uniform float4x4 WorldViewProj;
uniform float4 prmAtlasTilesHalfPixel;
uniform sampler2D filteredAtlasSampler : register(s0);
uniform sampler1D clusterWeightIndexSampler : register(s1);
uniform sampler2D clusterWeightSampler : register(s2);
uniform float allClusterCount; 
uniform float clusterCount; //clustercount / 4
//uniform float weights[32];

struct vsInputWalk
{
    float4 pos			: POSITION;
    float2 tex			: TEXCOORD0;
    float2 texAtlas		: TEXCOORD1;
};

struct vsOutputWalk
{
    float4 pos			: POSITION;
    float2 tex			: TEXCOORD0;
    float2 texAtlas		: TEXCOORD1;
};


vsOutputWalk vsWalk(vsInputWalk input)
{
	vsOutputWalk output = (vsOutputWalk)0;
	output.pos = mul(WorldViewProj, input.pos);
	output.tex = input.tex;
	output.texAtlas = input.texAtlas;	
	return output;
}

float4	psWalk(vsOutputWalk input) : COLOR
{
	int2 prmAtlasTiles = prmAtlasTilesHalfPixel.xy;
	float2 atlasHalfPixel = prmAtlasTilesHalfPixel.zw;
	
	float3 col = 0;
	for(int iCluster=0; iCluster<16; iCluster++)
	{
		float2 prmTexPos = float2(
			(input.texAtlas.x + (iCluster % prmAtlasTiles.x)) / prmAtlasTiles.x,
			1.0 - (input.texAtlas.y + (iCluster / prmAtlasTiles.x)) / prmAtlasTiles.y) + atlasHalfPixel;

		//float4 weight = 0.065;
		//weight = float4(weights[iCluster/4], weights[iCluster/4 + 1],weights[iCluster/4 + 2], weights[iCluster/4 + 3]);
		//float4 weight = weights[iCluster/4];
		float4 weightIndex = tex1D(clusterWeightIndexSampler, ((float)iCluster/4) / clusterCount).r / allClusterCount;
		float weight1 = tex2D(clusterWeightSampler, float2(weightIndex.r, 0.5)).r;
		float weight2 = tex2D(clusterWeightSampler, float2(weightIndex.g, 0.5)).r;
		float weight3 = tex2D(clusterWeightSampler, float2(weightIndex.b, 0.5)).r;
		float weight4 = tex2D(clusterWeightSampler, float2(weightIndex.a, 0.5)).r;
		float4 weight = float4(weight1, weight2, weight3, weight4);
		
		float3 val = tex2D(filteredAtlasSampler, prmTexPos);
		col += val.xyz * weight.x;
		iCluster++;
		
		prmTexPos.x += 1.0 / prmAtlasTiles.x;
		val = tex2D(filteredAtlasSampler, prmTexPos);
		col += val.xyz * weight.y;
		iCluster++;
		
		prmTexPos.x += 1.0 / prmAtlasTiles.x;
		val = tex2D(filteredAtlasSampler, prmTexPos);
		col += val.xyz * weight.z;
		iCluster++;
		
		prmTexPos.x += 1.0 / prmAtlasTiles.x;
		val = tex2D(filteredAtlasSampler, prmTexPos);
		col += val.xyz * weight.w;
	}
	
	//return col.xxxx *0.00001 + tex2D(filteredAtlasSampler, float2(input.texAtlas.x /32.0f, 1.0 - input.texAtlas.y));
//return float4(1,0,0,1)+col.xxxx*0.0000000001;
//return prmAtlasTilesHalfPixel.x*allClusterCount*clusterCount*0.000000000001 + tex2D(filteredAtlasSampler, float2(input.texAtlas.x /32.0f, 1.0 - input.texAtlas.y)).r;
	return float4(col, 1);
}