@yali-engine/render 0.0.0-stage → 0.4.0

package/dist/post-options.d.ts

@@ -0,0 +1,109 @@
+/** Bloom (ADR-0092): HDR light spilling over its surroundings. */
+export interface BloomOptions {
+    /** How much is added (default 0.6). */
+    readonly intensity?: number;
+    /** Linear brightness where bloom starts (default 1: only highlights; 0: everything, softly). */
+    readonly threshold?: number;
+    /** The threshold's soft knee, a share of it (default 0.5). */
+    readonly knee?: number;
+    /** The upsample's spread, in texels of each level (default 1). */
+    readonly radius?: number;
+    /** Half-size levels in the chain (default 5: the widest is 1/32 of the screen). */
+    readonly levels?: number;
+}
+/** Colour grading (ADR-0092): white balance, contrast and saturation before the tone curve, lift / gamma / gain after it. */
+export interface GradingOptions {
+    /** Cooler (−1) to warmer (+1); default 0. */
+    temperature?: number;
+    /** Green (−1) to magenta (+1); default 0. */
+    tint?: number;
+    /** About mid grey in log space; default 1. */
+    contrast?: number;
+    /** 0 grey, 1 as drawn (default), above 1 more saturated. */
+    saturation?: number;
+    /** Shadows, added after the curve; default [0, 0, 0]. */
+    lift?: [number, number, number];
+    /** Mid tones; default [1, 1, 1]. */
+    gamma?: [number, number, number];
+    /** Highlights; default [1, 1, 1]. */
+    gain?: [number, number, number];
+}
+/**
+ * The post chain's passes: `postPasses` from `@yali-engine/render/post`. A separate entry, so a page
+ * that never asks for bloom, TAA or FXAA carries none of their code or shaders.
+ */
+export interface PostPasses {
+    readonly Bloom: new (device: GPUDevice, o: BloomOptions) => import('./post.ts').Bloom;
+    readonly Taa: new (device: GPUDevice, blend?: number, upscale?: boolean) => import('./post.ts').Taa;
+    readonly Fxaa: new (device: GPUDevice, format: GPUTextureFormat) => import('./post.ts').Fxaa;
+}
+/**
+ * The post-processing chain (ADR-0092). Which passes exist is fixed at init; their knobs are live.
+ *
+ * @example
+ * import { postPasses } from '@yali-engine/render/post';
+ * renderPlugin({ canvas, post: { passes: postPasses, bloom: true, taa: true, grading: { contrast: 1.1 } } });
+ */
+export interface PostOptions {
+    /** Needed for bloom, TAA and FXAA (grading is built in). */
+    readonly passes?: PostPasses;
+    /** `true` for the defaults. */
+    readonly bloom?: BloomOptions | boolean;
+    readonly grading?: GradingOptions;
+    /** Fast approximate anti-aliasing on the tone-mapped image. */
+    readonly fxaa?: boolean;
+    /** Temporal anti-aliasing: a jittered camera, accumulated over frames. */
+    /**
+     * `motion: false` reprojects by the camera only (default: per-object motion where the GPU cull's
+     * storage path draws it, ADR-0098). `sharpen` (0–1) and `motionBlur` (the shutter, 0–1 of each
+     * pixel's motion; needs per-object motion) add a pass after TAA (ADR-0119; both live in `render.post`).
+     * `upscale` (ADR-0135): temporal upscaling, the scene drawn at `viewport.renderScale` of the canvas and
+     * TAA reconstructing to the canvas's size (TAAU); the governor's `render.scale` then drives it.
+     */
+    readonly taa?: boolean | {
+        readonly blend?: number;
+        readonly motion?: boolean;
+        readonly sharpen?: number;
+        readonly motionBlur?: number;
+        readonly upscale?: boolean;
+    };
+}
+/** The live post knobs (`render.post`). */
+export interface PostControls {
+    /** Which passes run (fixed at init; all false on WebGL2 and without the HDR target, where there is no chain). */
+    readonly bloom: boolean;
+    readonly fxaa: boolean;
+    readonly taa: boolean;
+    bloomIntensity: number;
+    bloomThreshold: number;
+    bloomKnee: number;
+    bloomRadius: number;
+    /** The current frame's weight in TAA's blend (default 0.1). */
+    taaBlend: number;
+    /** Sharpening after TAA, 0–1, and motion blur's shutter, 0–1 (ADR-0119; live when the pass was asked for at init). */
+    taaSharpen: number;
+    motionBlur: number;
+    /** Edit in place; `grading.enabled = false` turns it off. */
+    readonly grading: Required<GradingOptions> & {
+        enabled: boolean;
+    };
+    /** Drops TAA's history (call on a camera cut). */
+    resetHistory(): void;
+    /**
+     * What TAA's per-object motion held on the next frame (ADR-0124; reads back from the GPU): pixels drawn,
+     * those moving more than a twentieth of a pixel, the largest move in pixels, their mean view depth (last
+     * frame's). Null without TAA or per-object motion.
+     */
+    readMotion(): Promise<{
+        drawn: number;
+        moving: number;
+        maxPixels: number;
+        meanDepth: number;
+    } | null>;
+}
+/** The white-balance gains for `temperature` and `tint`, normalised to keep the luminance. */
+export declare function whiteBalance(temperature: number, tint: number): [number, number, number];
+/** Normalised grading options. */
+export declare function gradingDefaults(o: GradingOptions | undefined): Required<GradingOptions> & {
+    enabled: boolean;
+};

package/dist/post-options.js

@@ -0,0 +1,23 @@
+// The post chain's options and grading maths (ADR-0092), apart from its passes: the tone-mapping
+// pass and WebGL2 need these, while the passes (post.ts, and their shaders) load only when an app
+// asks for them.
+/** The white-balance gains for `temperature` and `tint`, normalised to keep the luminance. */
+export function whiteBalance(temperature, tint) {
+    const g = [1 + 0.2 * temperature, 1 - 0.2 * tint, 1 - 0.2 * temperature];
+    const y = 0.2126 * g[0] + 0.7152 * g[1] + 0.0722 * g[2];
+    return [g[0] / y, g[1] / y, g[2] / y];
+}
+/** Normalised grading options. */
+export function gradingDefaults(o) {
+    return {
+        enabled: o !== undefined,
+        temperature: o?.temperature ?? 0,
+        tint: o?.tint ?? 0,
+        contrast: o?.contrast ?? 1,
+        saturation: o?.saturation ?? 1,
+        lift: o?.lift ?? [0, 0, 0],
+        gamma: o?.gamma ?? [1, 1, 1],
+        gain: o?.gain ?? [1, 1, 1],
+    };
+}
+//# sourceMappingURL=post-options.js.map

package/dist/post-options.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"post-options.js","sourceRoot":"","sources":["../src/post-options.ts"],"names":[],"mappings":"AAAA,iGAAiG;AACjG,kGAAkG;AAClG,iBAAiB;AAiGjB,8FAA8F;AAC9F,MAAM,UAAU,YAAY,CAAC,WAAmB,EAAE,IAAY;IAC5D,MAAM,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,GAAG,WAAW,EAAE,CAAC,GAAG,GAAG,GAAG,IAAI,EAAE,CAAC,GAAG,GAAG,GAAG,WAAW,CAAC,CAAC;IACzE,MAAM,CAAC,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IACxD,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;AACxC,CAAC;AAED,kCAAkC;AAClC,MAAM,UAAU,eAAe,CAAC,CAA6B;IAC3D,OAAO;QACL,OAAO,EAAE,CAAC,KAAK,SAAS;QACxB,WAAW,EAAE,CAAC,EAAE,WAAW,IAAI,CAAC;QAChC,IAAI,EAAE,CAAC,EAAE,IAAI,IAAI,CAAC;QAClB,QAAQ,EAAE,CAAC,EAAE,QAAQ,IAAI,CAAC;QAC1B,UAAU,EAAE,CAAC,EAAE,UAAU,IAAI,CAAC;QAC9B,IAAI,EAAE,CAAC,EAAE,IAAI,IAAI,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;QAC1B,KAAK,EAAE,CAAC,EAAE,KAAK,IAAI,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;QAC5B,IAAI,EAAE,CAAC,EAAE,IAAI,IAAI,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;KAC3B,CAAC;AACJ,CAAC"}

package/dist/post-shaders.gen.d.ts

@@ -0,0 +1,40 @@
+/** Linked from shaders/post/bloom.wesl. */
+export declare const postBloom: {
+    readonly hash: "ce493f335d49f31e";
+    readonly wgsl: "struct BloomParams {\n/** Linear luminance where bloom starts (0: everything blooms, energy-conserving). */\nthreshold: f32,\n/** The soft knee's width, as a share of the threshold. */\nknee: f32,\n/** The upsample tent's radius, in texels of the smaller level. */\nradius: f32,\npad: f32,\n}\n@group(0) @binding(0) var src: texture_2d<f32>;\n@group(0) @binding(1) var smp: sampler;\n@group(0) @binding(2) var<uniform> bloom: BloomParams;\n@group(0) @binding(3) var low: texture_2d<f32>;\nstruct Out {\n@builtin(position) pos: vec4f,\n@location(0) uv: vec2f,\n}\n@vertex fn vs(@builtin(vertex_index) i: u32) -> Out {\nlet p = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nvar o: Out;\no.pos = vec4f(p * 2.0 - 1.0, 0.0, 1.0);\no.uv = vec2f(p.x, 1.0 - p.y);\nreturn o;\n}\nfn luma(c: vec3f) -> f32 {\nreturn dot(c, vec3f(0.2126, 0.7152, 0.0722));\n}\nfn tap(uv: vec2f) -> vec3f {\nreturn min(textureSampleLevel(src, smp, uv, 0.0).rgb, vec3f(65000.0));\n}\nfn downsample(uv: vec2f, karis: bool) -> vec3f {\nlet t = 1.0 / vec2f(textureDimensions(src));\nlet a = tap(uv + t * vec2f(-2.0, -2.0));\nlet b = tap(uv + t * vec2f(0.0, -2.0));\nlet c = tap(uv + t * vec2f(2.0, -2.0));\nlet d = tap(uv + t * vec2f(-1.0, -1.0));\nlet e = tap(uv + t * vec2f(1.0, -1.0));\nlet f = tap(uv + t * vec2f(-2.0, 0.0));\nlet g = tap(uv);\nlet h = tap(uv + t * vec2f(2.0, 0.0));\nlet i = tap(uv + t * vec2f(-1.0, 1.0));\nlet j = tap(uv + t * vec2f(1.0, 1.0));\nlet k = tap(uv + t * vec2f(-2.0, 2.0));\nlet l = tap(uv + t * vec2f(0.0, 2.0));\nlet m = tap(uv + t * vec2f(2.0, 2.0));\nlet boxes = array<vec3f, 5>((d + e + i + j) * 0.25, (a + b + f + g) * 0.25, (b + c + g + h) * 0.25, (f + g + k + l) * 0.25, (g + h + l + m) * 0.25);\nlet weights = array<f32, 5>(0.5, 0.125, 0.125, 0.125, 0.125);\nvar sum = vec3f(0.0);\nvar total = 0.0;\nfor (var n = 0; n < 5; n++) {\nvar w = weights[n];\nif (karis) {\nw = w / (1.0 + luma(boxes[n]));\n}\nsum += boxes[n] * w;\ntotal += w;\n}\nreturn sum / total;\n}\n@fragment fn fs_prefilter(in: Out) -> @location(0) vec4f {\nlet c = downsample(in.uv, true);\nif (bloom.threshold <= 0.0) {\nreturn vec4f(c, 1.0);\n}\nlet br = max(c.r, max(c.g, c.b));\nlet k = bloom.threshold * bloom.knee + 1e-5;\nvar soft = clamp(br - bloom.threshold + k, 0.0, 2.0 * k);\nsoft = soft * soft / (4.0 * k);\nlet w = max(soft, br - bloom.threshold) / max(br, 1e-5);\nreturn vec4f(c * w, 1.0);\n}\n@fragment fn fs_down(in: Out) -> @location(0) vec4f {\nreturn vec4f(downsample(in.uv, false), 1.0);\n}\n@fragment fn fs_up(in: Out) -> @location(0) vec4f {\nlet t = bloom.radius / vec2f(textureDimensions(low));\nvar s = textureSampleLevel(low, smp, in.uv, 0.0).rgb * 4.0;\ns += (textureSampleLevel(low, smp, in.uv + vec2f(-t.x, 0.0), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(t.x, 0.0), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(0.0, -t.y), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(0.0, t.y), 0.0).rgb) * 2.0;\ns += textureSampleLevel(low, smp, in.uv - t, 0.0).rgb + textureSampleLevel(low, smp, in.uv + t, 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(-t.x, t.y), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(t.x, -t.y), 0.0).rgb;\nreturn vec4f(textureSampleLevel(src, smp, in.uv, 0.0).rgb + s / 16.0, 1.0);\n}";
+};
+/** Linked from shaders/post/fxaa.wesl. */
+export declare const postFxaa: {
+    readonly hash: "362146f9cb191613";
+    readonly wgsl: "@group(0) @binding(0) var ldr: texture_2d<f32>;\n@group(0) @binding(1) var smp: sampler;\nstruct Out {\n@builtin(position) pos: vec4f,\n@location(0) uv: vec2f,\n}\n@vertex fn vs(@builtin(vertex_index) i: u32) -> Out {\nlet p = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nvar o: Out;\no.pos = vec4f(p * 2.0 - 1.0, 0.0, 1.0);\no.uv = vec2f(p.x, 1.0 - p.y);\nreturn o;\n}\nconst EDGE_MIN = 0.0312;\nconst EDGE_MAX = 0.125;\nconst SUBPIXEL = 0.75;\nconst STEPS = 10;\nfn luma_at(uv: vec2f) -> f32 {\nreturn dot(textureSampleLevel(ldr, smp, uv, 0.0).rgb, vec3f(0.299, 0.587, 0.114));\n}\nfn quality(i: i32) -> f32 {\nif (i < 5) {\nreturn 1.0;\n}\nif (i == 5) {\nreturn 1.5;\n}\nif (i < 9) {\nreturn 2.0;\n}\nreturn 4.0;\n}\n@fragment fn fs(in: Out) -> @location(0) vec4f {\nlet texel = 1.0 / vec2f(textureDimensions(ldr));\nlet uv = in.uv;\nlet centre = textureSampleLevel(ldr, smp, uv, 0.0);\nlet m = dot(centre.rgb, vec3f(0.299, 0.587, 0.114));\nlet n = luma_at(uv + vec2f(0.0, -texel.y));\nlet s = luma_at(uv + vec2f(0.0, texel.y));\nlet e = luma_at(uv + vec2f(texel.x, 0.0));\nlet w = luma_at(uv + vec2f(-texel.x, 0.0));\nlet lo = min(m, min(min(n, s), min(e, w)));\nlet hi = max(m, max(max(n, s), max(e, w)));\nlet range = hi - lo;\nif (range < max(EDGE_MIN, hi * EDGE_MAX)) {\nreturn centre;\n}\nlet nw = luma_at(uv + vec2f(-texel.x, -texel.y));\nlet ne = luma_at(uv + vec2f(texel.x, -texel.y));\nlet sw = luma_at(uv + vec2f(-texel.x, texel.y));\nlet se = luma_at(uv + vec2f(texel.x, texel.y));\nlet horizontal = abs(-2.0 * w + nw + sw) + 2.0 * abs(-2.0 * m + n + s) + abs(-2.0 * e + ne + se) >= abs(-2.0 * n + nw + ne) + 2.0 * abs(-2.0 * m + w + e) + abs(-2.0 * s + sw + se);\nlet l1 = select(w, n, horizontal);\nlet l2 = select(e, s, horizontal);\nlet g1 = l1 - m;\nlet g2 = l2 - m;\nlet steep1 = abs(g1) >= abs(g2);\nlet scaled = 0.25 * max(abs(g1), abs(g2));\nvar step_len = select(texel.x, texel.y, horizontal);\nvar avg = 0.5 * (l2 + m);\nif (steep1) {\nstep_len = -step_len;\navg = 0.5 * (l1 + m);\n}\nvar cur = uv;\nif (horizontal) {\ncur.y += step_len * 0.5;\n} else {\ncur.x += step_len * 0.5;\n}\nlet along = select(vec2f(0.0, texel.y), vec2f(texel.x, 0.0), horizontal);\nvar uv1 = cur - along;\nvar uv2 = cur + along;\nvar end1 = luma_at(uv1) - avg;\nvar end2 = luma_at(uv2) - avg;\nvar done1 = abs(end1) >= scaled;\nvar done2 = abs(end2) >= scaled;\nfor (var i = 1; i < STEPS && !(done1 && done2); i++) {\nif (!done1) {\nuv1 -= along * quality(i);\nend1 = luma_at(uv1) - avg;\ndone1 = abs(end1) >= scaled;\n}\nif (!done2) {\nuv2 += along * quality(i);\nend2 = luma_at(uv2) - avg;\ndone2 = abs(end2) >= scaled;\n}\n}\nlet d1 = select(uv.y - uv1.y, uv.x - uv1.x, horizontal);\nlet d2 = select(uv2.y - uv.y, uv2.x - uv.x, horizontal);\nlet nearer1 = d1 < d2;\nlet dist = min(d1, d2);\nlet pixel_offset = -dist / (d1 + d2) + 0.5;\nlet smaller = m < avg;\nlet right_way = (select(end2, end1, nearer1) < 0.0) != smaller;\nvar offset = select(0.0, pixel_offset, right_way);\nlet mean = (2.0 * (n + s + e + w) + nw + ne + sw + se) / 12.0;\nlet sub1 = clamp(abs(mean - m) / range, 0.0, 1.0);\nlet sub2 = (-2.0 * sub1 + 3.0) * sub1 * sub1;\noffset = max(offset, sub2 * sub2 * SUBPIXEL);\nvar at = uv;\nif (horizontal) {\nat.y += offset * step_len;\n} else {\nat.x += offset * step_len;\n}\nreturn vec4f(textureSampleLevel(ldr, smp, at, 0.0).rgb, 1.0);\n}";
+};
+/** Linked from shaders/post/taa.wesl. */
+export declare const postTaa: {
+    readonly hash: "05819cf45a86a800";
+    readonly wgsl: "struct TaaParams {\nreproject: mat4x4f,\nnear: f32,\nmotion: u32,\nblend: f32,\nreset: u32,\njitter_px: vec2f,\nupscale: u32,\npad: u32,\n}\n@group(0) @binding(0) var current: texture_2d<f32>;\n@group(0) @binding(1) var history: texture_2d<f32>;\n@group(0) @binding(2) var depth: texture_depth_2d;\n@group(0) @binding(3) var smp: sampler;\n@group(0) @binding(4) var<uniform> taa: TaaParams;\n@group(0) @binding(5) var motion_tex: texture_2d<f32>;\n@vertex fn vs(@builtin(vertex_index) i: u32) -> @builtin(position) vec4f {\nlet p = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nreturn vec4f(p * 2.0 - 1.0, 0.0, 1.0);\n}\nfn to_ycocg(c: vec3f) -> vec3f {\nreturn vec3f(dot(c, vec3f(0.25, 0.5, 0.25)), dot(c, vec3f(0.5, 0.0, -0.5)), dot(c, vec3f(-0.25, 0.5, -0.25)));\n}\nfn from_ycocg(c: vec3f) -> vec3f {\nreturn vec3f(c.x + c.y - c.z, c.x + c.z, c.x - c.y - c.z);\n}\nfn compress(c: vec3f) -> vec3f {\nreturn c / (1.0 + max(c.r, max(c.g, c.b)));\n}\nfn expand(c: vec3f) -> vec3f {\nreturn c / max(1.0 - max(c.r, max(c.g, c.b)), 1e-4);\n}\nconst FAR = 60000.0;\nfn view_depth(d: f32) -> f32 {\nreturn select(FAR, min(taa.near / d, FAR), d > 0.0);\n}\nfn same_surface(puv: vec2f, size: vec2f, expected: f32) -> bool {\nlet base = vec2i(floor(puv * size - 0.5));\nlet hi = vec2i(size) - 1;\nlet e = min(expected, FAR);\nfor (var k = 0; k < 4; k++) {\nlet q = clamp(base + vec2i(k & 1, k >> 1), vec2i(0), hi);\nlet h = textureLoad(history, q, 0).a;\nif (abs(h - e) <= 0.05 * e + 0.01) {\nreturn true;\n}\n}\nreturn false;\n}\nfn load(p: vec2i, size: vec2i) -> vec3f {\nlet c = textureLoad(current, clamp(p, vec2i(0), size - 1), 0).rgb;\nreturn select(vec3f(0.0), compress(min(c, vec3f(65000.0))), all(c == c));\n}\nfn history_at(uv: vec2f, size: vec2f) -> vec3f {\nlet pos = uv * size;\nlet centre = floor(pos - 0.5) + 0.5;\nlet f = pos - centre;\nlet w0 = f * (-0.5 + f * (1.0 - 0.5 * f));\nlet w1 = 1.0 + f * f * (-2.5 + 1.5 * f);\nlet w2 = f * (0.5 + f * (2.0 - 1.5 * f));\nlet w3 = f * f * (-0.5 + 0.5 * f);\nlet w12 = w1 + w2;\nlet t0 = (centre - 1.0) / size;\nlet t3 = (centre + 2.0) / size;\nlet t12 = (centre + w2 / w12) / size;\nvar c = textureSampleLevel(history, smp, vec2f(t12.x, t0.y), 0.0).rgb * (w12.x * w0.y);\nc += textureSampleLevel(history, smp, vec2f(t0.x, t12.y), 0.0).rgb * (w0.x * w12.y);\nc += textureSampleLevel(history, smp, vec2f(t12.x, t12.y), 0.0).rgb * (w12.x * w12.y);\nc += textureSampleLevel(history, smp, vec2f(t3.x, t12.y), 0.0).rgb * (w3.x * w12.y);\nc += textureSampleLevel(history, smp, vec2f(t12.x, t3.y), 0.0).rgb * (w12.x * w3.y);\nlet total = w12.x * w0.y + w0.x * w12.y + w12.x * w12.y + w3.x * w12.y + w12.x * w3.y;\nreturn max(c / total, vec3f(0.0));\n}\nfn upscaled(at: vec4f) -> vec4f {\nlet isize = vec2i(textureDimensions(current));\nlet size = vec2f(isize);\nlet osize = vec2f(textureDimensions(history));\nlet uv = at.xy / osize;\nlet ratio = osize / size;\nlet pin = uv * size;\nlet base = vec2i(floor(pin));\nvar m1 = vec3f(0.0);\nvar m2 = vec3f(0.0);\nvar sum = vec3f(0.0);\nvar wsum = 0.0;\nvar wmax = 0.0;\nfor (var y = -1; y <= 1; y++) {\nfor (var x = -1; x <= 1; x++) {\nlet q = base + vec2i(x, y);\nlet c = to_ycocg(load(q, isize));\nm1 += c;\nm2 += c * c;\nlet s = (vec2f(q) + 0.5 - taa.jitter_px - pin) * ratio;\nlet w = exp(-2.29 * dot(s, s));\nsum += c * w;\nwsum += w;\nwmax = max(wmax, w);\n}\n}\nlet here = sum / max(wsum, 1e-5);\nlet pd = clamp(base, vec2i(0), isize - 1);\nlet d = textureLoad(depth, pd, 0);\nlet fresh = vec4f(expand(from_ycocg(here)), view_depth(d));\nif (taa.reset != 0u) {\nreturn fresh;\n}\nvar puv: vec2f;\nvar expected: f32;\nlet mv = select(vec4f(0.0), textureLoad(motion_tex, pd, 0), taa.motion != 0u);\nif (mv.a > 0.5) {\npuv = uv - mv.xy;\nexpected = mv.z;\n} else {\nlet ndc = vec2f(uv.x * 2.0 - 1.0, 1.0 - uv.y * 2.0);\nlet prev = taa.reproject * vec4f(ndc, d, 1.0);\nif (prev.w <= 0.0 && d > 0.0) {\nreturn fresh;\n}\nlet pn = prev.xy / prev.w;\npuv = vec2f(pn.x * 0.5 + 0.5, 0.5 - pn.y * 0.5);\nexpected = select(FAR, prev.w * view_depth(d), d > 0.0);\n}\nif (any(puv < vec2f(0.0)) || any(puv > vec2f(1.0)) || !same_surface(puv, osize, expected)) {\nreturn fresh;\n}\nlet hist = to_ycocg(compress(history_at(puv, osize)));\nlet mean = m1 / 9.0;\nlet sigma = sqrt(max(m2 / 9.0 - mean * mean, vec3f(0.0)));\nlet half_box = 1.25 * sigma + 1e-5;\nlet off = hist - mean;\nlet units = abs(off / half_box);\nlet k = max(units.x, max(units.y, units.z));\nlet clipped = select(hist, mean + off / k, k > 1.0);\nlet out = mix(clipped, here, taa.blend * wmax);\nreturn vec4f(expand(from_ycocg(out)), fresh.a);\n}\n@fragment fn fs(@builtin(position) at: vec4f) -> @location(0) vec4f {\nif (taa.upscale != 0u) {\nreturn upscaled(at);\n}\nlet isize = vec2i(textureDimensions(current));\nlet size = vec2f(isize);\nlet p = vec2i(at.xy);\nvar m1 = vec3f(0.0);\nvar m2 = vec3f(0.0);\nvar here = vec3f(0.0);\nfor (var y = -1; y <= 1; y++) {\nfor (var x = -1; x <= 1; x++) {\nlet c = to_ycocg(load(p + vec2i(x, y), isize));\nm1 += c;\nm2 += c * c;\nif (x == 0 && y == 0) {\nhere = c;\n}\n}\n}\nlet d = textureLoad(depth, p, 0);\nlet fresh = vec4f(expand(from_ycocg(here)), view_depth(d));\nif (taa.reset != 0u) {\nreturn fresh;\n}\nvar puv: vec2f;\nvar expected: f32;\nlet mv = select(vec4f(0.0), textureLoad(motion_tex, p, 0), taa.motion != 0u);\nif (mv.a > 0.5) {\npuv = at.xy / size - mv.xy;\nexpected = mv.z;\n} else {\nlet ndc = vec2f(at.x / size.x * 2.0 - 1.0, 1.0 - at.y / size.y * 2.0);\nlet prev = taa.reproject * vec4f(ndc, d, 1.0);\nif (prev.w <= 0.0 && d > 0.0) {\nreturn fresh;\n}\nlet pn = prev.xy / prev.w;\npuv = vec2f(pn.x * 0.5 + 0.5, 0.5 - pn.y * 0.5);\nexpected = select(FAR, prev.w * view_depth(d), d > 0.0);\n}\nif (any(puv < vec2f(0.0)) || any(puv > vec2f(1.0)) || !same_surface(puv, size, expected)) {\nreturn fresh;\n}\nlet hist = to_ycocg(compress(history_at(puv, size)));\nlet mean = m1 / 9.0;\nlet sigma = sqrt(max(m2 / 9.0 - mean * mean, vec3f(0.0)));\nlet half_box = 1.25 * sigma + 1e-5;\nlet off = hist - mean;\nlet units = abs(off / half_box);\nlet k = max(units.x, max(units.y, units.z));\nlet clipped = select(hist, mean + off / k, k > 1.0);\nlet out = mix(clipped, here, taa.blend);\nreturn vec4f(expand(from_ycocg(out)), fresh.a);\n}";
+};
+/** Linked from shaders/post/taa_post.wesl. */
+export declare const postTaaPost: {
+    readonly hash: "5806c7af9e6201d4";
+    readonly wgsl: "@group(0) @binding(0) var color: texture_2d<f32>;\n@group(0) @binding(1) var motion: texture_2d<f32>;\n@group(0) @binding(2) var smp: sampler;\nstruct Params {\nsharpen: f32,\nblur: f32,\nsamples: f32,\nhas_motion: f32,\n}\n@group(0) @binding(3) var<uniform> p: Params;\nstruct Out {\n@builtin(position) pos: vec4f,\n@location(0) uv: vec2f,\n}\n@vertex fn vs(@builtin(vertex_index) i: u32) -> Out {\nlet q = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nvar o: Out;\no.pos = vec4f(q * 2.0 - 1.0, 0.0, 1.0);\no.uv = vec2f(q.x, 1.0 - q.y);\nreturn o;\n}\nfn load(px: vec2i, size: vec2i) -> vec3f {\nreturn textureLoad(color, clamp(px, vec2i(0), size - 1), 0).rgb;\n}\nfn tame(c: vec3f) -> vec3f {\nreturn c / (1.0 + c);\n}\n@fragment fn fs(in: Out) -> @location(0) vec4f {\nlet size = vec2i(textureDimensions(color));\nlet px = vec2i(in.pos.xy);\nlet c = load(px, size);\nvar out = c;\nif (p.sharpen > 0.0) {\nlet n = load(px + vec2i(0, -1), size);\nlet s = load(px + vec2i(0, 1), size);\nlet e = load(px + vec2i(1, 0), size);\nlet w = load(px + vec2i(-1, 0), size);\nlet tc = tame(c);\nlet mn = min(tc, min(min(tame(n), tame(s)), min(tame(e), tame(w))));\nlet mx = max(tc, max(max(tame(n), tame(s)), max(tame(e), tame(w))));\nlet amp = sqrt(clamp(min(mn, 1.0 - mx) / max(mx, vec3f(1e-4)), vec3f(0.0), vec3f(1.0)));\nlet k = amp * -mix(0.125, 0.2, clamp(p.sharpen, 0.0, 1.0));\nout = max((c + (n + s + e + w) * k) / (1.0 + 4.0 * k), vec3f(0.0));\n}\nif (p.blur > 0.0 && p.has_motion > 0.5) {\nlet msize = vec2i(textureDimensions(motion));\nlet m = textureLoad(motion, clamp(vec2i(vec2f(px) * vec2f(msize) / vec2f(size)), vec2i(0), msize - 1), 0);\nlet v = select(vec2f(0.0), m.rg * p.blur, m.a > 0.0);\nlet len_px = length(v * vec2f(size));\nif (len_px > 0.5) {\nlet n = i32(clamp(p.samples, 2.0, 32.0));\nvar sum = vec3f(0.0);\nfor (var k = 0; k < n; k++) {\nlet t = (f32(k) + 0.5) / f32(n) - 0.5;\nsum += textureSampleLevel(color, smp, in.uv - v * t, 0.0).rgb;\n}\nlet keep = clamp(1.0 - len_px / f32(n), 0.0, 1.0);\nout = mix(sum / f32(n), out, keep);\n}\n}\nreturn vec4f(out, 1.0);\n}";
+};
+/** Linked from shaders/scene/hzb.wesl. */
+export declare const sceneHzb: {
+    readonly hash: "b26b6bc789bda1de";
+    readonly wgsl: "@group(0) @binding(0) var depth: texture_depth_2d;\n@group(0) @binding(1) var src: texture_2d<f32>;\n@group(0) @binding(2) var dst: texture_storage_2d<r32float, write>;\n@compute @workgroup_size(8, 8) fn hzb_first(@builtin(global_invocation_id) id: vec3u) {\nlet size = textureDimensions(dst);\nif (id.x >= size.x || id.y >= size.y) {\nreturn;\n}\nlet hi = vec2i(textureDimensions(depth)) - 1;\nvar d = 1.0;\nfor (var y = 0; y < 3; y++) {\nfor (var x = 0; x < 3; x++) {\nd = min(d, textureLoad(depth, min(vec2i(id.xy) * 2 + vec2i(x, y), hi), 0));\n}\n}\ntextureStore(dst, id.xy, vec4f(d, 0.0, 0.0, 0.0));\n}\n@compute @workgroup_size(8, 8) fn hzb_down(@builtin(global_invocation_id) id: vec3u) {\nlet size = textureDimensions(dst);\nif (id.x >= size.x || id.y >= size.y) {\nreturn;\n}\nlet hi = vec2i(textureDimensions(src)) - 1;\nvar d = 1.0;\nfor (var y = 0; y < 3; y++) {\nfor (var x = 0; x < 3; x++) {\nd = min(d, textureLoad(src, min(vec2i(id.xy) * 2 + vec2i(x, y), hi), 0).r);\n}\n}\ntextureStore(dst, id.xy, vec4f(d, 0.0, 0.0, 0.0));\n}";
+};
+/** Linked from shaders/scene/motion.wesl. */
+export declare const postMotion: {
+    readonly hash: "407e14423b3c1d1b";
+    readonly wgsl: "struct Motion {\nview_proj: mat4x4f,\nprev_view_proj: mat4x4f,\nprev_origin: vec4f,\n}\n@group(0) @binding(0) var<uniform> camera: Camera;\n@group(0) @binding(1) var<storage, read> instances: array<vec4u>;\n@group(0) @binding(2) var<storage, read> visible: array<u32>;\n@group(0) @binding(3) var<storage, read> prev_instances: array<vec4u>;\n@group(0) @binding(4) var<uniform> motion: Motion;\n@group(0) @binding(5) var palette: texture_2d<f32>;\n@group(0) @binding(6) var prev_palette: texture_2d<f32>;\n@group(0) @binding(7) var<storage, read> prev_pool: array<f32>;\n@group(0) @binding(8) var<storage, read> prev_skinned: array<f32>;\nstruct Out {\n@builtin(position) clip: vec4f,\n@location(0) now: vec4f,\n@location(1) then: vec4f,\n}\nstruct Pose {\npos: vec3f,\nscale: f32,\nrot: vec4f,\n}\nfn pose_now(i: u32) -> Pose {\nlet a = bitcast<vec4f>(instances[2u * i]);\nreturn Pose(a.xyz - camera.origin.xyz, a.w, bitcast<vec4f>(instances[2u * i + 1u]));\n}\nfn pose_then(i: u32) -> Pose {\nlet a = bitcast<vec4f>(prev_instances[2u * i]);\nreturn Pose(a.xyz - motion.prev_origin.xyz, a.w, bitcast<vec4f>(prev_instances[2u * i + 1u]));\n}\nfn skin_row(i: u32) -> u32 {\nlet cap = arrayLength(&instances) / 3u;\nreturn instances[2u * cap + i].w;\n}\nfn place(now: Pose, then: Pose, p_now: vec3f, p_then: vec3f) -> Out {\nlet w_now = quat_rotate(now.rot, p_now * now.scale) + now.pos;\nlet w_then = quat_rotate(then.rot, p_then * then.scale) + then.pos;\nvar o: Out;\no.clip = camera.view_proj * vec4f(w_now, 1.0);\no.now = motion.view_proj * vec4f(w_now, 1.0);\no.then = motion.prev_view_proj * vec4f(w_then, 1.0);\nreturn o;\n}\n@vertex fn vs(@location(0) position: vec3f, @builtin(vertex_index) v: u32, @builtin(instance_index) slot: u32) -> Out {\nlet i = visible[slot];\nreturn place(pose_now(i), pose_then(i), position, vec3f(prev_pool[v * 8u], prev_pool[v * 8u + 1u], prev_pool[v * 8u + 2u]));\n}\n@vertex fn vs_skinned(@location(0) position: vec3f, @location(6) joints: vec4u, @location(7) weights: vec4f, @builtin(vertex_index) v: u32, @builtin(instance_index) slot: u32) -> Out {\nlet i = visible[slot];\nlet row = skin_row(i);\nlet n = vec3f(0.0, 0.0, 1.0);\nlet now = skin_vertex(palette, row, joints, weights, position, n);\nlet then = skin_vertex(prev_palette, row, joints, weights, vec3f(prev_skinned[v * 8u], prev_skinned[v * 8u + 1u], prev_skinned[v * 8u + 2u]), n);\nreturn place(pose_now(i), pose_then(i), now.position, then.position);\n}\n@fragment fn fs(in: Out) -> @location(0) vec4f {\nlet now = in.now.xy / in.now.w;\nlet then = in.then.xy / max(in.then.w, 1e-6);\nreturn vec4f((now - then) * vec2f(0.5, -0.5), in.then.w, 1.0);\n}\nstruct Camera {\nview_proj: mat4x4f,\nfrustum: array<vec4f, 6>,\nlight_dir: vec4f,\norigin: vec4f,\n}\nfn quat_rotate(q: vec4f, v: vec3f) -> vec3f {\nlet t = 2.0 * cross(q.xyz, v);\nreturn v + q.w * t + cross(q.xyz, t);\n}\nfn skin_vertex(palette: texture_2d<f32>, row: u32, joints: vec4u, weights: vec4f, position: vec3f, normal: vec3f) -> Skinned {\nif (row == 0u || dot(weights, vec4f(1.0)) <= 0.0) {\nreturn Skinned(position, normal);\n}\nlet base = row - 1u;\nlet t = (vec4u(base) + joints) * 3u;\nlet r0 = weights.x * palette_texel(palette, t.x) + weights.y * palette_texel(palette, t.y) + weights.z * palette_texel(palette, t.z) + weights.w * palette_texel(palette, t.w);\nlet r1 = weights.x * palette_texel(palette, t.x + 1u) + weights.y * palette_texel(palette, t.y + 1u) + weights.z * palette_texel(palette, t.z + 1u) + weights.w * palette_texel(palette, t.w + 1u);\nlet r2 = weights.x * palette_texel(palette, t.x + 2u) + weights.y * palette_texel(palette, t.y + 2u) + weights.z * palette_texel(palette, t.z + 2u) + weights.w * palette_texel(palette, t.w + 2u);\nlet p = vec4f(position, 1.0);\nreturn Skinned(vec3f(dot(r0, p), dot(r1, p), dot(r2, p)), vec3f(dot(r0.xyz, normal), dot(r1.xyz, normal), dot(r2.xyz, normal)));\n}\nfn palette_texel(palette: texture_2d<f32>, i: u32) -> vec4f {\nreturn textureLoad(palette, vec2u(i % PALETTE_WIDTH, i / PALETTE_WIDTH), 0);\n}\nstruct Skinned {\nposition: vec3f,\nnormal: vec3f,\n}\nconst PALETTE_WIDTH: u32 = 1024u;";
+};
+/** Linked from shaders/scene/clusters.wesl. */
+export declare const sceneClusters: {
+    readonly hash: "c7892829a492870a";
+    readonly wgsl: "struct Cluster {\ncenter: vec3f,\nradius: f32,\napex: vec3f,\ncutoff: f32,\naxis: vec3f,\nbase_vertex: u32,\nvert_offset: u32,\ntri_offset: u32,\ntri_count: u32,\nflags: u32,\nself_sphere: vec4f,\nparent_sphere: vec4f,\nself_error: f32,\nparent_error: f32,\nstream: u32,\nkids: u32,\npad1: vec4f,\n}\nstruct Job {\nargs_slot: u32,\nstart: u32,\ncluster_base: u32,\ncluster_count: u32,\nprefix: u32,\ndouble_sided: u32,\npad0: u32,\npad1: u32,\n}\nstruct Params {\njobs: u32,\ntotal: u32,\nhalf: u32,\nindex_half: u32,\n}\n@group(0) @binding(0) var<uniform> camera: Camera;\n@group(0) @binding(1) var<storage, read> instances: array<vec4u>;\n@group(0) @binding(2) var<storage, read> visible: array<u32>;\n@group(0) @binding(3) var<storage, read> args: array<u32>;\n@group(0) @binding(4) var<storage, read> clusters: array<Cluster>;\n@group(0) @binding(5) var<storage, read> jobs: array<Job>;\n@group(0) @binding(6) var<uniform> params: Params;\n@group(0) @binding(7) var<storage, read_write> out_args: array<atomic<u32>, 12>;\n@group(0) @binding(8) var<storage, read_write> out_visible: array<vec4u>;\n@group(0) @binding(9) var<storage, read_write> out_indices: array<u32>;\n@group(0) @binding(10) var<storage, read> triangles: array<u32>;\nfn load_instance(i: u32) -> Instance {\nlet cap = arrayLength(&instances) / 3u;\nlet a = bitcast<vec4f>(instances[2u * i]);\nlet b = bitcast<vec4f>(instances[2u * i + 1u]);\nlet s = instances[2u * cap + i];\nreturn Instance(a.xyz - camera.origin.xyz, a.w, b, s.x, s.y, s.z, s.w);\n}\nconst SLOT: u32 = 128u;\nstruct LodParams {\ncount: u32,\nlod_scale: f32,\nerror_px: f32,\nmin_radius_px: f32,\n}\n@group(0) @binding(16) var<uniform> lodp: LodParams;\nfn error_px(inst: Instance, s: vec4f, e: f32) -> f32 {\nlet c = inst.pos + quat_rotate(inst.rot, s.xyz * inst.scale);\nlet d = max(length(c) - s.w * inst.scale, 1e-3);\nreturn e * inst.scale * lodp.lod_scale / d;\n}\nconst STREAM_MISSING: u32 = 1u;\nconst STREAM_LEAF: u32 = 2u;\nfn on_cut(inst: Instance, cl: Cluster) -> bool {\nif ((cl.stream & STREAM_MISSING) != 0u) {\nreturn false;\n}\nif (cl.self_error <= 0.0 && cl.parent_error == 0.0) {\nreturn true;\n}\nif ((cl.stream & STREAM_LEAF) == 0u && error_px(inst, cl.self_sphere, cl.self_error) > lodp.error_px) {\nreturn false;\n}\nreturn cl.parent_error < 0.0 || error_px(inst, cl.parent_sphere, cl.parent_error) > lodp.error_px;\n}\nfn job_entry(id: vec3u) -> vec4u {\nlet x = id.x;\nif (x >= params.total) {\nreturn vec4u(0u);\n}\nvar j = 0u;\nfor (; j < params.jobs; j++) {\nif (x < jobs[j].prefix + jobs[j].cluster_count) {\nbreak;\n}\n}\nif (j == params.jobs) {\nreturn vec4u(0u);\n}\nlet job = jobs[j];\nif (id.y >= args[job.args_slot * 5u + 1u]) {\nreturn vec4u(0u);\n}\nreturn vec4u(visible[job.start + id.y], job.cluster_base + (x - job.prefix), 1u, 0u);\n}\nfn sphere_in_view(inst: Instance, cl: Cluster) -> vec4f {\nif (!on_cut(inst, cl)) {\nreturn vec4f(0.0, 0.0, 0.0, -1.0);\n}\nlet c = inst.pos + quat_rotate(inst.rot, cl.center * inst.scale);\nlet r = cl.radius * inst.scale;\nfor (var p = 0u; p < 6u; p++) {\nlet plane = camera.frustum[p];\nif (dot(plane.xyz, c) + plane.w < -r) {\nreturn vec4f(0.0, 0.0, 0.0, -1.0);\n}\n}\nif (cl.flags == 0u && cl.cutoff < 1.0) {\nlet apex = inst.pos + quat_rotate(inst.rot, cl.apex * inst.scale);\nif (dot(normalize(apex), quat_rotate(inst.rot, cl.axis)) >= cl.cutoff) {\nreturn vec4f(0.0, 0.0, 0.0, -1.0);\n}\n}\nreturn vec4f(c, r);\n}\n@group(0) @binding(21) var<storage, read_write> page_bits: array<atomic<u32>>;\nfn note_pages(index: u32, cl: Cluster) {\nif (cl.stream == 0u && cl.kids == 0u) {\nreturn;\n}\nlet half = arrayLength(&page_bits) / 2u;\nlet page = cl.stream >> 2u;\natomicOr(&page_bits[half + page / 32u], 1u << (page % 32u));\nif ((cl.stream & STREAM_LEAF) != 0u && cl.kids > 0u && error_px(load_instance(index), cl.self_sphere, cl.self_error) > lodp.error_px) {\nlet want = cl.kids - 1u;\natomicOr(&page_bits[want / 32u], 1u << (want % 32u));\n}\n}\nfn emit(index: u32, ci: u32, cl: Cluster) {\nnote_pages(index, cl);\nlet side = cl.flags;\nlet k = atomicAdd(&out_args[10u + side], 1u);\nlet at = side * params.index_half + atomicAdd(&out_args[side * 5u], cl.tri_count * 3u);\nout_visible[side * params.half + k] = vec4u(index, ci, at, k);\n}\n@compute @workgroup_size(64) fn cull_clusters(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet cl = clusters[e.y];\nif (sphere_in_view(load_instance(e.x), cl).w >= 0.0) {\nemit(e.x, e.y, cl);\n}\n}\n@group(0) @binding(12) var<uniform> occl: Occlusion;\n@group(0) @binding(13) var hzb: texture_2d<f32>;\n@group(0) @binding(14) var<storage, read_write> recheck: array<atomic<u32>>;\n@group(0) @binding(15) var<storage, read_write> recheck_dispatch: array<u32, 3>;\n@compute @workgroup_size(64) fn cull_clusters_first(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet cl = clusters[e.y];\nlet s = sphere_in_view(load_instance(e.x), cl);\nif (s.w < 0.0) {\nreturn;\n}\nif (hidden_in(hzb, occl.hzb_size, occl.hzb_levels, s.xyz, s.w, occl.prev)) {\nlet k = atomicAdd(&recheck[0], 1u);\nif (2u * k + 2u < arrayLength(&recheck)) {\natomicStore(&recheck[2u * k + 1u], e.x);\natomicStore(&recheck[2u * k + 2u], e.y);\n}\nreturn;\n}\nemit(e.x, e.y, cl);\n}\n@compute @workgroup_size(64) fn cull_clusters_second(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet cl = clusters[e.y];\nlet s = sphere_in_view(load_instance(e.x), cl);\nif (s.w >= 0.0 && !hidden_in(hzb, occl.hzb_size, occl.hzb_levels, s.xyz, s.w, camera.view_proj)) {\nemit(e.x, e.y, cl);\n}\n}\n@compute @workgroup_size(1) fn size_recheck() {\nlet n = min(atomicLoad(&recheck[0]), (arrayLength(&recheck) - 1u) / 2u);\nrecheck_dispatch[0] = (n + 63u) / 64u;\nrecheck_dispatch[1] = 1u;\nrecheck_dispatch[2] = 1u;\n}\n@compute @workgroup_size(64) fn recheck_clusters(@builtin(global_invocation_id) id: vec3u) {\nlet n = min(atomicLoad(&recheck[0]), (arrayLength(&recheck) - 1u) / 2u);\nif (id.x >= n) {\nreturn;\n}\nlet index = atomicLoad(&recheck[2u * id.x + 1u]);\nlet ci = atomicLoad(&recheck[2u * id.x + 2u]);\nlet cl = clusters[ci];\nlet s = sphere_in_view(load_instance(index), cl);\nif (s.w >= 0.0 && !hidden_in(hzb, occl.hzb_size, occl.hzb_levels, s.xyz, s.w, camera.view_proj)) {\nemit(index, ci, cl);\n}\n}\nstruct ShadowCull {\nplanes: array<vec4f, 20>,\nlight: vec4f,\n}\n@group(0) @binding(17) var<uniform> shc: ShadowCull;\n@group(0) @binding(18) var<storage, read_write> sh_args: array<atomic<u32>, 32>;\n@group(0) @binding(19) var<storage, read_write> sh_visible: array<vec4u>;\n@compute @workgroup_size(64) fn cull_clusters_shadow(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet c = id.z;\nlet cl = clusters[e.y];\nlet inst = load_instance(e.x);\nif (!on_cut(inst, cl)) {\nreturn;\n}\nlet center = inst.pos + quat_rotate(inst.rot, cl.center * inst.scale);\nlet r = cl.radius * inst.scale;\nfor (var p = 0u; p < 5u; p++) {\nlet plane = shc.planes[c * 5u + p];\nif (dot(plane.xyz, center) + plane.w < -r) {\nreturn;\n}\n}\nif (cl.flags == 0u && cl.cutoff < 1.0 && dot(-shc.light.xyz, quat_rotate(inst.rot, cl.axis)) >= cl.cutoff) {\nreturn;\n}\nlet side = cl.flags;\nlet b = c * 2u + side;\nlet k = atomicAdd(&sh_args[b * 4u + 1u], 1u);\nsh_visible[c * params.index_half + side * params.half + k] = vec4u(e.x, e.y, 0u, 0u);\n}\n@group(0) @binding(11) var<storage, read_write> dispatch: array<u32, 3>;\n@compute @workgroup_size(1) fn size_expand() {\ndispatch[0] = atomicLoad(&out_args[10]) + atomicLoad(&out_args[11]);\ndispatch[1] = 1u;\ndispatch[2] = 1u;\n}\n@compute @workgroup_size(128) fn expand_clusters(@builtin(workgroup_id) wg: vec3u, @builtin(local_invocation_index) t: u32) {\nlet single = atomicLoad(&out_args[10]);\nlet g = wg.x;\nlet rec = out_visible[select(params.half + g - single, g, g < single)];\nlet cl = clusters[rec.y];\nif (t >= cl.tri_count) {\nreturn;\n}\nfor (var c = 0u; c < 3u; c++) {\nlet b = cl.tri_offset + t * 3u + c;\nout_indices[rec.z + t * 3u + c] = rec.w * SLOT + ((triangles[b / 4u] >> ((b % 4u) * 8u)) & 0xffu);\n}\n}\nstruct Camera {\nview_proj: mat4x4f,\nfrustum: array<vec4f, 6>,\nlight_dir: vec4f,\norigin: vec4f,\n}\nstruct Instance {\npos: vec3f,\nscale: f32,\nrot: vec4f,\ncolor: u32,\nmesh: u32,\nmaterial: u32,\nskin: u32,\n}\nfn quat_rotate(q: vec4f, v: vec3f) -> vec3f {\nlet t = 2.0 * cross(q.xyz, v);\nreturn v + q.w * t + cross(q.xyz, t);\n}\nstruct Occlusion {\nprev: mat4x4f,\nfrustum: array<vec4f, 6>,\nhzb_size: vec2f,\nhzb_levels: u32,\nshadows: u32,\nslots: u32,\nregion: u32,\npad0: u32,\npad1: u32,\n}\nfn hidden_in(hzb: texture_2d<f32>, size: vec2f, levels: u32, c: vec3f, r: f32, view_proj: mat4x4f) -> bool {\nif (levels == 0u) {\nreturn false;\n}\nvar lo = vec2f(1e9);\nvar hi = vec2f(-1e9);\nvar nearest = 0.0;\nfor (var k = 0u; k < 8u; k++) {\nlet corner = c + r * vec3f(select(-1.0, 1.0, (k & 1u) != 0u), select(-1.0, 1.0, (k & 2u) != 0u), select(-1.0, 1.0, (k & 4u) != 0u));\nlet p = view_proj * vec4f(corner, 1.0);\nif (p.w <= 1e-4) {\nreturn false;\n}\nlet n = p.xyz / p.w;\nlo = min(lo, n.xy);\nhi = max(hi, n.xy);\nnearest = max(nearest, n.z);\n}\nif (any(hi < vec2f(-1.0)) || any(lo > vec2f(1.0))) {\nreturn false;\n}\nlet uv_lo = clamp(vec2f(lo.x, -hi.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet uv_hi = clamp(vec2f(hi.x, -lo.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet extent = (uv_hi - uv_lo) * size;\nlet level = min(u32(max(ceil(log2(max(max(extent.x, extent.y), 1.0))), 0.0)), levels - 1u);\nlet dims = vec2f(textureDimensions(hzb, level));\nlet t0 = vec2i(uv_lo * dims);\nlet t1 = min(vec2i(uv_hi * dims), vec2i(dims) - 1);\nvar farthest = 1.0;\nfor (var y = t0.y; y <= min(t1.y, t0.y + 2); y++) {\nfor (var x = t0.x; x <= min(t1.x, t0.x + 2); x++) {\nfarthest = min(farthest, textureLoad(hzb, vec2i(x, y), i32(level)).r);\n}\n}\nreturn nearest < farthest;\n}";
+};
+/** Linked from shaders/scene/occlusion_cull.wesl. */
+export declare const sceneOcclusionCull: {
+    readonly hash: "d66ce30369f82685";
+    readonly wgsl: "@group(0) @binding(7) var<uniform> occl: Occlusion;\n@group(0) @binding(8) var hzb: texture_2d<f32>;\n@group(0) @binding(9) var<storage, read_write> recheck: array<atomic<u32>>;\nfn hidden(c: vec3f, r: f32, view_proj: mat4x4f) -> bool {\nreturn hidden_in(hzb, occl.hzb_size, occl.hzb_levels, c, r, view_proj);\n}\n@compute @workgroup_size(64) fn cull_occlusion(@builtin(global_invocation_id) id: vec3u) {\nlet i = id.x;\nif (i >= params.count) {\nreturn;\n}\nlet inst = load_instance(i);\nlet m = meshes[inst.mesh];\nlet c = inst.pos + quat_rotate(inst.rot, m.center * inst.scale);\nlet r = m.radius * inst.scale;\nif (occl.shadows != 0u) {\nlet dst = place(inst);\nif (dst != 0xffffffffu) {\nvisible[dst] = i;\n}\n}\nfor (var p = 0u; p < 6u; p++) {\nlet plane = occl.frustum[p];\nif (dot(plane.xyz, c) + plane.w < -r) {\nreturn;\n}\n}\nif (hidden(c, r, occl.prev)) {\nlet k = atomicAdd(&recheck[0], 1u);\natomicStore(&recheck[k + 1u], i);\nreturn;\n}\nlet dst = place_lod(inst, c, r, occl.slots, occl.region);\nif (dst != 0xffffffffu) {\nvisible[dst] = i;\n}\n}\n@compute @workgroup_size(64) fn cull_recheck(@builtin(global_invocation_id) id: vec3u) {\nif (id.x >= atomicLoad(&recheck[0])) {\nreturn;\n}\nlet i = atomicLoad(&recheck[id.x + 1u]);\nlet inst = load_instance(i);\nlet m = meshes[inst.mesh];\nlet c = inst.pos + quat_rotate(inst.rot, m.center * inst.scale);\nlet r = m.radius * inst.scale;\nif (hidden(c, r, camera.view_proj)) {\nreturn;\n}\nlet dst = place_lod(inst, c, r, 2u * occl.slots, 2u * occl.region);\nif (dst != 0xffffffffu) {\nvisible[dst] = i;\n}\n}\nstruct Occlusion {\nprev: mat4x4f,\nfrustum: array<vec4f, 6>,\nhzb_size: vec2f,\nhzb_levels: u32,\nshadows: u32,\nslots: u32,\nregion: u32,\npad0: u32,\npad1: u32,\n}\nfn hidden_in(hzb: texture_2d<f32>, size: vec2f, levels: u32, c: vec3f, r: f32, view_proj: mat4x4f) -> bool {\nif (levels == 0u) {\nreturn false;\n}\nvar lo = vec2f(1e9);\nvar hi = vec2f(-1e9);\nvar nearest = 0.0;\nfor (var k = 0u; k < 8u; k++) {\nlet corner = c + r * vec3f(select(-1.0, 1.0, (k & 1u) != 0u), select(-1.0, 1.0, (k & 2u) != 0u), select(-1.0, 1.0, (k & 4u) != 0u));\nlet p = view_proj * vec4f(corner, 1.0);\nif (p.w <= 1e-4) {\nreturn false;\n}\nlet n = p.xyz / p.w;\nlo = min(lo, n.xy);\nhi = max(hi, n.xy);\nnearest = max(nearest, n.z);\n}\nif (any(hi < vec2f(-1.0)) || any(lo > vec2f(1.0))) {\nreturn false;\n}\nlet uv_lo = clamp(vec2f(lo.x, -hi.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet uv_hi = clamp(vec2f(hi.x, -lo.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet extent = (uv_hi - uv_lo) * size;\nlet level = min(u32(max(ceil(log2(max(max(extent.x, extent.y), 1.0))), 0.0)), levels - 1u);\nlet dims = vec2f(textureDimensions(hzb, level));\nlet t0 = vec2i(uv_lo * dims);\nlet t1 = min(vec2i(uv_hi * dims), vec2i(dims) - 1);\nvar farthest = 1.0;\nfor (var y = t0.y; y <= min(t1.y, t0.y + 2); y++) {\nfor (var x = t0.x; x <= min(t1.x, t0.x + 2); x++) {\nfarthest = min(farthest, textureLoad(hzb, vec2i(x, y), i32(level)).r);\n}\n}\nreturn nearest < farthest;\n}\n@group(0) @binding(4) var<uniform> params: CullParams;\nfn load_instance(i: u32) -> Instance {\nlet cap = arrayLength(&instances) / 3u;\nlet a = bitcast<vec4f>(instances[2u * i]);\nlet b = bitcast<vec4f>(instances[2u * i + 1u]);\nlet s = instances[2u * cap + i];\nreturn Instance(a.xyz - camera.origin.xyz, a.w, b, s.x, s.y, s.z, s.w);\n}\n@group(0) @binding(2) var<storage, read> meshes: array<MeshInfo>;\nfn quat_rotate(q: vec4f, v: vec3f) -> vec3f {\nlet t = 2.0 * cross(q.xyz, v);\nreturn v + q.w * t + cross(q.xyz, t);\n}\nfn place(inst: Instance) -> u32 {\nlet m = meshes[inst.mesh];\nlet c = inst.pos + quat_rotate(inst.rot, m.center * inst.scale);\nlet r = m.radius * inst.scale;\nfor (var p = 0u; p < 6u; p++) {\nlet plane = camera.frustum[p];\nif (dot(plane.xyz, c) + plane.w < -r) {\nreturn 0xffffffffu;\n}\n}\nreturn place_lod(inst, c, r, 0u, 0u);\n}\n@group(0) @binding(5) var<storage, read_write> visible: array<u32>;\nfn place_lod(inst: Instance, c: vec3f, r: f32, slots: u32, region: u32) -> u32 {\nlet m = meshes[inst.mesh];\nlet px = params.lod_scale / max(length(c) - r, 1e-3);\nif (r * px < params.min_radius_px) {\nreturn 0xffffffffu;\n}\nvar lod = 0u;\nfor (var l = 1u; l < m.lod_count; l++) {\nif (m.error[l / 4u][l % 4u] * inst.scale * px > params.error_px) {\nbreak;\n}\nlod = l;\n}\nlet k = atomicAdd(&args[m.draw_base + lod + slots].instance_count, 1u);\nreturn m.instance_base + lod * m.instance_count + k + region;\n}\nstruct Instance {\npos: vec3f,\nscale: f32,\nrot: vec4f,\ncolor: u32,\nmesh: u32,\nmaterial: u32,\nskin: u32,\n}\n@group(0) @binding(0) var<uniform> camera: Camera;\nstruct Camera {\nview_proj: mat4x4f,\nfrustum: array<vec4f, 6>,\nlight_dir: vec4f,\norigin: vec4f,\n}\n@group(0) @binding(3) var<storage, read_write> args: array<DrawArgs>;\nstruct DrawArgs {\nindex_count: u32,\ninstance_count: atomic<u32>,\nfirst_index: u32,\nbase_vertex: i32,\nfirst_instance: u32,\n}\nstruct CullParams {\ncount: u32,\nlod_scale: f32,\nerror_px: f32,\nmin_radius_px: f32,\n}\n@group(0) @binding(1) var<storage, read> instances: array<vec4u>;\nstruct MeshInfo {\ncenter: vec3f,\nradius: f32,\nlod_count: u32,\ndraw_base: u32,\ninstance_base: u32,\ninstance_count: u32,\nerror: array<vec4f, 2>,\n}";
+};

package/dist/post-shaders.gen.js

@@ -0,0 +1,43 @@
+// GENERATED by tools/shaders-gen.mjs from shaders/**/*.wesl. Do not edit.
+// Regenerate with `node tools/shaders-gen.mjs`; CI runs it with --check.
+/** Linked from shaders/post/bloom.wesl. */
+export const postBloom = {
+    hash: 'ce493f335d49f31e',
+    wgsl: "struct BloomParams {\n/** Linear luminance where bloom starts (0: everything blooms, energy-conserving). */\nthreshold: f32,\n/** The soft knee's width, as a share of the threshold. */\nknee: f32,\n/** The upsample tent's radius, in texels of the smaller level. */\nradius: f32,\npad: f32,\n}\n@group(0) @binding(0) var src: texture_2d<f32>;\n@group(0) @binding(1) var smp: sampler;\n@group(0) @binding(2) var<uniform> bloom: BloomParams;\n@group(0) @binding(3) var low: texture_2d<f32>;\nstruct Out {\n@builtin(position) pos: vec4f,\n@location(0) uv: vec2f,\n}\n@vertex fn vs(@builtin(vertex_index) i: u32) -> Out {\nlet p = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nvar o: Out;\no.pos = vec4f(p * 2.0 - 1.0, 0.0, 1.0);\no.uv = vec2f(p.x, 1.0 - p.y);\nreturn o;\n}\nfn luma(c: vec3f) -> f32 {\nreturn dot(c, vec3f(0.2126, 0.7152, 0.0722));\n}\nfn tap(uv: vec2f) -> vec3f {\nreturn min(textureSampleLevel(src, smp, uv, 0.0).rgb, vec3f(65000.0));\n}\nfn downsample(uv: vec2f, karis: bool) -> vec3f {\nlet t = 1.0 / vec2f(textureDimensions(src));\nlet a = tap(uv + t * vec2f(-2.0, -2.0));\nlet b = tap(uv + t * vec2f(0.0, -2.0));\nlet c = tap(uv + t * vec2f(2.0, -2.0));\nlet d = tap(uv + t * vec2f(-1.0, -1.0));\nlet e = tap(uv + t * vec2f(1.0, -1.0));\nlet f = tap(uv + t * vec2f(-2.0, 0.0));\nlet g = tap(uv);\nlet h = tap(uv + t * vec2f(2.0, 0.0));\nlet i = tap(uv + t * vec2f(-1.0, 1.0));\nlet j = tap(uv + t * vec2f(1.0, 1.0));\nlet k = tap(uv + t * vec2f(-2.0, 2.0));\nlet l = tap(uv + t * vec2f(0.0, 2.0));\nlet m = tap(uv + t * vec2f(2.0, 2.0));\nlet boxes = array<vec3f, 5>((d + e + i + j) * 0.25, (a + b + f + g) * 0.25, (b + c + g + h) * 0.25, (f + g + k + l) * 0.25, (g + h + l + m) * 0.25);\nlet weights = array<f32, 5>(0.5, 0.125, 0.125, 0.125, 0.125);\nvar sum = vec3f(0.0);\nvar total = 0.0;\nfor (var n = 0; n < 5; n++) {\nvar w = weights[n];\nif (karis) {\nw = w / (1.0 + luma(boxes[n]));\n}\nsum += boxes[n] * w;\ntotal += w;\n}\nreturn sum / total;\n}\n@fragment fn fs_prefilter(in: Out) -> @location(0) vec4f {\nlet c = downsample(in.uv, true);\nif (bloom.threshold <= 0.0) {\nreturn vec4f(c, 1.0);\n}\nlet br = max(c.r, max(c.g, c.b));\nlet k = bloom.threshold * bloom.knee + 1e-5;\nvar soft = clamp(br - bloom.threshold + k, 0.0, 2.0 * k);\nsoft = soft * soft / (4.0 * k);\nlet w = max(soft, br - bloom.threshold) / max(br, 1e-5);\nreturn vec4f(c * w, 1.0);\n}\n@fragment fn fs_down(in: Out) -> @location(0) vec4f {\nreturn vec4f(downsample(in.uv, false), 1.0);\n}\n@fragment fn fs_up(in: Out) -> @location(0) vec4f {\nlet t = bloom.radius / vec2f(textureDimensions(low));\nvar s = textureSampleLevel(low, smp, in.uv, 0.0).rgb * 4.0;\ns += (textureSampleLevel(low, smp, in.uv + vec2f(-t.x, 0.0), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(t.x, 0.0), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(0.0, -t.y), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(0.0, t.y), 0.0).rgb) * 2.0;\ns += textureSampleLevel(low, smp, in.uv - t, 0.0).rgb + textureSampleLevel(low, smp, in.uv + t, 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(-t.x, t.y), 0.0).rgb + textureSampleLevel(low, smp, in.uv + vec2f(t.x, -t.y), 0.0).rgb;\nreturn vec4f(textureSampleLevel(src, smp, in.uv, 0.0).rgb + s / 16.0, 1.0);\n}",
+};
+/** Linked from shaders/post/fxaa.wesl. */
+export const postFxaa = {
+    hash: '362146f9cb191613',
+    wgsl: "@group(0) @binding(0) var ldr: texture_2d<f32>;\n@group(0) @binding(1) var smp: sampler;\nstruct Out {\n@builtin(position) pos: vec4f,\n@location(0) uv: vec2f,\n}\n@vertex fn vs(@builtin(vertex_index) i: u32) -> Out {\nlet p = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nvar o: Out;\no.pos = vec4f(p * 2.0 - 1.0, 0.0, 1.0);\no.uv = vec2f(p.x, 1.0 - p.y);\nreturn o;\n}\nconst EDGE_MIN = 0.0312;\nconst EDGE_MAX = 0.125;\nconst SUBPIXEL = 0.75;\nconst STEPS = 10;\nfn luma_at(uv: vec2f) -> f32 {\nreturn dot(textureSampleLevel(ldr, smp, uv, 0.0).rgb, vec3f(0.299, 0.587, 0.114));\n}\nfn quality(i: i32) -> f32 {\nif (i < 5) {\nreturn 1.0;\n}\nif (i == 5) {\nreturn 1.5;\n}\nif (i < 9) {\nreturn 2.0;\n}\nreturn 4.0;\n}\n@fragment fn fs(in: Out) -> @location(0) vec4f {\nlet texel = 1.0 / vec2f(textureDimensions(ldr));\nlet uv = in.uv;\nlet centre = textureSampleLevel(ldr, smp, uv, 0.0);\nlet m = dot(centre.rgb, vec3f(0.299, 0.587, 0.114));\nlet n = luma_at(uv + vec2f(0.0, -texel.y));\nlet s = luma_at(uv + vec2f(0.0, texel.y));\nlet e = luma_at(uv + vec2f(texel.x, 0.0));\nlet w = luma_at(uv + vec2f(-texel.x, 0.0));\nlet lo = min(m, min(min(n, s), min(e, w)));\nlet hi = max(m, max(max(n, s), max(e, w)));\nlet range = hi - lo;\nif (range < max(EDGE_MIN, hi * EDGE_MAX)) {\nreturn centre;\n}\nlet nw = luma_at(uv + vec2f(-texel.x, -texel.y));\nlet ne = luma_at(uv + vec2f(texel.x, -texel.y));\nlet sw = luma_at(uv + vec2f(-texel.x, texel.y));\nlet se = luma_at(uv + vec2f(texel.x, texel.y));\nlet horizontal = abs(-2.0 * w + nw + sw) + 2.0 * abs(-2.0 * m + n + s) + abs(-2.0 * e + ne + se) >= abs(-2.0 * n + nw + ne) + 2.0 * abs(-2.0 * m + w + e) + abs(-2.0 * s + sw + se);\nlet l1 = select(w, n, horizontal);\nlet l2 = select(e, s, horizontal);\nlet g1 = l1 - m;\nlet g2 = l2 - m;\nlet steep1 = abs(g1) >= abs(g2);\nlet scaled = 0.25 * max(abs(g1), abs(g2));\nvar step_len = select(texel.x, texel.y, horizontal);\nvar avg = 0.5 * (l2 + m);\nif (steep1) {\nstep_len = -step_len;\navg = 0.5 * (l1 + m);\n}\nvar cur = uv;\nif (horizontal) {\ncur.y += step_len * 0.5;\n} else {\ncur.x += step_len * 0.5;\n}\nlet along = select(vec2f(0.0, texel.y), vec2f(texel.x, 0.0), horizontal);\nvar uv1 = cur - along;\nvar uv2 = cur + along;\nvar end1 = luma_at(uv1) - avg;\nvar end2 = luma_at(uv2) - avg;\nvar done1 = abs(end1) >= scaled;\nvar done2 = abs(end2) >= scaled;\nfor (var i = 1; i < STEPS && !(done1 && done2); i++) {\nif (!done1) {\nuv1 -= along * quality(i);\nend1 = luma_at(uv1) - avg;\ndone1 = abs(end1) >= scaled;\n}\nif (!done2) {\nuv2 += along * quality(i);\nend2 = luma_at(uv2) - avg;\ndone2 = abs(end2) >= scaled;\n}\n}\nlet d1 = select(uv.y - uv1.y, uv.x - uv1.x, horizontal);\nlet d2 = select(uv2.y - uv.y, uv2.x - uv.x, horizontal);\nlet nearer1 = d1 < d2;\nlet dist = min(d1, d2);\nlet pixel_offset = -dist / (d1 + d2) + 0.5;\nlet smaller = m < avg;\nlet right_way = (select(end2, end1, nearer1) < 0.0) != smaller;\nvar offset = select(0.0, pixel_offset, right_way);\nlet mean = (2.0 * (n + s + e + w) + nw + ne + sw + se) / 12.0;\nlet sub1 = clamp(abs(mean - m) / range, 0.0, 1.0);\nlet sub2 = (-2.0 * sub1 + 3.0) * sub1 * sub1;\noffset = max(offset, sub2 * sub2 * SUBPIXEL);\nvar at = uv;\nif (horizontal) {\nat.y += offset * step_len;\n} else {\nat.x += offset * step_len;\n}\nreturn vec4f(textureSampleLevel(ldr, smp, at, 0.0).rgb, 1.0);\n}",
+};
+/** Linked from shaders/post/taa.wesl. */
+export const postTaa = {
+    hash: '05819cf45a86a800',
+    wgsl: "struct TaaParams {\nreproject: mat4x4f,\nnear: f32,\nmotion: u32,\nblend: f32,\nreset: u32,\njitter_px: vec2f,\nupscale: u32,\npad: u32,\n}\n@group(0) @binding(0) var current: texture_2d<f32>;\n@group(0) @binding(1) var history: texture_2d<f32>;\n@group(0) @binding(2) var depth: texture_depth_2d;\n@group(0) @binding(3) var smp: sampler;\n@group(0) @binding(4) var<uniform> taa: TaaParams;\n@group(0) @binding(5) var motion_tex: texture_2d<f32>;\n@vertex fn vs(@builtin(vertex_index) i: u32) -> @builtin(position) vec4f {\nlet p = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nreturn vec4f(p * 2.0 - 1.0, 0.0, 1.0);\n}\nfn to_ycocg(c: vec3f) -> vec3f {\nreturn vec3f(dot(c, vec3f(0.25, 0.5, 0.25)), dot(c, vec3f(0.5, 0.0, -0.5)), dot(c, vec3f(-0.25, 0.5, -0.25)));\n}\nfn from_ycocg(c: vec3f) -> vec3f {\nreturn vec3f(c.x + c.y - c.z, c.x + c.z, c.x - c.y - c.z);\n}\nfn compress(c: vec3f) -> vec3f {\nreturn c / (1.0 + max(c.r, max(c.g, c.b)));\n}\nfn expand(c: vec3f) -> vec3f {\nreturn c / max(1.0 - max(c.r, max(c.g, c.b)), 1e-4);\n}\nconst FAR = 60000.0;\nfn view_depth(d: f32) -> f32 {\nreturn select(FAR, min(taa.near / d, FAR), d > 0.0);\n}\nfn same_surface(puv: vec2f, size: vec2f, expected: f32) -> bool {\nlet base = vec2i(floor(puv * size - 0.5));\nlet hi = vec2i(size) - 1;\nlet e = min(expected, FAR);\nfor (var k = 0; k < 4; k++) {\nlet q = clamp(base + vec2i(k & 1, k >> 1), vec2i(0), hi);\nlet h = textureLoad(history, q, 0).a;\nif (abs(h - e) <= 0.05 * e + 0.01) {\nreturn true;\n}\n}\nreturn false;\n}\nfn load(p: vec2i, size: vec2i) -> vec3f {\nlet c = textureLoad(current, clamp(p, vec2i(0), size - 1), 0).rgb;\nreturn select(vec3f(0.0), compress(min(c, vec3f(65000.0))), all(c == c));\n}\nfn history_at(uv: vec2f, size: vec2f) -> vec3f {\nlet pos = uv * size;\nlet centre = floor(pos - 0.5) + 0.5;\nlet f = pos - centre;\nlet w0 = f * (-0.5 + f * (1.0 - 0.5 * f));\nlet w1 = 1.0 + f * f * (-2.5 + 1.5 * f);\nlet w2 = f * (0.5 + f * (2.0 - 1.5 * f));\nlet w3 = f * f * (-0.5 + 0.5 * f);\nlet w12 = w1 + w2;\nlet t0 = (centre - 1.0) / size;\nlet t3 = (centre + 2.0) / size;\nlet t12 = (centre + w2 / w12) / size;\nvar c = textureSampleLevel(history, smp, vec2f(t12.x, t0.y), 0.0).rgb * (w12.x * w0.y);\nc += textureSampleLevel(history, smp, vec2f(t0.x, t12.y), 0.0).rgb * (w0.x * w12.y);\nc += textureSampleLevel(history, smp, vec2f(t12.x, t12.y), 0.0).rgb * (w12.x * w12.y);\nc += textureSampleLevel(history, smp, vec2f(t3.x, t12.y), 0.0).rgb * (w3.x * w12.y);\nc += textureSampleLevel(history, smp, vec2f(t12.x, t3.y), 0.0).rgb * (w12.x * w3.y);\nlet total = w12.x * w0.y + w0.x * w12.y + w12.x * w12.y + w3.x * w12.y + w12.x * w3.y;\nreturn max(c / total, vec3f(0.0));\n}\nfn upscaled(at: vec4f) -> vec4f {\nlet isize = vec2i(textureDimensions(current));\nlet size = vec2f(isize);\nlet osize = vec2f(textureDimensions(history));\nlet uv = at.xy / osize;\nlet ratio = osize / size;\nlet pin = uv * size;\nlet base = vec2i(floor(pin));\nvar m1 = vec3f(0.0);\nvar m2 = vec3f(0.0);\nvar sum = vec3f(0.0);\nvar wsum = 0.0;\nvar wmax = 0.0;\nfor (var y = -1; y <= 1; y++) {\nfor (var x = -1; x <= 1; x++) {\nlet q = base + vec2i(x, y);\nlet c = to_ycocg(load(q, isize));\nm1 += c;\nm2 += c * c;\nlet s = (vec2f(q) + 0.5 - taa.jitter_px - pin) * ratio;\nlet w = exp(-2.29 * dot(s, s));\nsum += c * w;\nwsum += w;\nwmax = max(wmax, w);\n}\n}\nlet here = sum / max(wsum, 1e-5);\nlet pd = clamp(base, vec2i(0), isize - 1);\nlet d = textureLoad(depth, pd, 0);\nlet fresh = vec4f(expand(from_ycocg(here)), view_depth(d));\nif (taa.reset != 0u) {\nreturn fresh;\n}\nvar puv: vec2f;\nvar expected: f32;\nlet mv = select(vec4f(0.0), textureLoad(motion_tex, pd, 0), taa.motion != 0u);\nif (mv.a > 0.5) {\npuv = uv - mv.xy;\nexpected = mv.z;\n} else {\nlet ndc = vec2f(uv.x * 2.0 - 1.0, 1.0 - uv.y * 2.0);\nlet prev = taa.reproject * vec4f(ndc, d, 1.0);\nif (prev.w <= 0.0 && d > 0.0) {\nreturn fresh;\n}\nlet pn = prev.xy / prev.w;\npuv = vec2f(pn.x * 0.5 + 0.5, 0.5 - pn.y * 0.5);\nexpected = select(FAR, prev.w * view_depth(d), d > 0.0);\n}\nif (any(puv < vec2f(0.0)) || any(puv > vec2f(1.0)) || !same_surface(puv, osize, expected)) {\nreturn fresh;\n}\nlet hist = to_ycocg(compress(history_at(puv, osize)));\nlet mean = m1 / 9.0;\nlet sigma = sqrt(max(m2 / 9.0 - mean * mean, vec3f(0.0)));\nlet half_box = 1.25 * sigma + 1e-5;\nlet off = hist - mean;\nlet units = abs(off / half_box);\nlet k = max(units.x, max(units.y, units.z));\nlet clipped = select(hist, mean + off / k, k > 1.0);\nlet out = mix(clipped, here, taa.blend * wmax);\nreturn vec4f(expand(from_ycocg(out)), fresh.a);\n}\n@fragment fn fs(@builtin(position) at: vec4f) -> @location(0) vec4f {\nif (taa.upscale != 0u) {\nreturn upscaled(at);\n}\nlet isize = vec2i(textureDimensions(current));\nlet size = vec2f(isize);\nlet p = vec2i(at.xy);\nvar m1 = vec3f(0.0);\nvar m2 = vec3f(0.0);\nvar here = vec3f(0.0);\nfor (var y = -1; y <= 1; y++) {\nfor (var x = -1; x <= 1; x++) {\nlet c = to_ycocg(load(p + vec2i(x, y), isize));\nm1 += c;\nm2 += c * c;\nif (x == 0 && y == 0) {\nhere = c;\n}\n}\n}\nlet d = textureLoad(depth, p, 0);\nlet fresh = vec4f(expand(from_ycocg(here)), view_depth(d));\nif (taa.reset != 0u) {\nreturn fresh;\n}\nvar puv: vec2f;\nvar expected: f32;\nlet mv = select(vec4f(0.0), textureLoad(motion_tex, p, 0), taa.motion != 0u);\nif (mv.a > 0.5) {\npuv = at.xy / size - mv.xy;\nexpected = mv.z;\n} else {\nlet ndc = vec2f(at.x / size.x * 2.0 - 1.0, 1.0 - at.y / size.y * 2.0);\nlet prev = taa.reproject * vec4f(ndc, d, 1.0);\nif (prev.w <= 0.0 && d > 0.0) {\nreturn fresh;\n}\nlet pn = prev.xy / prev.w;\npuv = vec2f(pn.x * 0.5 + 0.5, 0.5 - pn.y * 0.5);\nexpected = select(FAR, prev.w * view_depth(d), d > 0.0);\n}\nif (any(puv < vec2f(0.0)) || any(puv > vec2f(1.0)) || !same_surface(puv, size, expected)) {\nreturn fresh;\n}\nlet hist = to_ycocg(compress(history_at(puv, size)));\nlet mean = m1 / 9.0;\nlet sigma = sqrt(max(m2 / 9.0 - mean * mean, vec3f(0.0)));\nlet half_box = 1.25 * sigma + 1e-5;\nlet off = hist - mean;\nlet units = abs(off / half_box);\nlet k = max(units.x, max(units.y, units.z));\nlet clipped = select(hist, mean + off / k, k > 1.0);\nlet out = mix(clipped, here, taa.blend);\nreturn vec4f(expand(from_ycocg(out)), fresh.a);\n}",
+};
+/** Linked from shaders/post/taa_post.wesl. */
+export const postTaaPost = {
+    hash: '5806c7af9e6201d4',
+    wgsl: "@group(0) @binding(0) var color: texture_2d<f32>;\n@group(0) @binding(1) var motion: texture_2d<f32>;\n@group(0) @binding(2) var smp: sampler;\nstruct Params {\nsharpen: f32,\nblur: f32,\nsamples: f32,\nhas_motion: f32,\n}\n@group(0) @binding(3) var<uniform> p: Params;\nstruct Out {\n@builtin(position) pos: vec4f,\n@location(0) uv: vec2f,\n}\n@vertex fn vs(@builtin(vertex_index) i: u32) -> Out {\nlet q = vec2f(f32((i << 1u) & 2u), f32(i & 2u));\nvar o: Out;\no.pos = vec4f(q * 2.0 - 1.0, 0.0, 1.0);\no.uv = vec2f(q.x, 1.0 - q.y);\nreturn o;\n}\nfn load(px: vec2i, size: vec2i) -> vec3f {\nreturn textureLoad(color, clamp(px, vec2i(0), size - 1), 0).rgb;\n}\nfn tame(c: vec3f) -> vec3f {\nreturn c / (1.0 + c);\n}\n@fragment fn fs(in: Out) -> @location(0) vec4f {\nlet size = vec2i(textureDimensions(color));\nlet px = vec2i(in.pos.xy);\nlet c = load(px, size);\nvar out = c;\nif (p.sharpen > 0.0) {\nlet n = load(px + vec2i(0, -1), size);\nlet s = load(px + vec2i(0, 1), size);\nlet e = load(px + vec2i(1, 0), size);\nlet w = load(px + vec2i(-1, 0), size);\nlet tc = tame(c);\nlet mn = min(tc, min(min(tame(n), tame(s)), min(tame(e), tame(w))));\nlet mx = max(tc, max(max(tame(n), tame(s)), max(tame(e), tame(w))));\nlet amp = sqrt(clamp(min(mn, 1.0 - mx) / max(mx, vec3f(1e-4)), vec3f(0.0), vec3f(1.0)));\nlet k = amp * -mix(0.125, 0.2, clamp(p.sharpen, 0.0, 1.0));\nout = max((c + (n + s + e + w) * k) / (1.0 + 4.0 * k), vec3f(0.0));\n}\nif (p.blur > 0.0 && p.has_motion > 0.5) {\nlet msize = vec2i(textureDimensions(motion));\nlet m = textureLoad(motion, clamp(vec2i(vec2f(px) * vec2f(msize) / vec2f(size)), vec2i(0), msize - 1), 0);\nlet v = select(vec2f(0.0), m.rg * p.blur, m.a > 0.0);\nlet len_px = length(v * vec2f(size));\nif (len_px > 0.5) {\nlet n = i32(clamp(p.samples, 2.0, 32.0));\nvar sum = vec3f(0.0);\nfor (var k = 0; k < n; k++) {\nlet t = (f32(k) + 0.5) / f32(n) - 0.5;\nsum += textureSampleLevel(color, smp, in.uv - v * t, 0.0).rgb;\n}\nlet keep = clamp(1.0 - len_px / f32(n), 0.0, 1.0);\nout = mix(sum / f32(n), out, keep);\n}\n}\nreturn vec4f(out, 1.0);\n}",
+};
+/** Linked from shaders/scene/hzb.wesl. */
+export const sceneHzb = {
+    hash: 'b26b6bc789bda1de',
+    wgsl: "@group(0) @binding(0) var depth: texture_depth_2d;\n@group(0) @binding(1) var src: texture_2d<f32>;\n@group(0) @binding(2) var dst: texture_storage_2d<r32float, write>;\n@compute @workgroup_size(8, 8) fn hzb_first(@builtin(global_invocation_id) id: vec3u) {\nlet size = textureDimensions(dst);\nif (id.x >= size.x || id.y >= size.y) {\nreturn;\n}\nlet hi = vec2i(textureDimensions(depth)) - 1;\nvar d = 1.0;\nfor (var y = 0; y < 3; y++) {\nfor (var x = 0; x < 3; x++) {\nd = min(d, textureLoad(depth, min(vec2i(id.xy) * 2 + vec2i(x, y), hi), 0));\n}\n}\ntextureStore(dst, id.xy, vec4f(d, 0.0, 0.0, 0.0));\n}\n@compute @workgroup_size(8, 8) fn hzb_down(@builtin(global_invocation_id) id: vec3u) {\nlet size = textureDimensions(dst);\nif (id.x >= size.x || id.y >= size.y) {\nreturn;\n}\nlet hi = vec2i(textureDimensions(src)) - 1;\nvar d = 1.0;\nfor (var y = 0; y < 3; y++) {\nfor (var x = 0; x < 3; x++) {\nd = min(d, textureLoad(src, min(vec2i(id.xy) * 2 + vec2i(x, y), hi), 0).r);\n}\n}\ntextureStore(dst, id.xy, vec4f(d, 0.0, 0.0, 0.0));\n}",
+};
+/** Linked from shaders/scene/motion.wesl. */
+export const postMotion = {
+    hash: '407e14423b3c1d1b',
+    wgsl: "struct Motion {\nview_proj: mat4x4f,\nprev_view_proj: mat4x4f,\nprev_origin: vec4f,\n}\n@group(0) @binding(0) var<uniform> camera: Camera;\n@group(0) @binding(1) var<storage, read> instances: array<vec4u>;\n@group(0) @binding(2) var<storage, read> visible: array<u32>;\n@group(0) @binding(3) var<storage, read> prev_instances: array<vec4u>;\n@group(0) @binding(4) var<uniform> motion: Motion;\n@group(0) @binding(5) var palette: texture_2d<f32>;\n@group(0) @binding(6) var prev_palette: texture_2d<f32>;\n@group(0) @binding(7) var<storage, read> prev_pool: array<f32>;\n@group(0) @binding(8) var<storage, read> prev_skinned: array<f32>;\nstruct Out {\n@builtin(position) clip: vec4f,\n@location(0) now: vec4f,\n@location(1) then: vec4f,\n}\nstruct Pose {\npos: vec3f,\nscale: f32,\nrot: vec4f,\n}\nfn pose_now(i: u32) -> Pose {\nlet a = bitcast<vec4f>(instances[2u * i]);\nreturn Pose(a.xyz - camera.origin.xyz, a.w, bitcast<vec4f>(instances[2u * i + 1u]));\n}\nfn pose_then(i: u32) -> Pose {\nlet a = bitcast<vec4f>(prev_instances[2u * i]);\nreturn Pose(a.xyz - motion.prev_origin.xyz, a.w, bitcast<vec4f>(prev_instances[2u * i + 1u]));\n}\nfn skin_row(i: u32) -> u32 {\nlet cap = arrayLength(&instances) / 3u;\nreturn instances[2u * cap + i].w;\n}\nfn place(now: Pose, then: Pose, p_now: vec3f, p_then: vec3f) -> Out {\nlet w_now = quat_rotate(now.rot, p_now * now.scale) + now.pos;\nlet w_then = quat_rotate(then.rot, p_then * then.scale) + then.pos;\nvar o: Out;\no.clip = camera.view_proj * vec4f(w_now, 1.0);\no.now = motion.view_proj * vec4f(w_now, 1.0);\no.then = motion.prev_view_proj * vec4f(w_then, 1.0);\nreturn o;\n}\n@vertex fn vs(@location(0) position: vec3f, @builtin(vertex_index) v: u32, @builtin(instance_index) slot: u32) -> Out {\nlet i = visible[slot];\nreturn place(pose_now(i), pose_then(i), position, vec3f(prev_pool[v * 8u], prev_pool[v * 8u + 1u], prev_pool[v * 8u + 2u]));\n}\n@vertex fn vs_skinned(@location(0) position: vec3f, @location(6) joints: vec4u, @location(7) weights: vec4f, @builtin(vertex_index) v: u32, @builtin(instance_index) slot: u32) -> Out {\nlet i = visible[slot];\nlet row = skin_row(i);\nlet n = vec3f(0.0, 0.0, 1.0);\nlet now = skin_vertex(palette, row, joints, weights, position, n);\nlet then = skin_vertex(prev_palette, row, joints, weights, vec3f(prev_skinned[v * 8u], prev_skinned[v * 8u + 1u], prev_skinned[v * 8u + 2u]), n);\nreturn place(pose_now(i), pose_then(i), now.position, then.position);\n}\n@fragment fn fs(in: Out) -> @location(0) vec4f {\nlet now = in.now.xy / in.now.w;\nlet then = in.then.xy / max(in.then.w, 1e-6);\nreturn vec4f((now - then) * vec2f(0.5, -0.5), in.then.w, 1.0);\n}\nstruct Camera {\nview_proj: mat4x4f,\nfrustum: array<vec4f, 6>,\nlight_dir: vec4f,\norigin: vec4f,\n}\nfn quat_rotate(q: vec4f, v: vec3f) -> vec3f {\nlet t = 2.0 * cross(q.xyz, v);\nreturn v + q.w * t + cross(q.xyz, t);\n}\nfn skin_vertex(palette: texture_2d<f32>, row: u32, joints: vec4u, weights: vec4f, position: vec3f, normal: vec3f) -> Skinned {\nif (row == 0u || dot(weights, vec4f(1.0)) <= 0.0) {\nreturn Skinned(position, normal);\n}\nlet base = row - 1u;\nlet t = (vec4u(base) + joints) * 3u;\nlet r0 = weights.x * palette_texel(palette, t.x) + weights.y * palette_texel(palette, t.y) + weights.z * palette_texel(palette, t.z) + weights.w * palette_texel(palette, t.w);\nlet r1 = weights.x * palette_texel(palette, t.x + 1u) + weights.y * palette_texel(palette, t.y + 1u) + weights.z * palette_texel(palette, t.z + 1u) + weights.w * palette_texel(palette, t.w + 1u);\nlet r2 = weights.x * palette_texel(palette, t.x + 2u) + weights.y * palette_texel(palette, t.y + 2u) + weights.z * palette_texel(palette, t.z + 2u) + weights.w * palette_texel(palette, t.w + 2u);\nlet p = vec4f(position, 1.0);\nreturn Skinned(vec3f(dot(r0, p), dot(r1, p), dot(r2, p)), vec3f(dot(r0.xyz, normal), dot(r1.xyz, normal), dot(r2.xyz, normal)));\n}\nfn palette_texel(palette: texture_2d<f32>, i: u32) -> vec4f {\nreturn textureLoad(palette, vec2u(i % PALETTE_WIDTH, i / PALETTE_WIDTH), 0);\n}\nstruct Skinned {\nposition: vec3f,\nnormal: vec3f,\n}\nconst PALETTE_WIDTH: u32 = 1024u;",
+};
+/** Linked from shaders/scene/clusters.wesl. */
+export const sceneClusters = {
+    hash: 'c7892829a492870a',
+    wgsl: "struct Cluster {\ncenter: vec3f,\nradius: f32,\napex: vec3f,\ncutoff: f32,\naxis: vec3f,\nbase_vertex: u32,\nvert_offset: u32,\ntri_offset: u32,\ntri_count: u32,\nflags: u32,\nself_sphere: vec4f,\nparent_sphere: vec4f,\nself_error: f32,\nparent_error: f32,\nstream: u32,\nkids: u32,\npad1: vec4f,\n}\nstruct Job {\nargs_slot: u32,\nstart: u32,\ncluster_base: u32,\ncluster_count: u32,\nprefix: u32,\ndouble_sided: u32,\npad0: u32,\npad1: u32,\n}\nstruct Params {\njobs: u32,\ntotal: u32,\nhalf: u32,\nindex_half: u32,\n}\n@group(0) @binding(0) var<uniform> camera: Camera;\n@group(0) @binding(1) var<storage, read> instances: array<vec4u>;\n@group(0) @binding(2) var<storage, read> visible: array<u32>;\n@group(0) @binding(3) var<storage, read> args: array<u32>;\n@group(0) @binding(4) var<storage, read> clusters: array<Cluster>;\n@group(0) @binding(5) var<storage, read> jobs: array<Job>;\n@group(0) @binding(6) var<uniform> params: Params;\n@group(0) @binding(7) var<storage, read_write> out_args: array<atomic<u32>, 12>;\n@group(0) @binding(8) var<storage, read_write> out_visible: array<vec4u>;\n@group(0) @binding(9) var<storage, read_write> out_indices: array<u32>;\n@group(0) @binding(10) var<storage, read> triangles: array<u32>;\nfn load_instance(i: u32) -> Instance {\nlet cap = arrayLength(&instances) / 3u;\nlet a = bitcast<vec4f>(instances[2u * i]);\nlet b = bitcast<vec4f>(instances[2u * i + 1u]);\nlet s = instances[2u * cap + i];\nreturn Instance(a.xyz - camera.origin.xyz, a.w, b, s.x, s.y, s.z, s.w);\n}\nconst SLOT: u32 = 128u;\nstruct LodParams {\ncount: u32,\nlod_scale: f32,\nerror_px: f32,\nmin_radius_px: f32,\n}\n@group(0) @binding(16) var<uniform> lodp: LodParams;\nfn error_px(inst: Instance, s: vec4f, e: f32) -> f32 {\nlet c = inst.pos + quat_rotate(inst.rot, s.xyz * inst.scale);\nlet d = max(length(c) - s.w * inst.scale, 1e-3);\nreturn e * inst.scale * lodp.lod_scale / d;\n}\nconst STREAM_MISSING: u32 = 1u;\nconst STREAM_LEAF: u32 = 2u;\nfn on_cut(inst: Instance, cl: Cluster) -> bool {\nif ((cl.stream & STREAM_MISSING) != 0u) {\nreturn false;\n}\nif (cl.self_error <= 0.0 && cl.parent_error == 0.0) {\nreturn true;\n}\nif ((cl.stream & STREAM_LEAF) == 0u && error_px(inst, cl.self_sphere, cl.self_error) > lodp.error_px) {\nreturn false;\n}\nreturn cl.parent_error < 0.0 || error_px(inst, cl.parent_sphere, cl.parent_error) > lodp.error_px;\n}\nfn job_entry(id: vec3u) -> vec4u {\nlet x = id.x;\nif (x >= params.total) {\nreturn vec4u(0u);\n}\nvar j = 0u;\nfor (; j < params.jobs; j++) {\nif (x < jobs[j].prefix + jobs[j].cluster_count) {\nbreak;\n}\n}\nif (j == params.jobs) {\nreturn vec4u(0u);\n}\nlet job = jobs[j];\nif (id.y >= args[job.args_slot * 5u + 1u]) {\nreturn vec4u(0u);\n}\nreturn vec4u(visible[job.start + id.y], job.cluster_base + (x - job.prefix), 1u, 0u);\n}\nfn sphere_in_view(inst: Instance, cl: Cluster) -> vec4f {\nif (!on_cut(inst, cl)) {\nreturn vec4f(0.0, 0.0, 0.0, -1.0);\n}\nlet c = inst.pos + quat_rotate(inst.rot, cl.center * inst.scale);\nlet r = cl.radius * inst.scale;\nfor (var p = 0u; p < 6u; p++) {\nlet plane = camera.frustum[p];\nif (dot(plane.xyz, c) + plane.w < -r) {\nreturn vec4f(0.0, 0.0, 0.0, -1.0);\n}\n}\nif (cl.flags == 0u && cl.cutoff < 1.0) {\nlet apex = inst.pos + quat_rotate(inst.rot, cl.apex * inst.scale);\nif (dot(normalize(apex), quat_rotate(inst.rot, cl.axis)) >= cl.cutoff) {\nreturn vec4f(0.0, 0.0, 0.0, -1.0);\n}\n}\nreturn vec4f(c, r);\n}\n@group(0) @binding(21) var<storage, read_write> page_bits: array<atomic<u32>>;\nfn note_pages(index: u32, cl: Cluster) {\nif (cl.stream == 0u && cl.kids == 0u) {\nreturn;\n}\nlet half = arrayLength(&page_bits) / 2u;\nlet page = cl.stream >> 2u;\natomicOr(&page_bits[half + page / 32u], 1u << (page % 32u));\nif ((cl.stream & STREAM_LEAF) != 0u && cl.kids > 0u && error_px(load_instance(index), cl.self_sphere, cl.self_error) > lodp.error_px) {\nlet want = cl.kids - 1u;\natomicOr(&page_bits[want / 32u], 1u << (want % 32u));\n}\n}\nfn emit(index: u32, ci: u32, cl: Cluster) {\nnote_pages(index, cl);\nlet side = cl.flags;\nlet k = atomicAdd(&out_args[10u + side], 1u);\nlet at = side * params.index_half + atomicAdd(&out_args[side * 5u], cl.tri_count * 3u);\nout_visible[side * params.half + k] = vec4u(index, ci, at, k);\n}\n@compute @workgroup_size(64) fn cull_clusters(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet cl = clusters[e.y];\nif (sphere_in_view(load_instance(e.x), cl).w >= 0.0) {\nemit(e.x, e.y, cl);\n}\n}\n@group(0) @binding(12) var<uniform> occl: Occlusion;\n@group(0) @binding(13) var hzb: texture_2d<f32>;\n@group(0) @binding(14) var<storage, read_write> recheck: array<atomic<u32>>;\n@group(0) @binding(15) var<storage, read_write> recheck_dispatch: array<u32, 3>;\n@compute @workgroup_size(64) fn cull_clusters_first(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet cl = clusters[e.y];\nlet s = sphere_in_view(load_instance(e.x), cl);\nif (s.w < 0.0) {\nreturn;\n}\nif (hidden_in(hzb, occl.hzb_size, occl.hzb_levels, s.xyz, s.w, occl.prev)) {\nlet k = atomicAdd(&recheck[0], 1u);\nif (2u * k + 2u < arrayLength(&recheck)) {\natomicStore(&recheck[2u * k + 1u], e.x);\natomicStore(&recheck[2u * k + 2u], e.y);\n}\nreturn;\n}\nemit(e.x, e.y, cl);\n}\n@compute @workgroup_size(64) fn cull_clusters_second(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet cl = clusters[e.y];\nlet s = sphere_in_view(load_instance(e.x), cl);\nif (s.w >= 0.0 && !hidden_in(hzb, occl.hzb_size, occl.hzb_levels, s.xyz, s.w, camera.view_proj)) {\nemit(e.x, e.y, cl);\n}\n}\n@compute @workgroup_size(1) fn size_recheck() {\nlet n = min(atomicLoad(&recheck[0]), (arrayLength(&recheck) - 1u) / 2u);\nrecheck_dispatch[0] = (n + 63u) / 64u;\nrecheck_dispatch[1] = 1u;\nrecheck_dispatch[2] = 1u;\n}\n@compute @workgroup_size(64) fn recheck_clusters(@builtin(global_invocation_id) id: vec3u) {\nlet n = min(atomicLoad(&recheck[0]), (arrayLength(&recheck) - 1u) / 2u);\nif (id.x >= n) {\nreturn;\n}\nlet index = atomicLoad(&recheck[2u * id.x + 1u]);\nlet ci = atomicLoad(&recheck[2u * id.x + 2u]);\nlet cl = clusters[ci];\nlet s = sphere_in_view(load_instance(index), cl);\nif (s.w >= 0.0 && !hidden_in(hzb, occl.hzb_size, occl.hzb_levels, s.xyz, s.w, camera.view_proj)) {\nemit(index, ci, cl);\n}\n}\nstruct ShadowCull {\nplanes: array<vec4f, 20>,\nlight: vec4f,\n}\n@group(0) @binding(17) var<uniform> shc: ShadowCull;\n@group(0) @binding(18) var<storage, read_write> sh_args: array<atomic<u32>, 32>;\n@group(0) @binding(19) var<storage, read_write> sh_visible: array<vec4u>;\n@compute @workgroup_size(64) fn cull_clusters_shadow(@builtin(global_invocation_id) id: vec3u) {\nlet e = job_entry(id);\nif (e.z == 0u) {\nreturn;\n}\nlet c = id.z;\nlet cl = clusters[e.y];\nlet inst = load_instance(e.x);\nif (!on_cut(inst, cl)) {\nreturn;\n}\nlet center = inst.pos + quat_rotate(inst.rot, cl.center * inst.scale);\nlet r = cl.radius * inst.scale;\nfor (var p = 0u; p < 5u; p++) {\nlet plane = shc.planes[c * 5u + p];\nif (dot(plane.xyz, center) + plane.w < -r) {\nreturn;\n}\n}\nif (cl.flags == 0u && cl.cutoff < 1.0 && dot(-shc.light.xyz, quat_rotate(inst.rot, cl.axis)) >= cl.cutoff) {\nreturn;\n}\nlet side = cl.flags;\nlet b = c * 2u + side;\nlet k = atomicAdd(&sh_args[b * 4u + 1u], 1u);\nsh_visible[c * params.index_half + side * params.half + k] = vec4u(e.x, e.y, 0u, 0u);\n}\n@group(0) @binding(11) var<storage, read_write> dispatch: array<u32, 3>;\n@compute @workgroup_size(1) fn size_expand() {\ndispatch[0] = atomicLoad(&out_args[10]) + atomicLoad(&out_args[11]);\ndispatch[1] = 1u;\ndispatch[2] = 1u;\n}\n@compute @workgroup_size(128) fn expand_clusters(@builtin(workgroup_id) wg: vec3u, @builtin(local_invocation_index) t: u32) {\nlet single = atomicLoad(&out_args[10]);\nlet g = wg.x;\nlet rec = out_visible[select(params.half + g - single, g, g < single)];\nlet cl = clusters[rec.y];\nif (t >= cl.tri_count) {\nreturn;\n}\nfor (var c = 0u; c < 3u; c++) {\nlet b = cl.tri_offset + t * 3u + c;\nout_indices[rec.z + t * 3u + c] = rec.w * SLOT + ((triangles[b / 4u] >> ((b % 4u) * 8u)) & 0xffu);\n}\n}\nstruct Camera {\nview_proj: mat4x4f,\nfrustum: array<vec4f, 6>,\nlight_dir: vec4f,\norigin: vec4f,\n}\nstruct Instance {\npos: vec3f,\nscale: f32,\nrot: vec4f,\ncolor: u32,\nmesh: u32,\nmaterial: u32,\nskin: u32,\n}\nfn quat_rotate(q: vec4f, v: vec3f) -> vec3f {\nlet t = 2.0 * cross(q.xyz, v);\nreturn v + q.w * t + cross(q.xyz, t);\n}\nstruct Occlusion {\nprev: mat4x4f,\nfrustum: array<vec4f, 6>,\nhzb_size: vec2f,\nhzb_levels: u32,\nshadows: u32,\nslots: u32,\nregion: u32,\npad0: u32,\npad1: u32,\n}\nfn hidden_in(hzb: texture_2d<f32>, size: vec2f, levels: u32, c: vec3f, r: f32, view_proj: mat4x4f) -> bool {\nif (levels == 0u) {\nreturn false;\n}\nvar lo = vec2f(1e9);\nvar hi = vec2f(-1e9);\nvar nearest = 0.0;\nfor (var k = 0u; k < 8u; k++) {\nlet corner = c + r * vec3f(select(-1.0, 1.0, (k & 1u) != 0u), select(-1.0, 1.0, (k & 2u) != 0u), select(-1.0, 1.0, (k & 4u) != 0u));\nlet p = view_proj * vec4f(corner, 1.0);\nif (p.w <= 1e-4) {\nreturn false;\n}\nlet n = p.xyz / p.w;\nlo = min(lo, n.xy);\nhi = max(hi, n.xy);\nnearest = max(nearest, n.z);\n}\nif (any(hi < vec2f(-1.0)) || any(lo > vec2f(1.0))) {\nreturn false;\n}\nlet uv_lo = clamp(vec2f(lo.x, -hi.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet uv_hi = clamp(vec2f(hi.x, -lo.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet extent = (uv_hi - uv_lo) * size;\nlet level = min(u32(max(ceil(log2(max(max(extent.x, extent.y), 1.0))), 0.0)), levels - 1u);\nlet dims = vec2f(textureDimensions(hzb, level));\nlet t0 = vec2i(uv_lo * dims);\nlet t1 = min(vec2i(uv_hi * dims), vec2i(dims) - 1);\nvar farthest = 1.0;\nfor (var y = t0.y; y <= min(t1.y, t0.y + 2); y++) {\nfor (var x = t0.x; x <= min(t1.x, t0.x + 2); x++) {\nfarthest = min(farthest, textureLoad(hzb, vec2i(x, y), i32(level)).r);\n}\n}\nreturn nearest < farthest;\n}",
+};
+/** Linked from shaders/scene/occlusion_cull.wesl. */
+export const sceneOcclusionCull = {
+    hash: 'd66ce30369f82685',
+    wgsl: "@group(0) @binding(7) var<uniform> occl: Occlusion;\n@group(0) @binding(8) var hzb: texture_2d<f32>;\n@group(0) @binding(9) var<storage, read_write> recheck: array<atomic<u32>>;\nfn hidden(c: vec3f, r: f32, view_proj: mat4x4f) -> bool {\nreturn hidden_in(hzb, occl.hzb_size, occl.hzb_levels, c, r, view_proj);\n}\n@compute @workgroup_size(64) fn cull_occlusion(@builtin(global_invocation_id) id: vec3u) {\nlet i = id.x;\nif (i >= params.count) {\nreturn;\n}\nlet inst = load_instance(i);\nlet m = meshes[inst.mesh];\nlet c = inst.pos + quat_rotate(inst.rot, m.center * inst.scale);\nlet r = m.radius * inst.scale;\nif (occl.shadows != 0u) {\nlet dst = place(inst);\nif (dst != 0xffffffffu) {\nvisible[dst] = i;\n}\n}\nfor (var p = 0u; p < 6u; p++) {\nlet plane = occl.frustum[p];\nif (dot(plane.xyz, c) + plane.w < -r) {\nreturn;\n}\n}\nif (hidden(c, r, occl.prev)) {\nlet k = atomicAdd(&recheck[0], 1u);\natomicStore(&recheck[k + 1u], i);\nreturn;\n}\nlet dst = place_lod(inst, c, r, occl.slots, occl.region);\nif (dst != 0xffffffffu) {\nvisible[dst] = i;\n}\n}\n@compute @workgroup_size(64) fn cull_recheck(@builtin(global_invocation_id) id: vec3u) {\nif (id.x >= atomicLoad(&recheck[0])) {\nreturn;\n}\nlet i = atomicLoad(&recheck[id.x + 1u]);\nlet inst = load_instance(i);\nlet m = meshes[inst.mesh];\nlet c = inst.pos + quat_rotate(inst.rot, m.center * inst.scale);\nlet r = m.radius * inst.scale;\nif (hidden(c, r, camera.view_proj)) {\nreturn;\n}\nlet dst = place_lod(inst, c, r, 2u * occl.slots, 2u * occl.region);\nif (dst != 0xffffffffu) {\nvisible[dst] = i;\n}\n}\nstruct Occlusion {\nprev: mat4x4f,\nfrustum: array<vec4f, 6>,\nhzb_size: vec2f,\nhzb_levels: u32,\nshadows: u32,\nslots: u32,\nregion: u32,\npad0: u32,\npad1: u32,\n}\nfn hidden_in(hzb: texture_2d<f32>, size: vec2f, levels: u32, c: vec3f, r: f32, view_proj: mat4x4f) -> bool {\nif (levels == 0u) {\nreturn false;\n}\nvar lo = vec2f(1e9);\nvar hi = vec2f(-1e9);\nvar nearest = 0.0;\nfor (var k = 0u; k < 8u; k++) {\nlet corner = c + r * vec3f(select(-1.0, 1.0, (k & 1u) != 0u), select(-1.0, 1.0, (k & 2u) != 0u), select(-1.0, 1.0, (k & 4u) != 0u));\nlet p = view_proj * vec4f(corner, 1.0);\nif (p.w <= 1e-4) {\nreturn false;\n}\nlet n = p.xyz / p.w;\nlo = min(lo, n.xy);\nhi = max(hi, n.xy);\nnearest = max(nearest, n.z);\n}\nif (any(hi < vec2f(-1.0)) || any(lo > vec2f(1.0))) {\nreturn false;\n}\nlet uv_lo = clamp(vec2f(lo.x, -hi.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet uv_hi = clamp(vec2f(hi.x, -lo.y) * 0.5 + 0.5, vec2f(0.0), vec2f(1.0));\nlet extent = (uv_hi - uv_lo) * size;\nlet level = min(u32(max(ceil(log2(max(max(extent.x, extent.y), 1.0))), 0.0)), levels - 1u);\nlet dims = vec2f(textureDimensions(hzb, level));\nlet t0 = vec2i(uv_lo * dims);\nlet t1 = min(vec2i(uv_hi * dims), vec2i(dims) - 1);\nvar farthest = 1.0;\nfor (var y = t0.y; y <= min(t1.y, t0.y + 2); y++) {\nfor (var x = t0.x; x <= min(t1.x, t0.x + 2); x++) {\nfarthest = min(farthest, textureLoad(hzb, vec2i(x, y), i32(level)).r);\n}\n}\nreturn nearest < farthest;\n}\n@group(0) @binding(4) var<uniform> params: CullParams;\nfn load_instance(i: u32) -> Instance {\nlet cap = arrayLength(&instances) / 3u;\nlet a = bitcast<vec4f>(instances[2u * i]);\nlet b = bitcast<vec4f>(instances[2u * i + 1u]);\nlet s = instances[2u * cap + i];\nreturn Instance(a.xyz - camera.origin.xyz, a.w, b, s.x, s.y, s.z, s.w);\n}\n@group(0) @binding(2) var<storage, read> meshes: array<MeshInfo>;\nfn quat_rotate(q: vec4f, v: vec3f) -> vec3f {\nlet t = 2.0 * cross(q.xyz, v);\nreturn v + q.w * t + cross(q.xyz, t);\n}\nfn place(inst: Instance) -> u32 {\nlet m = meshes[inst.mesh];\nlet c = inst.pos + quat_rotate(inst.rot, m.center * inst.scale);\nlet r = m.radius * inst.scale;\nfor (var p = 0u; p < 6u; p++) {\nlet plane = camera.frustum[p];\nif (dot(plane.xyz, c) + plane.w < -r) {\nreturn 0xffffffffu;\n}\n}\nreturn place_lod(inst, c, r, 0u, 0u);\n}\n@group(0) @binding(5) var<storage, read_write> visible: array<u32>;\nfn place_lod(inst: Instance, c: vec3f, r: f32, slots: u32, region: u32) -> u32 {\nlet m = meshes[inst.mesh];\nlet px = params.lod_scale / max(length(c) - r, 1e-3);\nif (r * px < params.min_radius_px) {\nreturn 0xffffffffu;\n}\nvar lod = 0u;\nfor (var l = 1u; l < m.lod_count; l++) {\nif (m.error[l / 4u][l % 4u] * inst.scale * px > params.error_px) {\nbreak;\n}\nlod = l;\n}\nlet k = atomicAdd(&args[m.draw_base + lod + slots].instance_count, 1u);\nreturn m.instance_base + lod * m.instance_count + k + region;\n}\nstruct Instance {\npos: vec3f,\nscale: f32,\nrot: vec4f,\ncolor: u32,\nmesh: u32,\nmaterial: u32,\nskin: u32,\n}\n@group(0) @binding(0) var<uniform> camera: Camera;\nstruct Camera {\nview_proj: mat4x4f,\nfrustum: array<vec4f, 6>,\nlight_dir: vec4f,\norigin: vec4f,\n}\n@group(0) @binding(3) var<storage, read_write> args: array<DrawArgs>;\nstruct DrawArgs {\nindex_count: u32,\ninstance_count: atomic<u32>,\nfirst_index: u32,\nbase_vertex: i32,\nfirst_instance: u32,\n}\nstruct CullParams {\ncount: u32,\nlod_scale: f32,\nerror_px: f32,\nmin_radius_px: f32,\n}\n@group(0) @binding(1) var<storage, read> instances: array<vec4u>;\nstruct MeshInfo {\ncenter: vec3f,\nradius: f32,\nlod_count: u32,\ndraw_base: u32,\ninstance_base: u32,\ninstance_count: u32,\nerror: array<vec4f, 2>,\n}",
+};
+//# sourceMappingURL=post-shaders.gen.js.map

package/dist/post-shaders.gen.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"post-shaders.gen.js","sourceRoot":"","sources":["../src/post-shaders.gen.ts"],"names":[],"mappings":"AAAA,0EAA0E;AAC1E,yEAAyE;AAEzE,2CAA2C;AAC3C,MAAM,CAAC,MAAM,SAAS,GAAG;IACvB,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,mrGAAmrG;CACjrG,CAAC;AAEX,0CAA0C;AAC1C,MAAM,CAAC,MAAM,QAAQ,GAAG;IACtB,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,8xGAA8xG;CAC5xG,CAAC;AAEX,yCAAyC;AACzC,MAAM,CAAC,MAAM,OAAO,GAAG;IACrB,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,kmMAAkmM;CAChmM,CAAC;AAEX,8CAA8C;AAC9C,MAAM,CAAC,MAAM,WAAW,GAAG;IACzB,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,oiEAAoiE;CACliE,CAAC;AAEX,0CAA0C;AAC1C,MAAM,CAAC,MAAM,QAAQ,GAAG;IACtB,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,qhCAAqhC;CACnhC,CAAC;AAEX,6CAA6C;AAC7C,MAAM,CAAC,MAAM,UAAU,GAAG;IACxB,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,giIAAgiI;CAC9hI,CAAC;AAEX,+CAA+C;AAC/C,MAAM,CAAC,MAAM,aAAa,GAAG;IAC3B,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,qwTAAqwT;CACnwT,CAAC;AAEX,qDAAqD;AACrD,MAAM,CAAC,MAAM,kBAAkB,GAAG;IAChC,IAAI,EAAE,kBAAkB;IACxB,IAAI,EAAE,yoKAAyoK;CACvoK,CAAC"}

package/dist/post.d.ts

@@ -0,0 +1,101 @@
+import type { OrbitCamera } from './camera.ts';
+import type { FrameGraph } from './frame-graph.ts';
+import type { PipelineCache } from './pipeline-cache.ts';
+import type { BloomOptions } from './post-options.ts';
+import { type SceneRenderer } from './scene.ts';
+export type { BloomOptions, GradingOptions, PostControls, PostOptions } from './post-options.ts';
+export { gradingDefaults, whiteBalance } from './post-options.ts';
+export { halton, invert4 } from './matrix.ts';
+import type { PostPasses } from './post-options.ts';
+/**
+ * The bloom chain: `levels` half-size downsamples of the HDR scene (the first Karis-weighted
+ * and thresholded), then upsamples back to half size, each adding the level below through a
+ * tent. The tone-mapping pass samples `output`.
+ */
+export declare class Bloom {
+    #private;
+    readonly device: GPUDevice;
+    readonly levels: number;
+    intensity: number;
+    threshold: number;
+    knee: number;
+    radius: number;
+    constructor(device: GPUDevice, o: BloomOptions);
+    init(cache: PipelineCache): Promise<void>;
+    /** The chain's top level, which the tone-mapping pass reads. */
+    get output(): string;
+    /** Adds the chain's passes, reading `input` (the HDR scene). */
+    addPasses(graph: FrameGraph, input: string): void;
+    destroy(): void;
+}
+/** FXAA over the tone-mapped image (`input`, encoded) into `output`. */
+export declare class Fxaa {
+    #private;
+    readonly device: GPUDevice;
+    constructor(device: GPUDevice, format: GPUTextureFormat);
+    init(cache: PipelineCache): Promise<void>;
+    addPass(graph: FrameGraph, input: string, output: string): void;
+}
+/**
+ * Temporal anti-aliasing: owns two history targets it ping-pongs (`post.taa`, imported into
+ * the graph), jitters the camera (`jitter`, applied by the scene to its view-projection), and
+ * reprojects by depth and the camera's move.
+ */
+export declare class Taa {
+    #private;
+    readonly device: GPUDevice;
+    blend: number;
+    /**
+     * Sharpening after TAA (ADR-0119): 0 none to 1 strongest (live). Set before `init` to have the
+     * pass at all (with `motionBlur`); the history stays unsharpened, so it never accumulates.
+     */
+    sharpen: number;
+    /** Motion blur's shutter (ADR-0119): the share of each pixel's motion this frame it is smeared over, 0 to 1 (live; needs per-object motion). */
+    motionBlur: number;
+    /** Motion blur's taps (live). */
+    blurSamples: number;
+    /**
+     * What the motion target held on the next frame (ADR-0124; a diagnostic, reads back from the GPU):
+     * pixels drawn into it, those moving more than a twentieth of a pixel, the largest move in pixels, and
+     * the drawn pixels' mean view depth (last frame's).
+     * Null without per-object motion (the camera path).
+     *
+     * @example
+     * const m = await render.post.readMotion(); // { drawn: 18000, moving: 1200, maxPixels: 3.1, meanDepth: 7.4 }
+     */
+    readMotion(): Promise<{
+        drawn: number;
+        moving: number;
+        maxPixels: number;
+        meanDepth: number;
+    } | null>;
+    /** This frame's jitter in clip units (x, y), for the scene's view-projection. */
+    readonly jitter: Float32Array<ArrayBuffer>;
+    /** Temporal upscaling (ADR-0135): the history is the output's size, reconstructed from a smaller scene. */
+    readonly upscale: boolean;
+    constructor(device: GPUDevice, blend?: number, upscale?: boolean);
+    init(cache: PipelineCache): Promise<void>;
+    /** Draws per-object motion from `scene` (its GPU cull's storage path; elsewhere the camera's only). */
+    motionFrom(scene: SceneRenderer): void;
+    /** Whether per-object motion is drawn. */
+    get motion(): boolean;
+    /** This frame's camera offset from the records' anchor (ADR-0075), as the scene uploads it. */
+    setOrigin(origin: ArrayLike<number>): void;
+    /** Drops the history: the next frame starts over. */
+    reset(): void;
+    /**
+     * Before the camera upload, each frame: sizes the history to the output (the render target's size
+     * unless upscaling, ADR-0135), picks this frame's jitter (in the render target's pixels) and computes
+     * the reprojection from the camera (unjittered) and its eye.
+     */
+    prepare(camera: OrbitCamera, width: number, height: number, outputWidth?: number, outputHeight?: number): void;
+    /**
+     * Imports `post.taa` and adds the pass, reading `input` (the HDR scene) and `depth`; with
+     * sharpening or motion blur (ADR-0119), their pass after it. Returns the chain's next input.
+     */
+    addPass(graph: FrameGraph, input: string, depth: string): string;
+    get liveTextures(): number;
+    destroy(): void;
+}
+/** What `renderPlugin({ post: { passes } })` takes (ADR-0092). */
+export declare const postPasses: PostPasses;