// Native macOS panel and scene postprocess. Panel kernels consume the core's
// plain tracks; the final full-window kernel is the CRT. Panel easing/noise
// stays in step with panel_animation.zig.
//
// The three entry points are `extern "C" [[stitchable]]`. Both halves are
// load-bearing: this source is compiled at runtime by
// CIKernel.kernels(withMetalString:), which looks for stitchable functions and
// fails the whole compile with "cannot find a valid stitchable Metal function
// in the source" if it finds none — leaving the view with no postprocessor and
// every effect silently off.
#include <metal_stdlib>
#include <CoreImage/CoreImage.h>
using namespace metal;

constant float PARDES_TAU = 6.28318530718;

float pardesNoise(float2 p) {
    float3 q = fract(float3(p.x, p.y, p.x) * 0.1031);
    q += dot(q, q.yzx + 33.33);
    return fract((q.x + q.y) * q.z);
}

// Exact integer hash from panel_animation.cellNoise. Pane serial and logical
// cell coordinates, not destination pixels, keep TTY and GPU reveal decisions
// stable across scale factors and geometry transforms.
float pardesCellNoise(float2 serialParts, float colValue, float rowValue) {
    uint serial = uint(max(serialParts.x, 0.0)) |
        (uint(max(serialParts.y, 0.0)) << 16);
    uint col = uint(max(colValue, 0.0));
    uint row = uint(max(rowValue, 0.0));
    uint x = serial ^ (col * 0x9e3779b9u) ^ (row * 0x85ebca6bu);
    x ^= x >> 16;
    x *= 0x7feb352du;
    x ^= x >> 15;
    x *= 0x846ca68bu;
    x ^= x >> 16;
    return float(x & 0xffffu) / 65535.0;
}

float4 pardesSceneSample(
    coreimage::sampler source,
    float2 origin,
    float2 size,
    float2 uv
) {
    return coreimage::unpremultiply(
        source.sample(source.transform(origin + uv * size)));
}

float3 pardesSceneLinear(
    coreimage::sampler source,
    float2 origin,
    float2 size,
    float2 uv
) {
    return coreimage::srgb_to_linear(pardesSceneSample(source, origin, size, uv).rgb);
}

bool pardesInside(float2 p, float4 box) {
    return p.x >= box.x && p.y >= box.y &&
        p.x < box.x + box.z && p.y < box.y + box.w;
}

// Character effects (ASCII byte walks, glyph motion, per-cell locks) are
// composed in Pardes core, so this kernel never eases them: it clips their
// panel batch and draws the finished glyphs.
bool pardesComposedInCore(float effect) {
    return effect == 4.0 || effect >= 6.0;
}

float pardesPanelProgress(float effect, float frame) {
    float frames = effect == 1.0 ? 12.0 :
        effect == 2.0 ? 14.0 :
        effect == 3.0 ? 10.0 :
        effect == 5.0 ? 12.0 : 1.0;
    if (frames <= 1.0 || frame >= frames - 1.0) return 1.0;
    float t = clamp(frame / (frames - 1.0), 0.0, 1.0);
    if (effect == 1.0 || effect == 5.0) { // horizontal/vertical slide
        float u = 1.0 - t;
        return 1.0 - u * u * u;
    }
    if (effect == 2.0) { // zoom: ease-out back, including its overshoot
        const float c1 = 1.70158;
        const float c3 = c1 + 1.0;
        float u = t - 1.0;
        return 1.0 + c3 * u * u * u + c1 * u * u;
    }
    if (effect == 3.0) return t * t * (3.0 - 2.0 * t); // dissolve
    return t; // core-composed cells need no progress; retained for ABI completeness
}

// Prepare every geometric target before any translated panel is drawn. A
// vertical opener restores the old frame in its fixed clip; a closing pane is
// already absent from `canvas` and therefore needs no clear at all.
extern "C" [[stitchable]] float4 pardesPanelClear(
    coreimage::sampler canvas,
    coreimage::sampler previous,
    float4 target,
    float4 ground,
    float effect,
    float phase,
    coreimage::destination destination
) {
    float2 coord = destination.coord();
    if (!pardesInside(coord, target))
        return canvas.sample(canvas.transform(coord));
    if (effect == 1.0 || effect == 2.0) return ground;
    if (effect == 5.0 && phase == 0.0)
        return previous.sample(previous.transform(coord));
    return canvas.sample(canvas.transform(coord));
}

// One panel draw over the accumulated image. Current and previous are complete
// CoreText+attachment rasters. That makes a closing pane safe after its core
// allocation is gone and makes ASCII/dissolve true old-to-new data effects.
extern "C" [[stitchable]] float4 pardesPanel(
    coreimage::sampler current,
    coreimage::sampler previous,
    coreimage::sampler changedMask,
    coreimage::sampler canvas,
    float4 target,
    float4 from,
    float4 to,
    float effect,
    float phase,
    float frame,
    float2 serialParts,
    float2 cellSize,
    float2 gridOrigin,
    coreimage::destination destination
) {
    float2 coord = destination.coord();
    float4 underneath = canvas.sample(canvas.transform(coord));
    float progress = pardesPanelProgress(effect, frame);

    float2 safeCell = max(cellSize, float2(1.0));
    // ASCII bytes were already composed in Pardes core before CoreText built
    // `current`. Pixel attachments have no character value and intentionally
    // pass through unchanged, so this kernel only clips the pane batch.
    if (pardesComposedInCore(effect)) {
        if (!pardesInside(coord, target)) return underneath;
        return current.sample(current.transform(coord));
    }

    // Dissolve stays at canonical geometry. Unchanged cells bypass the effect
    // completely; changed cells retain the exact old raster until their
    // deterministic threshold is crossed.
    if (effect == 3.0) {
        if (!pardesInside(coord, target)) return underneath;
        // During a live resize the grid is still anchored at the view's top,
        // leaving a partial strip below its final logical row. Account for
        // that strip before indexing the compact cell mask.
        float2 absoluteCell = floor((coord - gridOrigin) / safeCell);
        float changed = changedMask.sample(changedMask.transform(
            absoluteCell + float2(0.5))).r;
        if (changed < 0.5) return underneath;

        float4 oldPixel = previous.sample(previous.transform(coord));
        float4 newPixel = current.sample(current.transform(coord));
        // Exact endpoint guards are not redundant: a noise hash can be zero
        // or one, and frame zero/last must still be wholly old/new.
        if (progress <= 0.0) return oldPixel;
        if (progress >= 1.0) return newPixel;

        // Core/TTY coordinates are relative to the target's TOP-left. Core
        // Image is bottom-up, hence the reversed row term.
        float2 cell = float2(
            floor((coord.x - target.x) / safeCell.x),
            floor((target.y + target.w - coord.y) / safeCell.y));
        float2 cellCount = max(round(target.zw / safeCell), float2(1.0));
        cell = clamp(cell, float2(0.0), cellCount - 1.0);
        float noise = pardesCellNoise(serialParts, cell.x, cell.y);
        return noise < progress ? newPixel : oldPixel;
    }

    float4 shown = mix(from, to, progress);
    shown.zw = max(shown.zw, float2(0.0));
    if (!pardesInside(coord, shown) || shown.z <= 0.0 || shown.w <= 0.0)
        return underneath;
    // Vertical lifecycle motion is deliberately confined to the pane's own
    // old/new box. It never travels across or clips a surviving neighbour.
    if (effect == 5.0 && !pardesInside(coord, target)) return underneath;

    float2 within = (coord - shown.xy) / shown.zw;
    float2 sourceCoord = target.xy + within * target.zw;
    if (!pardesInside(sourceCoord, target)) return underneath;
    return phase == 2.0
        ? previous.sample(previous.transform(sourceCoord))
        : current.sample(current.transform(sourceCoord));
}

extern "C" [[stitchable]] float4 pardesScene(
    coreimage::sampler source,
    float timeSeconds,
    float frame,
    float crtOn,
    float2 sceneSize,
    float2 sceneOrigin,
    coreimage::destination destination
) {
    const float barrel = 0.016;
    const float edgeSoftnessPx = 3.5;
    const float chromaPx = 0.28;
    const float bloomRadiusPx = 1.75;
    const float bloomThreshold = 0.60;
    const float bloomSoftness = 0.25;
    const float bloomStrength = 0.045;
    const float scanPeriodPx = 3.0;
    const float scanStrength = 0.055;
    const float maskPeriodPx = 3.0;
    const float maskStrength = 0.015;
    const float vignetteStrength = 0.11;
    const float humStrength = 0.0025;
    const float noiseStrength = 0.0020;
    const float exposure = 1.012;

    float2 size = sceneSize;
    float2 origin = sceneOrigin;
    float2 pixel = destination.coord() - origin;
    float2 texel = 1.0 / size;
    float2 cc = pixel / size * 2.0 - 1.0;
    float r2 = dot(cc, cc);
    float2 uv = cc * (1.0 + barrel * r2 * crtOn) * 0.5 + 0.5;
    float edge = min(min(uv.x, uv.y), min(1.0 - uv.x, 1.0 - uv.y));
    // Match GLSL's opaque black outside the tube. Transparent black would
    // reveal AppKit's backdrop and make the same CRT effect have no tube on a
    // transparent theme.
    if (crtOn > 0.5 && edge <= 0.0) return float4(0.0, 0.0, 0.0, 1.0);

    float4 center = pardesSceneSample(source, origin, size, uv);
    float3 scene = center.rgb;

    // Without the CRT the scene passes through, exact.
    if (crtOn <= 0.5) {
        return coreimage::premultiply(float4(clamp(scene, 0.0, 1.0), center.a));
    }

    float3 centerLinear = coreimage::srgb_to_linear(scene);
    float2 radialChroma = cc * (chromaPx * r2) * texel;
    float3 redSample = pardesSceneLinear(source, origin, size, uv + radialChroma);
    float3 blueSample = pardesSceneLinear(source, origin, size, uv - radialChroma);
    float3 col = float3(
        redSample.r,
        centerLinear.g,
        blueSample.b
    );

    float2 bloom = texel * bloomRadiusPx;
    float3 bloomSum = float3(0.0);
    float2 bloomOffsets[4] = {
        float2(bloom.x, 0.0), float2(-bloom.x, 0.0),
        float2(0.0, bloom.y), float2(0.0, -bloom.y)
    };
    for (uint i = 0; i < 4; i++) {
        // Deliberately raw, like GLSL brightPass.
        float4 sampleValue = pardesSceneSample(
            source, origin, size, uv + bloomOffsets[i]);
        float3 linearValue = coreimage::srgb_to_linear(sampleValue.rgb);
        float peak = max(linearValue.r, max(linearValue.g, linearValue.b));
        bloomSum += linearValue * smoothstep(bloomThreshold, bloomThreshold + bloomSoftness, peak);
    }
    col += bloomSum * (0.25 * bloomStrength);

    float scanWave = 0.5 + 0.5 * cos(pixel.y * PARDES_TAU / scanPeriodPx);
    col *= 1.0 - scanStrength * scanWave * scanWave;
    float3 mask = 1.0 + maskStrength * cos(
        pixel.x * PARDES_TAU / maskPeriodPx +
        float3(0.0, PARDES_TAU / 3.0, 2.0 * PARDES_TAU / 3.0));
    col *= mask;

    float2 edgePx = min(uv, 1.0 - uv) * size;
    float tube = smoothstep(0.0, edgeSoftnessPx, min(edgePx.x, edgePx.y));
    float vignette = 1.0 - vignetteStrength * pow(clamp(r2 * 0.5, 0.0, 1.0), 1.6);
    float hum = 1.0 + humStrength * sin((uv.y * 1.35 - timeSeconds * 0.11) * PARDES_TAU);
    col *= tube * vignette * hum * exposure;

    float luma = dot(col, float3(0.2126, 0.7152, 0.0722));
    float2 noiseSeed = pixel + float2(frame * 17.0, frame * 59.0);
    col += (pardesNoise(noiseSeed) - 0.5) * noiseStrength *
        (0.15 + 0.85 * clamp(luma, 0.0, 1.0));
    return coreimage::premultiply(float4(
        coreimage::linear_to_srgb(clamp(col, 0.0, 1.0)), center.a));
}