
 */
float4 SkyLightCaptureParameters;
float4 LowerHemisphereColor;

void CopySceneColorToCubeFaceColorPS(
	FCopyToCubeFaceVSOutput Input,
	out float4 OutColor : SV_Target0
	)
{
	float SceneDepth = SCENE_TEXTURES_DISABLED_SCENE_DEPTH_VALUE;
#if !SCENE_TEXTURES_DISABLED
	SceneDepth = ConvertFromDeviceZ(Texture2DSample(SceneDepthTexture, SceneDepthSampler, Input.UV).r);
#endif
	float3 SceneColor = Texture2DSample(SceneColorTexture, SceneColorSampler, Input.UV).rgb;

	// Convert INF's to valid values
	SceneColor = ClampToHalfFloatRange(SceneColor);
	float3 TranslatedWorldPosition = Input.ScreenVector * SceneDepth + View.TranslatedWorldCameraOrigin;

	float Alpha = 1;

	if (SkyLightCaptureParameters.x == 0)
	{
		// Assuming we're on a planet and no sky lighting is coming from below the horizon
		// This is important to avoid leaking from below since we are integrating incoming lighting and shadowing separately
		if (Input.ScreenVector.z < 0 && SkyLightCaptureParameters.z >= 1)
		{
			SceneColor = lerp(SceneColor, LowerHemisphereColor.rgb, LowerHemisphereColor.a);
		}
	}
	else if (SkyLightCaptureParameters.x == 1)
	{
		float RadialDistance = length(TranslatedWorldPosition - View.TranslatedWorldCameraOrigin);
		float MaxDistance = SkyLightCaptureParameters.y;

		// Setup alpha to fade out smoothly past the max distance
		// This allows a local reflection capture to only provide reflections where it has valid data, falls back to sky cubemap
		Alpha = 1 - smoothstep(.8f * MaxDistance, MaxDistance, RadialDistance);
	}

	// We need pre-multiplied alpha for correct filtering
	// However, we need to compute average brightness before masking out sky areas, so premultiplying happens later
	OutColor = float4(SceneColor, Alpha);
}

float3 GetCubemapVector(float2 ScaledUVs, int InCubeFace)
{
	float3 CubeCoordinates;

	//@todo - this could be a 3x3 matrix multiply
	if (InCubeFace == 0)
	{
		CubeCoordinates = float3(1, -ScaledUVs.y, -ScaledUVs.x);
	}
	else if (InCubeFace == 1)
	{
		CubeCoordinates = float3(-1, -ScaledUVs.y, ScaledUVs.x);
	}
	else if (InCubeFace == 2)
	{
		CubeCoordinates = float3(ScaledUVs.x, 1, ScaledUVs.y);
	}
	else if (InCubeFace == 3)
	{
		CubeCoordinates = float3(ScaledUVs.x, -1, -ScaledUVs.y);
	}
	else if (InCubeFace == 4)
	{
		CubeCoordinates = float3(ScaledUVs.x, -ScaledUVs.y, 1);
	}
	else
	{
		CubeCoordinates = float3(-ScaledUVs.x, -ScaledUVs.y, -1);
	}

	return CubeCoordinates;
}

float CubeTexelWeight( float3 N )
{
	uint Axis = 2;
	if( abs(N.x) >= abs(N.y) && abs(N.x) >= abs(N.z) )
	{
		Axis = 0;
	}
	else if( abs(N.y) > abs(N.z) )
	{
		Axis = 1;
	}

	N = Axis == 0 ? N.zyx : N;
	N = Axis == 1 ? N.xzy : N;

	float2 UV = N.xy / N.z;

	float VecLengthSqr = 1 + dot( UV, UV );
	return 4.0 / ( sqrt( VecLengthSqr ) * VecLengthSqr );
}


TextureCube SourceCubemapTexture;
SamplerState SourceCubemapSampler;
float2 SinCosSourceCubemapRotation;
float2 SvPositionToUVScale;

void CopyCubemapToCubeFaceColorPS(
	float4 SvPosition : SV_POSITION,
	out float4 OutColor : SV_Target0
	)
{
	float2 UV = SvPosition.xy * SvPositionToUVScale;
	float2 ScaledUVs = UV * 2 - 1;
	float3 CubeCoordinates = GetCubemapVector(ScaledUVs, CubeFace);
	// Rotate around Z axis
	CubeCoordinates.xy = float2(dot(CubeCoordinates.xy, float2(SinCosSourceCubemapRotation.y, -SinCosSourceCubemapRotation.x)), dot(CubeCoordinates.xy, SinCosSourceCubemapRotation));
	OutColor = TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, CubeCoordinates, 0);

	if (SkyLightCaptureParameters.x > 0)
	{
		// Assuming we're on a planet and no sky lighting is coming from below the horizon
		// This is important to avoid leaking from below since we are integrating incoming lighting and shadowing separately
		if (CubeCoordinates.z < 0 && SkyLightCaptureParameters.z >= 1)
		{
			OutColor.rgb = lerp(OutColor.rgb, LowerHemisphereColor.rgb, LowerHemisphereColor.a);
		}
	}

	OutColor.a = 1;
}

uint MipIndex;
uint NumMips;

#ifdef USE_COMPUTE
int FaceThreadGroupSize;
int2 ValidDispatchCoord;
RWTextureCube<float4> OutTextureMipColor;	// All slices, i.e. faces, of a cube map mip level
#endif 

#ifdef USE_COMPUTE
[numthreads(THREADGROUP_SIZE, THREADGROUP_SIZE, 1)]
void DownsampleCS(uint3 ThreadId : SV_DispatchThreadID)
{
	const uint2 FaceCoord = uint2(ThreadId.x % uint(FaceThreadGroupSize), ThreadId.y);
	if (any(FaceCoord >= uint2(ValidDispatchCoord)))
	{
		return;
	}
	const int SelectedCubeFace = int(ThreadId.x) / FaceThreadGroupSize;
	float2 ScaledUVs = ((float2(FaceCoord) + 0.5f) / float2(ValidDispatchCoord)) * 2.0f - 1.0f;
	float4 OutColor;
#else // USE_COMPUTE
void DownsamplePS(
	float4 SvPosition : SV_POSITION,
	out float4 OutColor : SV_Target0
	)
{ 
	float2 UV = SvPosition.xy * SvPositionToUVScale;
	float2 ScaledUVs = UV * 2 - 1;
	const int SelectedCubeFace = CubeFace;
#endif // USE_COMPUTE
	float3 CubeCoordinates = GetCubemapVector(ScaledUVs, SelectedCubeFace);

	uint MipSize = 1u << ( NumMips - MipIndex - 1 );

	float3 TangentZ = normalize( CubeCoordinates );
	float3 TangentX = normalize( cross( GetCubemapVector( ScaledUVs + float2(0,1), SelectedCubeFace), TangentZ ) );
	float3 TangentY = cross( TangentZ, TangentX );

	const float SampleOffset = 2.0 * 2 / MipSize;

	float2 Offsets[] =
	{
		float2(-1, -1) * 0.7,
		float2( 1, -1) * 0.7,
		float2(-1,  1) * 0.7,
		float2( 1,  1) * 0.7,
		
		float2( 0, -1),
		float2(-1,  0),
		float2( 1,  0),
		float2( 0,  1),
	};

	OutColor = SourceCubemapTexture.SampleLevel(SourceCubemapSampler, CubeCoordinates, 0 );

	UNROLL
	for( uint i = 0; i < 8; i++ )
	{
		float Weight = 0.375;

		float3 SampleDir = CubeCoordinates;
		SampleDir += TangentX * ( Offsets[i].x * SampleOffset );
		SampleDir += TangentY * ( Offsets[i].y * SampleOffset );
		OutColor += SourceCubemapTexture.SampleLevel(SourceCubemapSampler, SampleDir, 0 ) * Weight;
	}

	OutColor *= rcp( 1.0 + 1.0 + 2.0 );

#ifdef USE_COMPUTE
	OutTextureMipColor[uint3(FaceCoord, SelectedCubeFace)] = OutColor;
#endif
}

#ifdef USE_COMPUTE
float4 LowerHemisphereSolidColor;
[numthreads(THREADGROUP_SIZE, THREADGROUP_SIZE, 1)]
void ApplyLowerHemisphereColorCS(uint3 ThreadId : SV_DispatchThreadID)
{
	const uint2 FaceCoord = uint2(ThreadId.x % uint(FaceThreadGroupSize), ThreadId.y);
	if (any(FaceCoord >= uint2(ValidDispatchCoord)))
	{
		return;
	}
	const int SelectedCubeFace = int(ThreadId.x) / FaceThreadGroupSize;
	float2 ScaledUVs = ((float2(FaceCoord) + 0.5f) / float2(ValidDispatchCoord)) * 2.0f - 1.0f;
	float4 OutColor;

	float3 CubeCoordinates = GetCubemapVector(ScaledUVs, SelectedCubeFace);
	float3 N = normalize(CubeCoordinates);
	if (N.z < 0.0f)
	{
		OutTextureMipColor[uint3(FaceCoord, SelectedCubeFace)] = float4(LowerHemisphereSolidColor.rgb * LowerHemisphereSolidColor.a, 1.0f);
	}
}
#endif

int NumCaptureArrayMips;

/** Cube map array of reflection captures. */
TextureCube ReflectionEnvironmentColorTexture;
SamplerState ReflectionEnvironmentColorSampler;

#if COMPUTEBRIGHTNESS_PIXELSHADER

void ComputeBrightnessMain(out float4 OutColor : SV_Target0)
{ 
	// Sample the 6 1x1 cube faces and average
	float3 AverageColor = TextureCubeSampleLevel(ReflectionEnvironmentColorTexture, ReflectionEnvironmentColorSampler, float3(1, 0, 0), NumCaptureArrayMips - 1).rgb;
	AverageColor += TextureCubeSampleLevel(ReflectionEnvironmentColorTexture, ReflectionEnvironmentColorSampler, float3(-1, 0, 0), NumCaptureArrayMips - 1).rgb;
	AverageColor += TextureCubeSampleLevel(ReflectionEnvironmentColorTexture, ReflectionEnvironmentColorSampler, float3(0, 1, 0), NumCaptureArrayMips - 1).rgb;
	AverageColor += TextureCubeSampleLevel(ReflectionEnvironmentColorTexture, ReflectionEnvironmentColorSampler, float3(0, -1, 0), NumCaptureArrayMips - 1).rgb;
	AverageColor += TextureCubeSampleLevel(ReflectionEnvironmentColorTexture, ReflectionEnvironmentColorSampler, float3(0, 0, 1), NumCaptureArrayMips - 1).rgb;
	AverageColor += TextureCubeSampleLevel(ReflectionEnvironmentColorTexture, ReflectionEnvironmentColorSampler, float3(0, 0, -1), NumCaptureArrayMips - 1).rgb;

	OutColor = dot(AverageColor / 6, .3333f); 
}

#endif

float4 SampleCubemap(float3 Coordinates, uint InMipIndex)
{
	return TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, Coordinates, InMipIndex);
}

uint SourceMipIndex;
float4 SampleCubemap(float3 Coordinates)
{
	return SampleCubemap(Coordinates, SourceMipIndex);
}

void DownsamplePS_Mobile(
	float4 SvPosition : SV_POSITION,
	out float4 OutColor : SV_Target0
)
{
	float2 UV = SvPosition.xy * SvPositionToUVScale;
	float2 ScaledUVs = UV * 2 - 1;
	float3 CubeCoordinates = GetCubemapVector(ScaledUVs, CubeFace);

	OutColor = SampleCubemap(CubeCoordinates);
}

#ifdef USE_COMPUTE
int CubeFaceOffset;
[numthreads(THREADGROUP_SIZE, THREADGROUP_SIZE, 1)]
void FilterCS(uint3 ThreadId : SV_DispatchThreadID)
{
	const uint2 FaceCoord = uint2(ThreadId.x % uint(FaceThreadGroupSize), ThreadId.y);
	if (any(FaceCoord >= uint2(ValidDispatchCoord)))
	{
		return;
	}
	const int SelectedCubeFace = CubeFaceOffset + int(ThreadId.x) / FaceThreadGroupSize;
	float2 ScaledUVs = ((float2(FaceCoord) + 0.5f) / float2(ValidDispatchCoord)) * 2.0f - 1.0f;
	float4 OutColor;
#else // USE_COMPUTE
void FilterPS(
	float4 SvPosition : SV_POSITION,
	out float4 OutColor : SV_Target0
	)
{ 
	float2 UV = SvPosition.xy * SvPositionToUVScale;
	float2 ScaledUVs = UV * 2 - 1;
	const int SelectedCubeFace = CubeFace;
#endif // USE_COMPUTE
	float3 CubeCoordinates = GetCubemapVector(ScaledUVs, SelectedCubeFace);

	float3 N = normalize(CubeCoordinates);
	float3x3 TangentToWorld = GetTangentBasis( N );

	float Roughness = ComputeReflectionCaptureRoughnessFromMip( MipIndex, NumMips - 1 );

	if( Roughness < 0.01 )
	{
		OutColor = SourceCubemapTexture.SampleLevel(SourceCubemapSampler, CubeCoordinates, 0 );
	#ifdef USE_COMPUTE
		OutTextureMipColor[uint3(FaceCoord, SelectedCubeFace)] = OutColor;
	#endif
		return;
	}

	uint CubeSize = 1u << ( NumMips - 1 );
	const float SolidAngleTexel = 4.0f * PI / float(6 * CubeSize * CubeSize) * 2.0f;

	//const uint NumSamples = 1024;
	const uint NumSamples = Roughness < 0.1 ? 32 : 64;
	
	float4 FilteredColor = 0;
	BRANCH
	if( Roughness > 0.99 )
	{
		// Roughness=1, GGX is constant. Use cosine distribution instead

		LOOP
		for( uint i = 0; i < NumSamples; i++ )
		{
			float2 E = Hammersley( i, NumSamples, 0 );

			float3 L = CosineSampleHemisphere( E ).xyz;

			float NoL = L.z;

			float PDF = NoL / PI;
			float SolidAngleSample = 1.0 / ( NumSamples * PDF );
			float Mip = 0.5 * log2( SolidAngleSample / SolidAngleTexel );

			L = mul( L, TangentToWorld );
			FilteredColor += SourceCubemapTexture.SampleLevel(SourceCubemapSampler, L, Mip );
		}

		OutColor = FilteredColor / NumSamples;
	}
	else
	{
		float Weight = 0;

		LOOP
		for( uint i = 0; i < NumSamples; i++ )
		{
			float2 E = Hammersley( i, NumSamples, 0 );

			// 6x6 Offset rows. Forms uniform star pattern
			//uint2 Index = uint2( i % 6, i / 6 );
			//float2 E = ( Index + 0.5 ) / 5.8;
			//E.x = frac( E.x + (Index.y & 1) * (0.5 / 6.0) );

			E.y *= 0.995;

			float3 H = ImportanceSampleGGX( E, Pow4(Roughness) ).xyz;
			float3 L = 2 * H.z * H - float3(0,0,1);

			float NoL = L.z;
			float NoH = H.z;

			if( NoL > 0 )
			{
				//float TexelWeight = CubeTexelWeight( L );
				//float SolidAngleTexel = SolidAngleAvgTexel * TexelWeight;

				//float PDF = D_GGX( Pow4(Roughness), NoH ) * NoH / (4 * VoH);
				float PDF = D_GGX( Pow4(Roughness), NoH ) * 0.25;
				float SolidAngleSample = 1.0 / ( NumSamples * PDF );
				float Mip = 0.5 * log2( SolidAngleSample / SolidAngleTexel );

				float ConeAngle = acos( 1 - SolidAngleSample / (2*PI) );

				L = mul( L, TangentToWorld );
				FilteredColor += SourceCubemapTexture.SampleLevel(SourceCubemapSampler, L, Mip ) * NoL;
				Weight += NoL;
			}
		}

		OutColor = FilteredColor / Weight;
	}

#ifdef USE_COMPUTE
	OutTextureMipColor[uint3(FaceCoord, SelectedCubeFace)] = OutColor;
#endif
}

float4 CoefficientMask0;
float4 CoefficientMask1;
float CoefficientMask2;
int NumSamples;

void DiffuseIrradianceCopyPS(
	float4 SvPosition : SV_POSITION,
	out float4 OutColor : SV_Target0
	)
{ 
	float2 UV = SvPosition.xy * SvPositionToUVScale;
	float2 ScaledUVs = UV * 2 - 1;
	float3 CubeCoordinates = normalize(GetCubemapVector(ScaledUVs, CubeFace));

	float SquaredUVs = 1 + dot(ScaledUVs, ScaledUVs);
	// Dividing by NumSamples here to keep the sum in the range of fp16, once we get down to the 1x1 mip
	float TexelWeight = 4 / (sqrt(SquaredUVs) * SquaredUVs);

	FThreeBandSHVector SHCoefficients = SHBasisFunction3(CubeCoordinates);
	float CurrentSHCoefficient = dot(SHCoefficients.V0, CoefficientMask0) + dot(SHCoefficients.V1, CoefficientMask1) + SHCoefficients.V2 * CoefficientMask2;
	float3 TexelLighting = SampleCubemap(CubeCoordinates).rgb;

	OutColor = float4(TexelLighting * CurrentSHCoefficient * TexelWeight, TexelWeight);
}

float4 Sample01;
float4 Sample23;

void DiffuseIrradianceAccumulatePS(
	float4 SvPosition : SV_POSITION,
	out float4 OutColor : SV_Target0
	)
{ 
	float4 AccumulatedValue = 0;
	float2 UV = SvPosition.xy * SvPositionToUVScale;

	{
		float2 ScaledUVs = saturate(UV + Sample01.xy) * 2 - 1;
		float3 CubeCoordinates = GetCubemapVector(ScaledUVs, CubeFace);
		AccumulatedValue += SampleCubemap(CubeCoordinates);
	}
	
	{
		float2 ScaledUVs = saturate(UV + Sample01.zw) * 2 - 1;
		float3 CubeCoordinates = GetCubemapVector(ScaledUVs, CubeFace);
		AccumulatedValue += SampleCubemap(CubeCoordinates);
	}

	{
		float2 ScaledUVs = saturate(UV + Sample23.xy) * 2 - 1;
		float3 CubeCoordinates = GetCubemapVector(ScaledUVs, CubeFace);
		AccumulatedValue += SampleCubemap(CubeCoordinates);
	}

	{
		float2 ScaledUVs = saturate(UV + Sample23.zw) * 2 - 1;
		float3 CubeCoordinates = GetCubemapVector(ScaledUVs, CubeFace);
		AccumulatedValue += SampleCubemap(CubeCoordinates);
	}
	
	OutColor = float4(AccumulatedValue.rgb / 4.0f, AccumulatedValue.a / 4.0f);
}

void AccumulateCubeFacesPS(
	out float4 OutColor : SV_Target0
	)
{
	float4 AccumulatedValue; 
	AccumulatedValue  = TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, float3(1, 0, 0), SourceMipIndex);
	AccumulatedValue += TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, float3(-1, 0, 0), SourceMipIndex);
	AccumulatedValue += TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, float3(0, 1, 0), SourceMipIndex);
	AccumulatedValue += TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, float3(0, -1, 0), SourceMipIndex);
	AccumulatedValue += TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, float3(0, 0, 1), SourceMipIndex);
	AccumulatedValue += TextureCubeSampleLevel(SourceCubemapTexture, SourceCubemapSampler, float3(0, 0, -1), SourceMipIndex);
	OutColor = float4(4 * PI * AccumulatedValue.rgb / ( max(AccumulatedValue.a, .00001f)), 0);	
}


#ifdef SHADER_DIFFUSE_TO_SH

#include "SHCommon.ush"
#include "MonteCarlo.ush"

RWStructuredBuffer<float4> OutIrradianceEnvMapSH;
float UniformSampleSolidAngle;

#define THREADGROUP_SIZE (THREADGROUP_SIZE_X * THREADGROUP_SIZE_Y)
groupshared FThreeBandSHVectorRGB IrradianceSHShared[THREADGROUP_SIZE];

FThreeBandSHVectorRGB SampleSHRGB(in float3 SampleDirection, in float weight, in float MipLevel)
{
	const float3 SampleColor = SourceCubemapTexture.SampleLevel(SourceCubemapSampler, SampleDirection, MipLevel).rgb;

	FThreeBandSHVector Sh3Vector = SHBasisFunction3(SampleDirection);
	FThreeBandSHVectorRGB Result;
	Result.R = MulSH3(Sh3Vector, SampleColor.r * weight);
	Result.G = MulSH3(Sh3Vector, SampleColor.g * weight);
	Result.B = MulSH3(Sh3Vector, SampleColor.b * weight);
	return Result;
}

[numthreads(THREADGROUP_SIZE_X, THREADGROUP_SIZE_Y, 1)]
void ComputeSkyEnvMapDiffuseIrradianceCS(uint3 ThreadId : SV_DispatchThreadID)
{
	const uint LinearIndex = THREADGROUP_SIZE_X * ThreadId.y + ThreadId.x;

#if 1

	// For a 128x128 cubemap, sampling mip level 2 with only 8x8 samples matches closely the super sampled version.
	const float3 SampleDirection = UniformSampleSphere((float2(ThreadId.xy)+0.5f) / float2(THREADGROUP_SIZE_X, THREADGROUP_SIZE_Y)).xyz;
	IrradianceSHShared[LinearIndex] = SampleSHRGB(SampleDirection, UniformSampleSolidAngle, MipIndex);

#else

	FThreeBandSHVectorRGB IrradianceSHAcc;
	IrradianceSHAcc.R.V0 = IrradianceSHAcc.R.V1 = 0.0f; IrradianceSHAcc.R.V2 = 0.0f;
	IrradianceSHAcc.G.V0 = IrradianceSHAcc.G.V1 = 0.0f; IrradianceSHAcc.G.V2 = 0.0f;
	IrradianceSHAcc.B.V0 = IrradianceSHAcc.B.V1 = 0.0f; IrradianceSHAcc.B.V2 = 0.0f;

	// Uniform super sampling.
	// For a 128x128 cubemap, sampling mip level 0 with only 4x4 samples matches closely reference.
	const uint MipLevel = 0;
	const float SuperSampleAxisCount = 4.0f;
	const float SuperSampleCount = SuperSampleAxisCount * SuperSampleAxisCount;
	const float SuperSampleUniformSampleSolidAngle = UniformSampleSolidAngle / SuperSampleCount;
	for (float U = 0.0f; U < 1.0f; U += 1.0f / SuperSampleAxisCount)
	{
		for (float V = 0.0f; V < 1.0f; V += 1.0f / SuperSampleAxisCount)
		{
			const float3 SampleDirection = UniformSampleSphere((float2(ThreadId.xy) + float2(U,V)) / float2(THREADGROUP_SIZE_X, THREADGROUP_SIZE_Y)).xyz;
			IrradianceSHAcc = AddSH(IrradianceSHAcc, SampleSHRGB(SampleDirection, SuperSampleUniformSampleSolidAngle, MipLevel));
		}
	}

	IrradianceSHShared[LinearIndex] = IrradianceSHAcc;

#endif

#if THREADGROUP_SIZE != 64
#error That is the only reduction supported today
#endif

	// Wait for all group threads to be done
	GroupMemoryBarrierWithGroupSync();

	if (LinearIndex < 32)
	{
		IrradianceSHShared[LinearIndex] = AddSH(IrradianceSHShared[LinearIndex], IrradianceSHShared[LinearIndex + 32]);
	}
	GroupMemoryBarrierWithGroupSync();

	if (LinearIndex < 16)
	{
		IrradianceSHShared[LinearIndex] = AddSH(IrradianceSHShared[LinearIndex], IrradianceSHShared[LinearIndex + 16]);
	}
	GroupMemoryBarrierWithGroupSync();

	// The smallest wave size is 16 on Intel hardware. So now we can do simple math operations without group sync.

	if (LinearIndex < 8)
	{
		IrradianceSHShared[LinearIndex] = AddSH(IrradianceSHShared[LinearIndex], IrradianceSHShared[LinearIndex + 8]);
	}
	if (LinearIndex < 4)
	{
		IrradianceSHShared[LinearIndex] = AddSH(IrradianceSHShared[LinearIndex], IrradianceSHShared[LinearIndex + 4]);
	}
	if (LinearIndex < 2)
	{
		IrradianceSHShared[LinearIndex] = AddSH(IrradianceSHShared[LinearIndex], IrradianceSHShared[LinearIndex + 2]);
	}

	if (LinearIndex < 1)
	{
		FThreeBandSHVectorRGB SkyIrradiance = AddSH(IrradianceSHShared[LinearIndex], IrradianceSHShared[LinearIndex + 1]);

		// Pack the SH coefficients in a way that makes applying the lighting use the least shader instructions
		// This has the diffuse convolution coefficients baked in. See "Stupid Spherical Harmonics (SH) Tricks".
		// Also see UpdateSkyIrradianceGpuBuffer.
		const float SqrtPI = sqrt(PI);
		const float Coefficient0 = 1.0f / (2.0f * SqrtPI);
		const float Coefficient1 = sqrt(3.0f) / (3.0f * SqrtPI);
		const float Coefficient2 = sqrt(15.0f) / (8.0f * SqrtPI);
		const float Coefficient3 = sqrt(5.0f) / (16.0f * SqrtPI);
		const float Coefficient4 = 0.5f * Coefficient2;

		OutIrradianceEnvMapSH[0].x = -Coefficient1 * SkyIrradiance.R.V0[3];
		OutIrradianceEnvMapSH[0].y = -Coefficient1 * SkyIrradiance.R.V0[1];
		OutIrradianceEnvMapSH[0].z =  Coefficient1 * SkyIrradiance.R.V0[2];
		OutIrradianceEnvMapSH[0].w =  Coefficient0 * SkyIrradiance.R.V0[0] - Coefficient3 * SkyIrradiance.R.V1[2];//[6];

		OutIrradianceEnvMapSH[1].x = -Coefficient1 * SkyIrradiance.G.V0[3];
		OutIrradianceEnvMapSH[1].y = -Coefficient1 * SkyIrradiance.G.V0[1];
		OutIrradianceEnvMapSH[1].z =  Coefficient1 * SkyIrradiance.G.V0[2];
		OutIrradianceEnvMapSH[1].w =  Coefficient0 * SkyIrradiance.G.V0[0] - Coefficient3 * SkyIrradiance.G.V1[2];//[6];

		OutIrradianceEnvMapSH[2].x = -Coefficient1 * SkyIrradiance.B.V0[3];
		OutIrradianceEnvMapSH[2].y = -Coefficient1 * SkyIrradiance.B.V0[1];
		OutIrradianceEnvMapSH[2].z =  Coefficient1 * SkyIrradiance.B.V0[2];
		OutIrradianceEnvMapSH[2].w =  Coefficient0 * SkyIrradiance.B.V0[0] - Coefficient3 * SkyIrradiance.B.V1[2];//[6];

		OutIrradianceEnvMapSH[3].x =  Coefficient2 * SkyIrradiance.R.V1[0];//[4];
		OutIrradianceEnvMapSH[3].y = -Coefficient2 * SkyIrradiance.R.V1[1];//V[5];
		OutIrradianceEnvMapSH[3].z = 3.0f * Coefficient3 * SkyIrradiance.R.V1[2];//[6];
		OutIrradianceEnvMapSH[3].w = -Coefficient2 * SkyIrradiance.R.V1[3];//[7];

		OutIrradianceEnvMapSH[4].x =  Coefficient2 * SkyIrradiance.G.V1[0];//[4];
		OutIrradianceEnvMapSH[4].y = -Coefficient2 * SkyIrradiance.G.V1[1];//[5];
		OutIrradianceEnvMapSH[4].z = 3.0f * Coefficient3 * SkyIrradiance.G.V1[2];//[6];
		OutIrradianceEnvMapSH[4].w = -Coefficient2 * SkyIrradiance.G.V1[3];//[7];

		OutIrradianceEnvMapSH[5].x =  Coefficient2 * SkyIrradiance.B.V1[0];//[4];
		OutIrradianceEnvMapSH[5].y = -Coefficient2 * SkyIrradiance.B.V1[1];//[5];
		OutIrradianceEnvMapSH[5].z = 3.0f * Coefficient3 * SkyIrradiance.B.V1[2];//[6];
		OutIrradianceEnvMapSH[5].w = -Coefficient2 * SkyIrradiance.B.V1[3];//[7];

		OutIrradianceEnvMapSH[6].x = Coefficient4 * SkyIrradiance.R.V2;//[8];
		OutIrradianceEnvMapSH[6].y = Coefficient4 * SkyIrradiance.G.V2;//[8];
		OutIrradianceEnvMapSH[6].z = Coefficient4 * SkyIrradiance.B.V2;//[8];
		OutIrradianceEnvMapSH[6].w = 1.0f;

		// We inverse the weights applied onto the non directional SH component to recover the average sky irradiance.
		const float3 DummyVector = float3(0.0, 0.0, 1.0);
		const float SHBasisL0Weight = SHBasisFunction3(DummyVector).V0.x;
		const float InvSHBand0Weights = 1.0f / (SHBasisL0Weight * 4.0f * PI);
		const float SkyAverageIrradianceR = SkyIrradiance.R.V0.x * InvSHBand0Weights;
		const float SkyAverageIrradianceG = SkyIrradiance.G.V0.x * InvSHBand0Weights;
		const float SkyAverageIrradianceB = SkyIrradiance.B.V0.x * InvSHBand0Weights;
		const float SkyAverageIrradiance  = dot(float3(SkyAverageIrradianceR, SkyAverageIrradianceG, SkyAverageIrradianceB), .3333f);
		OutIrradianceEnvMapSH[7].x = SkyAverageIrradiance;
		OutIrradianceEnvMapSH[7].y = SkyAverageIrradiance;
		OutIrradianceEnvMapSH[7].z = SkyAverageIrradiance;
		OutIrradianceEnvMapSH[7].w = SkyAverageIrradiance;
	}
}
#endif



#ifdef REALTIME_REFLECTION_HEIGHT_FOG

#include "HeightFogCommon.ush"

#if PERMUTATION_DEPTHTEXTURE
Texture2D DepthTexture;
#endif

void RenderRealTimeReflectionHeightFogVS(
	in uint VertexId : SV_VertexID,
	out float4 Position : SV_POSITION,
	out float3 OutScreenVector : TEXCOORD0)
{
	float2 UV = -1.0f;
	UV = VertexId == 1 ? float2(-1.0f, 3.0f) : UV;
	UV = VertexId == 2 ? float2(3.0f, -1.0f) : UV;
	Position = float4(UV, 0.5f, 1.0f);
	OutScreenVector = mul(float4(Position.xy, 1, 0), View.ScreenToTranslatedWorld).xyz;
}


void RenderRealTimeReflectionHeightFogPS(
	in float4 SVPos : SV_POSITION,
	in float3 ScreenVector : TEXCOORD0,
	out float4 OutLuminance : SV_Target0
)
{
	ResolvedView = ResolveView();

	float4 WorldVector = mul(float4((SVPos.xy / 128.0f) * float2(2.0, -2.0) + float2(1.0, -1.0), 0.0, 0), ResolvedView.ScreenToTranslatedWorld);

#if PERMUTATION_DEPTHTEXTURE
	float DeviceZ = DepthTexture.Load(int3(SVPos.xy, 0)).x;
	const float HeightFogDistance = ConvertFromDeviceZ(DeviceZ);
#else
	const float HeightFogDistance = 100000.0f * 10.0f;	// 10 kilometers
#endif
	float4 HeightFogInscatteringAndOpacity = CalculateHeightFog(normalize(ScreenVector)* HeightFogDistance);

	const float OutputPreExposure = (ResolvedView.RealTimeReflectionCapture ? ResolvedView.RealTimeReflectionCapturePreExposure : ResolvedView.PreExposure);
	HeightFogInscatteringAndOpacity.rgb *= OutputPreExposure;

	OutLuminance = HeightFogInscatteringAndOpacity;
}

#endif