@driftengine/core 4.10.0 → 4.11.1

package/src/render/backend/webgpu/gpuClothSet.ts

@@ -0,0 +1,564 @@
+/**
+ * Skinned cloth solved on the device, a set of garments at a time: `ClothControl`'s schedule for
+ * each garment, over one set of buffers, in one compute pass and one submit a frame.
+ *
+ * **The CPU solver's steps, as dispatches.** A step is `clothBegin`, then per substep
+ * `clothPredict`, every colour of every kind of constraint per iteration, `clothLimit` and
+ * `clothFinish`; then `clothBlend` while a reset blends back in. The particles never leave the
+ * device: a frame ends with `clothPublish` writing each garment's, at the frame's `alpha`, into the
+ * particle texture its bound meshes' vertex stage reads — so a garment costs no upload a frame.
+ *
+ * **A round is one kind of work for every garment that has it next.** Each garment's control asks
+ * for poses, resets and steps as it always did; they wait in its queue (`clothOps.ts`) until the
+ * set records them a round at a time — the poses of every garment with a pose next, in one
+ * dispatch, then the rests, then a step for every garment with a step next, each colour of each kind
+ * one dispatch across the set (`clothSetPack.ts`). A garment with nothing of that kind sits the round
+ * out. So a step costs one garment's dispatches whatever the set holds: fifteen garments of nine
+ * colours at four iterations are about a hundred a step, where a pass a garment was some fourteen
+ * hundred.
+ *
+ * **One submit a frame, everything it needs written before it.** `queue.writeBuffer` lands ahead of
+ * the whole submit rather than between two dispatches, which is the trap `AGENTS.md` (2026-08-27)
+ * records for draws and is no different here. So every round's records take a slot of their own, a
+ * frame that runs out of slots submits what it has and goes on, and a garment posed while work
+ * against its last pose is still unsent sends that work first, since its matrices are one region.
+ *
+ * **What it gives up**: single precision; a colour dispatched over the most records any garment has
+ * in it; and garments of one set step together, so they share a step, substeps, iterations and most
+ * steps a frame, refused by name otherwise. Self-collision is accepted and not simulated, as on the
+ * CPU.
+ */
+import { ClothControl, exactExp, validateClothSetup } from '@driftengine/physics';
+import type { ClothDevice, SkinnedClothSetup } from '@driftengine/physics';
+
+import { clothKernelsFor } from './clothKernels.ts';
+import type { ClothEntry, ClothKernels } from './clothKernels.ts';
+import {
+  CLOTH_OP_POSE,
+  CLOTH_OP_REST,
+  CLOTH_OP_STEP,
+  ClothOps,
+  PACE_HEAD_WORDS,
+  PACE_WORDS,
+} from './clothOps.ts';
+import {
+  CLOTH_CONSTANT_BYTES,
+  clothBatchSlots,
+  clothConstants,
+  clothRoundOffsets,
+  writeClothCarry,
+} from './clothSetLayout.ts';
+import { packClothSet } from './clothSetPack.ts';
+import type { ClothGarmentRange, PackedClothSet } from './clothSetPack.ts';
+import { packCloth } from './clothSolvePack.ts';
+import { GpuClothParticles } from './clothTextures.ts';
+import { CLOTH_WORKGROUP } from './shaders/clothSolve.wgsl.ts';
+import { DYNAMIC_ALIGNMENT } from './uniformRing.ts';
+
+const STORAGE = 0x80;
+const UNIFORM = 0x40;
+const COPY_DST = 0x8;
+const COPY_SRC = 0x4;
+const MAP_READ = 0x1;
+
+/** Rounds a submit holds before it is sent and the slots start again. */
+const ROUNDS = 64;
+const KINDS = [CLOTH_OP_POSE, CLOTH_OP_REST, CLOTH_OP_STEP] as const;
+
+/** One garment of a set: its schedule, its queue and where it sits. */
+interface Garment {
+  readonly range: ClothGarmentRange;
+  readonly control: ClothControl;
+  readonly particles: GpuClothParticles;
+  readonly ops: ClothOps;
+  /** The air its steps drag toward: the set-up's until `setWind`. */
+  readonly wind: Float32Array;
+  readonly outputs: readonly [GPUBindGroup, GPUBindGroup];
+  /** Where the pose a step starts from and the one it moves toward sit in `pose`, in floats. */
+  fromAt: number;
+  toAt: number;
+  /** Work recorded against its pose that has not been sent. */
+  recorded: boolean;
+}
+
+export class GpuClothSet {
+  private readonly garments: Garment[] = [];
+  private readonly packed: PackedClothSet;
+  private readonly kernels: ClothKernels;
+  private readonly buffers: GPUBuffer[] = [];
+  private readonly state: GPUBuffer;
+  private readonly frameData: GPUBuffer;
+  private readonly paces: GPUBuffer;
+  private readonly group: GPUBindGroup;
+  /** One round's records, every garment's: the head an op's, the rest each garment's own. */
+  private readonly staging: ArrayBuffer;
+  private readonly stagingWords: Uint32Array;
+  private readonly stagingFloats: Float32Array;
+  /** Each kind's first batch slot, and the first garment's publishing slot. */
+  private readonly batchSlot: {
+    distance: number;
+    bending: number;
+    tether: number;
+    publish: number;
+  };
+  private readonly groups: number;
+  private readonly substeps: number;
+  private readonly iterations: number;
+
+  private encoder: GPUCommandEncoder | null = null;
+  private pass: GPUComputePassEncoder | null = null;
+  private round = 0;
+
+  constructor(
+    private readonly device: GPUDevice,
+    setups: readonly SkinnedClothSetup[],
+  ) {
+    if (setups.length === 0) throw new Error('skinned cloth set: no garments');
+    for (const setup of setups) validateClothSetup(setup);
+    const packed = packClothSet(setups.map(packCloth));
+    this.packed = packed;
+    const n = packed.count;
+    const count = setups.length;
+    this.kernels = clothKernelsFor(device);
+    this.groups = Math.ceil(n / CLOTH_WORKGROUP);
+
+    const buffer = (label: string, bytes: number, usage: number): GPUBuffer => {
+      const made = device.createBuffer({
+        label: `cloth.${label}`,
+        size: Math.max(16, bytes),
+        usage,
+      });
+      this.buffers.push(made);
+      return made;
+    };
+    this.staging = new ArrayBuffer(count * PACE_WORDS * 4);
+    this.stagingWords = new Uint32Array(this.staging);
+    this.stagingFloats = new Float32Array(this.staging);
+    setups.forEach((setup, g) => this.garments.push(this.garment(setup, g)));
+    const first = (this.garments[0] as Garment).control.parameters;
+    this.substeps = first.substeps;
+    this.iterations = first.iterations;
+    this.refuseStrangers(first);
+
+    const constants = buffer('constants', CLOTH_CONSTANT_BYTES, UNIFORM | COPY_DST);
+    const rounds = buffer('rounds', ROUNDS * DYNAMIC_ALIGNMENT, UNIFORM | COPY_DST);
+    this.paces = buffer('paces', ROUNDS * this.staging.byteLength, STORAGE | COPY_DST);
+    const batchWords = clothBatchSlots(packed);
+    const batches = buffer('batches', batchWords.byteLength, UNIFORM | COPY_DST);
+    this.state = buffer('state', n * 12 * 4, STORAGE | COPY_DST | COPY_SRC);
+    const pose = buffer('pose', n * 9 * 4, STORAGE);
+    const statics = buffer('statics', packed.statics.byteLength, STORAGE | COPY_DST);
+    const constraints = buffer('constraints', packed.constraints.byteLength, STORAGE | COPY_DST);
+    const lambda = buffer('lambda', (packed.distances + packed.bendings) * 4, STORAGE);
+    this.frameData = buffer(
+      'frame',
+      (packed.joints * 16 + packed.colliders * 8) * 4,
+      STORAGE | COPY_DST,
+    );
+
+    const queue = device.queue;
+    queue.writeBuffer(constants, 0, clothConstants(packed, first.step / first.substeps));
+    queue.writeBuffer(rounds, 0, clothRoundOffsets(ROUNDS, count));
+    queue.writeBuffer(batches, 0, batchWords);
+    queue.writeBuffer(statics, 0, packed.statics as Float32Array<ArrayBuffer>);
+    if (packed.constraints.byteLength > 0) {
+      queue.writeBuffer(constraints, 0, packed.constraints as Uint32Array<ArrayBuffer>);
+    }
+    /* The rest pose until the first pose, which resets each garment to its own. */
+    const state = new Float32Array(n * 12);
+    setups.forEach((setup, g) => {
+      const base = (this.garments[g] as Garment).range.particleBase * 3;
+      state.set(setup.positions, base);
+      state.set(setup.positions, n * 9 + base);
+    });
+    queue.writeBuffer(this.state, 0, state);
+
+    const slot = (of: GPUBuffer): GPUBufferBinding => ({ buffer: of, size: DYNAMIC_ALIGNMENT });
+    this.group = device.createBindGroup({
+      label: 'cloth.solve',
+      layout: this.kernels.layout,
+      entries: [
+        { binding: 0, resource: { buffer: constants } },
+        { binding: 1, resource: slot(rounds) },
+        { binding: 2, resource: slot(batches) },
+        { binding: 3, resource: { buffer: this.state } },
+        { binding: 4, resource: { buffer: pose } },
+        { binding: 5, resource: { buffer: statics } },
+        { binding: 6, resource: { buffer: constraints } },
+        { binding: 7, resource: { buffer: lambda } },
+        { binding: 8, resource: { buffer: this.frameData } },
+        { binding: 9, resource: { buffer: this.paces } },
+      ],
+    });
+    const colours = (batches: Uint32Array): number => (batches.length - 1) * 2;
+    this.batchSlot = {
+      distance: 0,
+      bending: colours(packed.distanceBatches),
+      tether: colours(packed.distanceBatches) + colours(packed.bendingBatches),
+      publish:
+        colours(packed.distanceBatches) +
+        colours(packed.bendingBatches) +
+        colours(packed.tetherBatches),
+    };
+  }
+
+  /** How many garments the set holds. */
+  get size(): number {
+    return this.garments.length;
+  }
+
+  /** Garment `g`'s particles, what `setCloth` draws its bound meshes by. */
+  particles(g: number): GpuClothParticles {
+    return this.at(g).particles;
+  }
+
+  /** Garment `g`'s rig for the steps to come; see `ClothControl.setPose`. */
+  setPose(g: number, globals: Float32Array, model: Float32Array): void {
+    this.at(g).control.setPose(globals, model);
+  }
+
+  /** The frame's sample of the scene's wind for garment `g`. See `SkinnedCloth.setWind`. */
+  setWind(g: number, x: number, y: number, z: number): void {
+    const wind = this.at(g).wind;
+    wind[0] = x;
+    wind[1] = y;
+    wind[2] = z;
+  }
+
+  /** Garment `g` back to its skinned pose, settled; the next `step` publishes it. */
+  reset(g: number): void {
+    this.at(g).control.reset();
+  }
+
+  /**
+   * `dt` of the caller's time in whole fixed steps for every garment, and each garment's particles
+   * at its `alpha` published for the draws that follow, as the particles of `frame`.
+   */
+  step(dt: number, frame: number): void {
+    const garments = this.garments;
+    for (let g = 0; g < garments.length; g++) (garments[g] as Garment).control.advance(dt);
+    this.run();
+    const at = this.takeRound();
+    for (let g = 0; g < garments.length; g++) {
+      const base = g * PACE_WORDS;
+      const garment = garments[g] as Garment;
+      this.stagingWords.fill(0, base, base + PACE_HEAD_WORDS);
+      this.stagingFloats[base + 2] = garment.control.alpha;
+      this.stagingFloats.set(garment.wind, base + 28);
+    }
+    this.device.queue.writeBuffer(this.paces, this.paceOffset(at), this.staging);
+    for (let g = 0; g < garments.length; g++) {
+      const garment = garments[g] as Garment;
+      const view = garment.particles.publish(frame);
+      const slot = (this.batchSlot.publish + g) * DYNAMIC_ALIGNMENT;
+      const groups = Math.ceil(garment.range.count / CLOTH_WORKGROUP);
+      this.dispatch('clothPublish', at, slot, garment.outputs[view] as GPUBindGroup, groups);
+    }
+    this.flush();
+  }
+
+  /**
+   * Garment `g`'s particles where its last step left them, three floats each: for the parity check,
+   * which compares them with the CPU solver's positions. Not for a frame — it waits on the device.
+   */
+  async read(g: number): Promise<Float32Array> {
+    this.run();
+    this.flush();
+    const { particleBase, count } = this.at(g).range;
+    const bytes = count * 12;
+    const readback = this.device.createBuffer({
+      label: 'cloth.readback',
+      size: bytes,
+      usage: MAP_READ | COPY_DST,
+    });
+    const encoder = this.device.createCommandEncoder({ label: 'cloth.read' });
+    encoder.copyBufferToBuffer(this.state, particleBase * 12, readback, 0, bytes);
+    this.device.queue.submit([encoder.finish()]);
+    await readback.mapAsync(MAP_READ);
+    const out = new Float32Array(readback.getMappedRange().slice(0));
+    readback.unmap();
+    readback.destroy();
+    return out;
+  }
+
+  /** Its buffers; the particle textures are the renderer's to retire with the frame. */
+  dispose(): void {
+    this.flush();
+    for (const made of this.buffers) made.destroy();
+  }
+
+  private at(g: number): Garment {
+    const garment = this.garments[g];
+    if (garment === undefined) throw new Error(`skinned cloth set: no garment ${g}`);
+    return garment;
+  }
+
+  /** Garment `g`: its schedule over a device that queues, its particles and its own numbers. */
+  private garment(setup: SkinnedClothSetup, g: number): Garment {
+    const range = this.packed.garments[g] as ClothGarmentRange;
+    const particles = new GpuClothParticles(this.device, range.count, true);
+    const ops = new ClothOps();
+    const solver: ClothDevice = {
+      targetsChanged: (change) => this.targetsChanged(garment, change),
+      rest: () => this.queueOp(garment, CLOTH_OP_REST),
+      step: (fraction, blend) => this.queueStep(garment, fraction, blend),
+    };
+    const control = new ClothControl(setup, solver, false);
+    const output = (k: number): GPUBindGroup =>
+      this.device.createBindGroup({
+        label: 'cloth.publish',
+        layout: this.kernels.outputLayout,
+        entries: [{ binding: 0, resource: particles.viewOf(k) }],
+      });
+    const garment: Garment = {
+      range,
+      control,
+      particles,
+      ops,
+      wind: new Float32Array(control.parameters.wind),
+      outputs: [output(0), output(1)],
+      fromAt: 0,
+      toAt: 0,
+      recorded: false,
+    };
+    this.writeOwn(garment, g);
+    return garment;
+  }
+
+  /**
+   * A garment's own words of its record, which every round carries: its colliders, its texture's
+   * width and its constants. Its limits ride in the head beside `live`, written as it joins a round.
+   */
+  private writeOwn(garment: Garment, g: number): void {
+    const p = garment.control.parameters;
+    const h = p.step / p.substeps;
+    const base = g * PACE_WORDS;
+    const u = this.stagingWords;
+    const f = this.stagingFloats;
+    u[base + 32] = garment.range.colliderBase;
+    u[base + 33] = garment.range.colliders;
+    u[base + 34] = garment.particles.width;
+    f[base + 36] = exactExp(p.damping * h);
+    f[base + 37] = exactExp(p.drag * h);
+    f[base + 38] = p.maxDistanceScale;
+    f[base + 39] = p.margin;
+    f.set(p.gravity, base + 40);
+  }
+
+  /** Garments that cannot step together, refused by the first number they disagree on. */
+  private refuseStrangers(first: ClothControl['parameters']): void {
+    this.garments.forEach((garment, g) => {
+      const p = garment.control.parameters;
+      for (const name of ['step', 'substeps', 'iterations', 'maxSteps'] as const) {
+        if (p[name] !== first[name]) {
+          throw new Error(
+            `skinned cloth set: garment ${g} has ${name} ${p[name]} where garment 0 has ` +
+              `${first[name]}; a set steps its garments together, so they must agree`,
+          );
+        }
+      }
+    });
+  }
+
+  private targetsChanged(garment: Garment, change: 'posed' | 'settled' | 'consumed'): void {
+    if (change !== 'posed') {
+      garment.fromAt = garment.toAt;
+      return;
+    }
+    /*
+     * The pose's matrices and colliders are written ahead of the next submit, so any work of this
+     * garment's still waiting or recorded — a reset's settle steps, which read the colliders — is
+     * sent first, or it would run against this pose rather than its own.
+     */
+    if (garment.ops.pending > 0 || garment.recorded) {
+      this.run();
+      this.flush();
+    }
+    this.writeFrameData(garment);
+    garment.fromAt = garment.toAt;
+    garment.toAt = garment.toAt === 0 ? this.packed.count * 3 : 0;
+    this.queueOp(garment, CLOTH_OP_POSE);
+  }
+
+  private queueOp(garment: Garment, kind: number): number {
+    const at = garment.ops.push(kind);
+    const words = garment.ops.headWords;
+    words[at + 4] = garment.fromAt;
+    words[at + 5] = garment.toAt;
+    garment.ops.headFloats.set(garment.wind, at + 28);
+    return at;
+  }
+
+  private queueStep(garment: Garment, fraction: number, blend: number): void {
+    const at = this.queueOp(garment, CLOTH_OP_STEP);
+    const f = garment.ops.headFloats;
+    f[at] = fraction;
+    f[at + 1] = blend;
+    const { parameters, targets } = garment.control;
+    writeClothCarry(
+      parameters,
+      targets.modelBefore,
+      targets.modelNow,
+      garment.ops.headWords,
+      f,
+      at,
+    );
+  }
+
+  /** Every op every garment is waiting on, a round of one kind at a time, in each garment's order. */
+  private run(): void {
+    for (;;) {
+      let ran = false;
+      for (let k = 0; k < KINDS.length; k++) {
+        const kind = KINDS[k] as number;
+        if (this.waiting(kind)) {
+          this.runRound(kind);
+          ran = true;
+        }
+      }
+      if (!ran) return;
+    }
+  }
+
+  /** Whether any garment has `kind` next. */
+  private waiting(kind: number): boolean {
+    for (let g = 0; g < this.garments.length; g++) {
+      if ((this.garments[g] as Garment).ops.next === kind) return true;
+    }
+    return false;
+  }
+
+  /** One round: every garment with `kind` next takes part, and the rest sit it out. */
+  private runRound(kind: number): void {
+    const at = this.takeRound();
+    const words = this.stagingWords;
+    let blends = false;
+    for (let g = 0; g < this.garments.length; g++) {
+      const garment = this.garments[g] as Garment;
+      const base = g * PACE_WORDS;
+      words.fill(0, base, base + PACE_HEAD_WORDS);
+      if (garment.ops.next !== kind) continue;
+      garment.ops.take(words, base);
+      words[base + 6] = 1;
+      words[base + 7] = garment.range.limits;
+      garment.recorded = true;
+      if ((this.stagingFloats[base + 1] as number) > 0) blends = true;
+    }
+    this.device.queue.writeBuffer(this.paces, this.paceOffset(at), this.staging);
+    if (kind === CLOTH_OP_POSE) this.dispatch('clothPose', at, 0, null);
+    else if (kind === CLOTH_OP_REST) this.dispatch('clothRest', at, 0, null);
+    else this.stepRound(at, blends);
+  }
+
+  private stepRound(at: number, blends: boolean): void {
+    this.dispatch('clothBegin', at, 0, null);
+    const { distanceBatches, bendingBatches, tetherBatches } = this.packed;
+    for (let s = 0; s < this.substeps; s++) {
+      this.dispatch('clothPredict', at, 0, null);
+      for (let pass = 0; pass < this.iterations; pass++) {
+        const first = pass === 0 ? 1 : 0;
+        this.batches('clothDistance', distanceBatches, this.batchSlot.distance, first, at);
+        this.batches('clothBending', bendingBatches, this.batchSlot.bending, first, at);
+        this.batches('clothTether', tetherBatches, this.batchSlot.tether, first, at);
+      }
+      this.dispatch('clothLimit', at, 0, null);
+      this.dispatch('clothFinish', at, 0, null);
+    }
+    if (blends) this.dispatch('clothBlend', at, 0, null);
+  }
+
+  /** Every colour of one kind across the set, in colour order, each a dispatch over its range. */
+  private batches(
+    entry: ClothEntry,
+    batches: Uint32Array,
+    firstSlot: number,
+    first: number,
+    roundAt: number,
+  ): void {
+    for (let b = 0; b + 1 < batches.length; b++) {
+      const size = (batches[b + 1] as number) - (batches[b] as number);
+      if (size === 0) continue;
+      const slot = firstSlot + b * 2 + first;
+      this.dispatch(
+        entry,
+        roundAt,
+        slot * DYNAMIC_ALIGNMENT,
+        null,
+        Math.ceil(size / CLOTH_WORKGROUP),
+      );
+    }
+  }
+
+  private dispatch(
+    entry: ClothEntry,
+    roundAt: number,
+    batchAt: number,
+    output: GPUBindGroup | null,
+    groups = this.groups,
+  ): void {
+    const pass = this.passFor();
+    pass.setPipeline(this.kernels.pipelines[entry]);
+    pass.setBindGroup(0, this.group, [roundAt, batchAt]);
+    if (output !== null) pass.setBindGroup(1, output);
+    pass.dispatchWorkgroups(groups);
+  }
+
+  private passFor(): GPUComputePassEncoder {
+    if (this.pass === null) {
+      this.encoder = this.device.createCommandEncoder({ label: 'cloth' });
+      this.pass = this.encoder.beginComputePass({ label: 'cloth' });
+    }
+    return this.pass;
+  }
+
+  /** A round's slot, sending what is recorded first where the slots are spent: its byte offset. */
+  private takeRound(): number {
+    if (this.round === ROUNDS) this.flush();
+    return this.round++ * DYNAMIC_ALIGNMENT;
+  }
+
+  /** Where the records of the round at `roundAt` sit in `paces`. */
+  private paceOffset(roundAt: number): number {
+    return (roundAt / DYNAMIC_ALIGNMENT) * this.staging.byteLength;
+  }
+
+  /** Send what has been recorded: every slot it reads was written ahead of it on the queue. */
+  private flush(): void {
+    const { pass, encoder } = this;
+    if (pass === null || encoder === null) return;
+    pass.end();
+    this.device.queue.submit([encoder.finish()]);
+    this.pass = null;
+    this.encoder = null;
+    this.round = 0;
+    for (let g = 0; g < this.garments.length; g++) (this.garments[g] as Garment).recorded = false;
+  }
+
+  /** A garment's skin matrices — or its model, with no rig — and its colliders, into the set's. */
+  private writeFrameData(garment: Garment): void {
+    const targets = garment.control.targets;
+    const { jointBase, joints, colliderBase, colliders, rigged } = garment.range;
+    const queue = this.device.queue;
+    const matrices = (
+      rigged && targets.skin !== null ? targets.skin : targets.model
+    ) as Float32Array<ArrayBuffer>;
+    queue.writeBuffer(this.frameData, jointBase * 64, matrices, 0, joints * 16);
+    if (colliders === 0) return;
+    const ends = this.packed.joints * 16;
+    const radii = ends + this.packed.colliders * 6;
+    queue.writeBuffer(
+      this.frameData,
+      (ends + colliderBase * 6) * 4,
+      targets.colliderEnds as Float32Array<ArrayBuffer>,
+      0,
+      colliders * 6,
+    );
+    queue.writeBuffer(
+      this.frameData,
+      (radii + colliderBase * 2) * 4,
+      targets.colliderRadii as Float32Array<ArrayBuffer>,
+      0,
+      colliders * 2,
+    );
+  }
+}

package/src/render/backend/webgpu/latePass.ts

@@ -126,6 +126,7 @@ export class LatePass {
     command.offsetCount = 0;
     command.vertexCount = 0;
     command.indexBuffer = null;
+    command.indexFormat = 'uint32';
     command.indexed = false;
     command.instances = 1;
     command.indirect = null;

package/src/render/backend/webgpu/motionPass.ts

@@ -254,6 +254,8 @@ export class MotionPass {
       readonly weights: number;
       readonly joints2: number;
       readonly weights2: number;
+      readonly weightsFormat: GPUVertexFormat;
+      readonly weights2Format: GPUVertexFormat;
     },
     cloth = false,
   ): GPURenderPipeline | null {
@@ -261,13 +263,18 @@ export class MotionPass {
     const eight = offsets.joints2 >= 0 && offsets.weights2 >= 0;
     /*
      * Five byte counts, each under 256 because a vertex is at most 172 bytes, as five base-256
-     * digits: exact below 2^40, so two layouts can never share a key.
+     * digits, and two bits for the weights' formats: exact below 2^42, so two layouts can never
+     * share a key.
      */
     const key =
-      (((stride * 256 + offsets.joints) * 256 + offsets.weights) * 256 + offsets.joints2 + 1) *
+      ((((stride * 256 + offsets.joints) * 256 + offsets.weights) * 256 + offsets.joints2 + 1) *
         256 +
-      offsets.weights2 +
-      1;
+        offsets.weights2 +
+        1) *
+        4 +
+      /* And whether each four weights are floats or fixed point, two bits: `vertexPacking.ts`. */
+      (offsets.weightsFormat === 'float32x4' ? 0 : 2) +
+      (offsets.weights2Format === 'float32x4' ? 0 : 1);
     const pipelines = cloth ? this.clothPipelines : this.skinnedPipelines;
     const held = pipelines.get(key);
     if (held !== undefined) return held;
@@ -287,11 +294,15 @@ export class MotionPass {
             attributes: [
               { shaderLocation: 0, offset: 0, format: 'float32x3' },
               { shaderLocation: 11, offset: offsets.joints, format: 'float32x4' },
-              { shaderLocation: 12, offset: offsets.weights, format: 'float32x4' },
+              { shaderLocation: 12, offset: offsets.weights, format: offsets.weightsFormat },
               ...(eight
                 ? ([
                     { shaderLocation: 14, offset: offsets.joints2, format: 'float32x4' },
-                    { shaderLocation: 15, offset: offsets.weights2, format: 'float32x4' },
+                    {
+                      shaderLocation: 15,
+                      offset: offsets.weights2,
+                      format: offsets.weights2Format,
+                    },
                   ] as const)
                 : []),
             ],
@@ -635,7 +646,7 @@ export class MotionPass {
         }
         pass.setBindGroup(0, this.group, [i * MOTION_DRAW_STRIDE]);
         pass.setVertexBuffer(0, mesh.vertexBuffers[0] as GPUBuffer);
-        pass.setIndexBuffer(mesh.indexBuffer, 'uint32');
+        pass.setIndexBuffer(mesh.indexBuffer, mesh.indexFormat);
         pass.drawIndexed(mesh.indexCount, instances);
       }
     }