float4x4 viewProj : register(c31);
float2 quadddx : register(c30);

struct QuadInput
{
	float4 p:position;
	float2 t:texcoord0;
};

struct QuadOutput
{
	float4 p:position;
	float2 t:texcoord0;
	float4 q:texcoord1;
};

struct IndexedQuadInput
{
	float4 p:position;
	float2 n:normal;
	float4 i:texcoord0;
	float k:texcoord1;
};

struct IndexedQuadOutput
{
	float4 q:position;
	float3 n:normal;
	float4 p:texcoord0;
	float4 k:texcoord1;
	float2 t:texcoord2;
	float size:texcoord3;
};

void calc(IndexedQuadInput i, out IndexedQuadOutput o, float sizeFactor)
{
	float3 eyeVector = viewProj._m02_m12_m22;
	float3 side;
    float3 up;
    
    side = normalize(cross(eyeVector, float3(0,1,0)));    
    up = normalize(cross(side,eyeVector));
    
    float3 finalPos = i.i.xyz;
    float size = i.i.w * sizeFactor;
    finalPos += (i.p.x) * side * size * (1 - i.k * 0.5);
    finalPos += (i.p.y) * up * size * (1 + i.k * 5);
    
    float4 finalPos4 = float4(finalPos, 1);    
    
    o.q = mul(finalPos4, viewProj);

	o.t = i.n;
	o.n.xy = i.n;
	o.n.z = 0.0;
	o.p = finalPos4;
	o.k.xyz = i.p.xyz;
	o.k.a = i.k;
	o.size = size;
}

void vsDeferred(IndexedQuadInput i, out IndexedQuadOutput o) { calc(i, o, 1.0); }

void vsDepth(IndexedQuadInput i, out IndexedQuadOutput o) { calc(i, o, 0.9); }

QuadOutput vsCompose(QuadInput i)
{
	QuadOutput o;
	o.p = i.p;
	o.q = i.p;
	o.t = i.t;
	o.t *= 1.0 + quadddx;
	
	return o;
}