#define SAMPLECUTDIST2 0.01
#define WEIGHTCUTOFF   0.01
#define DIST_BIAS      0.005

uniform int nRadionColumns;
uniform float3 lightPos;
uniform float3 lightDir;
uniform float lightPower;
uniform float clusterCount;
uniform float4x4 lightViewProj;
uniform float lightFarPlane;

struct vsInputComputeWeights
{
    float4	pos			: POSITION;
    float2	tex			: TEXCOORD0;
};

struct vsOutputComputeWeights
{
    float4	pos			: POSITION;
    float2	tex			: TEXCOORD0;
};

vsOutputComputeWeights
	vsComputeWeights(vsInputComputeWeights input)
{
    vsOutputComputeWeights output = (vsOutputComputeWeights)0;
    output.pos = input.pos;
    output.tex = (input.pos + 1.0) / 2.0;
    //output.tex = input.tex;
    output.tex.y = 1.0 - output.tex.y;
   
    return output;
}

float4 psComputeWeights(vsOutputComputeWeights input,
						uniform sampler2D radionSampler : register(s0),
						uniform sampler2D shadowMap : register(s1)) : COLOR0
{
	float dataColumnWidth = 1.0 / (float)nRadionColumns;
	int bushIndex = input.tex.x * 4096 + input.tex.y * nRadionColumns * 4096;
	int werx = bushIndex % nRadionColumns;
	int wery = bushIndex / nRadionColumns;
	float3 pos = tex2D(radionSampler, float2((werx + 0.25) * dataColumnWidth, (wery  + 0.5) / 4096.0) ).xyz;
	float3 dir = tex2D(radionSampler, float2((werx + 0.75) * dataColumnWidth, (wery  + 0.5) / 4096.0) ).xyz;
	dir = normalize(dir);
	
	float3 diff = lightPos - pos;
	float dist2 = dot(diff, diff);
	if(dist2 < SAMPLECUTDIST2)
		dist2 = SAMPLECUTDIST2;
	diff = normalize(diff);
	float cosa = max(dot(lightDir, -diff), 0);
	float cosb = max(dot(dir, diff), 0);	
	
	
	float4 lightVPos = mul(lightViewProj, float4(pos,1));
	float dist = sqrt(dist2) / lightFarPlane;
	lightVPos /= lightVPos.w;
	lightVPos.xy = (lightVPos.xy + 1.0) / 2.0;
	lightVPos.y = 1.0 - lightVPos.y;
	float storedDist = tex2D(shadowMap, lightVPos.xy).r;
	dist -= DIST_BIAS;
	float visibility = (dist <= storedDist);  	

	//float visibility = 1;
	//visibility = visibility * 0.00001 + 1.0;
	
	float4 ret = visibility * pow(cosa, 9) * cosb / dist2 * lightPower;
	return ret;
}

float4 psSumWeights(vsOutputComputeWeights input,
					uniform sampler2D radionSampler       : register(s0),
					uniform sampler2D entryPointCountSampler : register(s1),
					uniform sampler2D allWeightsSampler   : register(s2)) : COLOR0
{
	float halfPixel = 0.5 / clusterCount;
	float iCluster = input.tex.x + halfPixel;
	int entryPointCount = tex2D(entryPointCountSampler, float2(iCluster, 0.25));
	int currentEntryPoint = tex2D(entryPointCountSampler, float2(iCluster, 0.75));
	//sum entrypoint weights
	float weight = 0;
	float clusterRad = 0;
	for(int i = 0; i < entryPointCount; i++)
	{
		
		float dataColumnWidth = 1.0 / (float)nRadionColumns;
		int werx = currentEntryPoint % nRadionColumns;
		int wery = currentEntryPoint / nRadionColumns;
		float2 coord1 = float2((werx + 0.5) * dataColumnWidth, (wery  + 0.5) / 4096.0);
		float2 coord2 = coord1 + float2(0.25 * dataColumnWidth, 0);
		float w = tex2Dlod(allWeightsSampler, float4(coord1, 0, 0)).r;
		float radrad = tex2Dlod(radionSampler, float4(coord1, 0, 0)).a;		
		weight += w * radrad;		
		clusterRad += radrad;
		currentEntryPoint++;
	}	
	
	if(clusterRad >= 0)
		weight /= clusterRad;
	else
		weight = 0;
   
	return weight;
}

void EntryPointDisplayVS(float4 position :POSITION0,
						 float4 color :COLOR0,
						 uniform float4x4 worldViewProj,
						 out float4 hPos:POSITION,
						 out float2 texCoord: TEXCOORD0)
{
	hPos = mul(worldViewProj, position);
	texCoord = color.xy;
}

float4 EntryPointDisplayPS(float2 texCoord : TEXCOORD0,
							uniform sampler2D weightTexture : register(s0)):COLOR0
{
	//return texCoord.y;
	return texCoord.r;
}