struct v2p
{
	float2 c:  TEXCOORD0; // center
	float2 lt: TEXCOORD1; // left top
	float2 rb: TEXCOORD2; // right bottom
	float2 rt: TEXCOORD3; // right top
	float2 lb: TEXCOORD4; // left bottom
	float4 lr: TEXCOORD5; // left / right
	float4 tb: TEXCOORD6; // top / bottom
};

//uniform sampler2D s_distort;
uniform float4 e_barrier = float4(5e-5, 5e-5, 0, 0);
uniform float4 e_weights = float4(1.0f, 1.0f, 1.0f, 1.0f); // x = normal, y = depth
uniform float4 e_kernel = float4(0.3f, 1.0f, 1.0f, 1.0f); // x = normal, y = depth


float4 main(v2p IN, 
			uniform sampler2D colors,
			uniform sampler2D normals
			): COLOR
{
	// normal discontinuity filter
	float3 nc = (float3)tex2D(normals, IN.c.xy);

	float4 nd;
	nd.x = dot(nc, float3(tex2D(normals, IN.lt.xy)));
	nd.y = dot(nc, float3(tex2D(normals, IN.rb.xy)));
	nd.z = dot(nc, float3(tex2D(normals, IN.rt.xy)));
	nd.w = dot(nc, float3(tex2D(normals, IN.lb.xy)));

	nd -= e_barrier.x;
	nd = step(0.0f, nd);

	float ne = saturate(dot(nd, e_weights.x));

	// opposite coordinates
	float4 lrr = IN.lr.wzyx;
	float4 tbr = IN.tb.wzyx;

	// depth filter: compute gradient difference:
	// (c - sample) + (c - opposite sample)

	float4 dc = float4(tex2D(colors, IN.c).w);

	float pos_lt = (float)tex2D(colors, IN.lt.xy).w;
	float pos_rb = (float)tex2D(colors, IN.rb.xy).w;

	float pos_lb = (float)tex2D(colors, IN.lb.xy).w;
	float pos_rt = (float)tex2D(colors, IN.rt.xy).w;

	float pos_l = (float)tex2D(colors, IN.lr.xy).w;
	float pos_r = (float)tex2D(colors, lrr.xy).w;

	float pos_t = (float)tex2D(colors, IN.tb.xy).w;
	float pos_b = (float)tex2D(colors, tbr.xy).w;

	float4 dd;

	dd.x = pos_lt + pos_rb;
	dd.y = pos_lb + pos_rt;
	dd.z = pos_l + pos_r;
	dd.w = pos_t + pos_b;

	dd = abs(2.0f * dc - dd) - e_barrier.y;
	dd = step(dd, 0.0f);

	float de = saturate(dot(dd, e_weights.y));

	// weight: 0 = no aa, 1 = full antialiasing
	float w = (1.0f - de * ne) * e_kernel.x;

	// smoothed color
	// (a - c) * w + c = a * w + c * (1 - w)
	float2 offset = IN.c.xy * (1.0f - w);

	float4 s0 = tex2D(colors, offset + IN.lt.xy * w);
	float4 s1 = tex2D(colors, offset + IN.rb.xy * w);
	float4 s2 = tex2D(colors, offset + IN.rt.xy * w);
	float4 s3 = tex2D(colors, offset + IN.lb.xy * w);

	return (s0 + s1 + s2 + s3) / 4.0f;
	//return float4(w);
}