// The implementation is based on the demo on "fine pruned tiled lighting" published in GPU Pro 7.
// https://github.com/wolfgangfengel/GPU-Pro-7

#pragma kernel TileLightListGen					LIGHTLISTGEN=TileLightListGen
#pragma kernel TileLightListGen_SrcBigTile		LIGHTLISTGEN=TileLightListGen_SrcBigTile		USE_TWO_PASS_TILED_LIGHTING


#include "..\common\ShaderBase.h"
#include "LightDefinitions.cs.hlsl"

#include "LightingConvexHullUtils.hlsl"

#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
#include "SortingComputeUtils.hlsl"
#endif


#define FINE_PRUNING_ENABLED
#define PERFORM_SPHERICAL_INTERSECTION_TESTS


uniform int g_iNrVisibLights;
uniform uint2 g_viDimensions;
uniform float4x4 g_mInvScrProjection;
uniform float4x4 g_mScrProjection;


Texture2D g_depth_tex : register( t0 );
StructuredBuffer<float3> g_vBoundsBuffer : register( t1 );
StructuredBuffer<SFiniteLightData> g_vLightData : register( t2 );
StructuredBuffer<SFiniteLightBound> g_data : register( t3 );

#ifdef USE_TWO_PASS_TILED_LIGHTING
Buffer<uint> g_vBigTileLightList : register( t4 );
#endif

#define NR_THREADS			64

// output buffer
RWBuffer<uint> g_vLightList : register( u0 );


#define MAX_NR_COARSE_ENTRIES		64
#define MAX_NR_PRUNED_ENTRIES		24

groupshared unsigned int coarseList[MAX_NR_COARSE_ENTRIES];
groupshared unsigned int prunedList[MAX_NR_COARSE_ENTRIES];		// temporarily support room for all 64 while in LDS

groupshared uint ldsZMin;
groupshared uint ldsZMax;
groupshared uint lightOffs;
#ifdef FINE_PRUNING_ENABLED
groupshared uint ldsDoesLightIntersect[2];
#endif
groupshared int ldsNrLightsFinal;

groupshared int ldsModelListCount[NR_LIGHT_MODELS];		// since NR_LIGHT_MODELS is 2

#ifdef PERFORM_SPHERICAL_INTERSECTION_TESTS
groupshared uint lightOffsSph;
#endif


//float GetLinearDepth(float3 vP)
//{
//	float4 v4Pres = mul(g_mInvScrProjection, float4(vP,1.0));
//	return v4Pres.z / v4Pres.w;
//}

float GetLinearDepth(float zDptBufSpace)	// 0 is near 1 is far
{
	float3 vP = float3(0.0f,0.0f,zDptBufSpace);
	float4 v4Pres = mul(g_mInvScrProjection, float4(vP,1.0));
	return v4Pres.z / v4Pres.w;
}


float3 GetViewPosFromLinDepth(float2 v2ScrPos, float fLinDepth)
{
	float fSx = g_mScrProjection[0].x;
	float fCx = g_mScrProjection[0].z;
	float fSy = g_mScrProjection[1].y;
	float fCy = g_mScrProjection[1].z;

#if USE_LEFTHAND_CAMERASPACE
	return fLinDepth*float3( ((v2ScrPos.x-fCx)/fSx), ((v2ScrPos.y-fCy)/fSy), 1.0 );
#else
	return fLinDepth*float3( -((v2ScrPos.x+fCx)/fSx), -((v2ScrPos.y+fCy)/fSy), 1.0 );
#endif
}

float GetOnePixDiagWorldDistAtDepthOne()
{
	float fSx = g_mScrProjection[0].x;
	float fSy = g_mScrProjection[1].y;

	return length( float2(1.0/fSx,1.0/fSy) );
}

#ifdef PERFORM_SPHERICAL_INTERSECTION_TESTS
int SphericalIntersectionTests(uint threadID, int iNrCoarseLights, float2 screenCoordinate);
#endif

#ifdef FINE_PRUNING_ENABLED
void FinePruneLights(uint threadID, int iNrCoarseLights, uint2 viTilLL, float4 vLinDepths);
#endif


[numthreads(NR_THREADS, 1, 1)]
void LIGHTLISTGEN(uint threadID : SV_GroupIndex, uint3 u3GroupID : SV_GroupID)
{
	uint2 tileIDX = u3GroupID.xy;
	uint t=threadID;

	if(t<MAX_NR_COARSE_ENTRIES)
		prunedList[t]=0;
	
	uint iWidth = g_viDimensions.x;
	uint iHeight = g_viDimensions.y;
	uint nrTilesX = (iWidth+15)/16;
	uint nrTilesY = (iHeight+15)/16;

	// build tile scr boundary
	const uint uFltMax = 0x7f7fffff;  // FLT_MAX as a uint
	if(t==0)
	{
		ldsZMin = uFltMax;
		ldsZMax = 0;
		lightOffs = 0;
	}

#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
	GroupMemoryBarrierWithGroupSync();
#endif


	uint2 viTilLL = 16*tileIDX;

	// establish min and max depth first
	float dpt_mi=asfloat(uFltMax), dpt_ma=0.0;


	float4 vLinDepths;
	{
		// Fetch depths and calculate min/max
		[unroll]
		for(int i = 0; i < 4; i++)
		{
			int idx = i * NR_THREADS + t;
			uint2 uCrd = min( uint2(viTilLL.x+(idx&0xf), viTilLL.y+(idx>>4)), uint2(iWidth-1, iHeight-1) );
			const float fDepth = FetchDepth(g_depth_tex, uCrd);
			vLinDepths[i] = GetLinearDepth(fDepth);
			if(fDepth<VIEWPORT_SCALE_Z)		// if not skydome
			{
				dpt_mi = min(fDepth, dpt_mi);
				dpt_ma = max(fDepth, dpt_ma);
			}
		}

		InterlockedMax(ldsZMax, asuint(dpt_ma));
		InterlockedMin(ldsZMin, asuint(dpt_mi));


#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
		GroupMemoryBarrierWithGroupSync();
#endif
	}


	float3 vTileLL = float3(viTilLL.x/(float) iWidth, viTilLL.y/(float) iHeight, asfloat(ldsZMin));
	float3 vTileUR = float3((viTilLL.x+16)/(float) iWidth, (viTilLL.y+16)/(float) iHeight, asfloat(ldsZMax));
	vTileUR.xy = min(vTileUR.xy,float2(1.0,1.0)).xy;
	

	// build coarse list using AABB
#ifdef USE_TWO_PASS_TILED_LIGHTING
	int NrBigTilesX = (nrTilesX+3)>>2;
	const int bigTileIdx = (tileIDX.y>>2)*NrBigTilesX + (tileIDX.x>>2);		// map the idx to 64x64 tiles
	int nrBigTileLights = g_vBigTileLightList[MAX_NR_BIGTILE_LIGHTS_PLUSONE*bigTileIdx+0];
	for(int l0=(int) t; l0<(int) nrBigTileLights; l0 += NR_THREADS)
	{
		int l = g_vBigTileLightList[MAX_NR_BIGTILE_LIGHTS_PLUSONE*bigTileIdx+l0+1];
#else
	for(int l=(int) t; l<(int) g_iNrVisibLights; l += NR_THREADS)
	{
#endif
		const float3 vMi = g_vBoundsBuffer[l];
		const float3 vMa = g_vBoundsBuffer[l+g_iNrVisibLights];

		if( all(vMa>vTileLL) && all(vMi<vTileUR))
		{
			unsigned int uInc = 1;
			unsigned int uIndex;
			InterlockedAdd(lightOffs, uInc, uIndex);
			if(uIndex<MAX_NR_COARSE_ENTRIES) coarseList[uIndex] = l;		// add to light list
		}
	}

#ifdef FINE_PRUNING_ENABLED	
	if(t<2) ldsDoesLightIntersect[t] = 0;
#endif

#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
	GroupMemoryBarrierWithGroupSync();
#endif

	int iNrCoarseLights = min(lightOffs,MAX_NR_COARSE_ENTRIES);

#ifdef PERFORM_SPHERICAL_INTERSECTION_TESTS
	iNrCoarseLights = SphericalIntersectionTests( t, iNrCoarseLights, float2(min(viTilLL.xy+uint2(16/2,16/2), uint2(iWidth-1, iHeight-1))) );
#endif

#ifndef FINE_PRUNING_ENABLED	
	{
		if((int)t<iNrCoarseLights) prunedList[t] = coarseList[t];
		if(t==0) ldsNrLightsFinal=iNrCoarseLights;
	}
#else
	{
		// initializes ldsNrLightsFinal with the number of accepted lights.
		// all accepted entries delivered in prunedList[].
		FinePruneLights(t, iNrCoarseLights, viTilLL, vLinDepths);
	}
#endif

	//
	if(t<NR_LIGHT_MODELS) ldsModelListCount[t]=0;

#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
	GroupMemoryBarrierWithGroupSync();
#endif

	
	int nrLightsCombinedList = min(ldsNrLightsFinal,MAX_NR_COARSE_ENTRIES);
	for(int i=t; i<nrLightsCombinedList; i+=NR_THREADS) 
	{
		InterlockedAdd(ldsModelListCount[ g_vLightData[ prunedList[i] ].lightModel ], 1);
	}


	// sort lights (gives a more efficient execution in both deferred and tiled forward lighting).
#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
	SORTLIST(prunedList, nrLightsCombinedList, MAX_NR_COARSE_ENTRIES, t, NR_THREADS);
	//MERGESORTLIST(prunedList, coarseList, nrLightsCombinedList, t, NR_THREADS);
#endif

	// write lights to global buffers
	int localOffs=0;
	int offs = tileIDX.y*nrTilesX + tileIDX.x;
	for(int m=0; m<NR_LIGHT_MODELS; m++)
	{
		int nrLightsFinal = ldsModelListCount[ m ];
		int nrLightsFinalClamped = nrLightsFinal<MAX_NR_PRUNED_ENTRIES ? nrLightsFinal : MAX_NR_PRUNED_ENTRIES;
		

		const int nrDWords = ((nrLightsFinalClamped+1)+1)>>1;
		for(int l=(int) t; l<(int) nrDWords; l += NR_THREADS)
		{
			uint uLow = l==0 ? nrLightsFinalClamped : prunedList[2*l-1+localOffs];
			uint uHigh = prunedList[2*l+0+localOffs];

			g_vLightList[16*offs + l] = (uLow&0xffff) | (uHigh<<16);
		}

		localOffs += nrLightsFinal;
		offs += (nrTilesX*nrTilesY);
	}

}



#ifdef PERFORM_SPHERICAL_INTERSECTION_TESTS
int SphericalIntersectionTests(uint threadID, int iNrCoarseLights, float2 screenCoordinate)
{
	lightOffsSph = 0;

	// make a copy of coarseList in prunedList.
	for(int l=threadID; l<iNrCoarseLights; l+=NR_THREADS)
		prunedList[l]=coarseList[l];

#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
	GroupMemoryBarrierWithGroupSync();
#endif

#if USE_LEFTHAND_CAMERASPACE
	float3 V = GetViewPosFromLinDepth( screenCoordinate, 1.0);
#else
	float3 V = GetViewPosFromLinDepth( screenCoordinate, -1.0);
#endif

	float onePixDiagDist = GetOnePixDiagWorldDistAtDepthOne();
	float halfTileSizeAtZDistOne = 8*onePixDiagDist;		// scale by half a tile
	
	for(int l=threadID; l<iNrCoarseLights; l+=NR_THREADS)
	{
		SFiniteLightBound lightData = g_data[coarseList[l]];
	
		if( DoesSphereOverlapTile(V, halfTileSizeAtZDistOne, lightData.center.xyz, lightData.radius) )
		{
			unsigned int uIndex;
			InterlockedAdd(lightOffsSph, 1, uIndex);
			coarseList[uIndex]=prunedList[l];		// read from the original copy of coarseList which is backed up in prunedList
		}
	}

#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
	GroupMemoryBarrierWithGroupSync();
#endif

	return lightOffsSph;
}
#endif


#ifdef FINE_PRUNING_ENABLED
// initializes ldsNrLightsFinal with the number of accepted lights.
// all accepted entries delivered in prunedList[].
void FinePruneLights(uint threadID, int iNrCoarseLights, uint2 viTilLL, float4 vLinDepths)
{
	uint t = threadID;
	uint iWidth = g_viDimensions.x;
	uint iHeight = g_viDimensions.y;

	uint uLightsFlags[2] = {0,0};
	int l=0;
	// need this outer loop even on xb1 and ps4 since direct lights and
	// reflection lights are kept in separate regions.
	while(l<iNrCoarseLights)
	{
		// fetch light
		int idxCoarse = l<iNrCoarseLights ? coarseList[l] : 0;
		uint uLgtType = l<iNrCoarseLights ? g_vLightData[idxCoarse].lightType : 0;

		// spot
		while(l<iNrCoarseLights && uLgtType==SPOT_LIGHT)
		{
			SFiniteLightData lightData = g_vLightData[idxCoarse];
			const bool bIsSpotDisc = (lightData.flags&IS_CIRCULAR_SPOT_SHAPE)!=0;
				
			// serially check 4 pixels
			uint uVal = 0;
			for(int i=0; i<4; i++)
			{
				int idx = t + i*NR_THREADS;
	
				uint2 uPixLoc = min(uint2(viTilLL.x+(idx&0xf), viTilLL.y+(idx>>4)), uint2(iWidth-1, iHeight-1));
				float3 vVPos = GetViewPosFromLinDepth(uPixLoc + float2(0.5,0.5), vLinDepths[i]);
	
				// check pixel
				float3 fromLight = vVPos-lightData.lightPos.xyz;
				float distSq = dot(fromLight,fromLight);
				const float fSclProj = dot(fromLight, lightData.lightAxisZ.xyz);		// spotDir = lightData.lightAxisZ.xyz

				float2 V = abs( float2( dot(fromLight, lightData.lightAxisX.xyz), dot(fromLight, lightData.lightAxisY.xyz) ) );

				float fDist2D = bIsSpotDisc ? length(V) : max(V.x,V.y);
				if( all( float2(lightData.radiusSq, fSclProj) > float2(distSq, fDist2D*lightData.cotan) ) ) uVal = 1;
			}

			uLightsFlags[l<32 ? 0 : 1] |= (uVal<<(l&31));
			++l; idxCoarse = l<iNrCoarseLights ? coarseList[l] : 0;
			uLgtType = l<iNrCoarseLights ? g_vLightData[idxCoarse].lightType : 0;
		}

		// sphere
		while(l<iNrCoarseLights && uLgtType==SPHERE_LIGHT)
		{
			SFiniteLightData lightData = g_vLightData[idxCoarse];

			// serially check 4 pixels
			uint uVal = 0;
			for(int i=0; i<4; i++)
			{
				int idx = t + i*NR_THREADS;
	
				uint2 uPixLoc = min(uint2(viTilLL.x+(idx&0xf), viTilLL.y+(idx>>4)), uint2(iWidth-1, iHeight-1));
				float3 vVPos = GetViewPosFromLinDepth(uPixLoc + float2(0.5,0.5), vLinDepths[i]);
	
				// check pixel
				float3 vLp = lightData.lightPos.xyz;
				float3 toLight = vLp - vVPos; 
				float distSq = dot(toLight,toLight);
			
				if(lightData.radiusSq>distSq) uVal = 1;
			}

			uLightsFlags[l<32 ? 0 : 1] |= (uVal<<(l&31));
			++l; idxCoarse = l<iNrCoarseLights ? coarseList[l] : 0;
			uLgtType = l<iNrCoarseLights ? g_vLightData[idxCoarse].lightType : 0;
		}

		// Box
		while(l<iNrCoarseLights && uLgtType==BOX_LIGHT)
		{
			SFiniteLightData lightData = g_vLightData[idxCoarse];

			// serially check 4 pixels
			uint uVal = 0;
			for(int i=0; i<4; i++)
			{
				int idx = t + i*NR_THREADS;
	
				uint2 uPixLoc = min(uint2(viTilLL.x+(idx&0xf), viTilLL.y+(idx>>4)), uint2(iWidth-1, iHeight-1));
				float3 vVPos = GetViewPosFromLinDepth(uPixLoc + float2(0.5,0.5), vLinDepths[i]);

				// check pixel
				float3 toLight  = lightData.lightPos.xyz - vVPos;

				float3 dist = float3( dot(toLight, lightData.lightAxisX), dot(toLight, lightData.lightAxisY), dot(toLight, lightData.lightAxisZ) );
				dist = (abs(dist) - lightData.boxInnerDist) * lightData.boxInvRange;		// not as efficient as it could be
				if( max(max(dist.x, dist.y), dist.z)<1 ) uVal = 1;						// but allows us to not write out OuterDists
			}

			uLightsFlags[l<32 ? 0 : 1] |= (uVal<<(l&31));
			++l; idxCoarse = l<iNrCoarseLights ? coarseList[l] : 0;
			uLgtType = l<iNrCoarseLights ? g_vLightData[idxCoarse].lightType : 0;
		}

		// in case we have some corrupt data make sure we terminate
		if(uLgtType>=MAX_TYPES) ++l;
	}

	InterlockedOr(ldsDoesLightIntersect[0], uLightsFlags[0]);
	InterlockedOr(ldsDoesLightIntersect[1], uLightsFlags[1]);
	if(t==0) ldsNrLightsFinal = 0;

#if !defined(SHADER_API_XBOXONE) && !defined(SHADER_API_PSSL)
	GroupMemoryBarrierWithGroupSync();
#endif

	if(t<(uint) iNrCoarseLights && (ldsDoesLightIntersect[t<32 ? 0 : 1]&(1<<(t&31)))!=0 )
	{
		unsigned int uInc = 1;
		unsigned int uIndex;
		InterlockedAdd(ldsNrLightsFinal, uInc, uIndex);
		if(uIndex<MAX_NR_COARSE_ENTRIES) prunedList[uIndex] = coarseList[t];		// we allow up to 64 pruned lights while stored in LDS.
	}
}
#endif