// Shared glyph/path coverage; consumers supply textures, metrics and styling.
use mod.pod.*
use mod.math.*
use mod.shader.*
use mod.glyph_coverage

// Texture-slot adapters live in the consumer; texel addressing and packed
// index decoding are shared by glyph scans and path distance queries.
mod.glyph_coverage.CurveTexels = {
    fetch_curve_texel: fn(texel_idx: float) -> vec4 {
        let tex_size = self.curve_tex_size()
        let row = floor(texel_idx / tex_size.x)
        let col = texel_idx - row * tex_size.x
        let uv = vec2(
            (col + 0.5) / tex_size.x,
            (row + 0.5) / tex_size.y
        )
        return self.sample_curve_tex(uv)
    }

    fetch_band_texel: fn(texel_idx: float) -> vec4 {
        let tex_size = self.band_tex_size()
        let row = floor(texel_idx / tex_size.x)
        let col = texel_idx - row * tex_size.x
        let uv = vec2(
            (col + 0.5) / tex_size.x,
            (row + 0.5) / tex_size.y
        )
        return self.sample_band_tex(uv)
    }

    pick_channel: fn(v: vec4, channel: float) -> float {
        if channel < 0.5 {
            return v.x
        }
        if channel < 1.5 {
            return v.y
        }
        if channel < 2.5 {
            return v.z
        }
        return v.w
    }
}

mod.glyph_coverage.SlugCoverage = {
    ..mod.glyph_coverage.CurveTexels

    // ---- slug curve coverage (the outline's quadratic curves, scanned per pixel). The shader
    // declares `curve_texture` and `band_texture`, and the accessors `slug_curve_offset()`,
    // `slug_curve_count()`, `slug_band_offset()`, `slug_band_count()`, `slug_fill_flags()` (where in
    // the curve texture its glyph is), `aa_pad_px` and `slug_viewport_px`.

    saturate: fn(v: float) -> float {
        return clamp(v, 0.0, 1.0)
    }

    slug_dilate: fn(pos: vec2, tex: vec2, jac: vec4, normal: vec2) -> vec4 {
        let n = normalize(normal)
        // The camera and list transform the GPU renders with: baked in when
        // the text was drawn they were the previous paint's (identity for a
        // new pass), and the first draw came out different from later ones.
        let m = self.draw_pass.camera_projection * (self.draw_pass.camera_view * self.draw_list.view_transform)
        let hp = m * vec4(pos.x, pos.y, 0.0, 1.0)
        let hn = m * vec4(n.x, n.y, 0.0, 0.0)
        let s = hp.w
        let t = hn.w

        let u = (s * hn.x - t * hp.x) * self.slug_viewport_px.x
        let v = (s * hn.y - t * hp.y) * self.slug_viewport_px.y

        let s2 = s * s
        let st = s * t
        let uv = u * u + v * v
        let d = normal * (
            s2 * (st + sqrt(uv)) / max(uv - st * st, 0.0000001)
        ) * self.aa_pad_px

        let vpos = pos + d
        let vtex = vec2(tex.x + dot(d, jac.xy), tex.y + dot(d, jac.zw))
        return vec4(vtex.x, vtex.y, vpos.x, vpos.y)
    }

    curve_tex_size: fn() -> vec2 { return self.curve_texture.size() }
    band_tex_size: fn() -> vec2 { return self.band_texture.size() }
    sample_curve_tex: fn(uv: vec2) -> vec4 { return self.curve_texture.sample_nearest(uv) }
    sample_band_tex: fn(uv: vec2) -> vec4 { return self.band_texture.sample_nearest(uv) }

    calc_root_code: fn(y1: float, y2: float, y3: float) -> u32 {
        let i1 = asuint(y1) >> u32(31)
        let i2 = asuint(y2) >> u32(30)
        let i3 = asuint(y3) >> u32(29)

        let shift = (i1 & u32(1)) | (i2 & u32(2)) | (i3 & u32(4))
        return (u32(11892) >> shift) & u32(257)
    }

    solve_horiz_poly: fn(p12: vec4, p3: vec2) -> vec2 {
        let a = p12.xy - p12.zw * 2.0 + p3
        let b = p12.xy - p12.zw
        let ra = 1.0 / a.y
        let rb = 0.5 / b.y

        let d = sqrt(max(b.y * b.y - a.y * p12.y, 0.0))
        let mut t1 = (b.y - d) * ra
        let mut t2 = (b.y + d) * ra
        if abs(a.y) < 1.0 / 65536.0 {
            t1 = p12.y * rb
            t2 = t1
        }
        return vec2(
            (a.x * t1 - b.x * 2.0) * t1 + p12.x,
            (a.x * t2 - b.x * 2.0) * t2 + p12.x
        )
    }

    solve_vert_poly: fn(p12: vec4, p3: vec2) -> vec2 {
        let a = p12.xy - p12.zw * 2.0 + p3
        let b = p12.xy - p12.zw
        let ra = 1.0 / a.x
        let rb = 0.5 / b.x

        let d = sqrt(max(b.x * b.x - a.x * p12.x, 0.0))
        let mut t1 = (b.x - d) * ra
        let mut t2 = (b.x + d) * ra
        if abs(a.x) < 1.0 / 65536.0 {
            t1 = p12.x * rb
            t2 = t1
        }
        return vec2(
            (a.y * t1 - b.y * 2.0) * t1 + p12.y,
            (a.y * t2 - b.y * 2.0) * t2 + p12.y
        )
    }

    scan_horizontal_list: fn(list_offset: float, list_count: float, sample: vec2, px_size: float) -> vec2 {
        let limit = floor(list_count + 0.5)
        var coverage = 0.0
        var weight = 0.0

        var j = 0.0
        loop {
            if j >= limit { break }

            let packed_idx = floor(j * 0.25)
            let channel = j - packed_idx * 4.0
            let idx_data = self.fetch_band_texel(list_offset + packed_idx)
            let curve_idx = self.pick_channel(idx_data, channel)

            let p12 = self.fetch_curve_texel(curve_idx * 2.0) - vec4(sample.x, sample.y, sample.x, sample.y)
            let p3 = self.fetch_curve_texel(curve_idx * 2.0 + 1.0).xy - sample
            if max(max(p12.x, p12.z), p3.x) / px_size < -0.5 { break }

            let code = self.calc_root_code(p12.y, p12.w, p3.y)
            if code != u32(0) {
                let r = self.solve_horiz_poly(p12, p3) / px_size
                if (code & u32(1)) != u32(0) {
                    coverage = coverage + self.saturate(r.x + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.x) * 2.0))
                }
                if code > u32(1) {
                    coverage = coverage - self.saturate(r.y + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.y) * 2.0))
                }
            }

            j = j + 1.0
        }

        return vec2(coverage, weight)
    }

    scan_vertical_list: fn(list_offset: float, list_count: float, sample: vec2, px_size: float) -> vec2 {
        let limit = floor(list_count + 0.5)
        var coverage = 0.0
        var weight = 0.0

        var j = 0.0
        loop {
            if j >= limit { break }

            let packed_idx = floor(j * 0.25)
            let channel = j - packed_idx * 4.0
            let idx_data = self.fetch_band_texel(list_offset + packed_idx)
            let curve_idx = self.pick_channel(idx_data, channel)

            let p12 = self.fetch_curve_texel(curve_idx * 2.0) - vec4(sample.x, sample.y, sample.x, sample.y)
            let p3 = self.fetch_curve_texel(curve_idx * 2.0 + 1.0).xy - sample
            if max(max(p12.y, p12.w), p3.y) / px_size < -0.5 { break }

            let code = self.calc_root_code(p12.x, p12.z, p3.x)
            if code != u32(0) {
                let r = self.solve_vert_poly(p12, p3) / px_size
                if (code & u32(1)) != u32(0) {
                    coverage = coverage - self.saturate(r.x + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.x) * 2.0))
                }
                if code > u32(1) {
                    coverage = coverage + self.saturate(r.y + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.y) * 2.0))
                }
            }

            j = j + 1.0
        }

        return vec2(coverage, weight)
    }

    scan_horizontal_all: fn(sample: vec2, px_size: float) -> vec2 {
        let limit = floor(self.slug_curve_count() + 0.5)
        var coverage = 0.0
        var weight = 0.0

        var i = 0.0
        loop {
            if i >= limit { break }

            let curve_idx = self.slug_curve_offset() + i
            let p12 = self.fetch_curve_texel(curve_idx * 2.0) - vec4(sample.x, sample.y, sample.x, sample.y)
            let p3 = self.fetch_curve_texel(curve_idx * 2.0 + 1.0).xy - sample
            let code = self.calc_root_code(p12.y, p12.w, p3.y)
            if code != u32(0) {
                let r = self.solve_horiz_poly(p12, p3) / px_size
                if (code & u32(1)) != u32(0) {
                    coverage = coverage + self.saturate(r.x + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.x) * 2.0))
                }
                if code > u32(1) {
                    coverage = coverage - self.saturate(r.y + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.y) * 2.0))
                }
            }

            i = i + 1.0
        }

        return vec2(coverage, weight)
    }

    scan_vertical_all: fn(sample: vec2, px_size: float) -> vec2 {
        let limit = floor(self.slug_curve_count() + 0.5)
        var coverage = 0.0
        var weight = 0.0

        var i = 0.0
        loop {
            if i >= limit { break }

            let curve_idx = self.slug_curve_offset() + i
            let p12 = self.fetch_curve_texel(curve_idx * 2.0) - vec4(sample.x, sample.y, sample.x, sample.y)
            let p3 = self.fetch_curve_texel(curve_idx * 2.0 + 1.0).xy - sample
            let code = self.calc_root_code(p12.x, p12.z, p3.x)
            if code != u32(0) {
                let r = self.solve_vert_poly(p12, p3) / px_size
                if (code & u32(1)) != u32(0) {
                    coverage = coverage - self.saturate(r.x + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.x) * 2.0))
                }
                if code > u32(1) {
                    coverage = coverage + self.saturate(r.y + 0.5)
                    weight = max(weight, self.saturate(1.0 - abs(r.y) * 2.0))
                }
            }

            i = i + 1.0
        }

        return vec2(coverage, weight)
    }

    calc_coverage: fn(xcov: float, ycov: float, xwgt: float, ywgt: float) -> float {
        let coverage = max(
            abs(xcov * xwgt + ycov * ywgt) / max(xwgt + ywgt, 1.0 / 65536.0),
            min(abs(xcov), abs(ycov))
        )
        if self.slug_fill_flags() >= 4096.0 {
            return 1.0 - abs(1.0 - fract(coverage * 0.5) * 2.0)
        }
        return self.saturate(coverage)
    }

    // Text takes the full-curve scan: the band-accelerated path has shown correctness issues in
    // some GPU shader compilers, and text glyphs carry no bands. One walk over the curves serves
    // both rays (each curve's two texels are fetched once, not once per direction): the same sums
    // as scan_horizontal_all + scan_vertical_all, at half the texture reads.
    alpha_at_scan: fn(sample: vec2, px_x: float, px_y: float) -> float {
        let limit = floor(self.slug_curve_count() + 0.5)
        let first = self.slug_curve_offset() * 2.0
        // a curve's two texels sit side by side in one row (the row width is even)
        let tex_size = self.curve_tex_size()
        let texel = vec2(1.0 / tex_size.x, 1.0 / tex_size.y)
        let inv_px = vec2(1.0 / px_x, 1.0 / px_y)
        var xcov = 0.0
        var xwgt = 0.0
        var ycov = 0.0
        var ywgt = 0.0

        var i = 0.0
        loop {
            if i >= limit { break }

            let at = first + i * 2.0
            let row = floor(at * texel.x)
            let uv = vec2((at - row * tex_size.x + 0.5) * texel.x, (row + 0.5) * texel.y)
            let p12 = self.sample_curve_tex(uv) - vec4(sample.x, sample.y, sample.x, sample.y)
            let p3 = self.sample_curve_tex(vec2(uv.x + texel.x, uv.y)).xy - sample

            let hcode = self.calc_root_code(p12.y, p12.w, p3.y)
            if hcode != u32(0) {
                let r = self.solve_horiz_poly(p12, p3) * inv_px.x
                if (hcode & u32(1)) != u32(0) {
                    xcov = xcov + self.saturate(r.x + 0.5)
                    xwgt = max(xwgt, self.saturate(1.0 - abs(r.x) * 2.0))
                }
                if hcode > u32(1) {
                    xcov = xcov - self.saturate(r.y + 0.5)
                    xwgt = max(xwgt, self.saturate(1.0 - abs(r.y) * 2.0))
                }
            }

            let vcode = self.calc_root_code(p12.x, p12.z, p3.x)
            if vcode != u32(0) {
                let r = self.solve_vert_poly(p12, p3) * inv_px.y
                if (vcode & u32(1)) != u32(0) {
                    ycov = ycov - self.saturate(r.x + 0.5)
                    ywgt = max(ywgt, self.saturate(1.0 - abs(r.x) * 2.0))
                }
                if vcode > u32(1) {
                    ycov = ycov + self.saturate(r.y + 0.5)
                    ywgt = max(ywgt, self.saturate(1.0 - abs(r.y) * 2.0))
                }
            }

            i = i + 1.0
        }

        return self.calc_coverage(xcov, ycov, xwgt, ywgt)
    }

    // A glyph with bands scans only its band, one without scans every curve.
    alpha_at_banded: fn(sample: vec2, px_x: float, px_y: float) -> float {
        var coverage_x = 0.0
        var coverage_y = 0.0
        var weight_x = 0.0
        var weight_y = 0.0

        if self.slug_band_count() > 0.5 {
            let num_bands = max(floor(self.slug_band_count() + 0.5), 1.0)
            let h_band_idx = clamp(floor(sample.y * num_bands), 0.0, num_bands - 1.0)
            let v_band_idx = clamp(floor(sample.x * num_bands), 0.0, num_bands - 1.0)

            let h_band_info = self.fetch_band_texel(self.slug_band_offset() + h_band_idx)
            let h_band = self.scan_horizontal_list(
                floor(h_band_info.x + 0.5),
                h_band_info.y,
                sample,
                px_x,
            )
            coverage_x = h_band.x
            weight_x = h_band.y

            let v_band_info = self.fetch_band_texel(self.slug_band_offset() + num_bands + v_band_idx)
            let v_band = self.scan_vertical_list(
                floor(v_band_info.x + 0.5),
                v_band_info.y,
                sample,
                px_y,
            )
            coverage_y = v_band.x
            weight_y = v_band.y
        } else {
            let x_scan = self.scan_horizontal_all(sample, px_x)
            coverage_x = x_scan.x
            weight_x = x_scan.y

            let y_scan = self.scan_vertical_all(sample, px_y)
            coverage_y = y_scan.x
            weight_y = y_scan.y
        }

        return self.calc_coverage(coverage_x, coverage_y, weight_x, weight_y)
    }
}

mod.glyph_coverage.FeatherCover = {
    // Reciprocal-square-root approximation of the error function.
    erf7: fn(x: float) -> float {
        let s = x * 1.1283791671
        let ss = s * s
        let t = s + (0.24295 + (0.03395 + 0.0104 * ss) * ss) * (s * ss)
        return t / sqrt(1.0 + t * t)
    }

    // sd is positive inside, in device pixels; sigma is Gaussian sigma in
    // device pixels. Zero sigma gives the ordinary one-pixel AA ramp.
    feather_cover: fn(sd: float, sigma: float, side: float) -> float {
        if sigma <= 0.01 { return clamp(0.5 + sd, 0.0, 1.0) }
        // Saturating the far tail avoids overflow in erf7 for distant/empty
        // bands; at six erf units the Gaussian tail is already negligible.
        let g = 0.5 + 0.5 * self.erf7(clamp(sd * (sqrt(0.5) / sigma), -6.0, 6.0))
        // Both: Gaussian integral, 50% coverage on the path.
        // Inner: nothing outside, zero on the path, rising inward.
        // Outer: full coverage inside/on the path, fading outward.
        if side < 0.0 { return max(2.0 * g - 1.0, 0.0) }
        if side > 0.0 { return min(2.0 * g, 1.0) }
        return g
    }
}

mod.glyph_coverage.Feather = {
    ..mod.glyph_coverage.FeatherCover,
    ..mod.glyph_coverage.CurveTexels

    curve_tex_size: fn() -> vec2 { return self.feather_curve_texture.size() }
    band_tex_size: fn() -> vec2 { return self.feather_band_texture.size() }
    sample_curve_tex: fn(uv: vec2) -> vec4 { return self.feather_curve_texture.sample_nearest(uv) }
    sample_band_tex: fn(uv: vec2) -> vec4 { return self.feather_band_texture.sample_nearest(uv) }

    // Three shape-header texels (see draw/vector/feather.rs):
    // 0 = curve offset/count, band offset/count;
    // 1 = expanded hull/band bbox min.xy/max.xy, in path coordinates;
    // 2 = sigma, side, nonzero flag, stroke half width (-1 means fill).
    // The horizontal band's list contains every segment within the feather
    // reach, including horizontal segments. Walk it once for distance and
    // +x-ray winding. Return (signed distance in path units, winding).
    // Each line's second curve texel is (p2.xy, winding_only, 0): .z = 1
    // means both sides have the same fill state, so keep its winding but
    // skip its distance. Boundaries and all stroke segments have .z = 0.
    band_signed_distance: fn(sample: vec2, header: float) -> vec2 {
        let shape = self.fetch_band_texel(header)
        let bounds = self.fetch_band_texel(header + 1.0)
        let style = self.fetch_band_texel(header + 2.0)
        let count = max(floor(shape.w + 0.5), 1.0)
        let band = clamp(floor((sample.y - bounds.y) / max(bounds.w - bounds.y, 0.000001) * count), 0.0, count - 1.0)
        let list = self.fetch_band_texel(shape.z + band)
        var dist_squared = 1e30
        var winding = 0.0
        var j = 0.0
        loop {
            if j >= list.y { break }
            let packed = floor(j * 0.25)
            let indices = self.fetch_band_texel(list.x + packed)
            let curve = self.pick_channel(indices, j - packed * 4.0)
            let p0 = self.fetch_curve_texel(curve * 2.0).xy
            let endpoint = self.fetch_curve_texel(curve * 2.0 + 1.0)
            let p2 = endpoint.xy
            let delta = p2 - p0
            if endpoint.z < 0.5 {
                let relative = sample - p0
                let t = clamp(dot(relative, delta) / max(dot(delta, delta), 0.000000000001), 0.0, 1.0)
                let nearest = relative - delta * t
                dist_squared = min(dist_squared, dot(nearest, nearest))
            }
            // Half-open crossings count a shared endpoint once. Horizontal
            // boundaries contribute distance, but never winding.
            if p0.y <= sample.y && p2.y > sample.y {
                if p0.x + (sample.y - p0.y) * delta.x / delta.y > sample.x {
                    winding = winding + 1.0
                }
            } else if p2.y <= sample.y && p0.y > sample.y {
                if p0.x + (sample.y - p0.y) * delta.x / delta.y > sample.x {
                    winding = winding - 1.0
                }
            }
            j = j + 1.0
        }
        let dist = sqrt(dist_squared)
        if style.w >= 0.0 { return vec2(style.w - dist, winding) }
        var inside = abs(winding) > 0.5
        if style.z < 0.5 { inside = modf(abs(winding), 2.0) > 0.5 }
        if inside { return vec2(dist, winding) }
        return vec2(-dist, winding)
    }
}

mod.glyph_coverage.AtlasCoverage = {
    // ---- distance-field coverage from the glyph atlas. The shader declares `grayscale_texture` and
    // `msdf_texture`, `atlas_plane` (which channel holds the field), `cutoff`, `radius`,
    // `sdf_sharpness` and `sdf_luma_bias`.

    sdf: fn(scale, p, color) {
        let sampled = self.grayscale_texture.sample(p)
        let s = if self.atlas_plane < 0.5 {
            sampled.r
        } else if self.atlas_plane < 1.5 {
            sampled.g
        } else if self.atlas_plane < 2.5 {
            sampled.b
        } else {
            sampled.a
        }
        let safe_scale = max(scale, 0.0001)
        let luma = dot(color.rgb, vec3(0.299, 0.587, 0.114))
        var a = clamp(
            (s - (1.0 - self.cutoff)) * self.radius / safe_scale * self.sdf_sharpness + 0.5,
            0.0,
            1.0,
        )
        let bias = (0.5 - luma) * self.sdf_luma_bias
        // Light text lifts coverage by a constant: not where there is none, or
        // every glyph cell shows as a faint quad over the scene.
        if a > self.sdf_luma_bias * 0.5 {
            a = clamp(a - bias, 0.0, 1.0)
        }
        return a
    }

    msdf: fn(scale, p, color) {
        let s = self.msdf_texture.sample(p)
        let dist = s.a
        let safe_scale = max(scale, 0.0001)
        let luma = dot(color.rgb, vec3(0.299, 0.587, 0.114))
        var a = clamp(
            (dist - (1.0 - self.cutoff)) * self.radius / safe_scale * self.sdf_sharpness + 0.5,
            0.0,
            1.0,
        )
        let bias = (0.5 - luma) * self.sdf_luma_bias
        if a > self.sdf_luma_bias * 0.5 {
            a = clamp(a - bias, 0.0, 1.0)
        }
        return a
    }
}