// GPU brush: paints a square mask stamp seamlessly into the paint texture.
//
// In the vertex shader the mesh is "unwrapped" into the texture layout using its
// UVs (uv -> clip space), while the actual local 3D position of the vertex is
// passed through to the fragment shader. There, per texel, the offset to the brush
// center (_BrushCenter, in local mesh space) is projected onto the stamp's tangent
// frame: _BrushU (right), _BrushV (up) and _BrushNormal. The texel is painted only
// if it lies inside the box (|du|,|dv|,|dn| <= _BrushRadius); the brush mask
// (_MaskTex, sampled along du/dv) then provides the per-texel OPACITY:
//   opacity = luma(mask.rgb) * mask.a * _BrushColor.a
// and the color is alpha-blended over the existing paint (source-over). So a mask
// gray value of 0.5 paints at 50%, white = full, black/transparent = nothing.
//
// This bridges UV seams automatically: triangles on both sides of a seam share the
// same 3D edge and are captured by the same brush stamp, no matter where they lie in
// the UV layout (identical behavior to the CPU path).
Shader "Hidden/MeshPaintBrush"
{
    Properties
    {
        _BrushColor ("Brush Color", Color) = (1,0,0,1)
        _MaskTex ("Brush Mask", 2D) = "white" {}
        _ColorWriteMask ("Color Write Mask", Float) = 15
    }
    SubShader
    {
        Tags { "RenderType" = "Opaque" }
        Pass
        {
            ZTest Always
            ZWrite Off
            Cull Off
            // Per-channel write mask: lets the shared mask map (R=metallic, G=smoothness,
            // B=occlusion) be painted by metallic/smoothness and occlusion independently.
            // ColorWriteMask bits: R=8, G=4, B=2, A=1 (15 = all). Alpha (coverage) is always
            // written so the dilation pass still sees the freshly painted texels.
            ColorMask [_ColorWriteMask]
            // Source-over for color; alpha accumulates coverage toward 1 (painted marker
            // for the dilation pass). Soft edges keep partial alpha.
            Blend SrcAlpha OneMinusSrcAlpha, One OneMinusSrcAlpha

            CGPROGRAM
            #pragma vertex vert
            #pragma fragment frag
            #include "UnityCG.cginc"

            float4 _BrushColor;
            float3 _BrushCenter;   // brush center in local mesh space
            float3 _BrushU;        // stamp right axis (unit, local space)
            float3 _BrushV;        // stamp up axis (unit, local space)
            float3 _BrushNormal;   // surface normal at the hit (unit, local space)
            float  _BrushRadius;   // stamp half-extent in local mesh space
            sampler2D _MaskTex;

            struct appdata
            {
                float4 vertex : POSITION;   // local 3D position
                float2 uv     : TEXCOORD0;  // UV (target in the texture layout)
            };

            struct v2f
            {
                float4 pos   : SV_POSITION;
                float3 local : TEXCOORD0;   // interpolated local 3D position
            };

            v2f vert (appdata v)
            {
                v2f o;
                // UV (0..1) -> clip space (-1..1). This makes the GPU rasterize
                // each triangle exactly at its place in the texture.
                float2 p = v.uv * 2.0 - 1.0;
                #if UNITY_UV_STARTS_AT_TOP
                    p.y = -p.y; // compensate the RenderTexture Y-flip (D3D etc.)
                #endif
                o.pos = float4(p, 0.0, 1.0);
                o.local = v.vertex.xyz;
                return o;
            }

            fixed4 frag (v2f i) : SV_Target
            {
                // Offset from the brush center, projected onto the stamp frame.
                float3 d = i.local - _BrushCenter;
                float du = dot(d, _BrushU);
                float dv = dot(d, _BrushV);
                float dn = dot(d, _BrushNormal);

                // Square stamp in the tangent plane + depth bound (keeps the seam-
                // bridging 3D behavior while preventing bleed to far surfaces).
                clip(_BrushRadius - abs(du));
                clip(_BrushRadius - abs(dv));
                clip(_BrushRadius - abs(dn));

                // Mask -> opacity. du/dv in [-r,r] -> uv in [0,1].
                float2 muv = float2(du, dv) / _BrushRadius * 0.5 + 0.5;
                fixed4 m = tex2D(_MaskTex, muv);
                float luma = dot(m.rgb, fixed3(0.299, 0.587, 0.114));
                float opacity = luma * m.a * _BrushColor.a;
                clip(opacity - 1e-4); // skip fully transparent texels

                return fixed4(_BrushColor.rgb, opacity);
            }
            ENDCG
        }

        // Pass 1 - coverage: rasterizes the mesh in UV space (identical vertex transform to
        // the brush) and writes alpha = 1 on exactly the texels the brush can paint. Used to
        // seed the dilation guard (display RT alpha) so it matches the GPU rasterizer instead
        // of a CPU point-in-triangle approximation. Texels the brush can never reach (sub-pixel
        // island borders) stay alpha = 0 -> the gutter, so dilation can still overpaint them.
        Pass
        {
            ZTest Always
            ZWrite Off
            Cull Off
            Blend Off
            ColorMask A

            CGPROGRAM
            #pragma vertex vert
            #pragma fragment frag
            #include "UnityCG.cginc"

            struct appdata { float4 vertex : POSITION; float2 uv : TEXCOORD0; };
            struct v2f { float4 pos : SV_POSITION; };

            v2f vert (appdata v)
            {
                v2f o;
                float2 p = v.uv * 2.0 - 1.0;
                #if UNITY_UV_STARTS_AT_TOP
                    p.y = -p.y;
                #endif
                o.pos = float4(p, 0.0, 1.0);
                return o;
            }

            fixed4 frag (v2f i) : SV_Target { return fixed4(0, 0, 0, 1); }
            ENDCG
        }
    }
    Fallback Off
}