// DVD-style bouncing logo (image texture, e.g. assets/mxl.png) on studio black,
// Rec.709, 10-bit Y'CbCr, packed as v210 (4:2:2).
// One work-item per 16-byte block = 6 pixels (6 Y + 3 Cb + 3 Cr).
// Stateless bounce: logo position is a pure function of the frame tick
// (two triangle waves), so no velocity/position state is carried between frames.
//
// Sub-pixel motion: the logo plane is bilinearly resampled at the fractional
// position, so edges glide continuously instead of stepping whole pixels.
// The blend is built from mix() chains, which reproduce texel values exactly
// in solid regions (mix of equal values is exact in f32).
//
// Logo plane (binding 2, read-only), one u32 per logo pixel:
//   logo[0] = logo_w | logo_h << 16
//   logo[1 + dy*logo_w + dx] = opaque<<30 | Cr<<20 | Cb<<10 | Y

struct Params {
    width: u32,
    height: u32,
    frame: u32, // animation tick (0,1,2,... per generated frame), NOT the raw grain index
    _pad0: u32,
};

@group(0) @binding(0) var<storage, read_write> out: array<u32>;
@group(0) @binding(1) var<uniform> params: Params;
@group(0) @binding(2) var<storage, read> logo: array<u32>;

// 0 -> 1 -> 0 triangle wave; input is elapsed bounce cycles.
fn tri01(x: f32) -> f32 {
    let f = fract(x);
    return select(2.0 * f, 2.0 - 2.0 * f, f > 0.5);
}

// Deviation of a logo texel from the background (64/512/512); zero when the
// texel is transparent or out of the plane. u32 wraparound in the callers
// makes every lower/upper bound case fail the range test, so out-of-bounds
// needs no special handling.
fn tap(lw: u32, lh: u32, dx: u32, dy: u32) -> vec3<f32> {
    if (dx < lw && dy < lh) {
        let s = logo[1u + dy * lw + dx];
        if ((s >> 30u) & 1u == 1u) {
            return vec3<f32>(f32(s & 0x3FFu) - 64.0,
                             f32((s >> 10u) & 0x3FFu) - 512.0,
                             f32((s >> 20u) & 0x3FFu) - 512.0);
        }
    }
    return vec3<f32>(0.0, 0.0, 0.0);
}

@compute @workgroup_size(64)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let block = gid.x;
    let total = (params.width * params.height) / 6u;
    if (block >= total) {
        return;
    }

    let x = (block * 6u) % params.width;
    let py = (block * 6u) / params.width;

    // Logo geometry comes from the plane header.
    let hdr = logo[0];
    let lw = hdr & 0xFFFFu;
    let lh = hdr >> 16u;

    // Slow glide: ~8.6 px/frame horizontal, ~6.5 px/frame vertical at 25 fps.
    // Near-irrational speed ratio (~sqrt(2)) covers the whole frame and
    // rarely repeats the path, like the DVD logo.
    let t = f32(params.frame) / 25.0;
    let lx = tri01(t * 0.061) * max(f32(params.width) - f32(lw), 0.0);
    let ly = tri01(t * 0.088) * max(f32(params.height) - f32(lh), 0.0);
    let bx = u32(lx);      // integer part of the position
    let by = u32(ly);
    let fx = lx - f32(bx); // fractional part, [0,1)
    let fy = ly - f32(by);

    var y: array<u32, 6>;
    var cb: array<u32, 6>;
    var cr: array<u32, 6>;
    for (var i = 0u; i < 6u; i++) {
        // Screen pixel (x+i, py) covers logo-plane footprint [m-f, m+1-f) x
        // [n-f, n+1-f): texel m-1 gets weight fx, texel m gets 1-fx (same
        // vertically). Chroma is blended the same way and packed at even
        // columns, matching v210 4:2:2 co-siting.
        let m = (x + i) - bx;
        let n = py - by;
        let top = mix(tap(lw, lh, m, n - 1u), tap(lw, lh, m - 1u, n - 1u), fx);
        let bot = mix(tap(lw, lh, m, n), tap(lw, lh, m - 1u, n), fx);
        let d = mix(bot, top, fy);
        y[i]  = u32(64.0 + d.x + 0.5);
        cb[i] = u32(512.0 + d.y + 0.5);
        cr[i] = u32(512.0 + d.z + 0.5);
    }

    // v210 word layout, chroma co-sited with luma samples 0/2/4:
    // w0 = Cb0|Y0<<10|Cr0<<20; w1 = Y1|Cb2<<10|Y2<<20;
    // w2 = Cr2|Y3<<10|Cb4<<20; w3 = Y4|Cr4<<10|Y5<<20
    out[block * 4u + 0u] = (cb[0] & 0x3FFu) | ((y[0] & 0x3FFu) << 10u) | ((cr[0] & 0x3FFu) << 20u);
    out[block * 4u + 1u] = (y[1] & 0x3FFu) | ((cb[2] & 0x3FFu) << 10u) | ((y[2] & 0x3FFu) << 20u);
    out[block * 4u + 2u] = (cr[2] & 0x3FFu) | ((y[3] & 0x3FFu) << 10u) | ((cb[4] & 0x3FFu) << 20u);
    out[block * 4u + 3u] = (y[4] & 0x3FFu) | ((cr[4] & 0x3FFu) << 10u) | ((y[5] & 0x3FFu) << 20u);
}