Optimize viewer shaders and camera controls; sample GPU profiles asynchronously

Amp-Thread-ID: https://ampcode.com/threads/T-01a0b382-1e79-712b-abd1-d29201c9bb5e Co-authored-by: Heaust Azure <heaust.azure@gmail.com>
2026-09-18 11:52:09 +00:00
parent bc957cf40d
commit eb8aa0e9dc
21 changed files with 762 additions and 174 deletions
@@ -117,6 +117,15 @@ version = "0.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2"

+[[package]]
+name = "futures-channel"
+version = "0.3.34"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b1f9e3d69d39e4862ffed03ed071a76f9a13ba1d9109d355b0f0aa6b15e393c4"
+dependencies = [
+ "futures-core",
+]
+
 [[package]]
 name = "futures-core"
 version = "0.3.34"
@@ -717,6 +726,7 @@ dependencies = [
 name = "yawn-core"
 version = "0.1.0"
 dependencies = [
+ "futures-channel",
 "js-sys",
 "serde",
 "serde_json",
@@ -230,6 +230,9 @@ function pbrShader(
  transmission: boolean,
  textureInfo: ArrayLike<number>,
  extensionInfo: ArrayLike<number>,
+  sharedOcclusion: boolean,
+  lightMask: number,
+  directToCanvas: boolean,
 ) {
  const names = [
    "baseTexture",
@@ -306,11 +309,11 @@ struct VertexOutput {
@group(0) @binding(5) var<storage, read> meshInfo: array<u32>;
@group(0) @binding(6) var<storage, read> materials: array<vec4<f32>>;
@group(0) @binding(7) var<storage, read> cameraMatrices: array<vec4<f32>>;
-@group(0) @binding(8) var<storage, read> pointLights: array<vec4<f32>>;
-@group(0) @binding(9) var<storage, read> spotLights: array<vec4<f32>>;
-@group(0) @binding(10) var<storage, read> directionalLights: array<vec4<f32>>;
+${lightMask & 1 ? "@group(0) @binding(8) var<storage, read> pointLights: array<vec4<f32>>;" : ""}
+${lightMask & 2 ? "@group(0) @binding(9) var<storage, read> spotLights: array<vec4<f32>>;" : ""}
+${lightMask & 4 ? "@group(0) @binding(10) var<storage, read> directionalLights: array<vec4<f32>>;" : ""}
 ${mask ? "@group(1) @binding(0) var materialSampler: sampler;" : ""}
-${names.map((name, slot) => (texture(slot) ? `@group(1) @binding(${slot + 1}) var ${name}: texture_2d<f32>;` : "")).join("\n")}
+${names.map((name, slot) => (texture(slot) && !(slot === 4 && sharedOcclusion) ? `@group(1) @binding(${slot + 1}) var ${name}: texture_2d<f32>;` : "")).join("\n")}
 ${transmission ? "@group(2) @binding(0) var sceneColor: texture_2d<f32>;\n@group(2) @binding(1) var sceneSampler: sampler;" : ""}

 fn rotate(rotor: vec4<f32>, value: vec3<f32>) -> vec3<f32> {
@@ -357,11 +360,10 @@ fn pow5(value: f32) -> f32 {
  return squared * squared * value;
 }

-fn directContribution(normal: vec3<f32>, view: vec3<f32>, lightDirection: vec3<f32>, radiance: vec3<f32>, baseColor: vec3<f32>, metallic: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
+fn directContribution(normal: vec3<f32>, view: vec3<f32>, lightDirection: vec3<f32>, radiance: vec3<f32>, diffuseColor: vec3<f32>, nDotV: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
  let nDotL = max(dot(normal, lightDirection), 0.0);
  if (nDotL <= 0.0) { return vec3(0.0); }
  let halfVector = normalize(lightDirection + view);
-  let nDotV = max(dot(normal, view), 0.0001);
  let nDotH = max(dot(normal, halfVector), 0.0);
  let vDotH = max(dot(view, halfVector), 0.0);
  let fresnel = f0 + (vec3(1.0) - f0) * pow5(1.0 - vDotH);
@@ -372,11 +374,14 @@ fn directContribution(normal: vec3<f32>, view: vec3<f32>, lightDirection: vec3<f
  let geometryV = nDotV / mix(nDotV, 1.0, alpha * 0.5);
  let geometryL = nDotL / mix(nDotL, 1.0, alpha * 0.5);
  let specular = distribution * geometryV * geometryL * fresnel / max(4.0 * nDotV * nDotL, 0.0001);
-  let diffuse = baseColor * (vec3(1.0) - fresnel) * (1.0 - metallic) / 3.14159265;
+  let diffuse = diffuseColor * (vec3(1.0) - fresnel);
  return (diffuse + specular) * radiance * nDotL;
 }

-fn pointContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, world: vec3<f32>, baseColor: vec3<f32>, metallic: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
+${
+  lightMask & 1
+    ? `
+fn pointContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, world: vec3<f32>, diffuseColor: vec3<f32>, nDotV: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
  let colorIntensity = pointLights[index * 2u];
  let state = pointLights[index * 2u + 1u];
  if (state.y == 0.0 || state.x <= 0.0) { return vec3(0.0); }
@@ -385,10 +390,16 @@ fn pointContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, world: vec3
  let falloff = max(1.0 - distance / state.x, 0.0);
  let radiance = colorIntensity.rgb * colorIntensity.a * falloff * falloff * 0.08;
  let sourceRoughness = clamp(roughness + max(state.z, 0.0) / max(distance, 0.0001), roughness, 1.0);
-  return directContribution(normal, view, offset / max(distance, 0.0001), radiance, baseColor, metallic, sourceRoughness, f0);
+  return directContribution(normal, view, offset / max(distance, 0.0001), radiance, diffuseColor, nDotV, sourceRoughness, f0);
+}
+`
+    : ""
 }

-fn spotContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, world: vec3<f32>, baseColor: vec3<f32>, metallic: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
+${
+  lightMask & 2
+    ? `
+fn spotContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, world: vec3<f32>, diffuseColor: vec3<f32>, nDotV: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
  let colorIntensity = spotLights[index * 2u];
  let state = spotLights[index * 2u + 1u];
  if (state.w == 0.0 || state.x <= 0.0) { return vec3(0.0); }
@@ -398,16 +409,27 @@ fn spotContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, world: vec3<
  let cone = smoothstep(cos(state.z), cos(state.y), dot(direction, fromLight / max(distance, 0.0001)));
  let falloff = max(1.0 - distance / state.x, 0.0);
  let radiance = colorIntensity.rgb * colorIntensity.a * cone * falloff * falloff * 0.08;
-  return directContribution(normal, view, -fromLight / max(distance, 0.0001), radiance, baseColor, metallic, roughness, f0);
+  return directContribution(normal, view, -fromLight / max(distance, 0.0001), radiance, diffuseColor, nDotV, roughness, f0);
+}
+`
+    : ""
 }

-fn directionalContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, baseColor: vec3<f32>, metallic: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
+${
+  lightMask & 4
+    ? `
+fn directionalContribution(index: u32, normal: vec3<f32>, view: vec3<f32>, diffuseColor: vec3<f32>, nDotV: f32, roughness: f32, f0: vec3<f32>) -> vec3<f32> {
  let colorIntensity = directionalLights[index * 2u];
  let state = directionalLights[index * 2u + 1u];
  if (state.x == 0.0) { return vec3(0.0); }
  let lightDirection = normalize(rotate(rotors[index], vec3<f32>(0.0, 0.0, 1.0)));
-  return directContribution(normal, view, lightDirection, colorIntensity.rgb * colorIntensity.a * 0.1, baseColor, metallic, roughness, f0);
+  return directContribution(normal, view, lightDirection, colorIntensity.rgb * colorIntensity.a * 0.1, diffuseColor, nDotV, roughness, f0);
 }
+`
+    : ""
+}
+
+${directToCanvas ? canvasOutputShader : ""}

@fragment
 fn fragment(input: VertexOutput, @builtin(front_facing) frontFacing: bool) -> @location(0) vec4<f32> {
@@ -432,6 +454,7 @@ fn fragment(input: VertexOutput, @builtin(front_facing) frontFacing: bool) -> @l
  ${normalTexture ? `let tangent = normalize(input.tangent.xyz - normal * dot(input.tangent.xyz, normal)); let bitangent = cross(normal, tangent) * input.tangent.w; let mapped = ${sample(2)}.xyz * 2.0 - 1.0; normal = normalize(mat3x3<f32>(tangent, bitangent, normal) * vec3(mapped.xy * extra.y, mapped.z));` : ""}
  let view = normalize(cameraMatrices[4].xyz - input.world);
  let nDotV = max(dot(normal, view), 0.0001);
+  let diffuseColor = base.rgb * (1.0 - metallic) / 3.14159265;
  let specularWeight = extension2.z * ${sample(11)}.a;
  let dielectricColor = vec3(${number(dielectric)} * specularWeight) * extension6.rgb * ${sample(12)}.rgb;
  var f0 = ${specularGlossiness ? "packedMaterial.rgb * extension3.rgb" : "mix(dielectricColor, base.rgb, metallic)"};
@@ -448,30 +471,48 @@ fn fragment(input: VertexOutput, @builtin(front_facing) frontFacing: bool) -> @l
  let fresnelPower = pow5(1.0 - nDotV);
  let ambient = vec3<f32>(bitcast<f32>(clusters[${compactAmbientStart}]), bitcast<f32>(clusters[${compactAmbientStart + 1}]), bitcast<f32>(clusters[${compactAmbientStart + 2}]));
  var directLight = vec3<f32>(0.0);
+  ${
+    lightMask & 1
+      ? `
  let pointCount = clusters[10];
  if (pointCount <= ${compactLightLimit}u) {
    for (var light = 0u; light < pointCount; light++) {
-      directLight += pointContribution(clusters[${compactPointStart}u + light], normal, view, input.world, base.rgb, metallic, roughness, f0);
+      directLight += pointContribution(clusters[${compactPointStart}u + light], normal, view, input.world, diffuseColor, nDotV, roughness, f0);
    }
  } else {
    for (var light = 0u; light < arrayLength(&pointLights) / 2u; light++) {
-      directLight += pointContribution(light, normal, view, input.world, base.rgb, metallic, roughness, f0);
+      directLight += pointContribution(light, normal, view, input.world, diffuseColor, nDotV, roughness, f0);
    }
  }
+  `
+      : ""
+  }
+  ${
+    lightMask & 2
+      ? `
  let spotCount = clusters[${compactSpotCount}];
  if (spotCount <= ${compactLightLimit}u) {
    for (var light = 0u; light < spotCount; light++) {
-      directLight += spotContribution(clusters[${compactSpotStart}u + light], normal, view, input.world, base.rgb, metallic, roughness, f0);
+      directLight += spotContribution(clusters[${compactSpotStart}u + light], normal, view, input.world, diffuseColor, nDotV, roughness, f0);
    }
  } else {
    for (var light = 0u; light < arrayLength(&spotLights) / 2u; light++) {
-      directLight += spotContribution(light, normal, view, input.world, base.rgb, metallic, roughness, f0);
+      directLight += spotContribution(light, normal, view, input.world, diffuseColor, nDotV, roughness, f0);
    }
  }
+  `
+      : ""
+  }
+  ${
+    lightMask & 4
+      ? `
  for (var light = 0u; light < arrayLength(&directionalLights) / 2u; light++) {
-    directLight += directionalContribution(light, normal, view, base.rgb, metallic, roughness, f0);
+    directLight += directionalContribution(light, normal, view, diffuseColor, nDotV, roughness, f0);
  }
-  let occlusion = mix(1.0, ${sample(4)}.r, extension2.w);
+  `
+      : ""
+  }
+  let occlusion = mix(1.0, ${sharedOcclusion ? "packedMaterial" : sample(4)}.r, extension2.w);
  let emissive = vec3(properties.z, properties.w, extra.x) * ${sample(3, "vec4(1.0)")}.rgb * extension1.w;
  ${sheen ? "let sheen = extension3.rgb * pow(1.0 - nDotV, mix(2.0, 8.0, extension3.w));" : ""}
  ${
@@ -508,7 +549,7 @@ fn fragment(input: VertexOutput, @builtin(front_facing) frontFacing: bool) -> @l
  color = mix(color, transmitted, transmissionAmount);`
      : ""
  }
-  return vec4(color * accent.a, ${alphaMode === 2 ? "base.a" : "1.0"});
+  return ${directToCanvas ? "canvasOutput(vec4(color * accent.a, 1.0))" : `vec4(color * accent.a, ${alphaMode === 2 ? "base.a" : "1.0"})`};
 }`;
 }

@@ -559,13 +600,23 @@ function effectFragment(kind: string, options: Record<string, unknown>) {
 }`;
 }

-function presentShader(toneMap: string, composite: "scene" | undefined) {
-  const tone =
-    toneMap === "reinhard"
+function toneMapping(toneMap: string) {
+  return toneMap === "reinhard"
    ? "color / (color + vec3(1.0))"
    : toneMap === "linear"
      ? "clamp(color, vec3(0.0), vec3(1.0))"
      : "clamp((color * (2.51 * color + vec3(0.03))) / (color * (2.43 * color + vec3(0.59)) + vec3(0.14)), vec3(0.0), vec3(1.0))";
+}
+
+// Match the intermediate rgba16float rounding before applying the same presentation transform.
+const canvasOutputShader = /* wgsl */ `
+fn canvasOutput(value: vec4<f32>) -> vec4<f32> {
+  let rounded = vec4(unpack2x16float(pack2x16float(value.rg)), unpack2x16float(pack2x16float(value.ba)));
+  let color = max(rounded.rgb, vec3(0.0));
+  return vec4(${toneMapping("aces")}, rounded.a);
+}`;
+
+function presentShader(toneMap: string, composite: "scene" | undefined) {
  return `${fullscreenVertex}
@group(0) @binding(0) var source: texture_2d<f32>;
@group(0) @binding(1) var sourceSampler: sampler;
@@ -573,7 +624,7 @@ ${composite === "scene" ? "@group(0) @binding(2) var scene: texture_2d<f32>;" :
@fragment fn fragment(input: VertexOutput) -> @location(0) vec4<f32> {
  let value = textureSample(source, sourceSampler, input.uv);
  let color = max(${composite === "scene" ? "value.rgb + textureSample(scene, sourceSampler, input.uv).rgb * (1.0 - value.a)" : "value.rgb"}, vec3(0.0));
-  return vec4(${tone}, ${composite ? "1.0" : "value.a"});
+  return vec4(${toneMapping(toneMap)}, ${composite ? "1.0" : "value.a"});
 }`;
 }

@@ -611,6 +662,8 @@ export class Scene {
  readonly core: YawnCore;
  readonly ready: Promise<this>;
  readonly hdr: boolean;
+  /** Fuse tone mapping into opaque PBR draws when no linear-color consumers need HDR. */
+  readonly directToCanvas: boolean;
  readonly clearColor: readonly [number, number, number, number];
  #graphUpdates = Promise.resolve();
  #renderGraph?: object;
@@ -635,6 +688,11 @@ export class Scene {
  #signals?: Float32Array;
  #arrays = new WeakMap<object, object>();
  #views = new WeakMap<object, object>();
+  #lights = {
+    point: new Set<number>(),
+    spot: new Set<number>(),
+    directional: new Set<number>(),
+  };

  constructor(
    canvas: HTMLCanvasElement,
@@ -643,6 +701,7 @@ export class Scene {
      debug?: boolean;
      fps?: number;
      hdr?: boolean;
+      directToCanvas?: boolean;
      clearColor?: ArrayLike<number>;
    } = {},
  ) {
@@ -653,6 +712,7 @@ export class Scene {
    )
      throw new TypeError("Scene clearColor must contain four finite numbers");
    this.hdr = options.hdr ?? true;
+    this.directToCanvas = options.directToCanvas ?? false;
    this.clearColor = clearColor as [number, number, number, number];
    this.core = new YawnCore(canvas, {
      arenaBytes: options.arenaBytes,
@@ -1171,6 +1231,21 @@ export class Scene {
    return this.updateRenderGraph();
  }

+  /** Light handles register their lifetime, not intensity; hot edits never select a new variant. */
+  setLight(
+    kind: "point" | "spot" | "directional",
+    id: number,
+    present: boolean,
+  ) {
+    const lights = this.#lights[kind];
+    const hadType = lights.size > 0;
+    if (present) lights.add(id);
+    else lights.delete(id);
+    return hadType !== Boolean(lights.size)
+      ? this.updateRenderGraph()
+      : Promise.resolve();
+  }
+
  updateRenderGraph() {
    this.markDirty(true);
    if (this.#graphBatchDepth) {
@@ -1327,9 +1402,31 @@ export class Scene {
    const hasTransmission = forwardDraws.some(
      (draw) => this.array("materialExtensions").row(draw.material)[0] > 0,
    );
+    // Custom compute can author light rows directly, so retain the generic shader in that case.
+    const lightMask = this.#computePasses.size
+      ? 7
+      : (this.#lights.point.size ? 1 : 0) |
+        (this.#lights.spot.size ? 2 : 0) |
+        (this.#lights.directional.size ? 4 : 0);
+    const directToCanvas =
+      this.directToCanvas &&
+      this.hdr &&
+      !this.#computePasses.size &&
+      !this.#effects.size &&
+      !this.#shaders.size &&
+      !hasTransmission &&
+      forwardDraws.length > 0 &&
+      forwardDraws.every(
+        (draw) =>
+          (Math.round(this.array("materials").row(draw.material)[11]) & 3) !==
+            2 &&
+          draw.geometry.data.has("normals") &&
+          draw.geometry.data.has("uvs"),
+      );
    const compositeSceneColor =
      hasTransmission && this.#effects.size === 0 && this.#shaders.size === 0;
    const hdrFormat = this.hdr ? "rgba16float" : "rgba8unorm";
+    if (!directToCanvas)
      addTexture({
        id: "hdr",
        format: hdrFormat,
@@ -1374,6 +1471,31 @@ export class Scene {

    let previous = computeIds.length ? computeIds : ["cluster-lights"];
    const bucketRows = new Map<number, number[][]>();
+    if (directToCanvas) {
+      // Shares attachments with the forward draws, so core merges this constant background
+      // into their physical render pass. Use the same HDR rounding/tone mapping for the clear.
+      renderPipelines.push({
+        id: "canvas-background",
+        code: `${fullscreenVertex}\n${canvasOutputShader}\n@fragment fn fragment() -> @location(0) vec4<f32> { return canvasOutput(vec4<f32>(${this.clearColor.map((value) => (Number.isInteger(value) ? `${value}.0` : String(value))).join(", ")})); }`,
+        vertex: { entry: "vertex" },
+        fragment: { entry: "fragment", targets: [{ format: "canvas" }] },
+        depthStencil: {
+          format: "depth24plus",
+          depth_write_enabled: false,
+          depth_compare: "always",
+        },
+      });
+      passes.push({
+        id: "canvas-background",
+        type: "render",
+        pipeline: "canvas-background",
+        after: previous,
+        color: [{ resource: "canvas", clear: [0, 0, 0, 0] }],
+        depth: { resource: "depth", clear: 1 },
+        draw: { vertices: 3 },
+      });
+      previous = ["canvas-background"];
+    }
    if (!forwardDraws.length) {
      renderPipelines.push({
        id: "empty-forward",
@@ -1435,10 +1557,14 @@ export class Scene {
        sync: "manual",
      });
      const forwardPipelines = new Map<string, string>();
-      let firstColor = true;
-      let firstDepth = true;
+      let firstColor = !directToCanvas;
+      let firstDepth = !directToCanvas;
      let sceneColorReady = false;
-      let forwardTarget = compositeSceneColor ? "scene-color" : "hdr";
+      let forwardTarget = directToCanvas
+        ? "canvas"
+        : compositeSceneColor
+          ? "scene-color"
+          : "hdr";
      const attachments = () => {
        const color = firstColor
          ? { clear: [...this.clearColor] }
@@ -1518,6 +1644,13 @@ export class Scene {
        if (extensionInfo[11] <= 0) mask &= ~(1 << 4);
        if ((flags & 8) !== 0 && !transmitted) mask &= (1 << 0) | (1 << 3);
        const textureInfo = this.array("materialTextureInfo").row(material);
+        // All material samples currently use one sampler. Reuse only identical texture/UV reads.
+        const sharedOcclusion =
+          Boolean(mask & (1 << 1) && mask & (1 << 4)) &&
+          materialTextures[1] === materialTextures[4] &&
+          Array.from(textureInfo)
+            .slice(8, 15)
+            .every((value, lane) => value === textureInfo[32 + lane]);
        let clearCompositeLayer = false;
        if (transmitted || (flags & 3) === 2) flushBuckets();
        if (transmitted && !sceneColorReady) {
@@ -1558,7 +1691,16 @@ export class Scene {
          }
        }
        const code = detailed
-          ? pbrShader(mask, flags, transmitted, textureInfo, extensionInfo)
+          ? pbrShader(
+              mask,
+              flags,
+              transmitted,
+              textureInfo,
+              extensionInfo,
+              sharedOcclusion,
+              lightMask,
+              directToCanvas,
+            )
          : basicForwardShader;
        const pipelineKey = `${flags}-${transmitted ? 1 : 0}\n${code}`;
        let pipeline = forwardPipelines.get(pipelineKey);
@@ -1583,7 +1725,7 @@ export class Scene {
              entry: "fragment",
              targets: [
                {
-                  format: hdrFormat,
+                  format: directToCanvas ? "canvas" : hdrFormat,
                  ...((flags & 3) === 2
                    ? {
                        blend: {
@@ -1644,7 +1786,7 @@ export class Scene {
                  binding: 10,
                  resource: "directional-lights",
                },
-              ]
+              ].filter((_, index) => lightMask & (1 << index))
            : []),
          ...(detailed && mask
            ? [
@@ -1656,7 +1798,9 @@ export class Scene {
                    : "material-linear",
                },
                ...materialTextures.flatMap((value, lane) =>
-                  value && (mask & (1 << lane)) !== 0
+                  value &&
+                  (mask & (1 << lane)) !== 0 &&
+                  !(lane === 4 && sharedOcclusion)
                    ? [
                        {
                          group: 1,
@@ -1770,6 +1914,7 @@ export class Scene {
      previous = [pipeline];
    }

+    if (!directToCanvas) {
      const toneMap = String(
        [...this.#effects.values()].find(
          (effect) => effect.kind === "colorGrading",
@@ -1796,6 +1941,7 @@ export class Scene {
        color: [{ resource: "canvas", clear: [...this.clearColor] }],
        draw: { vertices: 3 },
      });
+    }

    const graph = {
      id: "scene",
@@ -25,6 +25,9 @@ export class ArcRotateCamera extends Camera {
  #controls?: ArcRotateControls;
  #pointer?: { x: number; y: number; button: number };
  #frame = 0;
+  #inertiaFrame = 0;
+  #lastFrame = 0;
+  #motion = [0, 0, 0, 0, 0]; // orbit x/y, pan x/y, logarithmic zoom

  constructor(scene: Scene, options: ArcRotateCameraOptions = {}) {
    super(scene, options);
@@ -116,6 +119,9 @@ export class ArcRotateCamera extends Camera {
    }
    if (this.#frame) cancelAnimationFrame(this.#frame);
    this.#frame = 0;
+    if (this.#inertiaFrame) cancelAnimationFrame(this.#inertiaFrame);
+    this.#inertiaFrame = 0;
+    this.#motion.fill(0);
    this.#pointer = undefined;
    this.#controls = undefined;
  }
@@ -161,7 +167,14 @@ export class ArcRotateCamera extends Camera {
    const y = event.clientY - this.#pointer.y;
    this.#pointer.x = event.clientX;
    this.#pointer.y = event.clientY;
-    if (this.#pointer.button === 2) {
+    const offset = this.#pointer.button === 2 ? 2 : 0;
+    this.#motion[offset] += x;
+    this.#motion[offset + 1] += y;
+    this.#scheduleInertia();
+  };
+
+  #applyDrag(x: number, y: number, pan: boolean) {
+    if (pan) {
      const row = this.cameraRow();
      this.#updateTransform(() => {
        const element = this.#controls!.element;
@@ -184,26 +197,46 @@ export class ArcRotateCamera extends Camera {
        row[18] -= horizontal * sinAlpha + vertical * cosBeta * cosAlpha;
      });
    } else {
-      const speed = this.#controls.orbitSpeed ?? 0.005;
+      const speed = this.#controls?.orbitSpeed ?? 0.005;
      const row = this.cameraRow();
      this.#updateTransform(() => {
        row[13] -= x * speed;
        row[14] = Math.min(
          Math.PI - 0.001,
-          Math.max(0.001, row[14] + y * speed),
+          Math.max(0.001, row[14] - y * speed),
        );
      });
    }
-  };
+  }
  #wheel = (event: WheelEvent) => {
    event.preventDefault();
-    const row = this.cameraRow();
-    this.#updateTransform(() => {
-      row[15] = Math.max(
-        0.01,
-        row[15] * Math.exp(event.deltaY * (this.#controls?.zoomSpeed ?? 0.001)),
-      );
+    this.#motion[4] += event.deltaY * (this.#controls?.zoomSpeed ?? 0.001);
+    this.#scheduleInertia();
+  };
+
+  #scheduleInertia() {
+    if (this.#inertiaFrame) return;
+    this.#lastFrame = performance.now();
+    this.#inertiaFrame = requestAnimationFrame(this.#tickInertia);
+  }
+
+  #tickInertia = (now: number) => {
+    this.#inertiaFrame = 0;
+    const fraction = -Math.expm1(-Math.max(0, now - this.#lastFrame) / 80);
+    this.#lastFrame = now;
+    const step = this.#motion.map((remaining, index) => {
+      const value =
+        Math.abs(remaining) < 0.00001 ? remaining : remaining * fraction;
+      this.#motion[index] -= value;
+      return value;
    });
+    this.batchTransformChanges(() => {
+      if (step[0] || step[1]) this.#applyDrag(step[0], step[1], false);
+      if (step[2] || step[3]) this.#applyDrag(step[2], step[3], true);
+      if (step[4]) this.radius *= Math.exp(step[4]);
+    });
+    if (this.#motion.some(Boolean))
+      this.#inertiaFrame = requestAnimationFrame(this.#tickInertia);
  };
  #pollController = () => {
    const pad = navigator.getGamepads?.().find(Boolean);
@@ -10,6 +10,7 @@ export class DirectionalLight extends Node {
    this.ready = nodeReady.then(async () => {
      const array = await scene.ensureRows("directionalLights", this.id + 1, 32, "f32");
      array.row(this.id).set([...Array.from(options.color ?? [1, 1, 1]), options.intensity ?? 1, 1, 0, 0, 0]);
+      await scene.setLight("directional", this.id, true);
      return this;
    });
  }
@@ -21,6 +22,7 @@ export class DirectionalLight extends Node {
    await this.ready;
    if (this.disposed) return;
    this.scene.array("directionalLights").row(this.id).fill(0);
+    await this.scene.setLight("directional", this.id, false);
    await super.dispose();
  }
 }
@@ -29,6 +29,7 @@ export class PointLight extends Node {
          options.radius ?? 0,
          0,
        ]);
+      await scene.setLight("point", this.id, true);
      return this;
    });
  }
@@ -62,6 +63,7 @@ export class PointLight extends Node {
    await this.ready;
    if (this.disposed) return;
    this.scene.array("pointLights").row(this.id).fill(0);
+    await this.scene.setLight("point", this.id, false);
    await super.dispose();
  }
 }
@@ -19,6 +19,7 @@ export class SpotLight extends Node {
        ...Array.from(options.color ?? [1, 1, 1]), options.intensity ?? 1,
        options.range ?? 10, options.innerAngle ?? 0.35, options.outerAngle ?? 0.7, 1,
      ]);
+      await scene.setLight("spot", this.id, true);
      return this;
    });
  }
@@ -32,6 +33,7 @@ export class SpotLight extends Node {
    await this.ready;
    if (this.disposed) return;
    this.scene.array("spotLights").row(this.id).fill(0);
+    await this.scene.setLight("spot", this.id, false);
    await super.dispose();
  }
 }
@@ -292,7 +292,31 @@ test("Node, ArcRotateCamera, and light handles update their shared rows", async
  assert.equal(scene.array("pointLights").row(light.id)[6], 3);
 });

-test("ArcRotateCamera pans one projected world pixel per pointer pixel", async () => {
+test("ArcRotateCamera damps pan, orbit and zoom with reversed vertical orbit", async (t) => {
+  let now = 0;
+  let next = 1;
+  const frames = new Map<number, FrameRequestCallback>();
+  const originalRequest = globalThis.requestAnimationFrame;
+  const originalCancel = globalThis.cancelAnimationFrame;
+  t.mock.method(performance, "now", () => now);
+  globalThis.requestAnimationFrame = (callback) => {
+    const id = next++;
+    frames.set(id, callback);
+    return id;
+  };
+  globalThis.cancelAnimationFrame = (id) => {
+    frames.delete(id);
+  };
+  t.after(() => {
+    globalThis.requestAnimationFrame = originalRequest;
+    globalThis.cancelAnimationFrame = originalCancel;
+  });
+  const advance = (ms: number) => {
+    now += ms;
+    const callbacks = [...frames.values()];
+    frames.clear();
+    for (const callback of callbacks) callback(now);
+  };
  const listeners = new Map<string, Set<(event: any) => void>>();
  const element = {
    clientWidth: 400,
@@ -329,7 +353,43 @@ test("ArcRotateCamera pans one projected world pixel per pointer pixel", async (
    pointerId: 1,
  });
  dispatch("pointermove", { clientX: 120, clientY: 110 });
-
+  dispatch("pointerup", {});
+  advance(80);
+  close(camera.position, [-2 * (1 - Math.exp(-1)), 1 - Math.exp(-1), 10]);
+  advance(1600);
  close(camera.position, [-2, 1, 10]);
+  advance(16);
+  assert.equal(frames.size, 0);
+  const orbit = () => {
+    dispatch("pointerdown", {
+      button: 0,
+      clientX: 100,
+      clientY: 100,
+      pointerId: 1,
+    });
+    dispatch("pointermove", { clientX: 140, clientY: 120 });
+    dispatch("pointerup", {});
+  };
+  orbit();
+  advance(40);
+  const released = camera.alpha;
+  advance(40);
+  assert.ok(camera.alpha < released);
+  const pose = [camera.alpha, camera.beta];
+  close(pose, [
+    -0.2 * (1 - Math.exp(-1)),
+    Math.PI / 2 - 0.1 * (1 - Math.exp(-1)),
+  ]);
  camera.detachControls();
+  camera.alpha = 0;
+  camera.beta = Math.PI / 2;
+  camera.attachControls({ element });
+  orbit();
+  advance(80);
+  close([camera.alpha, camera.beta], pose);
+  dispatch("wheel", { deltaY: 100 });
+  advance(80);
+  close([camera.radius], [10 * Math.exp(0.1 * (1 - Math.exp(-1)))]);
+  camera.detachControls();
+  assert.equal(frames.size, 0);
 });
@@ -4,6 +4,11 @@ import { Mesh } from "../src/Mesh";
 import { FXAA } from "../src/PostProcesses/FXAA";
 import { PBRMaterial } from "../src/materials/PBRMaterial";
 import { Texture } from "../src/materials/Texture";
+import { ShaderMaterial } from "../src/materials/ShaderMaterial";
+import { PointLight } from "../src/lights/PointLight";
+import { SpotLight } from "../src/lights/SpotLight";
+import { DirectionalLight } from "../src/lights/DirectionalLight";
+import { ComputePass } from "../src/ComputePass";
 import {
  bucketDraws,
  createScene,
@@ -34,6 +39,11 @@ test("empty scenes use neutral authored defaults without placeholder geometry",

 test("opaque PBR graph has no implicit environment and compacts sparse lights", async () => {
  const scene = await createScene();
+  await Promise.all([
+    new PointLight(scene).ready,
+    new SpotLight(scene).ready,
+    new DirectionalLight(scene).ready,
+  ]);
  const material = new PBRMaterial(scene);
  await material.ready;
  const mesh = detailedMesh(scene, material);
@@ -451,3 +461,151 @@ test("serialized render graphs compile into explicitly selectable loadouts", asy
  assert.equal(loadoutId, `loadout-${before + 1}`);
  assert.equal(core.loadouts.at(-1), loadoutId);
 });
+
+test("packed ORM reads are shared only for the same texture and UV transform", async () => {
+  const scene = await createScene();
+  const orm = new Texture(scene, { size: [1, 1, 1] });
+  const other = new Texture(scene, { size: [1, 1, 1] });
+  await Promise.all([orm.ready, other.ready]);
+  for (const [occlusion, info, shared] of [
+    [orm, {}, true],
+    [other, {}, false],
+    [orm, { texCoord: 1 }, false],
+    [orm, { offset: [0.25, -0.1] }, false],
+    [orm, { scale: [2, 1] }, false],
+    [orm, { rotation: 0.4 }, false],
+  ] as const) {
+    const material = new PBRMaterial(scene, {
+      metallicRoughnessTexture: orm,
+      occlusionTexture: occlusion,
+      textureInfos: [{}, {}, {}, {}, info],
+    });
+    await material.ready;
+    const mesh = await detailedMesh(scene, material).ready;
+    const graph = latestGraph(scene);
+    const pass = forwardPasses(graph).at(-1);
+    const code = renderPipeline(graph, pass.pipeline).code;
+    assert.equal(code.includes("mix(1.0, packedMaterial.r"), shared);
+    assert.equal(code.includes("textureSample(occlusionTexture"), !shared);
+    assert.equal(
+      pass.bindings.some((b: any) => b.group === 1 && b.binding === 5),
+      !shared,
+    );
+    await mesh.dispose();
+  }
+});
+
+test("light variants follow first/last light lifetimes and keep hot edits rebuild-free", async () => {
+  const scene = await createScene();
+  const material = await new PBRMaterial(scene).ready;
+  await detailedMesh(scene, material).ready;
+  const code = () => pbrPipelines(latestGraph(scene))[0].code;
+  assert.doesNotMatch(code(), /fn (point|spot|directional)Contribution/);
+  const first = await new PointLight(scene, { intensity: 0 }).ready;
+  assert.match(code(), /fn pointContribution/);
+  assert.doesNotMatch(code(), /fn (spot|directional)Contribution/);
+  assert.equal(code().match(/let nDotV =/g)?.length, 1);
+  assert.match(code(), /let diffuseColor = base.rgb \* \(1.0 - metallic\)/);
+  const before = testCore(scene).graphs.length;
+  first.intensity = 3;
+  first.radius = 0.4;
+  const second = await new PointLight(scene).ready;
+  await first.dispose();
+  assert.equal(testCore(scene).graphs.length, before);
+  assert.match(code(), /fn pointContribution/);
+  const spot = await new SpotLight(scene).ready;
+  const directional = await new DirectionalLight(scene).ready;
+  assert.match(code(), /fn spotContribution/);
+  assert.match(code(), /fn directionalContribution/);
+  await second.dispose();
+  assert.doesNotMatch(code(), /fn pointContribution/);
+  await spot.dispose();
+  await directional.dispose();
+  assert.doesNotMatch(code(), /fn (point|spot|directional)Contribution/);
+});
+
+test("direct canvas loadout preserves clear/depth and falls back for linear-color consumers", async () => {
+  const scene = await createScene({
+    directToCanvas: true,
+    clearColor: [0.02, 0.09, 0.3, 0.5],
+  });
+  const material = await new PBRMaterial(scene).ready;
+  const mesh = await detailedMesh(scene, material).ready;
+  const assertDirect = () => {
+    const graph = latestGraph(scene);
+    assert.equal(
+      graph.resources.textures.some((t: any) => t.id === "hdr"),
+      false,
+    );
+    assert.equal(
+      graph.passes.some((p: any) => p.id === "present"),
+      false,
+    );
+    const background = graph.passes.find(
+      (p: any) => p.id === "canvas-background",
+    );
+    assert.deepEqual(background.depth, { resource: "depth", clear: 1 });
+    assert.match(
+      renderPipeline(graph, "canvas-background").code,
+      /canvasOutput\(vec4<f32>\(0.02, 0.09, 0.3, 0.5\)\)/,
+    );
+    const forward = forwardPasses(graph)[0];
+    assert.deepEqual(forward.color, [{ resource: "canvas", load: "load" }]);
+    assert.deepEqual(forward.depth, { resource: "depth", load: "load" });
+    assert.equal(pbrPipelines(graph)[0].fragment.targets[0].format, "canvas");
+    assert.match(pbrPipelines(graph)[0].code, /unpack2x16float\(pack2x16float/);
+  };
+  const assertHdr = () =>
+    assert.ok(renderPipeline(latestGraph(scene), "present"));
+  assertDirect();
+  const effect = new FXAA(scene);
+  await effect.ready;
+  assertHdr();
+  await effect.dispose();
+  assertDirect();
+  for (const options of [{ flags: 2 }, { transmission: 0.5 }]) {
+    const special = await new PBRMaterial(scene, options).ready;
+    const extra = await detailedMesh(scene, special).ready;
+    assertHdr();
+    await extra.dispose();
+    assertDirect();
+  }
+  const compute = new ComputePass({
+    code: "@compute @workgroup_size(1) fn main() {}",
+  });
+  await scene.addComputePass(compute);
+  assertHdr();
+  assert.match(
+    pbrPipelines(latestGraph(scene))[0].code,
+    /fn directionalContribution/,
+  );
+  await scene.removeComputePass(compute.id);
+  assertDirect();
+  const custom = new ShaderMaterial(scene, { code: "custom WGSL" });
+  await custom.ready;
+  assertHdr();
+  await custom.dispose();
+  assertDirect();
+  const basic = await new Mesh(scene, {
+    vertexData: { positions: triangle.positions, indices: triangle.indices },
+  }).ready;
+  assertHdr();
+  await basic.dispose();
+  assertDirect();
+  const mask = await new PBRMaterial(scene, { flags: 1 }).ready;
+  const masked = await detailedMesh(scene, mask).ready;
+  assertDirect();
+  assert.match(
+    pbrPipelines(latestGraph(scene)).find((p: any) =>
+      p.code.includes("discard"),
+    ).code,
+    /canvasOutput/,
+  );
+  await masked.dispose();
+  await mesh.dispose();
+  assertHdr();
+  const ldr = await createScene({ directToCanvas: true, hdr: false });
+  const ldrMaterial = await new PBRMaterial(ldr).ready;
+  await detailedMesh(ldr, ldrMaterial).ready;
+  assert.ok(renderPipeline(latestGraph(ldr), "present"));
+});
@@ -9,6 +9,7 @@ crate-type = ["cdylib"]
 path = "lib.rs"

 [dependencies]
+futures-channel = "0.3"
 js-sys = "0.3"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
@@ -1,3 +1,5 @@
+import { SharedProfile } from "./profile.js";
+
 const views = Object.freeze({ f32: Float32Array, u32: Uint32Array, i32: Int32Array });

 export class SharedRows {
@@ -49,6 +51,8 @@ export class YawnCore {
  #arrays = new Map();
  #pending = new Map();
  #profileListeners = new Set();
+  #profile = new SharedProfile();
+  #profileTimer;
  #next = 1;
  /** Optional WebGPU features the device has (WebGPU spelling); populated once `ready` resolves. */
  features = new Set();
@@ -64,7 +68,7 @@ export class YawnCore {
    this.#worker.addEventListener("messageerror", () => this.#fail("WORKER_ERROR"));
    this.#worker.start?.();
    const offscreen = canvas.transferControlToOffscreen?.() ?? canvas;
-    this.ready = this.#request("init", { canvas: offscreen, arenaBytes }, [offscreen]).then(async result => {
+    this.ready = this.#request("init", { canvas: offscreen, arenaBytes, profileBuffer: this.#profile.buffer }, [offscreen]).then(async result => {
      this.#buffer = result.buffer;
      this.features = new Set(result.features);
      for (const descriptor of result.rows) this.#arrays.set(
@@ -195,7 +199,17 @@ export class YawnCore {
  onProfile(listener) {
    if (typeof listener !== "function") throw new TypeError("PROFILE_LISTENER");
    this.#profileListeners.add(listener);
-    return () => this.#profileListeners.delete(listener);
+    if (!this.#profileTimer) this.#profileTimer = setInterval(() => {
+      const stats = this.#profile.read();
+      if (stats) for (const listener of this.#profileListeners) listener(stats);
+    }, 100);
+    return () => {
+      this.#profileListeners.delete(listener);
+      if (!this.#profileListeners.size) {
+        clearInterval(this.#profileTimer);
+        this.#profileTimer = undefined;
+      }
+    };
  }

  array(name) {
@@ -213,10 +227,6 @@ export class YawnCore {
  }

  #message(message) {
-    if (message?.type === "profile") {
-      for (const listener of this.#profileListeners) listener(message.stats);
-      return;
-    }
    const pending = this.#pending.get(message?.request);
    if (!pending) return;
    this.#pending.delete(message.request);
@@ -231,6 +241,7 @@ export class YawnCore {

  dispose() {
    this.#fail("DISPOSED");
+    clearInterval(this.#profileTimer);
    this.#profileListeners.clear();
    this.#worker.terminate();
  }
@@ -302,7 +302,8 @@ impl Core {
        supported
    }

-    pub fn take_profile(&self) -> Option<String> {
-        self.render.take_profile()
+    /// Installs the worker-local sink that publishes completed GPU timings into shared memory.
+    pub fn on_profile(&self, callback: js_sys::Function) {
+        self.render.on_profile(callback);
    }
 }
@@ -4,5 +4,13 @@
  "description": "Rust/WASM render data and render graph core",
  "type": "module",
  "exports": "./index.js",
-  "files": ["index.js", "worker.js", "pkg"]
+  "scripts": {
+    "test": "node --test profile.test.js"
+  },
+  "files": [
+    "index.js",
+    "worker.js",
+    "profile.js",
+    "pkg"
+  ]
 }
@@ -0,0 +1,42 @@
+// Single-writer/single-reader latest-sample mailbox. Header: state, UTF-8 byte length.
+// State 0 = empty, 1 = owned by a reader/writer, 2 = complete sample.
+// Neither side waits: an unread sample can be replaced, and a busy publication is skipped.
+export class SharedProfile {
+  constructor(buffer = new SharedArrayBuffer(1024 * 1024)) {
+    this.buffer = buffer;
+    this.header = new Int32Array(buffer, 0, 2);
+    this.bytes = new Uint8Array(buffer, 8);
+    this.encoder = new TextEncoder();
+    this.decoder = new TextDecoder();
+  }
+
+  write(json) {
+    const bytes = this.encoder.encode(json);
+    if (bytes.length > this.bytes.length) return false;
+    const state = Atomics.load(this.header, 0);
+    if (
+      state === 1 ||
+      Atomics.compareExchange(this.header, 0, state, 1) !== state
+    )
+      return false;
+    this.bytes.set(bytes);
+    this.header[1] = bytes.length;
+    Atomics.store(this.header, 0, 2);
+    return true;
+  }
+
+  read() {
+    if (Atomics.compareExchange(this.header, 0, 2, 1) !== 2) return;
+    let bytes;
+    try {
+      bytes = this.bytes.slice(0, this.header[1]);
+    } finally {
+      Atomics.store(this.header, 0, 0);
+    }
+    return JSON.parse(this.decoder.decode(bytes));
+  }
+
+  clear() {
+    Atomics.compareExchange(this.header, 0, 2, 0);
+  }
+}
@@ -0,0 +1,122 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { Worker } from "node:worker_threads";
+import { SharedProfile } from "./profile.js";
+import { YawnCore } from "./index.js";
+
+test("shared profiles replace unread samples and exclude overlapping reads/writes", () => {
+  const writer = new SharedProfile(new SharedArrayBuffer(512));
+  const reader = new SharedProfile(writer.buffer);
+  assert.equal(reader.read(), undefined);
+  writer.write(JSON.stringify({ frame: 1, name: "GPU → SAB" }));
+  writer.write(JSON.stringify({ frame: 2, name: "light + opaque" }));
+  assert.deepEqual(reader.read(), { frame: 2, name: "light + opaque" });
+  assert.equal(reader.read(), undefined);
+  Atomics.store(writer.header, 0, 1);
+  assert.equal(writer.write('{"frame":3}'), false);
+  assert.equal(reader.read(), undefined);
+  Atomics.store(writer.header, 0, 0);
+  assert.equal(writer.write(" ".repeat(513)), false);
+  assert.equal(writer.write('{"frame":4}'), true);
+  writer.clear();
+  assert.equal(reader.read(), undefined);
+});
+
+test("concurrent shared profile publication never mixes samples", async () => {
+  const reader = new SharedProfile();
+  const done = new Int32Array(new SharedArrayBuffer(4));
+  const worker = new Worker(
+    `
+    const {workerData} = require('node:worker_threads');
+    (async () => {
+      const {SharedProfile} = await import(workerData.module);
+      const writer = new SharedProfile(workerData.buffer);
+      for (let frame = 1; frame <= 2000; frame++) {
+        const json = JSON.stringify({frame, values: Array(128).fill(frame * 3 + 7), name: 'GPU → SAB'});
+        while (!writer.write(json)) await new Promise(r => setImmediate(r));
+      }
+      Atomics.store(new Int32Array(workerData.done), 0, 1);
+    })();
+  `,
+    {
+      eval: true,
+      workerData: {
+        module: new URL("./profile.js", import.meta.url).href,
+        buffer: reader.buffer,
+        done: done.buffer,
+      },
+    },
+  );
+  const exit = new Promise((resolve, reject) => {
+    worker.once("error", reject);
+    worker.once("exit", (code) =>
+      code ? reject(new Error(`Worker exit ${code}`)) : resolve(),
+    );
+  });
+  try {
+    let last = 0;
+    while (!Atomics.load(done, 0) || Atomics.load(reader.header, 0) === 2) {
+      const sample = reader.read();
+      if (sample) {
+        assert.ok(sample.frame > last);
+        assert.deepEqual(sample.values, Array(128).fill(sample.frame * 3 + 7));
+        assert.equal(sample.name, "GPU → SAB");
+        last = sample.frame;
+      }
+      await new Promise((r) => setImmediate(r));
+    }
+    await exit;
+    assert.equal(last, 2000);
+  } finally {
+    await worker.terminate();
+  }
+});
+
+test("onProfile reads SAB without result messages and stops on unsubscribe/dispose", async (t) => {
+  t.mock.timers.enable({ apis: ["setInterval"] });
+  const messages = [];
+  let receive;
+  let writer;
+  const worker = {
+    addEventListener(type, listener) {
+      if (type === "message") receive = listener;
+    },
+    postMessage(message) {
+      messages.push(message.type);
+      writer = new SharedProfile(message.profileBuffer);
+      queueMicrotask(() =>
+        receive({
+          data: {
+            request: message.request,
+            result: {
+              buffer: new SharedArrayBuffer(64),
+              features: [],
+              rows: [],
+            },
+          },
+        }),
+      );
+    },
+    terminate() {},
+  };
+  const core = new YawnCore({}, { workerFactory: () => worker });
+  await core.ready;
+  const samples = [];
+  const stop = core.onProfile((sample) => samples.push(sample));
+  writer.write('{"frame":9,"milliseconds":1.25}');
+  t.mock.timers.tick(99);
+  assert.equal(samples.length, 0);
+  t.mock.timers.tick(1);
+  assert.deepEqual(samples, [{ frame: 9, milliseconds: 1.25 }]);
+  t.mock.timers.tick(100);
+  assert.equal(samples.length, 1);
+  assert.deepEqual(messages, ["init"]);
+  stop();
+  writer.write('{"frame":10}');
+  t.mock.timers.tick(100);
+  assert.equal(samples.length, 1);
+  core.onProfile((sample) => samples.push(sample));
+  core.dispose();
+  t.mock.timers.tick(100);
+  assert.equal(samples.length, 1);
+});
@@ -1,8 +1,8 @@
 use std::collections::{BTreeMap, HashMap, HashSet};
 use std::num::NonZeroU64;
-use std::sync::atomic::{AtomicU8, Ordering};
 use std::sync::Arc;

+use futures_channel::oneshot;
 use serde::de::DeserializeOwned;
 use serde_json::Value;

@@ -31,7 +31,7 @@ pub struct GpuProfiler {

 pub struct ProfileMap {
    readback: Arc<wgpu::Buffer>,
-    state: Arc<AtomicU8>,
+    completion: oneshot::Receiver<Result<(), wgpu::BufferAsyncError>>,
    labels: Vec<String>,
    timestamp_period: f32,
 }
@@ -417,16 +417,15 @@ impl GpuResources {

    pub fn map_profile(&self, timestamp_period: f32) -> Option<ProfileMap> {
        let profiler = self.profiler.as_ref()?;
-        let state = Arc::new(AtomicU8::new(0));
-        let callback = state.clone();
+        let (complete, completion) = oneshot::channel();
        profiler
            .readback
            .map_async(wgpu::MapMode::Read, .., move |result| {
-                callback.store(if result.is_ok() { 1 } else { 2 }, Ordering::Release);
+                let _ = complete.send(result);
            });
        Some(ProfileMap {
            readback: profiler.readback.clone(),
-            state,
+            completion,
            labels: self
                .passes
                .iter()
@@ -438,11 +437,8 @@ impl GpuResources {
 }

 impl ProfileMap {
-    pub fn state(&self) -> u8 {
-        self.state.load(Ordering::Acquire)
-    }
-
-    pub fn read(self) -> Vec<(String, f64)> {
+    pub async fn read(self) -> Option<Vec<(String, f64)>> {
+        self.completion.await.ok()?.ok()?;
        let bytes = self.readback.get_mapped_range(..);
        let values = bytes
            .chunks_exact(8)
@@ -463,7 +459,7 @@ impl ProfileMap {
            .collect();
        drop(bytes);
        self.readback.unmap();
-        profile
+        Some(profile)
    }
 }

@@ -10,26 +10,13 @@ use crate::render_data::RenderData;
 use crate::store::{Loadout, Store};

 #[wasm_bindgen(inline_js = r#"
-const channel = new MessageChannel();
-const ready = [];
-channel.port1.onmessage = () => ready.shift()?.();
-
 export function nextFrame() {
    return new Promise(resolve => requestAnimationFrame(resolve));
 }
-
-export function yieldTask() {
-    return new Promise(resolve => {
-        ready.push(resolve);
-        channel.port2.postMessage(0);
-    });
-}
 "#)]
 extern "C" {
    #[wasm_bindgen(js_name = nextFrame)]
    async fn next_frame();
-    #[wasm_bindgen(js_name = yieldTask)]
-    async fn yield_task();
 }

 pub struct RenderLoop {
@@ -42,7 +29,8 @@ pub struct RenderLoop {
    profiling: Cell<bool>,
    profile_pending: Cell<bool>,
    profile_after: Cell<f64>,
-    profile: RefCell<Option<String>>,
+    profile_callback: RefCell<Option<js_sys::Function>>,
+    profile_generation: Cell<u32>,
 }

 impl RenderLoop {
@@ -57,7 +45,8 @@ impl RenderLoop {
            profiling: Cell::new(false),
            profile_pending: Cell::new(false),
            profile_after: Cell::new(0.0),
-            profile: RefCell::new(None),
+            profile_callback: RefCell::new(None),
+            profile_generation: Cell::new(0),
        }
    }

@@ -81,13 +70,12 @@ impl RenderLoop {
    pub fn set_profiling(&self, enabled: bool) {
        self.profiling.set(enabled);
        self.profile_after.set(0.0);
-        if !enabled {
-            self.profile.borrow_mut().take();
-        }
+        self.profile_generation
+            .set(self.profile_generation.get().wrapping_add(1));
    }

-    pub fn take_profile(&self) -> Option<String> {
-        self.profile.borrow_mut().take()
+    pub fn on_profile(&self, callback: js_sys::Function) {
+        *self.profile_callback.borrow_mut() = Some(callback);
    }

    pub fn start(
@@ -152,15 +140,12 @@ impl RenderLoop {
                };
                if let Some((profile, adapter, width, height)) = submission {
                    control.profile_pending.set(true);
-                    control.profile_after.set(started + 250.0);
+                    control.profile_after.set(started + 100.0);
+                    let generation = control.profile_generation.get();
                    let control = control.clone();
                    wasm_bindgen_futures::spawn_local(async move {
-                        while profile.state() == 0 {
-                            yield_task().await;
-                        }
-                        if profile.state() == 1 {
-                            let passes = profile
-                                .read()
+                        if let Some(passes) = profile.read().await {
+                            let passes = passes
                                .into_iter()
                                .map(|(name, milliseconds)| {
                                    serde_json::json!({
@@ -173,9 +158,10 @@ impl RenderLoop {
                                .iter()
                                .filter_map(|pass| pass["milliseconds"].as_f64())
                                .sum::<f64>();
-                            if control.profiling.get() {
-                                *control.profile.borrow_mut() = Some(
-                                    serde_json::json!({
+                            if control.profiling.get()
+                                && control.profile_generation.get() == generation
+                            {
+                                let sample = serde_json::json!({
                                    "frame": frame,
                                    "milliseconds": milliseconds,
                                    "readbackMilliseconds": js_sys::Date::now() - started,
@@ -183,8 +169,11 @@ impl RenderLoop {
                                    "canvas": { "width": width, "height": height },
                                    "passes": passes,
                                })
-                                    .to_string(),
-                                );
+                                .to_string();
+                                if let Some(callback) = control.profile_callback.borrow().as_ref() {
+                                    let _ = callback
+                                        .call1(&wasm_bindgen::JsValue::NULL, &sample.into());
+                                }
                            }
                        }
                        control.profile_pending.set(false);
@@ -1,7 +1,8 @@
 import initWasm, { Core } from "./pkg/yawn_core.js";
+import { SharedProfile } from "./profile.js";

 let core;
-let profileTimer;
+let profile;
 // While a relocation is staged, every request except its commit waits here.
 let awaitingCommit = false;
 const queue = [];
@@ -15,6 +16,8 @@ async function handle(message) {
      if (core || !(message.canvas instanceof OffscreenCanvas)) fail("INIT");
      const wasm = await initWasm();
      core = new Core(message.arenaBytes);
+      profile = new SharedProfile(message.profileBuffer);
+      core.on_profile(json => profile.write(json));
      await core.initialize(message.canvas);
      const buffer = wasm.memory.buffer;
      if (!(buffer instanceof SharedArrayBuffer)) fail("WASM_MEMORY_NOT_SHARED");
@@ -70,11 +73,7 @@ async function handle(message) {
          break;
        case "set-profiler":
          result = core.set_profiler(Boolean(message.enabled));
-          clearInterval(profileTimer);
-          profileTimer = result && message.enabled ? setInterval(() => {
-            const stats = core.take_profile();
-            if (stats) postMessage({ type: "profile", stats: JSON.parse(stats) });
-          }, 250) : undefined;
+          profile.clear();
          break;
        default:
          fail("MESSAGE");
@@ -4,7 +4,7 @@ All fields below are top-level beside `type` and `request`. Except for `init` an

 | `type` | Fields | Transfer list | Success `result` | Constraints / effects |
 |---|---|---|---|---|
-| `init` | `canvas: OffscreenCanvas`, `arenaBytes: number` | `[canvas]` | `{ buffer: SharedArrayBuffer, rows: Descriptor[], features: string[] }` | Once only; initializes WASM/WebGPU. `features` lists optional device features that were available (`timestamp-query`, `indirect-first-instance`). |
+| `init` | `canvas: OffscreenCanvas`, `arenaBytes: number`, `profileBuffer: SharedArrayBuffer` | `[canvas]` | `{ buffer: SharedArrayBuffer, rows: Descriptor[], features: string[] }` | Once only; initializes WASM/WebGPU. `features` lists optional device features that were available (`timestamp-query`, `indirect-first-instance`). `profileBuffer` is shared, not transferred. |
 | `create-rows` | `name`, `rows`, `stride`, `format` | — | `{ rows: descriptor, commit: boolean }` | Nonempty name; rows > 0; stride ≥16 and multiple of 16; format f32/u32/i32. Existing name may only retain format/stride and grow. `commit: true` means the rows moved: copy the old bytes into the descriptor and send `commit-rows`. |
 | `create-rows-batch` | `rows: Array<{name,rows,stride,format}>` | — | `{ rows: descriptor[] (input order), commit: string[] }` | Nonempty. Sequential; in-place creations/growths survive a later failure, staged relocations are discarded. `commit` lists the arrays that moved. Active GPU resources refresh once for in-place growth. |
 | `commit-rows` | `names: string[]` | — | `undefined` | Publishes staged relocations after the caller copied them. Must name every staged array. Frees the old rows, refreshes bound GPU buffers once, sets `sabDirty`. While a relocation is staged the worker queues every other request until this arrives. |
@@ -23,14 +23,14 @@ All fields below are top-level beside `type` and `request`. Except for `init` an

 Rust/WASM numeric conversion applies to `u32` fields; callers should send finite nonnegative integers in range rather than rely on coercion.

-## Profiler events
+## Shared profiler results

-When enabled and supported, the worker polls every 250 ms and may send an unsolicited message with **no request field**:
+When enabled and supported, GPU timestamp queries are sampled at most once every 100 ms, with only one readback in flight. Completion of `map_async` wakes an awaiting future; neither the worker nor render loop polls for GPU completion. Rendering continues independently.
+
+Completed results are published into `profileBuffer`, not sent as worker messages. `YawnCore.onProfile` reads this buffer locally every 100 ms while listeners exist and delivers the same stats shape:

 ```js
 {
-  type: "profile",
-  stats: {
    frame: 42,
    milliseconds: 0.31,
    readbackMilliseconds: 4.8,
@@ -38,10 +38,11 @@ When enabled and supported, the worker polls every 250 ms and may send an unsoli
    canvas: { width: 1280, height: 720 },
    passes: [{ name: "triangle", milliseconds: 0.31 }]
 }
-}
 ```

-One pass timing corresponds to a compiled execution; compatible adjacent render passes are merged into one execution labelled `first … last (N draws)`. Samples are throttled and asynchronous, so they are diagnostics, not one event per frame. Disabling clears pending published statistics and the worker polling interval.
+One pass timing corresponds to a compiled execution; compatible adjacent render passes are merged into one execution labelled `first … last (N draws)`. `milliseconds` is GPU time; `readbackMilliseconds` includes the asynchronous wait. Disabling clears published statistics and suppresses results from older profiling sessions.
+
+The default mailbox is a separate 1 MiB SAB. Its first two `Int32` values are ownership state and UTF-8 JSON length; the payload starts at byte 8. State 0 means empty, 1 means owned by a reader or writer, and 2 means a complete sample. Both sides claim ownership using `Atomics.compareExchange` and never wait. An unread sample is replaced by the latest sample; a publication during a read, or larger than the payload capacity, is skipped. The reader copies the bytes before releasing ownership, so it cannot observe a partially written result. Initialization and profiler enable/disable remain ordinary control requests; results never use `postMessage`.

 ## Texture lifecycle

@@ -12,7 +12,7 @@
    "server": "go run ./server",
    "coredocs": "vitepress dev coredocs --host 0.0.0.0 --port ${PORT:-8080}",
    "build:coredocs": "vitepress build coredocs",
-    "test": "npm run test:handles && npm run test:fxnode-rg",
+    "test": "npm test --workspace @yawn/core && npm run test:handles && npm run test:fxnode-rg",
    "test:handles": "npm test --workspace @yawn/handles",
    "test:fxnode-rg": "npm test --workspace @yawn/fxnode-rg",
    "test:handles:screenshots": "npm run test:screenshots --workspace @yawn/handles"
@@ -34,12 +34,20 @@ const fxNodeImports = {
  },
 };

+const coreWorker = await build({
+  ...browser,
+  entryPoints: [path.join(root, "core", "worker.js")],
+  external: ["./pkg/yawn_core.js"],
+  write: false,
+});
+const coreWorkerSource = coreWorker.outputFiles[0].text;
+
 const workerFiles = new Map([
  [
    path.join(root, "core", "index.js"),
    {
      url: "./worker.js",
-      source: await fs.readFile(path.join(root, "core", "worker.js"), "utf8"),
+      source: coreWorkerSource,
    },
  ],
  [
@@ -113,15 +121,7 @@ await Promise.all([
    entryPoints: [path.join(fxNodeSource, "worker", "fxnode.worker.ts")],
    outfile: path.join(output, "fxnode-worker.js"),
  }),
-  build({
-    bundle: false,
-    entryPoints: [path.join(root, "core", "worker.js")],
-    format: "esm",
-    legalComments: "none",
-    minify: true,
-    outfile: path.join(output, "worker.js"),
-    target: "es2022",
-  }),
+  fs.writeFile(path.join(output, "worker.js"), coreWorkerSource),
  build({
    ...browser,
    entryPoints: [
@@ -224,6 +224,7 @@ async function openFile(file) {
    scene = new Scene(canvas, {
      arenaBytes: 384 * 1024 * 1024,
      hdr: true,
+      directToCanvas: true,
      clearColor: [0.025, 0.035, 0.045, 1],
    });
    await scene.ready;
@@ -266,9 +267,12 @@ async function openFile(file) {
    );
    await Promise.all([...lights.map((light) => light.ready), camera.ready]);

+    // Diagnostic A/B: omit queries and readback, not just the timing badge.
+    let stopProfile;
+    if (new URLSearchParams(location.search).get("profile") !== "off") {
      elements.renderValue.textContent = "— ms";
      elements.renderTime.hidden = false;
-    const stopProfile = scene.core.onProfile((profile) => {
+      stopProfile = scene.core.onProfile((profile) => {
        elements.renderValue.textContent = `${profile.milliseconds.toFixed(2)} ms`;
      });
      const profiling = await scene.core.setProfiler(true);
@@ -276,6 +280,7 @@ async function openFile(file) {
        elements.renderTime.title =
          "GPU timestamp queries are not supported by this adapter.";
      else elements.renderTime.removeAttribute("title");
+    }

    activeViewer = {
      camera,