@meimu/mutracer 0.1.5 → 0.3.0

package/lib/renderer/wgsl/jnlm.compute.js

@@ -0,0 +1,315 @@
+import shadowCatcherAccumulationWGSL from './chunks/shadowCatcherAccumulation.wgsl.js';
+import { linearToSrgbDefine } from './compose.frag.wgsl.js';
+import { JNLM_DEFAULT_COLOR_SENSITIVITY, JNLM_DEFAULT_SHAPE, JNLM_ALBEDO_FLOOR, JNLM_ALBEDO_REGULARIZATION, JNLM_COVERAGE_BANDWIDTH, jnlmPassDilation } from '../jnlm.js';
+function guideBinding(guides, binding, name) {
+    return guides === 'textures'
+        ? `@group(0) @binding(${binding}) var ${name}: texture_2d<f32>;`
+        : `@group(0) @binding(${binding}) var<storage, read> ${name}: array<vec4f>;`;
+}
+function loadGuide(guides, name, position) {
+    return guides === 'textures' ? `textureLoad(${name}, ${position}, 0)` : `${name}[gf(${position})]`;
+}
+// Shared uniform packing also fixes sensitivity and dilation for native callers.
+export function packJNLMParameters(width, height, strength, pass) {
+    const bytes = new ArrayBuffer(16);
+    new Uint32Array(bytes, 0, 2).set([width, height]);
+    new Float32Array(bytes, 8, 2).set([
+        JNLM_DEFAULT_COLOR_SENSITIVITY * Math.max(strength, 1e-3), jnlmPassDilation(pass)
+    ]);
+    return bytes;
+}
+export const JNLM_WORKGROUP_SIZE = 8;
+const paramsWGSL = /* wgsl */ `
+struct JNLMParams {
+size: vec2u,
+colorSensitivity: f32,
+
+dilation: f32,
+}
+@group(0) @binding(0) var<uniform> params: JNLMParams;
+`;
+/**
+ * Pass 1: the guide mean and the variance of that mean from the accumulated
+ * moments. Ray buffers are stored bottom-up; guides use texture rows.
+ */
+export const jnlmGuideWGSL = /* wgsl */ `
+${paramsWGSL}
+${shadowCatcherAccumulationWGSL}
+@group(0) @binding(1) var<storage, read> accumulation: array<ShadowCatcherAccumulation>;
+@group(0) @binding(2) var<storage, read> normalBuffer: array<vec4f>;
+@group(0) @binding(3) var<storage, read_write> rawGuide: array<vec2f>;
+
+@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)
+fn main(@builtin(global_invocation_id) id: vec3u) {
+let size = params.size;
+if (id.x >= size.x || id.y >= size.y) { return; }
+let qw = id.x + (size.y - 1u - id.y) * size.x;
+let n = normalBuffer[qw].a;
+let moments = accumulation[qw].moments;
+let mean = moments.x / max(n, 1.0);
+
+var variance = -1.0;
+if (n >= 2.0) {
+variance = max(moments.y - n * mean * mean, 0.0) / (n * (n - 1.0));
+}
+rawGuide[id.x + id.y * size.x] = vec2f(mean, variance);
+}
+`;
+/**
+ * Pass 2: smooth the noisy per-pixel variance over 3 x 3 pixels. Pixels with
+ * a single sample take the spatial variance of their neighborhood instead.
+ */
+export const jnlmVarianceWGSL = /* wgsl */ `
+${paramsWGSL}
+@group(0) @binding(1) var<storage, read> rawGuide: array<vec2f>;
+@group(0) @binding(2) var<storage, read_write> guide: array<vec2f>;
+
+@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)
+fn main(@builtin(global_invocation_id) id: vec3u) {
+let size = params.size;
+if (id.x >= size.x || id.y >= size.y) { return; }
+let center = rawGuide[id.x + id.y * size.x];
+var lM = 0.0;
+var je = 0.0;
+var nM = 0.0;
+var iV = 0.0;
+for (var dy = -1; dy <= 1; dy++) {
+for (var dx = -1; dx <= 1; dx++) {
+let q = clamp(vec2i(id.xy) + vec2i(dx, dy), vec2i(0), vec2i(size) - 1);
+let value = rawGuide[u32(q.x) + u32(q.y) * size.x];
+nM += value.x;
+iV += value.x * value.x;
+if (value.y >= 0.0) {
+lM += value.y;
+je += 1.0;
+}
+}
+}
+var variance: f32;
+if (center.y >= 0.0) {
+variance = lM / je;
+} else {
+let lL = nM / 9.0;
+variance = max(iV / 9.0 - lL * lL, 0.0) * 9.0 / 8.0;
+}
+guide[id.x + id.y * size.x] = vec2f(center.x, variance);
+}
+`;
+/** Copies the composed image into the filter's color buffer. */
+export function createJNLMLoadColorWGSL(guides = 'textures') {
+    return /* wgsl */ `
+${paramsWGSL}
+@group(0) @binding(1) var colorTex: texture_2d<f32>;
+@group(0) @binding(2) var<storage, read_write> outColor: array<vec4f>;
+${guides === 'buffers' ? `${shadowCatcherAccumulationWGSL}
+@group(0) @binding(3) var<storage, read> normalAccumulation: array<vec4f>;
+@group(0) @binding(4) var<storage, read_write> normal: array<vec4f>;
+@group(0) @binding(5) var<storage, read> shadowAccumulation: array<ShadowCatcherAccumulation>;
+@group(0) @binding(6) var<storage, read_write> shadow: array<vec4f>;
+@group(0) @binding(7) var albedoTex: texture_2d<f32>;
+@group(0) @binding(8) var<storage, read_write> albedo: array<vec4f>;` : ''}
+
+@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)
+fn main(@builtin(global_invocation_id) id: vec3u) {
+if (id.x >= params.size.x || id.y >= params.size.y) { return; }
+outColor[id.x + id.y * params.size.x] = textureLoad(colorTex, vec2i(id.xy), 0);
+${guides === 'buffers' ? `  let index = id.x + id.y * params.size.x;
+let qw = id.x + (params.size.y - 1u - id.y) * params.size.x;
+let n = normalAccumulation[qw];
+let samples = max(n.a, 1.0);
+normal[index] = vec4f(n.rgb / samples * 0.5 + 0.5, select(0.0, 1.0, dot(n.rgb, n.rgb) > 0.1));
+shadow[index] = shadowAccumulation[qw].mask / samples;
+albedo[index] = textureLoad(albedoTex, vec2i(id.xy), 0);` : ''}
+}
+`;
+}
+export const jnlmLoadColorWGSL = createJNLMLoadColorWGSL();
+/**
+ * Pass 3: the joint NL-means filter of the composed image. It also filters
+ * the guide with the same weights, with the variance of the weighted mean,
+ * for a following pass. The albedo regression corrects color only, keeping
+ * uncertain fitted texture responses out of the matching guide.
+ */
+export function createJNLMFilterWGSL({ searchRadius, patchRadius, albedoBandwidth, normalBandwidth } = JNLM_DEFAULT_SHAPE, guides = 'textures') {
+    const patchSize = (2 * patchRadius + 1) ** 2;
+    return /* wgsl */ `
+${paramsWGSL}
+@group(0) @binding(1) var<storage, read> guide: array<vec2f>;
+@group(0) @binding(2) var<storage, read> color: array<vec4f>;
+${guideBinding(guides, 3, 'albedoTex')}
+${guideBinding(guides, 4, 'normalTex')}
+@group(0) @binding(5) var<storage, read_write> outColor: array<vec4f>;
+@group(0) @binding(6) var<storage, read_write> outGuide: array<vec2f>;
+${guideBinding(guides, 7, 'shadowCatcherTex')}
+
+const SEARCH_RADIUS = ${searchRadius};
+const PATCH_RADIUS = ${patchRadius};
+const PATCH_SIZE = ${patchSize};
+
+const VARIANCE_CANCELLATION = 1.0;
+
+const INV_ALBEDO_BANDWIDTH2 = ${(1 / albedoBandwidth ** 2).toFixed(6)};
+const INV_COVERAGE_BANDWIDTH2 = ${(1 / JNLM_COVERAGE_BANDWIDTH ** 2).toFixed(6)};
+const INV_NORMAL_BANDWIDTH2 = ${(1 / normalBandwidth ** 2).toFixed(6)};
+
+
+const CENTER_WEIGHT_MIN = 0.01;
+
+fn gf(p: vec2i) -> u32 {
+let q = clamp(p, vec2i(0), vec2i(params.size) - 1);
+return u32(q.x) + u32(q.y) * params.size.x;
+}
+
+fn lP(q: vec2i, bo: vec4f, gO: vec3f) -> f32 {
+let albedo = ${loadGuide(guides, 'albedoTex', 'q')};
+
+
+let pY = max((albedo.rgb + bo.rgb) * 0.5, vec3f(${JNLM_ALBEDO_FLOOR}));
+let lu = (albedo.rgb - bo.rgb) / pY;
+let iQ = albedo.a - bo.a;
+let lG = ${loadGuide(guides, 'normalTex', 'q')}.rgb - gO;
+return max(
+dot(lu, lu) * INV_ALBEDO_BANDWIDTH2
++ iQ * iQ * INV_COVERAGE_BANDWIDTH2,
+dot(lG, lG) * INV_NORMAL_BANDWIDTH2
+);
+}
+
+fn premultiplied(c: vec4f) -> vec4f {
+return vec4f(c.rgb * c.a, c.a);
+}
+
+@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)
+fn main(@builtin(global_invocation_id) id: vec3u) {
+let size = vec2i(params.size);
+let p = vec2i(id.xy);
+if (p.x >= size.x || p.y >= size.y) { return; }
+let step = i32(params.dilation);
+let bo = ${loadGuide(guides, 'albedoTex', 'p')};
+let fC = ${loadGuide(guides, 'normalTex', 'p')};
+let gO = fC.rgb;
+let index = gf(p);
+
+
+if (fC.a <= 0.0 && ${loadGuide(guides, 'shadowCatcherTex', 'p')}.r <= 0.0) {
+outColor[index] = color[index];
+outGuide[index] = guide[index];
+return;
+}
+
+var lA: array<vec2f, PATCH_SIZE>;
+var k = 0;
+for (var dy = -PATCH_RADIUS; dy <= PATCH_RADIUS; dy++) {
+for (var dx = -PATCH_RADIUS; dx <= PATCH_RADIUS; dx++) {
+lA[k] = guide[gf(p + vec2i(dx, dy) * step)];
+k++;
+}
+}
+let k2 = params.colorSensitivity * params.colorSensitivity;
+
+var sum = vec4f(0.0);
+let gR = max(bo.rgb, vec3f(${JNLM_ALBEDO_FLOOR}));
+var mQ = vec3f(0.0);
+var gv = vec3f(0.0);
+var hm = vec3f(0.0);
+var qf = 0.0;
+var gx = 0.0;
+var oE = 0.0;
+var kX = 0.0;
+for (var sy = -SEARCH_RADIUS; sy <= SEARCH_RADIUS; sy++) {
+for (var sx = -SEARCH_RADIUS; sx <= SEARCH_RADIUS; sx++) {
+let q = p + vec2i(sx, sy) * step;
+if (
+(sx == 0 && sy == 0) ||
+q.x < 0 || q.y < 0 || q.x >= size.x || q.y >= size.y
+) { continue; }
+
+let features = lP(q, bo, gO);
+if (features > 16.0) { continue; }
+
+var fW = 0.0;
+k = 0;
+for (var dy = -PATCH_RADIUS; dy <= PATCH_RADIUS; dy++) {
+for (var dx = -PATCH_RADIUS; dx <= PATCH_RADIUS; dx++) {
+let a = lA[k];
+let b = guide[gf(q + vec2i(dx, dy) * step)];
+let rV = a.x - b.x;
+fW += (rV * rV
+- VARIANCE_CANCELLATION * (a.y + min(a.y, b.y)))
+/ (1e-10 + k2 * (a.y + b.y));
+k++;
+}
+}
+fW /= f32(PATCH_SIZE);
+
+let weight = exp(-max(max(fW, 0.0), features));
+let index = gf(q);
+let c = color[index];
+let g = guide[index];
+
+let a = (${loadGuide(guides, 'albedoTex', 'q')}.rgb - bo.rgb) / gR;
+let value = premultiplied(c);
+sum += value * weight;
+mQ += a * weight;
+gv += a * a * weight;
+hm += a * value.rgb * weight;
+qf += g.x * weight;
+gx += g.y * weight * weight;
+oE += weight;
+kX = max(kX, weight);
+}
+}
+
+
+
+let center = color[index];
+let lw = guide[index];
+let dE = max(kX, CENTER_WEIGHT_MIN);
+let totalWeight = oE + dE;
+let vs = premultiplied(center);
+let mean = (sum + vs * dE) / totalWeight;
+
+
+let er = mQ / totalWeight;
+let lQ = max(gv / totalWeight - er * er, vec3f(0.0));
+let rq = hm / totalWeight - er * mean.rgb;
+
+
+
+let rD = bo.rgb + er * gR;
+let pd = mean.rgb * gR / max(rD, vec3f(${JNLM_ALBEDO_FLOOR}));
+let xI = clamp(rq / (lQ + vec3f(${JNLM_ALBEDO_REGULARIZATION})), vec3f(0.0), pd);
+let sz = mean.rgb - xI * er;
+let filtered = vec4f(max(sz, vec3f(0.0)), max(mean.a, 0.0));
+let rgb = select(
+center.rgb,
+filtered.rgb / max(filtered.a, 1e-6),
+filtered.a > 1e-6
+);
+outColor[index] = vec4f(rgb, filtered.a);
+outGuide[index] = vec2f(
+(qf + lw.x * dE) / totalWeight,
+(gx + lw.y * dE * dE)
+/ (totalWeight * totalWeight)
+);
+}
+`;
+}
+// Compute presentation for hosts that consume the filtered buffer directly.
+// Use the same sRGB transfer as browser display composition.
+export const jnlmEncodeWGSL = /* wgsl */ `
+struct OutputParams { size: vec2u, premultiply: u32, reserved: u32 }
+@group(0) @binding(0) var<uniform> params: OutputParams;
+@group(0) @binding(1) var<storage, read> color: array<vec4f>;
+@group(0) @binding(2) var output: texture_storage_2d<rgba8unorm, write>;
+${linearToSrgbDefine}
+@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)
+fn main(@builtin(global_invocation_id) id: vec3u) {
+if (any(id.xy >= params.size)) { return; }
+let c = color[id.y * params.size.x + id.x];
+var rgb = linearToSrgb(c.rgb);
+if (params.premultiply != 0u) { rgb *= c.a; }
+textureStore(output, id.xy, vec4f(rgb, c.a));
+}
+`;
+//# sourceMappingURL=jnlm.compute.js.map

package/lib/renderer/wgsl/jnlm.compute.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"jnlm.compute.js","sourceRoot":"","sources":["../../../src/renderer/wgsl/jnlm.compute.ts"],"names":[],"mappings":"AAAA,OAAO,6BAA6B,MAAM,4CAA4C,CAAC;AACvF,OAAO,EAAC,kBAAkB,EAAC,MAAM,wBAAwB,CAAC;AAC1D,OAAO,EACL,8BAA8B,EAAE,kBAAkB,EAAE,iBAAiB,EACrE,0BAA0B,EAAE,uBAAuB,EAAE,gBAAgB,EAEtE,MAAM,YAAY,CAAC;AAIpB,SAAS,YAAY,CAAC,MAAyB,EAAE,OAAe,EAAE,IAAY;IAC5E,OAAO,MAAM,KAAK,UAAU;QAC1B,CAAC,CAAC,sBAAsB,OAAO,SAAS,IAAI,oBAAoB;QAChE,CAAC,CAAC,sBAAsB,OAAO,wBAAwB,IAAI,iBAAiB,CAAC;AACjF,CAAC;AACD,SAAS,SAAS,CAAC,MAAyB,EAAE,IAAY,EAAE,QAAgB;IAC1E,OAAO,MAAM,KAAK,UAAU,CAAC,CAAC,CAAC,eAAe,IAAI,KAAK,QAAQ,MAAM,CAAC,CAAC,CAAC,GAAG,IAAI,OAAe,QAAQ,IAAI,CAAC;AAC7G,CAAC;AAED,iFAAiF;AACjF,MAAM,UAAU,kBAAkB,CAAC,KAAa,EAAE,MAAc,EAAE,QAAgB,EAAE,IAAY;IAC9F,MAAM,KAAK,GAAG,IAAI,WAAW,CAAC,EAAE,CAAC,CAAC;IAClC,IAAI,WAAW,CAAC,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC;IAClD,IAAI,YAAY,CAAC,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC;QAChC,8BAA8B,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,EAAE,IAAI,CAAC,EAAE,gBAAgB,CAAC,IAAI,CAAC;KAClF,CAAC,CAAC;IACH,OAAO,KAAK,CAAC;AACf,CAAC;AAED,MAAM,CAAC,MAAM,mBAAmB,GAAG,CAAC,CAAC;AAErC,MAAM,UAAU,GAAG,UAAU,CAAC;;;;;;;;CAQ7B,CAAC;AAEF;;;GAGG;AACH,MAAM,CAAC,MAAM,aAAa,GAAG,UAAU,CAAC;EACtC,UAAU;EACV,6BAA6B;;;;;2BAKJ,mBAAmB,KAAK,mBAAmB;;;;;;;;;;;;;;;CAerE,CAAC;AAEF;;;GAGG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAG,UAAU,CAAC;EACzC,UAAU;;;;2BAIe,mBAAmB,KAAK,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8BrE,CAAC;AAEF,gEAAgE;AAChE,MAAM,UAAU,uBAAuB,CAAC,SAA4B,UAAU;IAC5E,OAAO,UAAU,CAAC;EAClB,UAAU;;;EAGV,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,GAAG,6BAA6B;;;;;;qEAMY,CAAC,CAAC,CAAC,EAAE;;2BAE/C,mBAAmB,KAAK,mBAAmB;;;;EAIpE,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC;;;;;;yDAMkC,CAAC,CAAC,CAAC,EAAE;;CAE/D,CAAC;AACF,CAAC;AACD,MAAM,CAAC,MAAM,iBAAiB,GAAG,uBAAuB,EAAE,CAAC;AAE3D;;;;;GAKG;AACH,MAAM,UAAU,oBAAoB,CAAC,EACnC,YAAY,EAAE,WAAW,EAAE,eAAe,EAAE,eAAe,KACxC,kBAAkB,EAAE,SAA4B,UAAU;IAC7E,MAAM,SAAS,GAAG,CAAC,CAAC,GAAG,WAAW,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC;IAC7C,OAAO,UAAU,CAAC;EAClB,UAAU;;;EAGV,YAAY,CAAC,MAAM,EAAE,CAAC,EAAE,WAAW,CAAC;EACpC,YAAY,CAAC,MAAM,EAAE,CAAC,EAAE,WAAW,CAAC;;;EAGpC,YAAY,CAAC,MAAM,EAAE,CAAC,EAAE,kBAAkB,CAAC;;wBAErB,YAAY;uBACb,WAAW;qBACb,SAAS;;;;gCAIE,CAAC,CAAC,GAAG,eAAe,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC;kCACnC,CAAC,CAAC,GAAG,uBAAuB,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC;gCAC/C,CAAC,CAAC,GAAG,eAAe,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC;;;;;;;;;;;eAWpD,SAAS,CAAC,MAAM,EAAE,WAAW,EAAE,GAAG,CAAC;;;kDAGmB,iBAAiB;;;WAGlE,SAAS,CAAC,MAAM,EAAE,WAAW,EAAE,GAAG,CAAC;;;;;;;;;;;;2BAY9B,mBAAmB,KAAK,mBAAmB;;;;;;WAM/C,SAAS,CAAC,MAAM,EAAE,WAAW,EAAE,GAAG,CAAC;WAC9B,SAAS,CAAC,MAAM,EAAE,WAAW,EAAE,GAAG,CAAC;;;;;qBAKzB,SAAS,CAAC,MAAM,EAAE,kBAAkB,EAAE,GAAG,CAAC;;;;;;;;;;;;;;;;;6BAiB7B,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;WAuCnD,SAAS,CAAC,MAAM,EAAE,WAAW,EAAE,GAAG,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;yCA8BmB,iBAAiB;kCAC5B,0BAA0B;;;;;;;;;;;;;;;CAerF,CAAC;AACF,CAAC;AAED,4EAA4E;AAC5E,6DAA6D;AAC7D,MAAM,CAAC,MAAM,cAAc,GAAG,UAAU,CAAC;;;;;EAKvC,kBAAkB;2BACO,mBAAmB,KAAK,mBAAmB;;;;;;;;CAQrE,CAAC","sourcesContent":["import shadowCatcherAccumulationWGSL from './chunks/shadowCatcherAccumulation.wgsl.js';\nimport {linearToSrgbDefine} from './compose.frag.wgsl.js';\nimport {\n  JNLM_DEFAULT_COLOR_SENSITIVITY, JNLM_DEFAULT_SHAPE, JNLM_ALBEDO_FLOOR,\n  JNLM_ALBEDO_REGULARIZATION, JNLM_COVERAGE_BANDWIDTH, jnlmPassDilation,\n  type JNLMFilterShape\n} from '../jnlm.js';\n\n// Web textures and native normalized guide buffers use the same filter body.\nexport type JNLMGuideBindings = 'textures' | 'buffers';\nfunction guideBinding(guides: JNLMGuideBindings, binding: number, name: string): string {\n  return guides === 'textures'\n    ? `@group(0) @binding(${binding}) var ${name}: texture_2d<f32>;`\n    : `@group(0) @binding(${binding}) var<storage, read> ${name}: array<vec4f>;`;\n}\nfunction loadGuide(guides: JNLMGuideBindings, name: string, position: string): string {\n  return guides === 'textures' ? `textureLoad(${name}, ${position}, 0)` : `${name}[pixelIndex(${position})]`;\n}\n\n// Shared uniform packing also fixes sensitivity and dilation for native callers.\nexport function packJNLMParameters(width: number, height: number, strength: number, pass: number): ArrayBuffer {\n  const bytes = new ArrayBuffer(16);\n  new Uint32Array(bytes, 0, 2).set([width, height]);\n  new Float32Array(bytes, 8, 2).set([\n    JNLM_DEFAULT_COLOR_SENSITIVITY * Math.max(strength, 1e-3), jnlmPassDilation(pass)\n  ]);\n  return bytes;\n}\n\nexport const JNLM_WORKGROUP_SIZE = 8;\n\nconst paramsWGSL = /* wgsl */ `\nstruct JNLMParams {\n  size: vec2u,\n  colorSensitivity: f32,\n  // Pixel spacing of the search window and patches in this pass.\n  dilation: f32,\n}\n@group(0) @binding(0) var<uniform> params: JNLMParams;\n`;\n\n/**\n * Pass 1: the guide mean and the variance of that mean from the accumulated\n * moments. Ray buffers are stored bottom-up; guides use texture rows.\n */\nexport const jnlmGuideWGSL = /* wgsl */ `\n${paramsWGSL}\n${shadowCatcherAccumulationWGSL}\n@group(0) @binding(1) var<storage, read> accumulation: array<ShadowCatcherAccumulation>;\n@group(0) @binding(2) var<storage, read> normalBuffer: array<vec4f>;\n@group(0) @binding(3) var<storage, read_write> rawGuide: array<vec2f>;\n\n@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)\nfn main(@builtin(global_invocation_id) id: vec3u) {\n  let size = params.size;\n  if (id.x >= size.x || id.y >= size.y) { return; }\n  let rayIndex = id.x + (size.y - 1u - id.y) * size.x;\n  let n = normalBuffer[rayIndex].a;\n  let moments = accumulation[rayIndex].moments;\n  let mean = moments.x / max(n, 1.0);\n  // Variance of the mean, or -1 when one sample cannot estimate it.\n  var variance = -1.0;\n  if (n >= 2.0) {\n    variance = max(moments.y - n * mean * mean, 0.0) / (n * (n - 1.0));\n  }\n  rawGuide[id.x + id.y * size.x] = vec2f(mean, variance);\n}\n`;\n\n/**\n * Pass 2: smooth the noisy per-pixel variance over 3 x 3 pixels. Pixels with\n * a single sample take the spatial variance of their neighborhood instead.\n */\nexport const jnlmVarianceWGSL = /* wgsl */ `\n${paramsWGSL}\n@group(0) @binding(1) var<storage, read> rawGuide: array<vec2f>;\n@group(0) @binding(2) var<storage, read_write> guide: array<vec2f>;\n\n@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)\nfn main(@builtin(global_invocation_id) id: vec3u) {\n  let size = params.size;\n  if (id.x >= size.x || id.y >= size.y) { return; }\n  let center = rawGuide[id.x + id.y * size.x];\n  var varianceSum = 0.0;\n  var varianceCount = 0.0;\n  var meanSum = 0.0;\n  var meanSquareSum = 0.0;\n  for (var dy = -1; dy <= 1; dy++) {\n    for (var dx = -1; dx <= 1; dx++) {\n      let q = clamp(vec2i(id.xy) + vec2i(dx, dy), vec2i(0), vec2i(size) - 1);\n      let value = rawGuide[u32(q.x) + u32(q.y) * size.x];\n      meanSum += value.x;\n      meanSquareSum += value.x * value.x;\n      if (value.y >= 0.0) {\n        varianceSum += value.y;\n        varianceCount += 1.0;\n      }\n    }\n  }\n  var variance: f32;\n  if (center.y >= 0.0) {\n    variance = varianceSum / varianceCount;\n  } else {\n    let spatialMean = meanSum / 9.0;\n    variance = max(meanSquareSum / 9.0 - spatialMean * spatialMean, 0.0) * 9.0 / 8.0;\n  }\n  guide[id.x + id.y * size.x] = vec2f(center.x, variance);\n}\n`;\n\n/** Copies the composed image into the filter's color buffer. */\nexport function createJNLMLoadColorWGSL(guides: JNLMGuideBindings = 'textures'): string {\n  return /* wgsl */ `\n${paramsWGSL}\n@group(0) @binding(1) var colorTex: texture_2d<f32>;\n@group(0) @binding(2) var<storage, read_write> outColor: array<vec4f>;\n${guides === 'buffers' ? `${shadowCatcherAccumulationWGSL}\n@group(0) @binding(3) var<storage, read> normalAccumulation: array<vec4f>;\n@group(0) @binding(4) var<storage, read_write> normal: array<vec4f>;\n@group(0) @binding(5) var<storage, read> shadowAccumulation: array<ShadowCatcherAccumulation>;\n@group(0) @binding(6) var<storage, read_write> shadow: array<vec4f>;\n@group(0) @binding(7) var albedoTex: texture_2d<f32>;\n@group(0) @binding(8) var<storage, read_write> albedo: array<vec4f>;` : ''}\n\n@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)\nfn main(@builtin(global_invocation_id) id: vec3u) {\n  if (id.x >= params.size.x || id.y >= params.size.y) { return; }\n  outColor[id.x + id.y * params.size.x] = textureLoad(colorTex, vec2i(id.xy), 0);\n${guides === 'buffers' ? `  let index = id.x + id.y * params.size.x;\n  let rayIndex = id.x + (params.size.y - 1u - id.y) * params.size.x;\n  let n = normalAccumulation[rayIndex];\n  let samples = max(n.a, 1.0);\n  normal[index] = vec4f(n.rgb / samples * 0.5 + 0.5, select(0.0, 1.0, dot(n.rgb, n.rgb) > 0.1));\n  shadow[index] = shadowAccumulation[rayIndex].mask / samples;\n  albedo[index] = textureLoad(albedoTex, vec2i(id.xy), 0);` : ''}\n}\n`;\n}\nexport const jnlmLoadColorWGSL = createJNLMLoadColorWGSL();\n\n/**\n * Pass 3: the joint NL-means filter of the composed image. It also filters\n * the guide with the same weights, with the variance of the weighted mean,\n * for a following pass. The albedo regression corrects color only, keeping\n * uncertain fitted texture responses out of the matching guide.\n */\nexport function createJNLMFilterWGSL({\n  searchRadius, patchRadius, albedoBandwidth, normalBandwidth\n}: JNLMFilterShape = JNLM_DEFAULT_SHAPE, guides: JNLMGuideBindings = 'textures') {\n  const patchSize = (2 * patchRadius + 1) ** 2;\n  return /* wgsl */ `\n${paramsWGSL}\n@group(0) @binding(1) var<storage, read> guide: array<vec2f>;\n@group(0) @binding(2) var<storage, read> color: array<vec4f>;\n${guideBinding(guides, 3, 'albedoTex')}\n${guideBinding(guides, 4, 'normalTex')}\n@group(0) @binding(5) var<storage, read_write> outColor: array<vec4f>;\n@group(0) @binding(6) var<storage, read_write> outGuide: array<vec2f>;\n${guideBinding(guides, 7, 'shadowCatcherTex')}\n\nconst SEARCH_RADIUS = ${searchRadius};\nconst PATCH_RADIUS = ${patchRadius};\nconst PATCH_SIZE = ${patchSize};\n// Variance cancellation strength alpha of the paper.\nconst VARIANCE_CANCELLATION = 1.0;\n// Inverse squared feature bandwidths. Normals are encoded as n * 0.5 + 0.5.\nconst INV_ALBEDO_BANDWIDTH2 = ${(1 / albedoBandwidth ** 2).toFixed(6)};\nconst INV_COVERAGE_BANDWIDTH2 = ${(1 / JNLM_COVERAGE_BANDWIDTH ** 2).toFixed(6)};\nconst INV_NORMAL_BANDWIDTH2 = ${(1 / normalBandwidth ** 2).toFixed(6)};\n// Lower bound of the center weight, so a pixel unlike all of its neighbors\n// (a thin highlight) is kept rather than replaced by them.\nconst CENTER_WEIGHT_MIN = 0.01;\n\nfn pixelIndex(p: vec2i) -> u32 {\n  let q = clamp(p, vec2i(0), vec2i(params.size) - 1);\n  return u32(q.x) + u32(q.y) * params.size.x;\n}\n\nfn featureDistance(q: vec2i, centerAlbedo: vec4f, centerNormal: vec3f) -> f32 {\n  let albedo = ${loadGuide(guides, 'albedoTex', 'q')};\n  // Absolute differences in linear albedo erase low-contrast dark textures.\n  // Measure RGB contrast relative to both pixels; keep coverage absolute.\n  let albedoScale = max((albedo.rgb + centerAlbedo.rgb) * 0.5, vec3f(${JNLM_ALBEDO_FLOOR}));\n  let albedoDelta = (albedo.rgb - centerAlbedo.rgb) / albedoScale;\n  let coverageDelta = albedo.a - centerAlbedo.a;\n  let normalDelta = ${loadGuide(guides, 'normalTex', 'q')}.rgb - centerNormal;\n  return max(\n    dot(albedoDelta, albedoDelta) * INV_ALBEDO_BANDWIDTH2\n      + coverageDelta * coverageDelta * INV_COVERAGE_BANDWIDTH2,\n    dot(normalDelta, normalDelta) * INV_NORMAL_BANDWIDTH2\n  );\n}\n\nfn premultiplied(c: vec4f) -> vec4f {\n  return vec4f(c.rgb * c.a, c.a);\n}\n\n@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)\nfn main(@builtin(global_invocation_id) id: vec3u) {\n  let size = vec2i(params.size);\n  let p = vec2i(id.xy);\n  if (p.x >= size.x || p.y >= size.y) { return; }\n  let step = i32(params.dilation);\n  let centerAlbedo = ${loadGuide(guides, 'albedoTex', 'p')};\n  let centerNormalTexel = ${loadGuide(guides, 'normalTex', 'p')};\n  let centerNormal = centerNormalTexel.rgb;\n  let index = pixelIndex(p);\n  // Primary misses show the background directly. It has no features that\n  // protect its detail and no noise beyond antialiasing, so keep it.\n  if (centerNormalTexel.a <= 0.0 && ${loadGuide(guides, 'shadowCatcherTex', 'p')}.r <= 0.0) {\n    outColor[index] = color[index];\n    outGuide[index] = guide[index];\n    return;\n  }\n\n  var centerPatch: array<vec2f, PATCH_SIZE>;\n  var k = 0;\n  for (var dy = -PATCH_RADIUS; dy <= PATCH_RADIUS; dy++) {\n    for (var dx = -PATCH_RADIUS; dx <= PATCH_RADIUS; dx++) {\n      centerPatch[k] = guide[pixelIndex(p + vec2i(dx, dy) * step)];\n      k++;\n    }\n  }\n  let k2 = params.colorSensitivity * params.colorSensitivity;\n\n  var sum = vec4f(0.0);\n  let featureScale = max(centerAlbedo.rgb, vec3f(${JNLM_ALBEDO_FLOOR}));\n  var featureSum = vec3f(0.0);\n  var featureSquareSum = vec3f(0.0);\n  var featureColorSum = vec3f(0.0);\n  var guideSum = 0.0;\n  var guideVarianceSum = 0.0;\n  var weightSum = 0.0;\n  var maxWeight = 0.0;\n  for (var sy = -SEARCH_RADIUS; sy <= SEARCH_RADIUS; sy++) {\n    for (var sx = -SEARCH_RADIUS; sx <= SEARCH_RADIUS; sx++) {\n      let q = p + vec2i(sx, sy) * step;\n      if (\n        (sx == 0 && sy == 0) ||\n        q.x < 0 || q.y < 0 || q.x >= size.x || q.y >= size.y\n      ) { continue; }\n\n      let features = featureDistance(q, centerAlbedo, centerNormal);\n      if (features > 16.0) { continue; }\n\n      var colorDistance = 0.0;\n      k = 0;\n      for (var dy = -PATCH_RADIUS; dy <= PATCH_RADIUS; dy++) {\n        for (var dx = -PATCH_RADIUS; dx <= PATCH_RADIUS; dx++) {\n          let a = centerPatch[k];\n          let b = guide[pixelIndex(q + vec2i(dx, dy) * step)];\n          let delta = a.x - b.x;\n          colorDistance += (delta * delta\n              - VARIANCE_CANCELLATION * (a.y + min(a.y, b.y)))\n            / (1e-10 + k2 * (a.y + b.y));\n          k++;\n        }\n      }\n      colorDistance /= f32(PATCH_SIZE);\n\n      let weight = exp(-max(max(colorDistance, 0.0), features));\n      let index = pixelIndex(q);\n      let c = color[index];\n      let g = guide[index];\n      // Filter premultiplied color so coverage and color stay consistent.\n      let a = (${loadGuide(guides, 'albedoTex', 'q')}.rgb - centerAlbedo.rgb) / featureScale;\n      let value = premultiplied(c);\n      sum += value * weight;\n      featureSum += a * weight;\n      featureSquareSum += a * a * weight;\n      featureColorSum += a * value.rgb * weight;\n      guideSum += g.x * weight;\n      guideVarianceSum += g.y * weight * weight;\n      weightSum += weight;\n      maxWeight = max(maxWeight, weight);\n    }\n  }\n\n  // Weight the center like its most similar neighbor (Buades et al. 2005)\n  // instead of 1, so isolated outliers are not kept unfiltered.\n  let center = color[index];\n  let centerGuide = guide[index];\n  let centerWeight = max(maxWeight, CENTER_WEIGHT_MIN);\n  let totalWeight = weightSum + centerWeight;\n  let centerValue = premultiplied(center);\n  let mean = (sum + centerValue * centerWeight) / totalWeight;\n  // Fit each color channel against relative albedo with the same JNLM weights.\n  // Centering at this pixel avoids cancellation when the feature is constant.\n  let meanFeature = featureSum / totalWeight;\n  let featureVariance = max(featureSquareSum / totalWeight - meanFeature * meanFeature, vec3f(0.0));\n  let covariance = featureColorSum / totalWeight - meanFeature * mean.rgb;\n  // A 1% contrast ridge stabilizes flat guides. Bound the response between no\n  // texture dependence and multiplicative albedo modulation; specular color\n  // with an unrelated albedo falls back to the weighted mean.\n  let meanAlbedo = centerAlbedo.rgb + meanFeature * featureScale;\n  let maximumSlope = mean.rgb * featureScale / max(meanAlbedo, vec3f(${JNLM_ALBEDO_FLOOR}));\n  let slope = clamp(covariance / (featureVariance + vec3f(${JNLM_ALBEDO_REGULARIZATION})), vec3f(0.0), maximumSlope);\n  let corrected = mean.rgb - slope * meanFeature;\n  let filtered = vec4f(max(corrected, vec3f(0.0)), max(mean.a, 0.0));\n  let rgb = select(\n    center.rgb,\n    filtered.rgb / max(filtered.a, 1e-6),\n    filtered.a > 1e-6\n  );\n  outColor[index] = vec4f(rgb, filtered.a);\n  outGuide[index] = vec2f(\n    (guideSum + centerGuide.x * centerWeight) / totalWeight,\n    (guideVarianceSum + centerGuide.y * centerWeight * centerWeight)\n      / (totalWeight * totalWeight)\n  );\n}\n`;\n}\n\n// Compute presentation for hosts that consume the filtered buffer directly.\n// Use the same sRGB transfer as browser display composition.\nexport const jnlmEncodeWGSL = /* wgsl */ `\nstruct OutputParams { size: vec2u, premultiply: u32, reserved: u32 }\n@group(0) @binding(0) var<uniform> params: OutputParams;\n@group(0) @binding(1) var<storage, read> color: array<vec4f>;\n@group(0) @binding(2) var output: texture_storage_2d<rgba8unorm, write>;\n${linearToSrgbDefine}\n@compute @workgroup_size(${JNLM_WORKGROUP_SIZE}, ${JNLM_WORKGROUP_SIZE}, 1)\nfn main(@builtin(global_invocation_id) id: vec3u) {\n  if (any(id.xy >= params.size)) { return; }\n  let c = color[id.y * params.size.x + id.x];\n  var rgb = linearToSrgb(c.rgb);\n  if (params.premultiply != 0u) { rgb *= c.a; }\n  textureStore(output, id.xy, vec4f(rgb, c.a));\n}\n`;\n"]}

package/lib/renderer/wgsl/offlineAdaptive.compute.js

@@ -27,8 +27,8 @@ _padding: f32,
 const LUMINANCE = vec3f(0.2126, 0.7152, 0.0722);
 
 @compute @workgroup_size(64)
-fn main(@builtin(global_invocation_id) uq: vec3u) {
-let tile = uq.x;
+fn main(@builtin(global_invocation_id) eA: vec3u) {
+let tile = eA.x;
 let tileCount = params.columns * params.rows;
 if (tile >= tileCount) {
 return;
@@ -36,18 +36,18 @@ return;
 
 let tileX = tile % params.columns;
 let tileY = tile / params.columns;
-let ux = tileX * params.tileSize;
-let uy = tileY * params.tileSize;
-let wl = min(ux + params.tileSize, params.width);
-let wm = min(uy + params.tileSize, params.height);
-var mY = 0.0;
-var je = 0.0;
+let ws = tileX * params.tileSize;
+let wt = tileY * params.tileSize;
+let yF = min(ws + params.tileSize, params.width);
+let yG = min(wt + params.tileSize, params.height);
+var nM = 0.0;
+var jh = 0.0;
 var count = 0u;
 
-for (var y = uy; y < wm; y++) {
-for (var x = ux; x < wl; x++) {
-let fQ = y * params.width + x;
-let offset = fQ * 4u;
+for (var y = wt; y < yG; y++) {
+for (var x = ws; x < yF; x++) {
+let pixel = y * params.width + x;
+let offset = pixel * 4u;
 let samples = pilotMoments[offset + 3u];
 if (samples <= 0.0) {
 continue;
@@ -60,28 +60,28 @@ pilotColor[offset + 2u]
 ),
 LUMINANCE
 ) / samples;
-let tK = pilotMoments[offset] / samples;
-let pi = max(tK - mean * mean, 0.0);
-var variance = pi;
+let vj = pilotMoments[offset] / samples;
+let qb = max(vj - mean * mean, 0.0);
+var variance = qb;
 if (samples > 1.0) {
-variance = pi * samples / (samples - 1.0);
+variance = qb * samples / (samples - 1.0);
 }
-mY += mean;
-je = max(je, variance);
+nM += mean;
+jh = max(jh, variance);
 count += 1u;
 }
 }
 
-let mean = mY / f32(max(count, 1u));
-var oX = 0.0;
+let mean = nM / f32(max(count, 1u));
+var pK = 0.0;
 if (params.hasPrevious != 0u) {
-oX = previousVariance[tile];
+pK = previousVariance[tile];
 }
-let sj = max(je, oX);
-let tf = params.absoluteError + params.relativeError * abs(mean);
+let tu = max(jh, pK);
+let uH = params.absoluteError + params.relativeError * abs(mean);
 let required = ceil(
-params.confidence * params.confidence * sj /
-max(tf * tf, 1e-16)
+params.confidence * params.confidence * tu /
+max(uH * uH, 1e-16)
 );
 let targetSamples = u32(min(
 f32(params.maxSamples),
@@ -91,11 +91,11 @@ max(f32(params.minSamples), required)
 tileMeans[tile] = mean;
 
 
-tileVariances[tile] = je;
+tileVariances[tile] = jh;
 tileTargets[tile] = targetSamples;
 
-for (var y = uy; y < wm; y++) {
-for (var x = ux; x < wl; x++) {
+for (var y = wt; y < yG; y++) {
+for (var x = ws; x < yF; x++) {
 textureStore(
 targetTexture,
 vec2i(i32(x), i32(y)),

package/lib/renderer/wgsl/oidnAuxiliary.wgsl.d.ts

@@ -0,0 +1 @@
+export declare const oidnAuxiliaryDefine = "\nfn oidnEncodeAlbedo(value: vec3f) -> vec3f {\n  let linear = clamp(value, vec3f(0.0), vec3f(1.0));\n  return select(1.055 * pow(linear, vec3f(1.0 / 2.4)) - 0.055,\n    12.92 * linear, linear <= vec3f(0.0031308));\n}\nfn oidnDecodeAlbedo(value: vec3f) -> vec3f {\n  let encoded = clamp(value, vec3f(0.0), vec3f(1.0));\n  return select(pow((encoded + 0.055) / 1.055, vec3f(2.4)),\n    encoded / 12.92, encoded <= vec3f(0.04045));\n}\nfn oidnSanitizeNormal(value: vec3f) -> vec3f {\n  return clamp(value, vec3f(0.0), vec3f(1.0));\n}\n";

package/lib/renderer/wgsl/oidnAuxiliary.wgsl.js

@@ -0,0 +1,17 @@
+// OIDN's LDR auxiliary networks use sRGB albedo and normals in [0,1].
+export const oidnAuxiliaryDefine = /* wgsl */ `
+fn sQ(value: vec3f) -> vec3f {
+let linear = clamp(value, vec3f(0.0), vec3f(1.0));
+return select(1.055 * pow(linear, vec3f(1.0 / 2.4)) - 0.055,
+12.92 * linear, linear <= vec3f(0.0031308));
+}
+fn sP(value: vec3f) -> vec3f {
+let encoded = clamp(value, vec3f(0.0), vec3f(1.0));
+return select(pow((encoded + 0.055) / 1.055, vec3f(2.4)),
+encoded / 12.92, encoded <= vec3f(0.04045));
+}
+fn rG(value: vec3f) -> vec3f {
+return clamp(value, vec3f(0.0), vec3f(1.0));
+}
+`;
+//# sourceMappingURL=oidnAuxiliary.wgsl.js.map

package/lib/renderer/wgsl/oidnAuxiliary.wgsl.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"oidnAuxiliary.wgsl.js","sourceRoot":"","sources":["../../../src/renderer/wgsl/oidnAuxiliary.wgsl.ts"],"names":[],"mappings":"AAAA,sEAAsE;AACtE,MAAM,CAAC,MAAM,mBAAmB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;CAc7C,CAAC","sourcesContent":["// OIDN's LDR auxiliary networks use sRGB albedo and normals in [0,1].\nexport const oidnAuxiliaryDefine = /* wgsl */ `\nfn oidnEncodeAlbedo(value: vec3f) -> vec3f {\n  let linear = clamp(value, vec3f(0.0), vec3f(1.0));\n  return select(1.055 * pow(linear, vec3f(1.0 / 2.4)) - 0.055,\n    12.92 * linear, linear <= vec3f(0.0031308));\n}\nfn oidnDecodeAlbedo(value: vec3f) -> vec3f {\n  let encoded = clamp(value, vec3f(0.0), vec3f(1.0));\n  return select(pow((encoded + 0.055) / 1.055, vec3f(2.4)),\n    encoded / 12.92, encoded <= vec3f(0.04045));\n}\nfn oidnSanitizeNormal(value: vec3f) -> vec3f {\n  return clamp(value, vec3f(0.0), vec3f(1.0));\n}\n`;\n"]}

package/lib/renderer/wgsl/photonMap.compute.d.ts

@@ -0,0 +1,29 @@
+export declare const PHOTON_BYTES = 20;
+export declare const PHOTON_GRID_SIZE = 262144;
+export declare const PHOTON_SCAN_BLOCK = 256;
+export declare const PHOTON_REFERENCE_COUNT: number;
+export type PhotonRadiusScaling = 'none' | 'ppm' | 'constant-count';
+/** Internal density A/B, with the same progressive schedule in all modes. */
+export declare function photonRadius(initial: number, batch: number, alpha?: number, photons?: number, scaling?: PhotonRadiusScaling): number;
+/** Default photon-map batch: pixel count / 4, capped at 2^18. */
+export declare function defaultPhotonMapPhotons(width: number, height: number, divisor?: number): number;
+/** Power-of-two hash capacity, bounded independently of large photon batches. */
+export declare function photonBucketCount(photons: number): number;
+/** Shared texture address contract; payload is stored once, indices twice. */
+export declare function photonLayout(count: number, perCasterRadius?: boolean): {
+    buckets: number;
+    coarseHeaders: number;
+    fineHeaders: number;
+    coarseIndices: number;
+    fineIndices: number;
+    directions: number;
+    texels: number;
+};
+/** Batch-local hash buckets: count, two-level parallel prefix sum, then parallel scatter. */
+export declare const photonGridCompute = "\n  struct Bucket {\n    count: atomic<u32>,\n    offset: u32,\n    cursor: atomic<u32>,\n    padding: u32,\n  }\n  struct Parameters {\n    radius: f32,\n    count: u32,\n    screenPixels: f32,\n    skipQuery: f32,\n    buckets: u32,\n    fineRadius: f32,\n    experiments: u32,\n    padding2: u32,\n  }\n  @group(0)@binding(0)var photons:texture_2d<u32>;\n  @group(0)@binding(1)var<storage,read_write>grid:array<Bucket>;\n  @group(0)@binding(2)var<uniform>parameters:Parameters;\n  @group(0)@binding(4)var headers:texture_storage_2d<r32uint,write>;\n  fn hashCell(cell:vec3i)->u32 {\n    let p=bitcast<vec3u>(cell);\n    return ((p.x*73856093u)^(p.y*19349663u)^(p.z*83492791u))&(parameters.buckets-1u);\n  }\n  fn hashPhoton(p:vec3f,layer:u32)->u32 {\n    let radius=select(parameters.radius,parameters.fineRadius,layer==1u);\n    let cellSize=select(radius,2.0*radius,layer==1u);\n    return hashCell(vec3i(floor(p/cellSize)));\n  }\n  fn gridStart(layer:u32)->u32 {\n    return layer*(parameters.buckets+parameters.buckets/256u);\n  }\n  fn storeIndex(i:u32,value:u32) {\n    let width=textureDimensions(headers).x;\n    textureStore(headers,vec2i(vec2u(i%width,i/width)),vec4u(value));\n  }\n  fn photonRecord(i:u32)->vec4u {\n    let width=textureDimensions(photons).x;\n    return textureLoad(photons,vec2i(vec2u(i%width,i/width)),0);\n  }\n  fn photonPosition(i:u32)->vec3f {\n    return bitcast<vec3f>(photonRecord(i).xyz);\n  }\n  @compute @workgroup_size(64)\n  fn count(@builtin(global_invocation_id)id:vec3u) {\n    if(id.x+id.y*4194240u>=parameters.count) {\n      return;\n    }\n    let index=id.x+id.y*4194240u;\n    if(photonRecord(index).w==0u) {\n      return;\n    }\n    for(var layer=0u;layer<2u;layer++) {\n      atomicAdd(&grid[gridStart(layer)+hashPhoton(photonPosition(index),layer)].count,1u);\n    }\n  }\n  // Each block scans 256 buckets; the next workgroup scans chunks of block totals.\n  var<workgroup> scan:array<u32,256>;\n  fn inclusiveScan(local:u32,value:u32)->u32 {\n    scan[local]=value;\n    workgroupBarrier();\n    for(var stride=1u;stride<256u;stride*=2u) {\n      var previous=0u;\n      if(local>=stride) {\n        previous=scan[local-stride];\n      }\n      workgroupBarrier();\n      scan[local]+=previous;\n      workgroupBarrier();\n    }\n    return scan[local];\n  }\n  @compute @workgroup_size(256)\n  fn scanBlocks(@builtin(local_invocation_index)local:u32,@builtin(workgroup_id)group:vec3u) {\n    let index=gridStart(group.y)+group.x*256u+local;\n    let count=atomicLoad(&grid[index].count);\n    let prefix=inclusiveScan(local,count);\n    grid[index].offset=prefix-count;\n    if(local==255u) {\n      atomicStore(&grid[gridStart(group.y)+parameters.buckets+group.x].count,prefix);\n    }\n  }\n  @compute @workgroup_size(256)\n  fn scanSums(@builtin(local_invocation_index)local:u32,@builtin(workgroup_id)group:vec3u) {\n    let blocks=parameters.buckets/256u;\n    let chunk=(blocks+255u)/256u;\n    let first=local*chunk;\n    let start=gridStart(group.y)+parameters.buckets;\n    var count=0u;\n    for(var i=first;i<min(first+chunk,blocks);i++) {\n      count+=atomicLoad(&grid[start+i].count);\n    }\n    let prefix=inclusiveScan(local,count);\n    var offset=prefix-count;\n    for(var i=first;i<min(first+chunk,blocks);i++) {\n      grid[start+i].offset=offset;\n      offset+=atomicLoad(&grid[start+i].count);\n    }\n    if(local==255u&&group.y==0u) {\n      storeIndex(0u,bitcast<u32>(parameters.radius));\n      storeIndex(1u,prefix);\n      storeIndex(2u,bitcast<u32>(parameters.screenPixels));\n      storeIndex(3u,bitcast<u32>(parameters.skipQuery));\n      storeIndex(4u,parameters.count);\n      storeIndex(5u,parameters.buckets);\n      storeIndex(6u,bitcast<u32>(parameters.fineRadius));\n      storeIndex(7u,parameters.experiments);\n    }\n  }\n  @compute @workgroup_size(64)\n  fn scatter(@builtin(global_invocation_id)id:vec3u) {\n    let invocation=id.x+id.y*4194240u;\n    for(var layer=0u;layer<2u;layer++) {\n      let start=gridStart(layer);\n      if(invocation<parameters.buckets) {\n        let offset=grid[start+invocation].offset+grid[start+parameters.buckets+invocation/256u].offset;\n        storeIndex(8u+layer*parameters.buckets*2u+invocation*2u,offset);\n        storeIndex(9u+layer*parameters.buckets*2u+invocation*2u,atomicLoad(&grid[start+invocation].count));\n      }\n      if(invocation<parameters.count&&photonRecord(invocation).w!=0u) {\n        let bucket=hashPhoton(photonPosition(invocation),layer);\n        let index=grid[start+bucket].offset+grid[start+parameters.buckets+bucket/256u].offset+atomicAdd(&grid[start+bucket].cursor,1u);\n        storeIndex(8u+parameters.buckets*4u+layer*parameters.count+index,invocation);\n      }\n    }\n  }\n";
+/** Packed photon payload: position f32 x3, RGBE8 flux, two octahedral directions. */
+export declare const photonPackingWgsl = "\n  fn photonOctSign(v:vec2f)->vec2f {\n    return select(vec2f(-1.0),vec2f(1.0),v>=vec2f(0.0));\n  }\n  fn photonEncodeOct(n:vec3f)->vec2f {\n    let p=n/(abs(n.x)+abs(n.y)+abs(n.z));\n    let uv=select(p.xy,(vec2f(1.0)-abs(p.yx))*photonOctSign(p.xy),p.z<0.0);\n    return uv*.5+.5;\n  }\n  fn photonDecodeOct(uv:vec2f)->vec3f {\n    let p=uv*2.0-1.0;\n    var n=vec3f(p,1.0-abs(p.x)-abs(p.y));\n    if(n.z<0.0) {\n      n=vec3f((vec2f(1.0)-abs(n.yx))*photonOctSign(n.xy),n.z);\n    }\n    return normalize(n);\n  }\n  // An eight-bit exponent avoids RGB9E5's limited range for very dim/bright lights.\n  // Flux is stored before dividing by emitted count, preserving tiny large batches.\n  fn photonEncodeFlux(value:vec3f)->u32 {\n    let peak=max(max(value.x,value.y),value.z);\n    if(!(peak>0.0)) {\n      return 0u;\n    }\n    let exponent=clamp(i32(floor(log2(peak)))+1,-126,127);\n    let rgb=vec3u(clamp(round(value*exp2(-f32(exponent))*255.0),vec3f(0.0),vec3f(255.0)));\n    return rgb.x|(rgb.y<<8u)|(rgb.z<<16u)|(u32(exponent+127)<<24u);\n  }\n  fn photonDecodeFlux(value:u32)->vec3f {\n    if(value==0u) {\n      return vec3f(0.0);\n    }\n    return vec3f(f32(value&255u),f32((value>>8u)&255u),f32((value>>16u)&255u))*exp2(f32(i32(value>>24u)-127))/255.0;\n  }\n";
+export declare const photonQueryWgsl = "\n  fn photonOctSign(v:vec2f)->vec2f {\n    return select(vec2f(-1.0),vec2f(1.0),v>=vec2f(0.0));\n  }\n  fn photonEncodeOct(n:vec3f)->vec2f {\n    let p=n/(abs(n.x)+abs(n.y)+abs(n.z));\n    let uv=select(p.xy,(vec2f(1.0)-abs(p.yx))*photonOctSign(p.xy),p.z<0.0);\n    return uv*.5+.5;\n  }\n  fn photonDecodeOct(uv:vec2f)->vec3f {\n    let p=uv*2.0-1.0;\n    var n=vec3f(p,1.0-abs(p.x)-abs(p.y));\n    if(n.z<0.0) {\n      n=vec3f((vec2f(1.0)-abs(n.yx))*photonOctSign(n.xy),n.z);\n    }\n    return normalize(n);\n  }\n  // An eight-bit exponent avoids RGB9E5's limited range for very dim/bright lights.\n  // Flux is stored before dividing by emitted count, preserving tiny large batches.\n  fn photonEncodeFlux(value:vec3f)->u32 {\n    let peak=max(max(value.x,value.y),value.z);\n    if(!(peak>0.0)) {\n      return 0u;\n    }\n    let exponent=clamp(i32(floor(log2(peak)))+1,-126,127);\n    let rgb=vec3u(clamp(round(value*exp2(-f32(exponent))*255.0),vec3f(0.0),vec3f(255.0)));\n    return rgb.x|(rgb.y<<8u)|(rgb.z<<16u)|(u32(exponent+127)<<24u);\n  }\n  fn photonDecodeFlux(value:u32)->vec3f {\n    if(value==0u) {\n      return vec3f(0.0);\n    }\n    return vec3f(f32(value&255u),f32((value>>8u)&255u),f32((value>>16u)&255u))*exp2(f32(i32(value>>24u)-127))/255.0;\n  }\n\n  @group(1)@binding(11)var photonMapData:texture_2d<u32>;\n  @group(1)@binding(12)var photonMapGrid:texture_2d<u32>;\n  fn photonLoad(i:u32)->vec4u {\n    let width=textureDimensions(photonMapData).x;\n    return textureLoad(photonMapData,vec2i(vec2u(i%width,i/width)),0);\n  }\n  fn photonIndexWord(i:u32)->u32 {\n    let width=textureDimensions(photonMapGrid).x;\n    return textureLoad(photonMapGrid,vec2i(vec2u(i%width,i/width)),0).x;\n  }\n  fn photonHash(cell:vec3i)->u32 {\n    let p=bitcast<vec3u>(cell);\n    return ((p.x*73856093u)^(p.y*19349663u)^(p.z*83492791u))&(photonIndexWord(5u)-1u);\n  }\n  // Approximate a specular prefix by the camera-space depth of its first diffuse\n  // receiver. Orthographic footprints do not grow with distance. Grazing surfaces\n  // expand the disk conservatively; the progressive world radius remains the cap.\n  fn photonPixelFootprint(si:SurfaceInteraction)->f32 {\n    let inverse=cameraUniform.viewProjectionInv;\n    let a=inverse*vec4f(0.0,0.0,-1.0,1.0);\n    let b=inverse*vec4f(2.0/rtUniform.outputSize.x,0.0,-1.0,1.0);\n    let origin=a.xyz/a.w;\n    let offset=b.xyz/b.w;\n    let eye=cameraUniform.viewInv[3].xyz;\n    let forward=normalize(cameraUniform.viewInv[2].xyz);\n    let depth=abs(dot(si.position-eye,forward));\n    let nearDepth=max(abs(dot(origin-eye,forward)),1e-8);\n    let perspective=abs(inverse[2].w)>1e-8;\n    let footprint=length(offset-origin)*select(1.0,depth/nearDepth,perspective);\n    let cosine=abs(dot(si.faceNormal,normalize(si.position-eye)));\n    return max(1e-8,footprint/sqrt(max(.0625,cosine)));\n  }\n  fn photonContribution(si:SurfaceInteraction,prepared:OpenPbrPreparedBsdf,primary:bool,visible:bool)->vec3f {\n    if(samplingBounce>CAUSTIC_PHOTON_MAX_DEPTH||isCausticSpecular(prepared.material)||prepared.material.materialType==SHADOW_CATCHER) {\n      return vec3f(0.0);\n    }\n    if(CAUSTIC_LIGHT_TRACING&&primary) {\n      return vec3f(0.0);\n    }\n    let globalRadius=bitcast<f32>(photonIndexWord(0u));\n    let screenPixels=bitcast<f32>(photonIndexWord(2u));\n    if(!(globalRadius>0.0)||bitcast<f32>(photonIndexWord(3u))==1.0||(bitcast<f32>(photonIndexWord(3u))==2.0&&visible)) {\n      return vec3f(0.0);\n    }\n    var radius=globalRadius;\n    let experiments=photonIndexWord(7u);\n    let deepPixels=f32(experiments>>8u);\n    let queryPixels=select(deepPixels,screenPixels,visible);\n    if(queryPixels>0.0) {\n      radius=min(radius,photonPixelFootprint(si)*queryPixels);\n    }\n    let layer=select(0u,1u,visible&&screenPixels>0.0);\n    let cellSize=select(globalRadius,2.0*bitcast<f32>(photonIndexWord(6u)),layer==1u);\n    let firstCell=vec3i(floor((si.position-vec3f(radius))/cellSize));\n    let lastCell=vec3i(floor((si.position+vec3f(radius))/cellSize));\n    let buckets=photonIndexWord(5u);\n    let headerStart=8u+layer*buckets*2u;\n    let indexStart=8u+buckets*4u+layer*photonIndexWord(4u);\n    let emitted=f32(photonIndexWord(4u));\n    var sum=vec3f(0.0);\n    for(var z=firstCell.z;z<=lastCell.z;z++) {\n      for(var y=firstCell.y;y<=lastCell.y;y++) {\n        for(var x=firstCell.x;x<=lastCell.x;x++) {\n          let queryCell=vec3i(x,y,z);\n          // Reject cube corners before loading their photon lists. Padding keeps\n          // the bound conservative when cell multiplication rounds in f32.\n          let cellMin=vec3f(queryCell)*cellSize;\n          let padding=max(vec3f(cellSize*1e-5),abs(cellMin)*1e-6);\n          let closest=clamp(si.position,cellMin-padding,cellMin+vec3f(cellSize)+padding);\n          if(distance(closest,si.position)>=radius) { continue; }\n          let bucket=photonHash(queryCell);\n          let offset=photonIndexWord(headerStart+bucket*2u);\n          let count=photonIndexWord(headerStart+1u+bucket*2u);\n          for(var i=offset;i<offset+count;i++) {\n            let index=photonIndexWord(indexStart+i);\n            let record=photonLoad(index);\n            let position=bitcast<vec3f>(record.xyz);\n            // Hash collisions and neighboring buckets never count a photon twice.\n            if(any(vec3i(floor(position/cellSize))!=queryCell)||distance(position,si.position)>=radius) {\n              continue;\n            }\n            var kernelRadius=radius;\n            if((experiments&1u)!=0u) {\n              let scale=bitcast<f32>(photonIndexWord(8u+buckets*4u+u32(emitted)*3u+index));\n              kernelRadius=min(radius,globalRadius*scale);\n              if(!(kernelRadius>0.0)||distance(position,si.position)>=kernelRadius) { continue; }\n            }\n            let directions=photonIndexWord(8u+buckets*4u+u32(emitted)*2u+index);\n            let normal=photonDecodeOct(vec2f(f32(directions&255u)/255.0,f32((directions>>8u)&127u)/127.0));\n            if(dot(normal,si.faceNormal)<0.9) {\n              continue;\n            }\n            if(CAUSTIC_MNEE&&samplingBounce<=CAUSTIC_MNEE_MAX_DEPTH&&(directions&32768u)!=0u&&prepared.material.baseWeight*(1.0-prepared.material.baseMetalness)*(1.0-prepared.material.transmissionWeight)>0.0&&prepared.material.subsurfaceWeight==0.0) {\n              continue;\n            }\n            let incoming=photonDecodeOct(vec2f(f32((directions>>16u)&255u),f32(directions>>24u))/255.0);\n            let cosine=abs(dot(si.faceNormal,incoming));\n            if(cosine<=0.0) {\n              continue;\n            }\n            var flux=photonDecodeFlux(record.w)/emitted;\n            if(SPECTRAL) {\n              flux=sampling_spectral_rgb_filter(flux,si.spectralChannel);\n            }\n            var pdf:f32;\n            let response=openpbr_eval_prepared(prepared,incoming,&pdf);\n            sum+=response*flux/cosine*(radius*radius)/(kernelRadius*kernelRadius);\n          }\n        }\n      }\n    }\n    return sum/(PI*radius*radius);\n  }\n";
+export declare function photonDirectQuerySource(enabled?: boolean): string;
+/** Evidence-only per-query counters; no extra bindings or public render options. */
+export declare function photonQuerySource(debugGroup?: number, directSubset?: boolean): string;