struct pixelShaderInput
{
    float4 Diffuse             : COLOR0;
    float2 ShadowMapProjection : TEXCOORD0;
    float  LightSpaceDepth     : TEXCOORD1;
};

sampler2D ShadowMapSampler;

#define SM_SIZE 2048

float4 ps_LightPassBilin4x4PCF( pixelShaderInput Input ) : COLOR0
{
    float4   shadowMapVals[4];
    float4   shadowMapWeights[4];

    float2  shadowFractMap[4]={
        { -(1/SM_SIZE), -(1/SM_SIZE) },
        {  (1/SM_SIZE), -(1/SM_SIZE) },
        { -(1/SM_SIZE),  (1/SM_SIZE) },
        {  (1/SM_SIZE),  (1/SM_SIZE) }
    };

    //fetch 4 2x2 quads
    shadowMapVals[0] = tex2D(ShadowMapSampler, Input.ShadowMapProjection + shadowFractMap[0] );
    shadowMapVals[1] = tex2D(ShadowMapSampler, Input.ShadowMapProjection + shadowFractMap[1] );
    shadowMapVals[2] = tex2D(ShadowMapSampler, Input.ShadowMapProjection + shadowFractMap[2] );
    shadowMapVals[3] = tex2D(ShadowMapSampler, Input.ShadowMapProjection + shadowFractMap[3] );

    //fraction component of projected coordinates; here 1024 represents the shadowmap size
    // float4 pFrac = frac( Input.ShadowMapProjection * float4( SM_SIZE, SM_SIZE, 0, 0 ) );
    float2 pFrac = frac( Input.ShadowMapProjection.xy * float2( SM_SIZE, SM_SIZE ) ); // shaves 2 instructions

//	/*
    shadowMapWeights[0]  = float4( 1,          1-pFrac.x, 1,       1-pFrac.x );
    shadowMapWeights[0] *= float4( 1-pFrac.y,  1,         1,       1-pFrac.y );
    shadowMapWeights[1]  = float4( pFrac.x,    1,         pFrac.x, 1         );
    shadowMapWeights[1] *= float4( 1-pFrac.y,  1,         1,       1-pFrac.y );
    shadowMapWeights[2]  = float4( 1,          1-pFrac.x, 1,       1-pFrac.x );
    shadowMapWeights[2] *= float4( 1,          pFrac.y,   pFrac.y, 1         );
    shadowMapWeights[3]  = float4( pFrac.x,    1,         pFrac.x, 1         );
    shadowMapWeights[3] *= float4( 1,          pFrac.y,   pFrac.y, 1         );
//	*/
	/* compiler already opts to
    shadowMapWeights[0]  = float4( 1-pFrac.y,             1-pFrac.x,             1,                 (1-pFrac.x) * (1-pFrac.y) );
    shadowMapWeights[1]  = float4( pFrac.x * (1-pFrac.y), 1,                     pFrac.x,           1-pFrac.y                 );
    shadowMapWeights[2]  = float4( 1,                     (1-pFrac.x) * pFrac.y, pFrac.y,           1-pFrac.x                 );
    shadowMapWeights[3]  = float4( pFrac.x,               pFrac.y,               pFrac.x * pFrac.y, 1                         );
	*/

    // projective distance from z plane in view coords
    //float dist = Input.LightSpaceDepth - 0.005;
	float fDepth = Input.LightSpaceDepth - 0.005;

    float4 inLight;
	float percentInLight;
	inLight = ( fDepth < shadowMapVals[0] );
    percentInLight  = dot(inLight, shadowMapWeights[0]);
    inLight = ( fDepth < shadowMapVals[1] );
    percentInLight += dot(inLight, shadowMapWeights[1]);
    inLight = ( fDepth < shadowMapVals[2] );
    percentInLight += dot(inLight, shadowMapWeights[2]);
    inLight = ( fDepth < shadowMapVals[3] );
    percentInLight += dot(inLight, shadowMapWeights[3]);

    // (goboTexCoord, lightDist, lightVector, surfaceNormal)    
    //sum of weights is 9  since border taps are bilinearly filtered
    //return ( (1.0f/9.0f) * percentInLight) * Input.Diffuse + float4( 0.25, 0.25, 0.25, 1.0 );
	return ( (1.0f/9.0f) * percentInLight) * Input.Diffuse + float4( 0.25, 0.25, 0.25, 1.0 );
}