ScriptableRenderPipeline/ScriptableRenderPipeline/HDRenderPipeline/HDRP/Lighting/LightLoop/scrbound.compute

// 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 ScreenBoundsAABB

#include "CoreRP/ShaderLibrary/common.hlsl"
#include "LightLoop.cs.hlsl"
#include "LightCullUtils.hlsl"

#pragma only_renderers d3d11 ps4 xboxone vulkan metal

uniform int g_isOrthographic;
uniform int g_iNrVisibLights;

uniform float4x4 g_mInvProjectionArr[2];
uniform float4x4 g_mProjectionArr[2];

StructuredBuffer<SFiniteLightBound> g_data : register( t0 );

#define NR_THREADS			64

// output buffer
RWStructuredBuffer<float3> g_vBoundsBuffer : register( u0 );

#define MAX_PNTS		9		// strictly this should be 10=6+4 but we get more wavefronts and 10 seems to never hit (fingers crossed)
								// However, worst case the plane that would be skipped if such an extreme case ever happened would be backplane
								// clipping gets skipped which doesn't cause any errors.


// LDS (2496 bytes)
groupshared float posX[MAX_PNTS*8*2];
groupshared float posY[MAX_PNTS*8*2];
groupshared float posZ[MAX_PNTS*8*2];
groupshared float posW[MAX_PNTS*8*2];
groupshared unsigned int clipFlags[48];


unsigned int GetClip(const float4 P);
int ClipAgainstPlane(const int iSrcIndex, const int iNrSrcVerts, const int subLigt, const int p);
void CalcBound(out bool2 bIsMinValid, out bool2 bIsMaxValid, out float2 vMin, out float2 vMax, float4x4 InvProjection, float3 pos_view_space, float r);

#include "LightingConvexHullUtils.hlsl"


[numthreads(NR_THREADS, 1, 1)]
void ScreenBoundsAABB(uint threadID : SV_GroupIndex, uint3 u3GroupID : SV_GroupID)
{
	uint groupID = u3GroupID.x;
    uint eyeIndex = u3GroupID.y; // currently, can only be 0 or 1

    // The g_ is preserved in order to make cross-pipeline (FPTL) updates easier
    float4x4 g_mInvProjection = g_mInvProjectionArr[eyeIndex];
    float4x4 g_mProjection = g_mProjectionArr[eyeIndex];

	//uint vindex = groupID * NR_THREADS + threadID;
	unsigned int g = groupID;
	unsigned int t = threadID;

	const int subLigt = (int) (t/8);
	const int lgtIndex = subLigt+(int) g*8;
	const int sideIndex = (int) (t%8);

    const int eyeAdjustedLgtIndex = GenerateLightCullDataIndex(lgtIndex, g_iNrVisibLights, eyeIndex);
    SFiniteLightBound lgtDat = g_data[eyeAdjustedLgtIndex];

	const float3 boxX = lgtDat.boxAxisX.xyz;
	const float3 boxY = lgtDat.boxAxisY.xyz;
	const float3 boxZ = -lgtDat.boxAxisZ.xyz;           // flip axis (so it points away from the light direction for a spot-light)
	const float3 center = lgtDat.center.xyz;
	const float radius = lgtDat.radius;
	const float2 scaleXY = lgtDat.scaleXY;

	{
		if(sideIndex<6 && lgtIndex<(int) g_iNrVisibLights)		// mask 2 out of 8 threads
		{
			float3 q0, q1, q2, q3;
			GetHullQuad(q0, q1, q2, q3, boxX, boxY, boxZ, center, scaleXY, sideIndex);


			const float4 vP0 = mul(g_mProjection, float4(q0, 1));
			const float4 vP1 = mul(g_mProjection, float4(q1, 1));
			const float4 vP2 = mul(g_mProjection, float4(q2, 1));
			const float4 vP3 = mul(g_mProjection, float4(q3, 1));

			// test vertices of one quad (of the convex hull) for intersection
			const unsigned int uFlag0 = GetClip(vP0);
			const unsigned int uFlag1 = GetClip(vP1);
			const unsigned int uFlag2 = GetClip(vP2);
			const unsigned int uFlag3 = GetClip(vP3);

			const float4 vPnts[] = {vP0, vP1, vP2, vP3};

			// screen-space AABB of one quad (assuming no intersection)
			float3 vMin, vMax;
			for(int k=0; k<4; k++)
			{
				float fW = vPnts[k].w;
				float fS = fW<0 ? -1 : 1;
				float fWabs = fW<0 ? (-fW) : fW;
				fW = fS * (fWabs<FLT_EPS ? FLT_EPS : fWabs);
				float3 vP = float3(vPnts[k].x/fW, vPnts[k].y/fW, vPnts[k].z/fW);
				if(k==0) { vMin=vP; vMax=vP; }

				vMax = max(vMax, vP); vMin = min(vMin, vP);
			}

			clipFlags[subLigt*6+sideIndex] = (uFlag0<<0) | (uFlag1<<6) | (uFlag2<<12) | (uFlag3<<18);

			// store in clip buffer (only use these vMin and vMax if light is 100% visible in which case clipping isn't needed)
			posX[subLigt*MAX_PNTS*2 + sideIndex] = vMin.x;
			posY[subLigt*MAX_PNTS*2 + sideIndex] = vMin.y;
			posZ[subLigt*MAX_PNTS*2 + sideIndex] = vMin.z;

			posX[subLigt*MAX_PNTS*2 + sideIndex + 6] = vMax.x;
			posY[subLigt*MAX_PNTS*2 + sideIndex + 6] = vMax.y;
			posZ[subLigt*MAX_PNTS*2 + sideIndex + 6] = vMax.z;
		}
	}

	// if not XBONE and not PLAYSTATION4 we need a memorybarrier here
	// since we can't rely on the gpu cores being 64 wide.
	// We need a pound define around this.
	GroupMemoryBarrierWithGroupSync();


	{
		int f=0;

		if(sideIndex==0 && lgtIndex<(int) g_iNrVisibLights)
		{
			// quick acceptance or rejection
			unsigned int uCollectiveAnd = (unsigned int) -1;
			unsigned int uCollectiveOr = 0;
			for(f=0; f<6; f++)
			{
				unsigned int uFlagAnd = clipFlags[subLigt*6+f]&0x3f;
				unsigned int uFlagOr = uFlagAnd;
				for(int i=1; i<4; i++)
				{
					unsigned int uClipBits = (clipFlags[subLigt*6+f]>>(i*6))&0x3f;
					uFlagAnd &= uClipBits;
					uFlagOr |= uClipBits;
				}

				uCollectiveAnd &= uFlagAnd;
				uCollectiveOr |= uFlagOr;
			}

			bool bSetBoundYet = false;
			float3 vMin=0.0, vMax=0.0;
			if(uCollectiveAnd!=0 || uCollectiveOr==0)		// all invisible or all visible (early out)
			{
				if(uCollectiveOr==0)	// all visible
				{
					for(f=0; f<6; f++)
					{
						const int sideIndex = f;

						float3 vFaceMi = float3(posX[subLigt*MAX_PNTS*2 + sideIndex + 0], posY[subLigt*MAX_PNTS*2 + sideIndex + 0], posZ[subLigt*MAX_PNTS*2 + sideIndex + 0]);
						float3 vFaceMa = float3(posX[subLigt*MAX_PNTS*2 + sideIndex + 6], posY[subLigt*MAX_PNTS*2 + sideIndex + 6], posZ[subLigt*MAX_PNTS*2 + sideIndex + 6]);

						for(int k=0; k<2; k++)
						{
							float3 vP = k==0 ? vFaceMi : vFaceMa;
							if(f==0 && k==0) { vMin=vP; vMax=vP; }

							vMax = max(vMax, vP); vMin = min(vMin, vP);
						}
					}
					bSetBoundYet=true;
				}
			}
			else		// :( need true clipping
			{

				for(f=0; f<6; f++)
				{
					float3 q0, q1, q2, q3;
					GetHullQuad(q0, q1, q2, q3, boxX, boxY, boxZ, center, scaleXY, f);

					// 4 vertices to a quad of the convex hull in post projection space
					const float4 vP0 = mul(g_mProjection, float4(q0, 1));
					const float4 vP1 = mul(g_mProjection, float4(q1, 1));
					const float4 vP2 = mul(g_mProjection, float4(q2, 1));
					const float4 vP3 = mul(g_mProjection, float4(q3, 1));


					int iSrcIndex = 0;

					int offs = iSrcIndex*MAX_PNTS+subLigt*MAX_PNTS*2;

					// fill up source clip buffer with the quad
					posX[offs+0]=vP0.x; posX[offs+1]=vP1.x; posX[offs+2]=vP2.x; posX[offs+3]=vP3.x;
					posY[offs+0]=vP0.y; posY[offs+1]=vP1.y; posY[offs+2]=vP2.y; posY[offs+3]=vP3.y;
					posZ[offs+0]=vP0.z; posZ[offs+1]=vP1.z; posZ[offs+2]=vP2.z; posZ[offs+3]=vP3.z;
					posW[offs+0]=vP0.w; posW[offs+1]=vP1.w; posW[offs+2]=vP2.w; posW[offs+3]=vP3.w;

					int iNrSrcVerts = 4;

					// do true clipping
					for(int p=0; p<6; p++)
					{
						const int nrVertsDst = ClipAgainstPlane(iSrcIndex, iNrSrcVerts, subLigt, p);

						iSrcIndex = 1-iSrcIndex;
						iNrSrcVerts = nrVertsDst;

						if(iNrSrcVerts<3 || iNrSrcVerts>=MAX_PNTS) break;
					}

					// final clipped convex primitive is in src buffer
					if(iNrSrcVerts>2)
					{
						int offs_src = iSrcIndex*MAX_PNTS+subLigt*MAX_PNTS*2;
						for(int k=0; k<iNrSrcVerts; k++)
						{
							float4 vCur = float4(posX[offs_src+k], posY[offs_src+k], posZ[offs_src+k], posW[offs_src+k]);

							// project and apply toward AABB
							float3 vP = float3(vCur.x/vCur.w, vCur.y/vCur.w, vCur.z/vCur.w);
							if(!bSetBoundYet) { vMin=vP; vMax=vP; bSetBoundYet=true; }

							vMax = max(vMax, vP); vMin = min(vMin, vP);
						}
					}

				}

				////////////////////// look for camera frustum verts that need to be included. That is frustum vertices inside the convex hull for the light
				int i=0;
				for(i=0; i<8; i++)	// establish 8 camera frustum vertices
				{
					float3 vVertPSpace = float3((i&1)!=0 ? 1 : (-1), (i&2)!=0 ? 1 : (-1), (i&4)!=0 ? 1 : 0);

					float4 v4ViewSpace = mul(g_mInvProjection, float4(vVertPSpace,1));
					float3 vViewSpace = float3(v4ViewSpace.x/v4ViewSpace.w, v4ViewSpace.y/v4ViewSpace.w, v4ViewSpace.z/v4ViewSpace.w);

					posX[subLigt*MAX_PNTS*2 + i] = vViewSpace.x;
					posY[subLigt*MAX_PNTS*2 + i] = vViewSpace.y;
					posZ[subLigt*MAX_PNTS*2 + i] = vViewSpace.z;
				}

				// determine which camera frustum vertices are inside the convex hull
				uint uVisibFl = 0xff;
				for(f=0; f<6; f++)
				{
					float3 vP0, vN;
					GetHullPlane(vP0, vN, boxX, boxY, boxZ, center, scaleXY, f);

					for(i=0; i<8; i++)
					{
						float3 vViewSpace = float3(posX[subLigt*MAX_PNTS*2 + i], posY[subLigt*MAX_PNTS*2 + i], posZ[subLigt*MAX_PNTS*2 + i]);
						uVisibFl &= ( dot(vViewSpace-vP0, vN)<0 ? 0xff : (~(1<<i)) );
					}
				}

				// apply camera frustum vertices inside the convex hull to the AABB
				for(i=0; i<8; i++)
				{
					if((uVisibFl&(1<<i))!=0)
					{
						float3 vP = float3((i&1)!=0 ? 1 : (-1), (i&2)!=0 ? 1 : (-1), (i&4)!=0 ? 1 : 0);

						if(!bSetBoundYet) { vMin=vP; vMax=vP; bSetBoundYet=true; }

						vMax = max(vMax, vP); vMin = min(vMin, vP);
					}
				}
			}





			// determine AABB bound in [-1;1]x[-1;1] screen space using bounding sphere.
			// Use the result to make our already established AABB from the convex hull
			// potentially tighter.
			if(!bSetBoundYet)
			{
				// set the AABB off-screen
				vMin = float3(-3,-3,-3);
				vMax = float3(-2,-2,-2);
			}
			else
			{
				//if((center.z+radius)<0.0)
				if(g_isOrthographic==0 && length(center)>radius)
				{
					float2 vMi, vMa;
					bool2 bMi, bMa;
					CalcBound(bMi, bMa, vMi, vMa, g_mInvProjection, center, radius);

					vMin.xy = bMi ? max(vMin.xy, vMi) : vMin.xy;
					vMax.xy = bMa ? min(vMax.xy, vMa) : vMax.xy;
				}
				else if(g_isOrthographic!=0)
				{
					float2 vMi = mul(g_mProjection, float4(center.xyz-radius,1)).xy;	 // no division needed for ortho
					float2 vMa = mul(g_mProjection, float4(center.xyz+radius,1)).xy;	 // no division needed for ortho
					vMin.xy = max(vMin.xy, vMi);
					vMax.xy = min(vMax.xy, vMa);
				}

#if USE_LEFT_HAND_CAMERA_SPACE
				if((center.z-radius)>0.0)
				{
					float4 vPosF = mul(g_mProjection, float4(0,0,center.z-radius,1));
					vMin.z = max(vMin.z, vPosF.z/vPosF.w);
				}
				if((center.z+radius)>0.0)
				{
					float4 vPosB = mul(g_mProjection, float4(0,0,center.z+radius,1));
					vMax.z = min(vMax.z, vPosB.z/vPosB.w);
				}
#else
				if((center.z+radius)<0.0)
				{
					float4 vPosF = mul(g_mProjection, float4(0,0,center.z+radius,1));
					vMin.z = max(vMin.z, vPosF.z/vPosF.w);
				}
				if((center.z-radius)<0.0)
				{
					float4 vPosB = mul(g_mProjection, float4(0,0,center.z-radius,1));
					vMax.z = min(vMax.z, vPosB.z/vPosB.w);
				}
#endif
				else
				{
					vMin = float3(-3,-3,-3);
					vMax = float3(-2,-2,-2);
				}
			}


			// we should consider doing a look-up here into a max depth mip chain
			// to see if the light is occluded: vMin.z*VIEWPORT_SCALE_Z > MipTexelMaxDepth
			//g_vBoundsBuffer[lgtIndex+0] = float3(0.5*vMin.x+0.5, -0.5*vMax.y+0.5, vMin.z*VIEWPORT_SCALE_Z);
			//g_vBoundsBuffer[lgtIndex+g_iNrVisibLights] = float3(0.5*vMax.x+0.5, -0.5*vMin.y+0.5, vMax.z*VIEWPORT_SCALE_Z);

			// changed for unity

            // Each light's AABB is represented by two float3s, the min and max of the box.
            // And for stereo, we have two sets of lights.  Therefore, each eye has a set of mins, followed by
            // a set of maxs, and each set is equal to g_iNrVisibLights.
            const ScreenSpaceBoundsIndices boundsIndices = GenerateScreenSpaceBoundsIndices(lgtIndex, g_iNrVisibLights, eyeIndex);

            g_vBoundsBuffer[boundsIndices.min] = float3(0.5*vMin.x + 0.5, 0.5*vMin.y + 0.5, vMin.z*VIEWPORT_SCALE_Z);
            g_vBoundsBuffer[boundsIndices.max] = float3(0.5*vMax.x + 0.5, 0.5*vMax.y + 0.5, vMax.z*VIEWPORT_SCALE_Z);
		}
	}
}


float4 GenNewVert(const float4 vVisib, const float4 vInvisib, const int p);

int ClipAgainstPlane(const int iSrcIndex, const int iNrSrcVerts, const int subLigt, const int p)
{
	int offs_src = iSrcIndex*MAX_PNTS+subLigt*MAX_PNTS*2;
	int offs_dst = (1-iSrcIndex)*MAX_PNTS+subLigt*MAX_PNTS*2;

	float4 vPrev = float4(posX[offs_src+(iNrSrcVerts-1)], posY[offs_src+(iNrSrcVerts-1)], posZ[offs_src+(iNrSrcVerts-1)], posW[offs_src+(iNrSrcVerts-1)]);

	int nrVertsDst = 0;

	unsigned int uMask = (1<<p);
	bool bIsPrevVisib = (GetClip(vPrev)&uMask)==0;
	for(int i=0; i<iNrSrcVerts; i++)
	{
		float4 vCur = float4(posX[offs_src+i], posY[offs_src+i], posZ[offs_src+i], posW[offs_src+i]);
		bool bIsCurVisib = (GetClip(vCur)&uMask)==0;
		if( (bIsCurVisib && !bIsPrevVisib) || (!bIsCurVisib && bIsPrevVisib) )
		{
			//assert(nrVertsDst<MAX_PNTS);
			if(nrVertsDst<MAX_PNTS)
			{
				// generate new vertex
				float4 vNew = GenNewVert(bIsCurVisib ? vCur : vPrev, bIsCurVisib ? vPrev : vCur, p);
				posX[offs_dst+nrVertsDst]=vNew.x; posY[offs_dst+nrVertsDst]=vNew.y; posZ[offs_dst+nrVertsDst]=vNew.z; posW[offs_dst+nrVertsDst]=vNew.w;
				++nrVertsDst;
			}
		}

		if(bIsCurVisib)
		{
			//assert(nrVertsDst<MAX_PNTS);
			if(nrVertsDst<MAX_PNTS)
			{
				posX[offs_dst+nrVertsDst]=vCur.x; posY[offs_dst+nrVertsDst]=vCur.y; posZ[offs_dst+nrVertsDst]=vCur.z; posW[offs_dst+nrVertsDst]=vCur.w;
				++nrVertsDst;
			}
		}

		vPrev = vCur;
		bIsPrevVisib = bIsCurVisib;
	}

	return nrVertsDst;
}



unsigned int GetClip(const float4 P)
{
	//-P.w <= P.x <= P.w
	return ((P.x<-P.w)?1:0) | ((P.x>P.w)?2:0) | ((P.y<-P.w)?4:0) | ((P.y>P.w)?8:0) | ((P.z<0)?16:0) | ((P.z>P.w)?32:0);
}

float4 GenNewVert(const float4 vVisib, const float4 vInvisib, const int p)
{
	const float fS = p==4 ? 0 : ((p&1)==0 ? -1 : 1);
	const int index = ((uint) p)/2;
	float x1 = index==0 ? vVisib.x : (index==1 ? vVisib.y : vVisib.z);
	float x0 = index==0 ? vInvisib.x : (index==1 ? vInvisib.y : vInvisib.z);

	//fS*((vVisib.w-vInvisib.w)*t + vInvisib.w) = (x1-x0)*t + x0;

	const float fT = (fS*vInvisib.w-x0)/((x1-x0) - fS*(vVisib.w-vInvisib.w));
	float4 vNew = vVisib*fT + vInvisib*(1-fT);

	// just to be really anal we make sure the clipped against coordinate is precise
	if(index==0) vNew.x = fS*vNew.w;
	else if(index==1) vNew.y = fS*vNew.w;
	else vNew.z = fS*vNew.w;

	return vNew;
}


float4 TransformPlaneToPostSpace(float4x4 InvProjection, float4 plane)
{
	return mul(plane, InvProjection);
}

float4 EvalPlanePair(out bool validPlanes, float2 posXY_in, float r)
{
	// rotate by 90 degrees to avoid potential division by zero
	bool bMustFlip = abs(posXY_in.y)<abs(posXY_in.x);
	float2 posXY = bMustFlip ? float2(-posXY_in.y, posXY_in.x) : posXY_in;

	float fLenSQ = dot(posXY, posXY);
	float diffSq = fLenSQ - r*r;
	float D = posXY.y * sqrt(max(0.0, diffSq));

	float4 res;
	res.x = (-r*posXY.x - D) / fLenSQ;
	res.z = (-r*posXY.x + D) / fLenSQ;
	res.y = (-r-res.x*posXY.x) / posXY.y;
	res.w = (-r-res.z*posXY.x) / posXY.y;

	// rotate back by 90 degrees
	res = bMustFlip ? float4(res.y, -res.x, res.w, -res.z) : res;

	validPlanes = diffSq>0.0;

	return res;
}

void CalcBound(out bool2 bIsMinValid, out bool2 bIsMaxValid, out float2 vMin, out float2 vMax, float4x4 InvProjection, float3 pos_view_space, float r)
{
	bool validX, validY;
	float4 planeX = EvalPlanePair(validX, float2(pos_view_space.x, pos_view_space.z), r);
	float4 planeY = EvalPlanePair(validY, float2(pos_view_space.y, pos_view_space.z), r);


#if USE_LEFT_HAND_CAMERA_SPACE
	planeX = planeX.zwxy;		// need to swap left/right and top/bottom planes when using left hand system
	planeY = planeY.zwxy;
#endif

	bIsMinValid = bool2(planeX.z<0, planeY.z<0) && bool2(validX,validY);
	bIsMaxValid = bool2((-planeX.x)<0, (-planeY.x)<0) && bool2(validX,validY);

	// hopefully the compiler takes zeros into account
	// should be the case since the transformation in TransformPlaneToPostSpace()
	// is done using multiply-adds and not dot product instructions.
	float4 planeX0 = TransformPlaneToPostSpace(InvProjection, float4(planeX.x, 0, planeX.y, 0));
	float4 planeX1 = TransformPlaneToPostSpace(InvProjection, float4(planeX.z, 0, planeX.w, 0));
	float4 planeY0 = TransformPlaneToPostSpace(InvProjection, float4(0, planeY.x, planeY.y, 0));
	float4 planeY1 = TransformPlaneToPostSpace(InvProjection, float4(0, planeY.z, planeY.w, 0));


	// convert planes to the forms (1,0,0,D) and (0,1,0,D)
	// 2D bound is given by -D components
	float2 A = -float2(planeX0.w / planeX0.x, planeY0.w / planeY0.y);
	float2 B = -float2(planeX1.w / planeX1.x, planeY1.w / planeY1.y);

	// Bound is complete
	vMin = B;
	vMax = A;
}