#pragma kernel TileLightListGen

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

#define FINE_PRUNING_ENABLED


uniform int g_iNrVisibLights;
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 );


#define NR_THREADS			64

// output buffer
//RWBuffer<uint4> g_vLightList : register( u0 );
RWStructuredBuffer<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


//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;

#ifdef LEFT_HAND_COORDINATES
	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
}

void sortLightList(int localThreadID, int n);


[numthreads(NR_THREADS, 1, 1)]
void TileLightListGen(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;
	uint iHeight;
	g_depth_tex.GetDimensions(iWidth, iHeight);
	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(XBONE) && !defined(PLAYSTATION4)
	GroupMemoryBarrierWithGroupSync();
#endif


	uint2 viTilLL = 16*tileIDX;

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

	for(int idx=t; idx<256; idx+=NR_THREADS)
	{
		uint2 uCrd = min( uint2(viTilLL.x+(idx&0xf), viTilLL.y+(idx>>4)), uint2(iWidth-1, iHeight-1) );
		const float fDpth = FetchDepth(g_depth_tex, uCrd);
		if(fDpth<VIEWPORT_SCALE_Z)		// if not skydome
		{
			dpt_mi = min(fDpth, dpt_mi);
			dpt_ma = max(fDpth, dpt_ma);
		}
	}

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


#if !defined(XBONE) && !defined(PLAYSTATION4)
	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
	for(int l=(int) t; l<(int) g_iNrVisibLights; l += NR_THREADS)
	{
		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(XBONE) && !defined(PLAYSTATION4)
	GroupMemoryBarrierWithGroupSync();
#endif

	int iNrCoarseLights = lightOffs<MAX_NR_COARSE_ENTRIES ? lightOffs : MAX_NR_COARSE_ENTRIES;


#ifndef FINE_PRUNING_ENABLED	
	{
		int iNrLightsOut = iNrCoarseLights<MAX_NR_PRUNED_ENTRIES ? iNrCoarseLights : MAX_NR_PRUNED_ENTRIES;
		if((int)t<iNrLightsOut) prunedList[t] = coarseList[t];
		if(t==0) ldsNrLightsFinal=iNrLightsOut;
	}
#else
	{
		float4 vLinDepths;
		[unroll]for(int i=0; i<4; i++)
		{
			int idx = t + i*NR_THREADS;
			uint2 uCrd = min( uint2(viTilLL.x+(idx&0xf), viTilLL.y+(idx>>4)), uint2(iWidth-1, iHeight-1) );
			float3 v3ScrPos = float3(uCrd.x+0.5, uCrd.y+0.5, FetchDepth(g_depth_tex, uCrd));
			vLinDepths[i] = GetLinearDepth(v3ScrPos.z);
		}

		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].uLightType : 0;

			// spot
			while(l<iNrCoarseLights && uLgtType==SPOT_LIGHT)
			{
				SFiniteLightData lgtDat = g_vLightData[idxCoarse];
				const bool bIsSpotDisc = (lgtDat.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-lgtDat.vLpos.xyz;
					float distSq = dot(fromLight,fromLight);
					const float fSclProj = dot(fromLight, lgtDat.vLaxisZ.xyz);		// spotDir = lgtDat.vLaxisZ.xyz

					float2 V = abs( float2( dot(fromLight, lgtDat.vLaxisX.xyz), dot(fromLight, lgtDat.vLaxisY.xyz) ) );

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

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

			// sphere
			while(l<iNrCoarseLights && uLgtType==SPHERE_LIGHT)
			{
				SFiniteLightData lgtDat = 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 = lgtDat.vLpos.xyz;
					float3 toLight = vLp - vVPos; 
					float distSq = dot(toLight,toLight);
			
					if(lgtDat.fSphRadiusSq>distSq) uVal = 1;
				}

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

			// Box
			while(l<iNrCoarseLights && uLgtType==BOX_LIGHT)
			{
				SFiniteLightData lgtDat = 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  = lgtDat.vLpos.xyz - vVPos;

					float3 dist = float3( dot(toLight, lgtDat.vLaxisX), dot(toLight, lgtDat.vLaxisY), dot(toLight, lgtDat.vLaxisZ) );
					dist = (abs(dist) - lgtDat.vBoxInnerDist) * lgtDat.vBoxInvRange;		// 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].uLightType : 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(XBONE) && !defined(PLAYSTATION4)
		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

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

#if !defined(XBONE) && !defined(PLAYSTATION4)
	GroupMemoryBarrierWithGroupSync();
#endif

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


	// sort lights
#if !defined(XBONE) && !defined(PLAYSTATION4)
	sortLightList((int) t, nrLightsCombinedList);
#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(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);
	}

}


// original version
//float2 vRay2D = float2(max(V.x,V.y), fSclProj);
//float distSqB = bIsSpotDisc ? distSq : dot(vRay2D,vRay2D);
//if( all( float3(lgtDat.fSphRadiusSq, fSclProj, fSclProj) > float3(distSq, sqrt(distSqB)*lgtDat.fPenumbra, 0.0) ) ) uVal = 1;



// previous new version
//float fDist2DSqr = bIsSpotDisc ? dot(V,V) : (maC*maC);
//if( all( float3(lgtDat.fSphRadiusSq, (fSclProj*fSclProj), fSclProj) > float3(distSq, fDist2DSqr*cotaSqr, fSpotNearPlane) ) ) uVal = 1;

#if 0
void merge(int l, int m, int r);

void sortLightList(int localThreadID, int n)
{
   for(int curr_size=1; curr_size<=n-1; curr_size = 2*curr_size)
   {
		for(int left_start=localThreadID*(2*curr_size); left_start<(n-1); left_start+=NR_THREADS*(2*curr_size))
		{
			int mid = left_start + curr_size - 1;
			int right_end = min(left_start + 2*curr_size - 1, n-1);
			merge(left_start, mid, right_end);
		}

	   GroupMemoryBarrierWithGroupSync();
   }
}

//groupshared unsigned int tmpBuffer[MAX_NR_COARSE_ENTRIES];

void merge(int l, int m, int r)
{
    int i, j, k;
    
	int ol = l;		
	int or = m+1;	
	int sl = m - l + 1;		// capacity is size of left list = m - l + 1;
    int sr =  r - m;		// capacity is size of right list = r - m

	unsigned int tmpBuffer[] = coarseList;		// re use coarse list buffer as temp buffer.

	// could do this copy more efficiently before the if-statement
	// in sortLightList() but this requires another GroupMemoryBarrierWithGroupSync()
	for(int i=l; i<=r; i++) tmpBuffer[i] = prunedList[i];
 
    i = 0;
    j = 0;
    k = l;
    while (i < sl && j < sr)
    {
		const uint lVal = tmpBuffer[ol+i];
		const uint rVal = tmpBuffer[or+j];
		bool pickLeft = lVal <= rVal;
		i = pickLeft ? (i+1) : i;
		j = pickLeft ? j : (j+1);
		prunedList[k] = pickLeft ? lVal : rVal;
        k++;
    }
 
    while (i < sl)
    {
        prunedList[k] = tmpBuffer[ol+i];
        i++; k++;
    }
 
    while (j < sr)
    {
        prunedList[k] = tmpBuffer[or+j];
        j++; k++;
    }
}
 
#else

// NOTE! returns 1 when value_in==0
unsigned int LimitPow2AndClamp(unsigned int value_in, unsigned int maxValue)
{
	unsigned int value = 1;
	
	while(value<value_in && (value<<1)<=maxValue)
		value<<=1;

	return value;
}


void sortLightList(int localThreadID, int length)
{
	// closest pow2 integer greater than or equal to length
	const int N = (const int) LimitPow2AndClamp((unsigned int) length, MAX_NR_COARSE_ENTRIES);			// N is 1 when length is zero but will still not enter first for-loop

	// bitonic sort can only handle arrays with a power of two length. Fill remaining entries with greater than possible index.
	for(int t=length+localThreadID; t<N; t+=NR_THREADS) { prunedList[t]=MAX_NR_COARSE_ENTRIES; }		// impossible index
	GroupMemoryBarrierWithGroupSync();

	for(int k=2; k<=N; k=2*k)
	{
		for(int j=k>>1; j>0; j=j>>1)
		{
			for(int i=localThreadID; i<N; i+=NR_THREADS)
			{
				int ixj=i^j;
				if((ixj)>i)
				{
					const unsigned int Avalue = prunedList[i];
					const unsigned int Bvalue = prunedList[ixj];

					const bool mustSwap = ((i&k)!=0^(Avalue>Bvalue)) && Avalue!=Bvalue;
					if(mustSwap)
					{
						prunedList[i]=Bvalue;
						prunedList[ixj]=Avalue;
					}
				}
			}

			GroupMemoryBarrierWithGroupSync();
		}
	}
}

#endif