@driftengine/core 4.10.0 → 4.11.0

package/src/render/mesh.ts

@@ -3,6 +3,13 @@
 import { MorphTexture } from './backend/webgl2/morphTexture.ts';
 import { ABSENT_ATTRIBUTE } from './vertexDefaults.ts';
 import { UPLOAD_BYTES_PER_STEP } from './uploadStep.ts';
+import { compactIndices } from './indexWidth.ts';
+import {
+  fixedArray,
+  sharedValue,
+  vertexPackingOf,
+  type AttributePacking,
+} from './vertexPacking.ts';
 
 /*
  * `MeshData` and its validator live in `@driftengine/drft`, because the container's whole
@@ -92,7 +99,14 @@ export class Mesh {
   private readonly vao: WebGLVertexArrayObject;
   private readonly buffers: WebGLBuffer[] = [];
   /** Every attribute bound from a buffer, so a batch can build a vertex array of its own. */
-  private readonly layout: { location: number; buffer: WebGLBuffer; size: number }[] = [];
+  private readonly layout: {
+    location: number;
+    buffer: WebGLBuffer;
+    size: number;
+    /** `FLOAT`, or a sixteen-bit type read normalised: `vertexPacking.ts`. */
+    type: number;
+    normalized: boolean;
+  }[] = [];
   private indexBuffer: WebGLBuffer | null = null;
   /**
    * The position and normal buffers, kept so `update` can rewrite them.
@@ -110,6 +124,8 @@ export class Mesh {
    * same frame.
    */
   readonly indexCount: number;
+  /** `UNSIGNED_SHORT` or `UNSIGNED_INT`, as `compactIndices` stored them: every draw says which. */
+  private readonly indexType: number;
   /**
    * The constant values this mesh's *absent* attributes are shaded with, re-applied on
    * every draw.
@@ -226,11 +242,13 @@ export class Mesh {
      * that has been told a lie every frame is the shape of a stall nobody can find.
      */
     const usage = dynamic ? gl.DYNAMIC_DRAW : gl.STATIC_DRAW;
+    /* How each attribute travels: sixteen-bit fixed point, or one value as a constant. */
+    const packing = vertexPackingOf(data, dynamic);
     const positions = this.attachAttribute(gl, ATTR_POSITION, data.positions, 3, usage);
-    const normals = this.attachAttribute(gl, ATTR_NORMAL, data.normals, 3, usage);
+    const normals = this.attachAttribute(gl, ATTR_NORMAL, data.normals, 3, usage, packing.normals);
     if (dynamic) this.deformable = { positions, normals };
-    this.attachAttribute(gl, ATTR_COLOR, data.colors, 3);
-    this.attachAttribute(gl, ATTR_EMISSIVE, data.emissive, 1);
+    this.attachShared(gl, ATTR_COLOR, data.colors, 3, packing.colors);
+    this.attachShared(gl, ATTR_EMISSIVE, data.emissive, 1, packing.emissive);
     /*
      * Absent means "not shiny", supplied as a constant rather than as a buffer of
      * zeroes. `vertexAttrib1f` sets the value a *disabled* attribute reads, so this
@@ -304,7 +322,7 @@ export class Mesh {
       this.constants.push({ location: ATTR_WEIGHTS, value: ABSENT_ATTRIBUTE['weights'] });
     } else {
       this.attachAttribute(gl, ATTR_JOINTS, data.joints, 4);
-      this.attachAttribute(gl, ATTR_WEIGHTS, data.weights, 4);
+      this.attachAttribute(gl, ATTR_WEIGHTS, data.weights, 4, gl.STATIC_DRAW, packing.weights);
     }
     /* The fifth to eighth influences, or zeros, which add nothing. See `isSkinnedEight`. */
     this.isSkinnedEight = this.isSkinned && data.joints2 !== undefined;
@@ -313,7 +331,7 @@ export class Mesh {
       this.constants.push({ location: ATTR_WEIGHTS2, value: ABSENT_ATTRIBUTE['weights2'] });
     } else {
       this.attachAttribute(gl, ATTR_JOINTS2, data.joints2, 4);
-      this.attachAttribute(gl, ATTR_WEIGHTS2, data.weights2, 4);
+      this.attachAttribute(gl, ATTR_WEIGHTS2, data.weights2, 4, gl.STATIC_DRAW, packing.weights2);
     }
 
     this.hasTangents = data.tangents !== undefined;
@@ -321,7 +339,7 @@ export class Mesh {
       gl.disableVertexAttribArray(ATTR_TANGENT);
       this.constants.push({ location: ATTR_TANGENT, value: ABSENT_ATTRIBUTE['tangents'] });
     } else {
-      this.attachAttribute(gl, ATTR_TANGENT, data.tangents, 4);
+      this.attachAttribute(gl, ATTR_TANGENT, data.tangents, 4, gl.STATIC_DRAW, packing.tangents);
     }
 
     this.hasGrain = data.grain !== undefined;
@@ -351,7 +369,9 @@ export class Mesh {
     this.buffers.push(indexBuffer);
     gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, indexBuffer);
     this.indexBuffer = indexBuffer;
-    this.allocate(gl, gl.ELEMENT_ARRAY_BUFFER, indexBuffer, data.indices, gl.STATIC_DRAW);
+    const indices = compactIndices(data.indices, this.vertexCount);
+    this.indexType = indices instanceof Uint16Array ? gl.UNSIGNED_SHORT : gl.UNSIGNED_INT;
+    this.allocate(gl, gl.ELEMENT_ARRAY_BUFFER, indexBuffer, indices, gl.STATIC_DRAW);
 
     gl.bindVertexArray(null);
   }
@@ -370,7 +390,7 @@ export class Mesh {
       const value = constant.value;
       applyConstant(gl, constant.location, value);
     }
-    gl.drawElements(gl.TRIANGLES, this.indexCount, gl.UNSIGNED_INT, 0);
+    gl.drawElements(gl.TRIANGLES, this.indexCount, this.indexType, 0);
   }
 
   /**
@@ -394,7 +414,14 @@ export class Mesh {
     for (const attribute of this.layout) {
       gl.bindBuffer(gl.ARRAY_BUFFER, attribute.buffer);
       gl.enableVertexAttribArray(attribute.location);
-      gl.vertexAttribPointer(attribute.location, attribute.size, gl.FLOAT, false, 0, 0);
+      gl.vertexAttribPointer(
+        attribute.location,
+        attribute.size,
+        attribute.type,
+        attribute.normalized,
+        0,
+        0,
+      );
     }
     gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, this.indexBuffer);
     gl.bindBuffer(gl.ARRAY_BUFFER, instances);
@@ -424,7 +451,7 @@ export class Mesh {
       const value = constant.value;
       applyConstant(gl, constant.location, value);
     }
-    gl.drawElementsInstanced(gl.TRIANGLES, this.indexCount, gl.UNSIGNED_INT, 0, count);
+    gl.drawElementsInstanced(gl.TRIANGLES, this.indexCount, this.indexType, 0, count);
   }
 
   dispose(gl: WebGL2RenderingContext): void {
@@ -468,24 +495,49 @@ export class Mesh {
     }
   }
 
+  /**
+   * One attribute's buffer, as floats or in its sixteen-bit fixed point (`vertexPacking.ts`), which
+   * WebGL2 reads back through a normalised `SHORT` or `UNSIGNED_SHORT`: three lanes wide where
+   * WebGPU needs four, so a normal costs six bytes here.
+   */
   private attachAttribute(
     gl: WebGL2RenderingContext,
     location: number,
     data: Float32Array,
     size: number,
     usage: number = gl.STATIC_DRAW,
+    packing: AttributePacking = 'float',
   ): WebGLBuffer {
     const buffer = gl.createBuffer();
     if (buffer === null) throw new Error('createBuffer failed');
     this.buffers.push(buffer);
     gl.bindBuffer(gl.ARRAY_BUFFER, buffer);
-    this.allocate(gl, gl.ARRAY_BUFFER, buffer, data, usage);
+    const fixed = packing === 'snorm16' || packing === 'unorm16';
+    this.allocate(gl, gl.ARRAY_BUFFER, buffer, fixed ? fixedArray(data, packing) : data, usage);
+    const type =
+      packing === 'snorm16' ? gl.SHORT : packing === 'unorm16' ? gl.UNSIGNED_SHORT : gl.FLOAT;
     gl.enableVertexAttribArray(location);
-    gl.vertexAttribPointer(location, size, gl.FLOAT, false, 0, 0);
-    this.layout.push({ location, buffer, size });
+    gl.vertexAttribPointer(location, size, type, fixed, 0, 0);
+    this.layout.push({ location, buffer, size, type, normalized: fixed });
     return buffer;
   }
 
+  /** A buffer, or the one value every vertex shares as a constant, as the packing says. */
+  private attachShared(
+    gl: WebGL2RenderingContext,
+    location: number,
+    data: Float32Array,
+    size: number,
+    packing: AttributePacking,
+  ): void {
+    if (packing !== 'constant') {
+      this.attachAttribute(gl, location, data, size, gl.STATIC_DRAW, packing);
+      return;
+    }
+    gl.disableVertexAttribArray(location);
+    this.constants.push({ location, value: sharedValue(data, size) });
+  }
+
   /**
    * Give a bound buffer its storage, and its contents now or later.
    *

package/src/render/renderQuality.ts

@@ -50,6 +50,9 @@ export type ShadowFilterTaps = 4 | 8 | 12;
 export type DirectionalShadowDepthLayers = 1 | 2;
 export type WaterReflectionFilterTaps = 1 | 5 | 9;
 
+/** What a draw does when its pipeline has not compiled yet: see `RenderQuality.pipelineCompile`. */
+export type PipelineCompile = 'wait' | 'skip';
+
 /** How light through glass is coloured: see `RenderQuality.glassShadows`. */
 export type GlassShadows = 'full' | 'half' | 'off';
 
@@ -220,6 +223,34 @@ export interface RenderQuality {
    * lot of glass; that is what `'half'` is for. See `glassShadow.ts`.
    */
   readonly glassShadows: GlassShadows;
+  /**
+   * What a lit draw does when the pipeline it needs has not compiled: one first needed inside a
+   * frame, by a material's own state — two-sided, cut out, a surface model — or a draw state a
+   * mesh's creation did not warm.
+   *
+   * **`'wait'`, the default**: the pipeline is built where the draw asks for it, and on WebGPU the
+   * browser compiles it on its GPU process's main thread, holding the frame — 130 to 400 ms a
+   * pipeline was reported on a phone, for a material first drawn mid-scene. **`'skip'`** starts the
+   * compile off that thread and leaves the draw out until it lands, a few frames to a few dozen,
+   * and the draw then appears; nothing waits. Static draws always wait, since a recorded list keeps
+   * what it recorded, and a compile that fails is built where it is next asked for, so the device
+   * says why rather than the draw vanishing for good.
+   *
+   * **The way to have neither** is to draw what a scene will use once, behind a loading screen, and
+   * await `ready()`, which waits for every compile started — whichever this says. What would make
+   * `'skip'` wrong is a frame that must hold everything it was given, such as a capture of the first
+   * frame of a scene not awaited.
+   *
+   * **Measured on a desktop GPU** (`demo/dev/pipelineCompile.html`, eight new variants at once): on
+   * WebGPU the frame that met them took 700 to 3,700 ms waiting and 17 ms skipping, the draws
+   * landing 21 to 26 frames later. WebGL2 compiles its lit programs when it is made, all but a
+   * surface model's, which it compiles at the first draw that wants one: there `'skip'` asks the
+   * driver through `KHR_parallel_shader_compile` and waits where a context does not offer it — which
+   * ANGLE on Vulkan does not. Where it is offered, ANGLE on OpenGL, the longest frame went from 820 to
+   * 1,120 ms to 630 to 700 ms: the driver still finishes its own compile at the program's first
+   * draw, which nothing above it can move.
+   */
+  readonly pipelineCompile: PipelineCompile;
   /**
    * How light that scatters beneath skin is drawn, for a material whose model is `skinModel`.
    *
@@ -973,6 +1004,7 @@ export const DEFAULT_RENDER_QUALITY: Readonly<RenderQuality> = Object.freeze({
   lightVolumeSamples: 32,
   directionalShadowDepthLayers: 2,
   glassShadows: 'full',
+  pipelineCompile: 'wait',
   skinScattering: 'pre-integrated',
   directionalShadowMaxDistance: 6,
   directionalShadowMaxSlope: 3,
@@ -1140,6 +1172,7 @@ export function resolveRenderQuality(options: RenderQualityOptions = {}): Readon
     directionalShadowDepthLayers:
       options.directionalShadowDepthLayers ?? DEFAULT_RENDER_QUALITY.directionalShadowDepthLayers,
     glassShadows: options.glassShadows ?? DEFAULT_RENDER_QUALITY.glassShadows,
+    pipelineCompile: options.pipelineCompile ?? DEFAULT_RENDER_QUALITY.pipelineCompile,
     skinScattering: options.skinScattering ?? DEFAULT_RENDER_QUALITY.skinScattering,
     directionalShadowMaxDistance:
       options.directionalShadowMaxDistance ?? DEFAULT_RENDER_QUALITY.directionalShadowMaxDistance,
@@ -1298,6 +1331,11 @@ export function resolveRenderQuality(options: RenderQualityOptions = {}): Readon
       `RenderQuality.glassShadows must be 'full', 'half' or 'off', got ${String(quality.glassShadows)}`,
     );
   }
+  if (quality.pipelineCompile !== 'wait' && quality.pipelineCompile !== 'skip') {
+    throw new Error(
+      `RenderQuality.pipelineCompile must be 'wait' or 'skip', got ${String(quality.pipelineCompile)}`,
+    );
+  }
   if (quality.skinScattering !== 'pre-integrated' && quality.skinScattering !== 'screen-space') {
     throw new Error(
       `RenderQuality.skinScattering must be 'pre-integrated' or 'screen-space', got ${String(quality.skinScattering)}`,

package/src/render/rendererHarness.ts

@@ -58,8 +58,14 @@ export function recordingGl(
      * recording.
      */
     readonly uniforms?: readonly string[];
+    /**
+     * Whether a program started under `KHR_parallel_shader_compile` has finished, which the test
+     * flips: `COMPLETION_STATUS_KHR` answers `done`. Finished by default, as every program here is.
+     */
+    readonly compiling?: { done: boolean };
   } = {},
 ) {
+  const compiling = options.compiling ?? { done: true };
   const extensions = new Set(options.extensions ?? []);
   const uniforms = options.uniforms ?? [];
   /* What `clearDepth` last set, so a module that borrows the value can be seen to put it back. */
@@ -136,6 +142,13 @@ export function recordingGl(
     HALF_FLOAT: 0x140b,
     DEPTH_COMPONENT: 0x1902,
     UNSIGNED_INT: 0x1405,
+    /* Named so a mesh's index width and each attribute's type can be read back: the buffer its
+       indices fill, the type every draw of it names, and how each attribute is read. See
+       `mesh.test.ts`. */
+    UNSIGNED_SHORT: 0x1403,
+    SHORT: 0x1402,
+    FLOAT: 0x1406,
+    ELEMENT_ARRAY_BUFFER: 0x8893,
     LINEAR: 0x2601,
     COLOR_BUFFER_BIT: 0x00004000,
     /* Named so a test can tell a mesh that declared itself deforming from one that did not. */
@@ -173,6 +186,7 @@ export function recordingGl(
       if (name === 'EXT_clip_control') {
         return { LOWER_LEFT_EXT: 0x8ca1, ZERO_TO_ONE_EXT: 0x935f, clipControlEXT: () => undefined };
       }
+      if (name === 'KHR_parallel_shader_compile') return { COMPLETION_STATUS_KHR: 0x91b1 };
       return {};
     },
     getShaderParameter: () => true,
@@ -200,6 +214,7 @@ export function recordingGl(
     /* Linked yes; zero active uniforms and attributes, which ends every reflection loop at once
        and is why this harness does not have to know a single uniform name. */
     getProgramParameter: (program: unknown, pname: unknown) => {
+      if (pname === 0x91b1) return compiling.done;
       if (pname === constants.ACTIVE_UNIFORMS) return uniforms.length;
       if (pname === constants.ACTIVE_ATTRIBUTES) return 0;
       if (pname !== constants.LINK_STATUS) return true;

package/src/render/shader.ts

@@ -26,6 +26,65 @@ export function compileProgram(
   return program;
 }
 
+/**
+ * A program compiled and linked without asking whether that worked, since asking waits for the
+ * driver to finish: for a context offering `KHR_parallel_shader_compile`, whose
+ * `COMPLETION_STATUS_KHR` says when asking no longer waits. `finishProgram` asks then, and fails
+ * as loudly as `compileProgram` does; the stages are kept until it does, for their logs.
+ */
+export interface StartedProgram {
+  readonly program: WebGLProgram;
+  readonly vertex: WebGLShader;
+  readonly fragment: WebGLShader;
+  readonly label: string;
+}
+
+export function startProgram(
+  gl: WebGL2RenderingContext,
+  vertexSource: string,
+  fragmentSource: string,
+  label: string,
+): StartedProgram {
+  const stage = (type: number, source: string, name: string): WebGLShader => {
+    const shader = gl.createShader(type);
+    if (shader === null) throw new Error(`[${label}.${name}] createShader failed`);
+    gl.shaderSource(shader, source);
+    gl.compileShader(shader);
+    return shader;
+  };
+  const vertex = stage(gl.VERTEX_SHADER, vertexSource, 'vert');
+  const fragment = stage(gl.FRAGMENT_SHADER, fragmentSource, 'frag');
+  const program = gl.createProgram();
+  if (program === null) throw new Error(`[${label}] createProgram failed`);
+  gl.attachShader(program, vertex);
+  gl.attachShader(program, fragment);
+  gl.linkProgram(program);
+  return { program, vertex, fragment, label };
+}
+
+/** A started program, checked as `compileProgram` checks one: its stages, then its link. */
+export function finishProgram(gl: WebGL2RenderingContext, started: StartedProgram): WebGLProgram {
+  const { program, vertex, fragment, label } = started;
+  const fail = (what: string, log: string | null): never => {
+    gl.deleteShader(vertex);
+    gl.deleteShader(fragment);
+    gl.deleteProgram(program);
+    throw new Error(`[${label}${what}] ${log ?? 'no log'}`);
+  };
+  if (!gl.getShaderParameter(vertex, gl.COMPILE_STATUS)) {
+    fail('.vert compile failed:', gl.getShaderInfoLog(vertex));
+  }
+  if (!gl.getShaderParameter(fragment, gl.COMPILE_STATUS)) {
+    fail('.frag compile failed:', gl.getShaderInfoLog(fragment));
+  }
+  if (!gl.getProgramParameter(program, gl.LINK_STATUS)) {
+    fail(' link failed:', gl.getProgramInfoLog(program));
+  }
+  gl.deleteShader(vertex);
+  gl.deleteShader(fragment);
+  return program;
+}
+
 /**
  * Whether a uniform lookup that cannot reach the GPU is reported.
  *

package/src/render/shaders/flat/directionalShadow.ts

@@ -36,13 +36,56 @@ export const DIRECTIONALSHADOW_GLSL = `vec2 receiverPlaneDepthGradient(vec3 p) {
   );
 }
 
-float directionalVisibility(float receiverDepth, float compareDepth, float storedDepth) {
+/* One stored depth against a receiver, in a map spanning \`span\` metres of depth. */
+float directionalVisibilityAt(float receiverDepth, float compareDepth, float storedDepth, float span) {
   if (compareDepth <= storedDepth) return 1.0;
-  float rayDistance = max(receiverDepth - storedDepth, 0.0) * uShadowDepthSpan;
+  float rayDistance = max(receiverDepth - storedDepth, 0.0) * span;
   float groundDistance = rayDistance * length(uDirectionalDir.xz);
   return 1.0 - shadowReach(groundDistance, uShadowMaxDistance);
 }
 
+float directionalVisibility(float receiverDepth, float compareDepth, float storedDepth) {
+  return directionalVisibilityAt(receiverDepth, compareDepth, storedDepth, uShadowDepthSpan);
+}
+
+/* How far inside its square a receiver stands in a layer's map: 1 well inside, 0 at its border. */
+float sunMapFade(vec3 p) {
+  if (p.z > 1.0 || p.x < 0.0 || p.x > 1.0 || p.y < 0.0 || p.y > 1.0) return 0.0;
+  vec2 fromCentre = abs(p.xy - 0.5) * 2.0;
+  return (1.0 - smoothstep(0.72, 0.98, max(fromCentre.x, fromCentre.y))) *
+    (1.0 - smoothstep(0.90, 1.0, p.z));
+}
+
+/*
+ * The moving layer at its own place, where it has a matrix of its own (movingSun.ts): the coarse
+ * test of shadowFactor, and its filter, each against the moving map's own depth span, and lit by
+ * how far inside the moving map's square the receiver stands rather than cut at its border.
+ */
+float movingSunCoarse(vec3 pm, float span) {
+  float inside = sunMapFade(pm);
+  if (inside <= 0.0) return 1.0;
+  float stored = textureLod(uSunShadows, vec3(pm.xy, ${SUN_DYNAMIC_LAYER}.0), 0.0).r;
+  return mix(1.0, directionalVisibilityAt(pm.z, pm.z - 1.0 / span, stored, span), inside);
+}
+
+float movingSunFiltered(vec3 pm, vec2 gradient, float span) {
+  float inside = sunMapFade(pm);
+  if (inside <= 0.0) return 1.0;
+  float texel = 1.35 / uShadowMapSize;
+  float slopeLimit = 9.1 / uShadowMapSize;
+  float lit = 0.0;
+  for (int i = 0; i < ${MAX_SHADOW_FILTER_TAPS}; i++) {
+    if (i >= uShadowFilterTaps) break;
+    vec2 sampleUv = pm.xy + DIRECTIONAL_PCF_OFFSETS[i] * texel;
+    vec2 texelCentre = (floor(sampleUv * uShadowMapSize) + 0.5) / uShadowMapSize;
+    float slopeOffset = clamp(dot(gradient, texelCentre - pm.xy), -slopeLimit, slopeLimit);
+    float tapReceiverDepth = pm.z + slopeOffset;
+    float stored = textureLod(uSunShadows, vec3(sampleUv, ${SUN_DYNAMIC_LAYER}.0), 0.0).r;
+    lit += directionalVisibilityAt(tapReceiverDepth, tapReceiverDepth - 0.14 / span, stored, span);
+  }
+  return mix(1.0, lit / float(max(uShadowFilterTaps, 1)), inside);
+}
+
 float shadowFactor(float directionalNdl) {
   if (uShadowStrength <= 0.0) return 1.0;
 
@@ -72,6 +115,14 @@ float shadowFactor(float directionalNdl) {
   vec3 p = vLightPos.xyz / vLightPos.w;
   p = p * 0.5 + 0.5;
   vec2 depthGradient = receiverPlaneDepthGradient(p);
+  /* The moving layer's own place and plane where it has a matrix of its own (movingSun.ts), taken
+     here beside the static layer's for the same reason: above every varying-dependent return. */
+  vec3 pm = p;
+  vec2 movingGradient = depthGradient;
+  if (MOVING_SUN) {
+    pm = movingSunAt();
+    movingGradient = receiverPlaneDepthGradient(pm);
+  }
 
   // Near a surface's own light terminator its direct-light term is already
   // approaching zero, while the orthographic projection of that surface
@@ -136,7 +187,8 @@ float shadowFactor(float directionalNdl) {
   float coarseCompare = p.z - 1.0 / uShadowDepthSpan;
   float coarse =
     directionalVisibility(p.z, coarseCompare, textureLod(uSunShadows, vec3(p.xy, ${SUN_STATIC_LAYER}.0), 0.0).r) *
-    directionalVisibility(p.z, coarseCompare, textureLod(uSunShadows, vec3(p.xy, ${SUN_DYNAMIC_LAYER}.0), 0.0).r);
+    (MOVING_SUN ? movingSunCoarse(pm, movingSunSpan()) :
+      directionalVisibility(p.z, coarseCompare, textureLod(uSunShadows, vec3(p.xy, ${SUN_DYNAMIC_LAYER}.0), 0.0).r));
   if (uPeeledShadowEnabled != 0) {
     coarse *= directionalVisibility(p.z, coarseCompare, textureLod(uSunShadows, vec3(p.xy, ${SUN_PEELED_LAYER}.0), 0.0).r);
   }
@@ -200,7 +252,6 @@ float shadowFactor(float directionalNdl) {
      * there is no mip to select and level zero is the only thing \`texture\` could have read.
      */
     float staticDepth = textureLod(uSunShadows, vec3(sampleUv, ${SUN_STATIC_LAYER}.0), 0.0).r;
-    float dynamicDepth = textureLod(uSunShadows, vec3(sampleUv, ${SUN_DYNAMIC_LAYER}.0), 0.0).r;
     staticLit += directionalVisibility(tapReceiverDepth, compare, staticDepth);
     if (uPeeledShadowEnabled != 0) {
       float peeledDepth = textureLod(uSunShadows, vec3(sampleUv, ${SUN_PEELED_LAYER}.0), 0.0).r;
@@ -208,12 +259,18 @@ float shadowFactor(float directionalNdl) {
     } else {
       peeledLit += 1.0;
     }
-    dynamicLit += directionalVisibility(tapReceiverDepth, compare, dynamicDepth);
+    /* The moving layer here where it shares the static one's place; at its own below otherwise.
+       Branched on the switch itself, never its negation: see overridableConstants in transform.mjs. */
+    if (MOVING_SUN) {
+    } else {
+      float dynamicDepth = textureLod(uSunShadows, vec3(sampleUv, ${SUN_DYNAMIC_LAYER}.0), 0.0).r;
+      dynamicLit += directionalVisibility(tapReceiverDepth, compare, dynamicDepth);
+    }
   }
   float tapCount = float(max(uShadowFilterTaps, 1));
   staticLit /= tapCount;
   peeledLit /= tapCount;
-  dynamicLit /= tapCount;
+  dynamicLit = MOVING_SUN ? movingSunFiltered(pm, movingGradient, movingSunSpan()) : dynamicLit / tapCount;
 
   // Independent layers preserve every recorded fact at an overlap: each
   // caster keeps its own distance fade, and their transmissions compose
@@ -234,11 +291,11 @@ float shadowFactor(float directionalNdl) {
  * behind the receiver lending it colour within the spread, which only a receiver between two panes
  * meets. One function for both layers, because two copies were most of the lit shaders' growth.
  */
-vec3 sunGlassLayer(vec3 p, float paneLayer, float tintLayer, float bias, float texel, float uvPerMetre) {
+vec3 sunGlassLayer(vec3 p, float paneLayer, float tintLayer, float bias, float texel, float uvPerMetre, float span) {
   float tintSize = float(textureSize(uSunGlassTints, 0).x);
   float clarity = textureLod(uSunGlassTints, vec3(p.xy, tintLayer), 0.0).a;
   float behind =
-    max(p.z - textureLod(uSunShadows, vec3(p.xy, paneLayer), 0.0).r, 0.0) * uShadowDepthSpan;
+    max(p.z - textureLod(uSunShadows, vec3(p.xy, paneLayer), 0.0).r, 0.0) * span;
   float radius = clamp(
     (1.0 - clarity) * ${FROST_SPREAD.toFixed(6)} * behind * uvPerMetre,
     texel,
@@ -278,6 +335,13 @@ vec3 sunGlassLookup(float directionalNdl) {
   /* Map units per metre along the receiver, from derivatives taken first in uniform control flow,
      so a frost spread stated in metres lands the same whatever size of light matrix a consumer built. */
   float uvPerMetre = length(dFdx(p.xy)) / max(length(dFdx(vWorldPos)), 1e-6);
+  /* The moving glass at its own place where its layer has a matrix of its own: movingSun.ts. */
+  vec3 pm = p;
+  float movingPerMetre = uvPerMetre;
+  if (MOVING_SUN) {
+    pm = movingSunAt();
+    movingPerMetre = length(dFdx(pm.xy)) / max(length(dFdx(vWorldPos)), 1e-6);
+  }
   if (uShadowStrength <= 0.0 || textureSize(uSunGlassTints, 0).x <= 1) return vec3(1.0);
   if (directionalNdl <= 0.0) return vec3(1.0);
   if (p.z > 1.0 || p.x < 0.0 || p.x > 1.0 || p.y < 0.0 || p.y > 1.0) return vec3(1.0);
@@ -295,8 +359,18 @@ vec3 sunGlassLookup(float directionalNdl) {
 
   float bias = 0.14 / uShadowDepthSpan;
   float texel = 1.35 / uShadowMapSize;
-  vec3 staticTint = sunGlassLayer(p, ${SUN_GLASS_STATIC_LAYER}.0, 0.0, bias, texel, uvPerMetre);
-  vec3 movingTint = sunGlassLayer(p, ${SUN_GLASS_MOVING_LAYER}.0, 1.0, bias, texel, uvPerMetre);
+  vec3 staticTint = sunGlassLayer(p, ${SUN_GLASS_STATIC_LAYER}.0, 0.0, bias, texel, uvPerMetre, uShadowDepthSpan);
+  vec3 movingTint = MOVING_SUN && sunMapFade(pm) <= 0.0
+    ? vec3(1.0)
+    : sunGlassLayer(
+      pm,
+      ${SUN_GLASS_MOVING_LAYER}.0,
+      1.0,
+      MOVING_SUN ? 0.14 / movingSunSpan() : bias,
+      texel,
+      movingPerMetre,
+      MOVING_SUN ? movingSunSpan() : uShadowDepthSpan
+    );
   return mix(vec3(1.0), staticTint * movingTint, strength);
 }
 

package/src/render/shaders/flat/glassTint.ts

@@ -36,7 +36,10 @@ export type LitSwitch =
   | 'DRIFT_LIGHT'
   | 'PHYSICAL_SPECULAR'
   | 'SURFACE_OVERLAY'
-  | 'REFLECTION_MAPS';
+  | 'REFLECTION_MAPS'
+  | 'WORLD_UVS'
+  | 'MOVING_SUN'
+  | 'LAYERED';
 
 /** Every switch off: what a renderer starts from before its profile and its content say otherwise. */
 export function noLitSwitches(): Record<LitSwitch, boolean> {
@@ -49,6 +52,9 @@ export function noLitSwitches(): Record<LitSwitch, boolean> {
     PHYSICAL_SPECULAR: false,
     SURFACE_OVERLAY: false,
     REFLECTION_MAPS: false,
+    WORLD_UVS: false,
+    MOVING_SUN: false,
+    LAYERED: false,
   };
 }
 

package/src/render/shaders/flat/index.ts

@@ -21,6 +21,10 @@ import { DRIFT_LIGHT_GLSL } from './driftLight.ts';
 import { SURFACE_EFFECTS_GLSL } from './surfaceEffects.ts';
 import { overlayGlsl } from './overlay.ts';
 import { reflectionSurfaceGlsl } from './reflectionSurface.ts';
+import { worldUvGlsl } from './worldUv.ts';
+import { movingSunGlsl } from './movingSun.ts';
+import { layeredGlsl } from './layered.ts';
+import { MAP_READS_GLSL } from './mapReads.ts';
 import { CHANNEL_BEND, LIT_CHANNEL_ATTRIBUTE } from '../vertexChannel.ts';
 import { BONE_ANIMATION_GLSL } from '../boneAnimation.ts';
 import { FULL_LIGHT_BUDGET, type LightBudget } from '../../uniformVectorBudget.ts';
@@ -508,6 +512,21 @@ export interface FlatShaderOptions {
    * the frame's materials' reflections. True when absent. See `reflectionSurfaceGlsl`.
    */
   readonly reflectionMaps?: boolean;
+  /**
+   * Whether it places a material's maps by the world, `uWorldUv` and all (`SurfaceMaterial
+   * .projection`). True when absent. See `worldUvGlsl`.
+   */
+  readonly worldUvs?: boolean;
+  /**
+   * Whether it reads the sun's moving layer through a matrix of its own, \`uMovingLightViewProj\`
+   * and all. True when absent. See \`movingSunGlsl\`.
+   */
+  readonly movingSun?: boolean;
+  /**
+   * Whether it blends a material's layers by a mask, \`uLayers\` and all (\`SurfaceMaterial.layers\`).
+   * True when absent. See \`layeredGlsl\`.
+   */
+  readonly layered?: boolean;
   /**
    * How many point lights this build declares room for. `MAX_POINT_LIGHTS` when absent.
    *
@@ -619,6 +638,8 @@ export function flatFrag(options: FlatShaderOptions): string {
         surfaceEffects: options.surfaceEffects ?? true,
         driftLight: options.driftLight ?? true,
       }),
+      /* Depth zero and ahead of the sun's lookup, which calls it; its switch takes the next id. */
+      movingSunGlsl(options.movingSun ?? true),
       LOBES_GLSL,
       POINTSHADOW_GLSL,
       DIRECTIONALSHADOW_GLSL,
@@ -644,8 +665,6 @@ export function flatFrag(options: FlatShaderOptions): string {
        * and `main` is unconditional by construction.
        */
       TANGENT_FRAME_GLSL,
-      /* After the frame it calls, unconditional for the same reason. See normalMap.ts. */
-      NORMAL_MAP_GLSL,
       /* Unconditional for the same reason, and called only where a cutoff was asked for. */
       CUTOUT_COVERAGE_GLSL,
       CUTOUT_DITHER_GLSL,
@@ -657,6 +676,15 @@ export function flatFrag(options: FlatShaderOptions): string {
       overlayGlsl(options.surfaceOverlay ?? true),
       /* After the overlay, so its switch takes the next id. See reflectionSurface.ts. */
       reflectionSurfaceGlsl(options.reflectionMaps ?? true),
+      /* After the reflection maps, so its switch takes the next id; main calls it. worldUv.ts. */
+      worldUvGlsl(options.worldUvs ?? true),
+      /* After the projection, so its switch takes the next id; the normal map below calls it. */
+      layeredGlsl(options.layered ?? true),
+      /* Where each map is read, after both of the ways it can be read otherwise. See mapReads.ts. */
+      MAP_READS_GLSL,
+      /* After the frame and the projection it calls, unconditional for the same reason as the
+         frame. See normalMap.ts. */
+      NORMAL_MAP_GLSL,
       /* After the models, whose answer to a rectangle it asks for. See areaLight.ts. */
       AREA_LIGHT_GLSL,
       MAIN_GLSL,