@tinypoly/gpu-renderer 0.0.4 → 0.1.0

package/src/bvh.ts

@@ -1,15 +1,20 @@
 /**
- * Binned SAH bounding volume hierarchy, built straight into the layout the shader reads: eight words per node, the
- * first four a minimum corner and a left child or leaf start, the last four a maximum corner and a right child or
- * triangle count, matching `struct BvhNode` in `wgsl/bvh.wgsl.ts`.
+ * Binned SAH bounding volume hierarchy. It is built as a binary tree of boxes, eight words a node, and then laid out
+ * the way the shader reads it: each node holds the boxes of its two children, so one read tells the traversal both
+ * what it may enter and where to go, and a leaf is tested without being read at all. That is sixteen words a node,
+ * matching `struct BvhNode` in `wgsl/common.wgsl.ts`: per child a minimum corner and a link, then a maximum corner and
+ * a triangle count. A link of zero or more is the node to go to; a negative one is a leaf, `-1 - start` into `order`.
  *
  * The bounds are computed from the same `Float32Array` the shader traverses, so every corner is already a float and
  * needs no widening. Building from double precision positions, as the CPU renderer does, would need each minimum
  * rounded down and each maximum rounded up before they reach the GPU.
  */
 
-/** Words per node in the packed buffer. */
-export const NODE_WORDS = 8;
+/** Words per node in the packed buffer: both children's corners, links and counts. */
+export const NODE_WORDS = 16;
+
+/** Words per node of the binary tree the builder works on: its own corners, and its children or its leaf. */
+const BINARY_WORDS = 8;
 
 /** Triangles a leaf may hold. */
 const MAX_LEAF_SIZE = 4;
@@ -27,7 +32,10 @@ const TRAVERSAL_COST = 0.125;
 export const MAX_BVH_DEPTH = 32;
 
 export interface PackedBvh {
-  /** `nodeCount * NODE_WORDS` words, read as floats for the corners and as signed integers for the links. */
+  /**
+   * `nodeCount * NODE_WORDS` words, read as floats for the corners and as signed integers for the links. Node zero
+   * holds the root's two children, or the root itself beside an empty child when the whole scene is one leaf.
+   */
   nodes: ArrayBuffer;
   nodeCount: number;
   /** Triangle indices in leaf order: a leaf covers `order[start] .. order[start + count - 1]`. */
@@ -66,15 +74,15 @@ export function buildBvh(positions: Float32Array, triangleCount: number): Packed
 
   const { min, max, centroid } = triangleBounds(positions, triangleCount);
   // The heuristic may split down to one triangle per leaf, so the pool starts at the common case and grows.
-  let nodes = new ArrayBuffer(Math.max(4, Math.ceil(triangleCount / 2)) * NODE_WORDS * 4);
+  let nodes = new ArrayBuffer(Math.max(4, Math.ceil(triangleCount / 2)) * BINARY_WORDS * 4);
   let floats = new Float32Array(nodes);
   let ints = new Int32Array(nodes);
 
   const reserve = (needed: number) => {
-    if (needed * NODE_WORDS * 4 <= nodes.byteLength) return;
+    if (needed * BINARY_WORDS * 4 <= nodes.byteLength) return;
 
-    const grown = new ArrayBuffer(Math.max(needed, Math.ceil(nodes.byteLength / (NODE_WORDS * 4) * 1.6))
-      * NODE_WORDS * 4);
+    const grown = new ArrayBuffer(Math.max(needed, Math.ceil(nodes.byteLength / (BINARY_WORDS * 4) * 1.6))
+      * BINARY_WORDS * 4);
 
     new Float32Array(grown).set(floats);
     nodes = grown;
@@ -132,7 +140,7 @@ export function buildBvh(positions: Float32Array, triangleCount: number): Packed
       if (value > cMaxZ) cMaxZ = value;
     }
 
-    const word = node * NODE_WORDS;
+    const word = node * BINARY_WORDS;
 
     floats[word] = minX;
     floats[word + 1] = minY;
@@ -180,7 +188,68 @@ export function buildBvh(positions: Float32Array, triangleCount: number): Packed
     top += 2;
   }
 
-  return { nodes: nodes.slice(0, nodeCount * NODE_WORDS * 4), nodeCount, order, maxDepth };
+  return { ...widen(floats, ints, nodeCount), order, maxDepth };
+}
+
+/**
+ * The binary tree laid out for the shader: one node per interior node of it, holding both children's boxes. Interior
+ * nodes are numbered depth first, left before right, so a node and the child it leads to tend to share a cache line.
+ */
+function widen(floats: Float32Array, ints: Int32Array, binaryCount: number): { nodes: ArrayBuffer; nodeCount: number } {
+  const leaf = (node: number) => ints[node * BINARY_WORDS + 3] < 0;
+  const rootLeaf = leaf(0);
+  // A full binary tree has one fewer interior node than leaves; a single leaf still needs a node to stand in.
+  const nodeCount = rootLeaf ? 1 : (binaryCount - 1) / 2;
+  const nodes = new ArrayBuffer(nodeCount * NODE_WORDS * 4);
+  const outFloats = new Float32Array(nodes);
+  const outInts = new Int32Array(nodes);
+  const index = new Int32Array(binaryCount);
+
+  const writeChild = (word: number, child: number) => {
+    const from = child * BINARY_WORDS;
+
+    for (let axis = 0; axis < 3; axis++) {
+      outFloats[word + axis] = floats[from + axis];
+      outFloats[word + 4 + axis] = floats[from + 4 + axis];
+    }
+
+    outInts[word + 3] = leaf(child) ? ints[from + 3] : index[child];
+    outInts[word + 7] = leaf(child) ? ints[from + 7] : 0;
+  };
+
+  if (rootLeaf) {
+    writeChild(0, 0);
+
+    // A box at the far corner of the float range, which no ray reaches, around a leaf of nothing.
+    for (let axis = 0; axis < 3; axis++) outFloats[8 + axis] = outFloats[12 + axis] = 3.4028234663852886e38;
+
+    outInts[11] = -1;
+    outInts[15] = 0;
+
+    return { nodes, nodeCount };
+  }
+
+  const interior: number[] = [];
+  const pending = [0];
+
+  while (pending.length) {
+    const node = pending.pop() as number;
+
+    index[node] = interior.length;
+    interior.push(node);
+
+    const left = ints[node * BINARY_WORDS + 3], right = ints[node * BINARY_WORDS + 7];
+
+    if (!leaf(right)) pending.push(right);
+    if (!leaf(left)) pending.push(left);
+  }
+
+  interior.forEach((node, at) => {
+    writeChild(at * NODE_WORDS, ints[node * BINARY_WORDS + 3]);
+    writeChild(at * NODE_WORDS + 8, ints[node * BINARY_WORDS + 7]);
+  });
+
+  return { nodes, nodeCount };
 }
 
 /** Surface area, the cost the split heuristic weighs each side by. Empty bounds cost nothing. */

package/src/denoise.ts

@@ -15,6 +15,16 @@ export interface DenoiserSettings {
   onlyNoisy: boolean;
   /** Clamp a pixel brighter than every one of its neighbours to the brightest of them, before filtering. */
   antiFirefly: boolean;
+  /**
+   * How far the filter lets go of the normal and depth tests over a material that scatters light inside, from 0 to 1.
+   * Light leaves such a surface after wandering under it, so its shading hardly follows the normal at each point.
+   */
+  subsurface: number;
+  /**
+   * The smallest footprint a smooth reflection is filtered over, in pixels; 0 leaves a mirror alone. Within it, taps
+   * also have to agree on how far away the reflected thing is, so the edge of a reflected object stays sharp.
+   */
+  mirrorRadius: number;
 }
 
 export const DEFAULT_DENOISER: Readonly<DenoiserSettings> = Object.freeze({
@@ -23,6 +33,8 @@ export const DEFAULT_DENOISER: Readonly<DenoiserSettings> = Object.freeze({
   edge: 2,
   onlyNoisy: true,
   antiFirefly: true,
+  subsurface: 1,
+  mirrorRadius: 3,
 });
 
 /** The settings with `input` applied, each field kept in range. */
@@ -35,6 +47,8 @@ export function mergeDenoiser(base: DenoiserSettings, input: Partial<DenoiserSet
     edge: Math.max(0.01, Number(merged.edge) || DEFAULT_DENOISER.edge),
     onlyNoisy: Boolean(merged.onlyNoisy),
     antiFirefly: Boolean(merged.antiFirefly),
+    subsurface: Math.min(1, Math.max(0, Number(merged.subsurface) || 0)),
+    mirrorRadius: Math.min(16, Math.max(0, Number(merged.mirrorRadius) || 0)),
   };
 }
 
@@ -77,7 +91,8 @@ struct Params {
   converged: f32,
   pixelScale: f32,
   orthographic: u32,
-  padding: vec2<u32>,
+  subsurface: f32,
+  mirrorRadius: f32,
 };
 
 @group(0) @binding(0) var<uniform> params: Params;
@@ -86,6 +101,10 @@ struct Params {
 @group(0) @binding(3) var<storage, read> specularIn: array<vec4<f32>>;
 @group(0) @binding(4) var<storage, read_write> diffuseOut: array<vec4<f32>>;
 @group(0) @binding(5) var<storage, read_write> specularOut: array<vec4<f32>>;
+// The guides each tap compares, worked out once per pixel by prepare: the normal and the depth, then the roughness, or
+// -1 where no sample has landed. The same numbers the accumulation gives, two reads a tap instead of four.
+@group(0) @binding(6) var<storage, read_write> guides: array<vec4<f32>>;
+@group(0) @binding(7) var<storage, read_write> roughnesses: array<f32>;
 
 fn luminance(color: vec3<f32>) -> f32 {
   return dot(color, vec3<f32>(0.2126, 0.7152, 0.0722));
@@ -128,6 +147,11 @@ fn guideRoughness(index: u32) -> f32 {
   return accumulation[2u * params.count + index].w / samples(index);
 }
 
+/** How much of the first surface scatters light inside, averaged over the pixel's samples. */
+fn guideSubsurface(index: u32) -> f32 {
+  return accumulation[5u * params.count + index].w / samples(index);
+}
+
 fn specularDistance(index: u32) -> f32 {
   return accumulation[4u * params.count + index].w / samples(index);
 }
@@ -153,10 +177,15 @@ fn prepare(@builtin(global_invocation_id) id: vec3<u32>) {
   if (count <= 0.0) {
     diffuseOut[index] = vec4<f32>(0.0);
     specularOut[index] = vec4<f32>(0.0);
+    guides[index] = vec4<f32>(0.0);
+    roughnesses[index] = -1.0;
 
     return;
   }
 
+  guides[index] = vec4<f32>(guideNormal(index), guideDepth(index));
+  roughnesses[index] = guideRoughness(index);
+
   let squares = accumulation[params.count + index];
   let diffuse = diffuseMean(index);
   let specular = specularMean(index);
@@ -269,14 +298,26 @@ fn atrous(@builtin(global_invocation_id) id: vec3<u32>) {
     return;
   }
 
-  let normal = guideNormal(index);
-  let depth = guideDepth(index);
-  let roughness = guideRoughness(index);
-  let diffuseSigma = params.edge * sqrt(max(localVariance(pixel, false), 0.0)) + 1e-6;
+  let guide = guides[index];
+  let normal = guide.xyz;
+  let depth = guide.w;
+  let roughness = roughnesses[index];
+  let diffuseSigma = params.edge * sqrt(max(localVariance(pixel, false), 0.0))
+    * (1.0 + guideSubsurface(index) * params.subsurface) + 1e-6;
   let specularSigma = params.edge * sqrt(max(localVariance(pixel, true), 0.0)) + 1e-6;
-  let radius = specularRadius(index, depth);
-  // A smoother surface's reflection turns faster with its normal: the normal test sharpens with it.
-  let specularPower = min(128.0 / max(roughness * roughness, 0.03), 4096.0);
+  // A smooth reflection is filtered over at least the mirror radius, by taps that see the reflected thing at the same
+  // distance: that test keeps the edges of what it reflects, which lets the normal test relax over a curved mirror,
+  // where neighbouring normals part fast and would otherwise leave the reflection unfiltered.
+  let mirror = params.mirrorRadius > 0.0 && roughness < 0.3;
+  let radius = select(specularRadius(index, depth), max(specularRadius(index, depth), params.mirrorRadius), mirror);
+  let reflected = specularDistance(index);
+  // A smoother surface's reflection turns faster with its normal: the normal test sharpens with it, short of shutting
+  // out the neighbours of a curved mirror altogether once the reflected distance guards it.
+  let specularPower = min(128.0 / max(roughness * roughness, 0.03), select(4096.0, 64.0, mirror));
+  // Under a surface that scatters light inside, the normal and the depth say little about the light that leaves it.
+  let scattering = guideSubsurface(index) * params.subsurface;
+  let diffusePower = mix(128.0, 1.0, scattering);
+  let depthSlack = mix(1.0, 4.0, scattering);
   let step = i32(params.step);
   var diffuse = vec4<f32>(0.0);
   var specular = vec4<f32>(0.0);
@@ -299,21 +340,30 @@ fn atrous(@builtin(global_invocation_id) id: vec3<u32>) {
       var ws = base;
 
       if (dx != 0 || dy != 0) {
-        if (accumulation[other].w <= 0.0) {
+        let otherRoughness = roughnesses[other];
+
+        // No sample has landed there.
+        if (otherRoughness < 0.0) {
           continue;
         }
 
+        let otherGuide = guides[other];
         let offset = length(vec2<f32>(f32(dx * step), f32(dy * step)));
-        let alignment = max(dot(normal, guideNormal(other)), 0.0);
-        let depthWeight = exp(-abs(depth - guideDepth(other)) / (0.02 * depth * offset + 1e-3));
+        let alignment = max(dot(normal, otherGuide.xyz), 0.0);
+        let depthWeight = exp(-abs(depth - otherGuide.w) / (0.02 * depth * offset + 1e-3));
+        let diffuseDepth = exp(-abs(depth - otherGuide.w) / (0.02 * depthSlack * depth * offset + 1e-3));
 
-        wd = wd * pow(alignment, 128.0) * depthWeight
+        wd = wd * pow(alignment, diffusePower) * diffuseDepth
           * exp(-abs(luminance(d.rgb) - luminance(diffuseCenter.rgb)) / diffuseSigma);
 
         ws = ws * pow(alignment, specularPower) * depthWeight
-          * exp(-abs(guideRoughness(other) - roughness) * 10.0)
+          * exp(-abs(otherRoughness - roughness) * 10.0)
           * exp(-abs(luminance(s.rgb) - luminance(specularCenter.rgb)) / specularSigma)
           * select(0.0, exp(-offset * offset / (radius * radius)), radius >= 0.5);
+
+        if (mirror) {
+          ws = ws * exp(-abs(specularDistance(other) - reflected) / (0.05 * reflected + 1e-3));
+        }
       }
 
       diffuse = diffuse + vec4<f32>(d.rgb * wd, d.w * wd * wd);
@@ -381,8 +431,21 @@ export class DenoisePass {
   private readonly finish: GPUComputePipeline;
   /** One uniform per stage, so a whole denoise is a single submit. */
   private readonly params: GPUBuffer[];
-  /** Two ping-pong pairs, diffuse and specular, and the output. */
+  /** Two ping-pong pairs, diffuse and specular, the output, then the guides and the roughness. */
   private buffers: GPUBuffer[] = [];
+  /** Bind groups by the buffers they bind, and a number for each buffer seen, which is what the key is made of. */
+  private readonly groups = new Map<string, GPUBindGroup>();
+  private readonly ids = new WeakMap<GPUBuffer, number>();
+  private nextId = 0;
+
+  private idOf(buffer: GPUBuffer): number {
+    let id = this.ids.get(buffer);
+
+    if (id === undefined) this.ids.set(buffer, id = this.nextId++);
+
+    return id;
+  }
+
   private width = 0;
   private height = 0;
 
@@ -402,6 +465,8 @@ export class DenoisePass {
         { binding: 3, visibility, buffer: storage("read-only-storage") },
         { binding: 4, visibility, buffer: storage("storage") },
         { binding: 5, visibility, buffer: storage("storage") },
+        { binding: 6, visibility, buffer: storage("storage") },
+        { binding: 7, visibility, buffer: storage("storage") },
       ],
     });
 
@@ -423,9 +488,12 @@ export class DenoisePass {
 
     const size = Math.max(16, width * height * 16);
 
-    this.buffers = Array.from({ length: 5 }, () =>
+    this.buffers = Array.from({ length: 6 }, () =>
       this.device.createBuffer({ size, usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC }));
+    this.buffers.push(
+      this.device.createBuffer({ size: Math.max(16, width * height * 4), usage: GPUBufferUsage.STORAGE }));
     this.output = this.buffers[4];
+    this.groups.clear();
     this.width = width;
     this.height = height;
   }
@@ -445,6 +513,8 @@ export class DenoisePass {
       floats[7] = CONVERGED_ERROR;
       floats[8] = view.pixelScale;
       words[9] = view.orthographic ? 1 : 0;
+      floats[10] = settings.subsurface;
+      floats[11] = settings.mirrorRadius;
       this.device.queue.writeBuffer(this.params[k], 0, values);
     });
   }
@@ -460,22 +530,34 @@ export class DenoisePass {
     this.resize(width, height);
     this.writeParams(width, height, settings, view);
 
-    const [diffuseA, diffuseB, specularA, specularB, output] = this.buffers;
+    const [diffuseA, diffuseB, specularA, specularB, output, guides, roughnesses] = this.buffers;
     const groups = [Math.ceil(width / 8), Math.ceil(height / 8)] as const;
 
     const run = (pipeline: GPUComputePipeline, params: GPUBuffer,
       sources: [GPUBuffer, GPUBuffer], targets: [GPUBuffer, GPUBuffer]) => {
-      const group = this.device.createBindGroup({
-        layout: this.layout,
-        entries: [
-          { binding: 0, resource: { buffer: params } },
-          { binding: 1, resource: { buffer: accumulation } },
-          { binding: 2, resource: { buffer: sources[0] } },
-          { binding: 3, resource: { buffer: sources[1] } },
-          { binding: 4, resource: { buffer: targets[0] } },
-          { binding: 5, resource: { buffer: targets[1] } },
-        ],
-      });
+      // The same few groups serve every denoise of an image size: they are made once and kept.
+      const key = [params, accumulation, ...sources, ...targets].map(buffer => this.idOf(buffer)).join();
+      let group = this.groups.get(key);
+
+      if (!group) {
+        group = this.device.createBindGroup({
+          layout: this.layout,
+          entries: [
+            { binding: 0, resource: { buffer: params } },
+            { binding: 1, resource: { buffer: accumulation } },
+            { binding: 2, resource: { buffer: sources[0] } },
+            { binding: 3, resource: { buffer: sources[1] } },
+            { binding: 4, resource: { buffer: targets[0] } },
+            { binding: 5, resource: { buffer: targets[1] } },
+            { binding: 6, resource: { buffer: guides } },
+            { binding: 7, resource: { buffer: roughnesses } },
+          ],
+        });
+        // A new accumulation buffer comes with a new size, which clears these; the cap is only a guard.
+        if (this.groups.size > 64) this.groups.clear();
+
+        this.groups.set(key, group);
+      }
 
       const pass = encoder.beginComputePass({ label: "denoise" });
 
@@ -509,6 +591,7 @@ export class DenoisePass {
     for (const buffer of [...this.buffers, ...this.params]) buffer.destroy();
 
     this.buffers = [];
+    this.groups.clear();
     this.output = null;
   }
 }

package/src/dispatchScheduler.ts

@@ -13,6 +13,8 @@ export interface DispatchSchedulerOptions {
   initialRows?: number;
   /** Rows in the whole image. */
   maxRows: number;
+  /** The most paths per pixel one submit may take, which is how far past the whole image a band may grow. */
+  maxSamples?: number;
 }
 
 /** The most one measurement may widen the band, so an unmeasured machine meets the watchdog gradually. */
@@ -24,20 +26,26 @@ export class DispatchScheduler {
   private readonly targetMs: number;
   private readonly initialRows: number;
   private readonly maxRows: number;
+  private readonly maxSamples: number;
+  /** The work of the next submit, in rows times paths per pixel. */
   private rows: number;
-  /** Rows a millisecond, as far as anything has shown. Zero until the first submit says something. */
+  /** Rows a millisecond, one path each, as far as anything has shown. Zero until the first submit says something. */
   private throughput = 0;
 
   constructor(options: DispatchSchedulerOptions) {
     this.targetMs = options.targetMs ?? 12;
     this.maxRows = Math.max(1, Math.floor(options.maxRows));
+    this.maxSamples = Math.max(1, Math.floor(options.maxSamples ?? 1));
     this.initialRows = Math.max(1, Math.min(this.maxRows, Math.floor(options.initialRows ?? 16)));
     this.rows = this.initialRows;
   }
 
-  /** Rows the next submit should cover, never more than what is left. */
-  next(remaining: number): number {
-    return Math.max(1, Math.min(Math.floor(this.rows), Math.floor(remaining)));
+  /**
+   * Rows the next submit should cover, never more than what is left. A submit that takes several paths per pixel
+   * covers that many times fewer rows, so it still takes about the target time.
+   */
+  next(remaining: number, samples = 1): number {
+    return Math.max(1, Math.min(Math.floor(this.rows / Math.max(1, samples)), Math.floor(remaining)));
   }
 
   /**
@@ -52,15 +60,16 @@ export class DispatchScheduler {
    * The rate is measured from the rows actually submitted rather than from the current guess, so one band clamped by
    * what was left of the image does not stop the next one from growing.
    */
-  record(durationMs: number, rows: number) {
+  record(durationMs: number, rows: number, samples = 1) {
     if (!Number.isFinite(durationMs) || durationMs <= 0 || rows <= 0) return;
 
-    const observed = rows / durationMs;
+    const work = rows * Math.max(1, samples);
+    const observed = work / durationMs;
 
     // Good news is believed at once, since it can only mean the band was cheaper than the clock suggested. Bad news
     // is taken gradually, so one submit that happened to land behind something else does not collapse the band.
     this.throughput = observed > this.throughput ? observed : this.throughput * DECAY + observed * (1 - DECAY);
-    this.rows = Math.max(1, Math.min(this.maxRows, rows * GROW, this.targetMs * this.throughput));
+    this.rows = Math.max(1, Math.min(this.maxRows * this.maxSamples, work * GROW, this.targetMs * this.throughput));
   }
 
   /** Back to the unmeasured band, for a new scene or a new size. */

package/src/frameLayout.ts

@@ -30,11 +30,18 @@ export const FRAME_LAYOUT = layoutStruct("Frame", [
   { name: "triangleCount", type: "u32" },
   { name: "lightCount", type: "u32" },
   { name: "areaLightCount", type: "u32", doc: "How many leading lights are emissive triangles." },
+  { name: "lightCdf", type: "u32", doc: "Where every light's cumulative probability starts in the texture buffer." },
+  { name: "lightTriangles", type: "u32", doc: "Where every emissive light's triangle starts in the texture buffer." },
+  { name: "localLightCount", type: "u32", doc: "How many leading lights sit in the scene, which the light tree picks." },
+  { name: "localLightPick", type: "f32", doc: "The chance of picking one of them, the last one's cumulative probability." },
+  { name: "lightTree", type: "u32", doc: "Where the light tree's nodes start in the texture buffer; zero picks by power alone." },
+  { name: "lightLeaves", type: "u32", doc: "Where the leaf of every local light starts in the texture buffer." },
   { name: "clampIndirect", type: "f32", doc: "Maximum luminance one indirect contribution may add; zero leaves it." },
   { name: "shadowSpread", type: "f32", doc: "Tangent of half the angular diameter shadow rays spread over." },
   { name: "shadowsEnabled", type: "u32" },
   { name: "uvOffset", type: "u32", doc: "Where the UV sets start inside the attribute buffer." },
   { name: "orderOffset", type: "u32", doc: "Where the hierarchy's triangle order starts inside the index buffer." },
+  { name: "leafOffset", type: "u32", doc: "Where each triangle's position in that order starts inside the index buffer." },
   { name: "colorOffset", type: "u32", doc: "Where vertex colours start inside the attribute buffer." },
   { name: "hasColors", type: "u32" },
   { name: "varyingOffset", type: "u32", doc: "Where shader varyings start inside the attribute buffer." },
@@ -53,6 +60,17 @@ export const FRAME_LAYOUT = layoutStruct("Frame", [
   { name: "volumetricData", type: "u32" },
   { name: "adaptiveMin", type: "u32", doc: "Samples a pixel takes before it may stop as converged; 0 never stops." },
   { name: "adaptiveThreshold", type: "f32", doc: "The error under which a pixel and its neighbours count as converged." },
+  { name: "lensDistortion", type: "f32", doc: "Radial distortion of the camera ray: positive is barrel, negative pincushion." },
+  { name: "rouletteThreshold", type: "f32", doc: "Throughput under which a path may be ended by chance; zero never." },
+  { name: "rouletteStart", type: "u32", doc: "Bounces every path takes before the roulette may end it." },
+  { name: "subsurfaceEnabled", type: "u32", doc: "Whether materials that scatter light inside do; 0 renders them opaque." },
+  { name: "bevelEnabled", type: "u32", doc: "Whether materials with a bevel radius round their edges." },
+  { name: "volumesEnabled", type: "u32", doc: "Whether volumes scatter light; 0 leaves their meshes out altogether." },
+  { name: "cameraVolume", type: "u32", doc: "The volume material the camera stands inside, or NO_MATERIAL." },
+  { name: "atmosphereColor", type: "vec4f", doc: "rgb is the radiance the air scatters towards the camera, w 1 when it is on." },
+  { name: "atmosphereExtinction", type: "vec4f", doc: "rgb is how much of each colour the air takes per scene unit, w the haze." },
+  { name: "atmosphereSun", type: "vec4f", doc: "xyz is the direction towards the brightest directional light." },
+  { name: "atmosphereSunColor", type: "vec4f", doc: "rgb is that light's irradiance, which the air scatters." },
   { name: "clipPlane", type: "vec4f", doc: "A plane the camera ray drops surfaces behind, for planar captures; w = 0 and xyz = 0 is none." },
 ]);
 

package/src/geometry.ts

@@ -2,6 +2,7 @@ import { clipToWorld, transform4, worldToClip, type CameraMatrices } from "./cam
 import {
   isMaterialVisible,
   MATERIAL_KIND,
+  MATERIAL_UNLIT_LIGHT,
   materialProperties,
   NO_PLACEMENT,
   NO_PROGRAM,
@@ -525,9 +526,12 @@ export function packGeometry(
     materials: new Uint32Array(triangleMaterials),
     triangleCount: triangleMaterials.length,
     materialColors: new Float32Array(properties.flatMap(m => [...m.color, m.opacity])),
-    // Only a lit or plainly emissive surface is a light: a shader's colour is not known until it runs.
+    // Only a lit or plainly emissive surface is a light: a shader's colour is not known until it runs. An unlit
+    // material of a fixed colour emits that colour, so it is one too.
     materialEmissive: new Float32Array(properties.flatMap(m =>
-      m.kind === MATERIAL_KIND.shader ? [0, 0, 0, 0] : [...m.emissive, m.emissiveIntensity])),
+      m.kind === MATERIAL_KIND.shader ? [0, 0, 0, 0]
+        : (m.flags & MATERIAL_UNLIT_LIGHT) !== 0 ? [...m.color, 1]
+          : [...m.emissive, m.emissiveIntensity])),
     materialProperties: properties,
     materialPlacements: entries.map(entry => entry.placement),
     materialCount: properties.length,

package/src/index.ts

@@ -21,6 +21,19 @@ export { attachCanvas } from "./canvas.js";
 export { DEFAULT_DENOISER } from "./denoise.js";
 export type { DenoiserSettings } from "./denoise.js";
 
+// The settings only this renderer reads, beside the CPU renderer's, and the defaults it renders with.
+export {
+  DEFAULT_GPU_RENDER_SETTINGS,
+  effectiveExposure,
+  mergeGpuSettings,
+  physicalExposure,
+  whiteBalanceMatrix,
+} from "./settings.js";
+
+export type { GpuRenderSettings, GpuRenderSettingsInput } from "./settings.js";
+export { subsurfaceMedium } from "./subsurface.js";
+export type { SubsurfaceInput, SubsurfaceMaterial, SubsurfaceMedium } from "./subsurface.js";
+
 export type { Bucket, Listener } from "./upstream.js";
 
 // Scene packing, published because the sizes it produces are what a host checks against a device's limits.