SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillswebgpu-impl-async-patterns

webgpu-impl-async-patterns

General & other

View repository

Updated 10-15 22:18

About this skill

WebGPU Async Patterns

WebGPU is asynchronous by design: the GPU runs on its own timeline while JavaScript runs on the content timeline. This skill structures the frame loop, maps buffers safely, and loads pipelines without blocking on WebGPU 1.0-stable (Chrome 113+, Safari 26+, Firefox 141+).

Quick Reference

API Returns Timeline Use for
buffer.mapAsync(mode, offset, size) Promise<undefined> resolves when buffer is CPU-owned CPU access to a MAP_READ/MAP_WRITE buffer
buffer.mapState "unmapped" / "pending" / "mapped" read synchronously guard before calling mapAsync
buffer.getMappedRange(offset, size) ArrayBuffer valid only while mapped read or write the mapped bytes
buffer.unmap() undefined returns buffer to "unmapped" release the buffer back to the GPU
queue.onSubmittedWorkDone() Promise<undefined> resolves after submitted work completes one-time GPU completion checkpoint
device.createRenderPipelineAsync(desc) Promise<GPURenderPipeline> compiles off the content timeline load-time pipeline creation
device.createComputePipelineAsync(desc) Promise<GPUComputePipeline> compiles off the content timeline load-time pipeline creation
requestAnimationFrame(cb) frame handle content timeline the per-frame render driver

Alignment for mapAsync and getMappedRange: offset MUST be a multiple of 8, size MUST be a multiple of 4.

Promise ordering: mapAsync and onSubmittedWorkDone promises settle in submission order.

Decision Tree

Creating a pipeline?
├── At app load / level load, no frame is rendering yet
│     -> ALWAYS createRenderPipelineAsync / createComputePipelineAsync.
│        await it before the first frame. Compilation runs off the content
│        timeline, so no shader-compilation hitch.
└── Already inside the frame loop and need a pipeline RIGHT NOW
      -> This is the anti-pattern. NEVER create pipelines per frame.
         Pre-create every pipeline at load time. See webgpu-impl-performance.

Reading GPU data back to the CPU?
├── One-off readback (screenshot, test, debug), no frame loop running
│     -> await queue.onSubmittedWorkDone(), then await stagingBuffer.mapAsync(READ).
│        A stall here is acceptable because no frame is in flight.
└── Continuous readback inside the render loop (picking, histogram, stats)
      -> NEVER await mapAsync or onSubmittedWorkDone in the render path.
         Use a rotating staging buffer; read results one or two frames late.
         See Pattern 4 and references/examples.md.

Where to await?
├── Initialization (adapter, device, pipelines, initial uploads) -> await freely.
└── Inside requestAnimationFrame -> NEVER await GPU completion. Encode and submit
      only. Let the GPU run ahead.

Core Patterns

Pattern 1: The requestAnimationFrame frame loop

ALWAYS drive rendering with requestAnimationFrame. Each frame: update uniforms via queue.writeBuffer, create a FRESH GPUCommandEncoder, encode the pass, call queue.submit. NEVER await GPU completion inside the callback.

function frame() {
  // 1. Update CPU-side state, push uniforms.
  queue.writeBuffer(uniformBuffer, 0, uniformData);

  // 2. Fresh encoder every frame. Command buffers are single-use.
  const encoder = device.createCommandEncoder({ label: "frame-encoder" });
  const pass = encoder.beginRenderPass({
    colorAttachments: [{
      view: context.getCurrentTexture().createView(), // fresh view per frame
      loadOp: "clear", storeOp: "store", clearValue: { r: 0, g: 0, b: 0, a: 1 },
    }],
  });
  pass.setPipeline(pipeline);
  pass.setBindGroup(0, bindGroup);
  pass.draw(3);
  pass.end();

  // 3. Submit. Do NOT await anything here.
  queue.submit([encoder.finish()]);

  requestAnimationFrame(frame);
}
requestAnimationFrame(frame);

NEVER cache getCurrentTexture() or its view across frames: the swap-chain rotates textures. See webgpu-core-pipeline-architecture for pipeline reuse.

Pattern 2: The mapAsync map-state lifecycle

A GPUBuffer moves through three states exposed by buffer.mapState: "unmapped" (GPU-owned) -> "pending" (after mapAsync, before resolve) -> "mapped" (CPU-owned). ALWAYS map only when mapState === "unmapped".

if (buffer.mapState !== "unmapped") return; // already pending or mapped
await buffer.mapAsync(GPUMapMode.READ);     // offset multiple of 8, size multiple of 4
const data = new Float32Array(buffer.getMappedRange().slice(0)); // copy out NOW
buffer.unmap();                              // ArrayBuffer is now detached

ALWAYS copy data out of getMappedRange() before calling unmap(). After unmap() the ArrayBuffer is detached and any typed view over it throws TypeError.

Pattern 3: Async pipeline creation at load time

ALWAYS use createRenderPipelineAsync / createComputePipelineAsync during loading. They compile shaders off the content timeline, so no compilation hitch reaches a frame.

const [opaquePipeline, shadowPipeline] = await Promise.all([
  device.createRenderPipelineAsync(opaqueDesc),
  device.createRenderPipelineAsync(shadowDesc),
]);
// Only start the frame loop after every pipeline resolves.
requestAnimationFrame(frame);

The synchronous createRenderPipeline is correct ONLY before the frame loop starts and for a small number of pipelines. NEVER call it for many heavy shaders during a frame.

Pattern 4: Non-blocking readback with a rotating staging buffer

For continuous GPU-to-CPU readback, ALWAYS use a ring of staging buffers and read each result one or two frames late. NEVER await mapAsync inside the frame.

const RING = 3;
const staging = Array.from({ length: RING }, (_, i) =>
  device.createBuffer({ label: `readback-${i}`, size: 256,
    usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ }));
let frameIndex = 0;

function frame() {
  const writeBuf = staging[frameIndex % RING];
  if (writeBuf.mapState === "unmapped") {
    const encoder = device.createCommandEncoder();
    encoder.copyBufferToBuffer(gpuResult, 0, writeBuf, 0, 256);
    queue.submit([encoder.finish()]);
    // Fire-and-forget map. The promise resolves on a LATER frame.
    writeBuf.mapAsync(GPUMapMode.READ).then(() => {
      const data = new Float32Array(writeBuf.getMappedRange().slice(0));
      writeBuf.unmap();
      consumeReadback(data);
    });
  }
  frameIndex++;
  requestAnimationFrame(frame);
}

Pattern 5: onSubmittedWorkDone as a one-time checkpoint

queue.onSubmittedWorkDone() resolves after all previously submitted work completes. Use it ONLY outside the render loop: a one-off screenshot, a test, or a clean shutdown.

queue.submit([encoder.finish()]);
await queue.onSubmittedWorkDone(); // GPU has finished. Acceptable stall: no frame in flight.
await stagingBuffer.mapAsync(GPUMapMode.READ);

NEVER place await queue.onSubmittedWorkDone() in a requestAnimationFrame callback.

Pattern 6: Synchronous mapSync in workers (experimental)

Chrome 145 (Jan 2026) adds experimental synchronous buffer.mapSync() for Web Workers behind --enable-features=WebGPUMapSyncOnWorkers. It is NOT in the WebGPU 1.0-stable spec and is unavailable on the main thread, in Safari, and in Firefox. ALWAYS feature-gate it (typeof buffer.mapSync === "function") and keep mapAsync as the only path on 1.0-stable.

Common Anti-Patterns

  1. Awaiting mapAsync or onSubmittedWorkDone inside the render loop. This forces a full CPU-GPU synchronization: the CPU idles until the GPU drains every submitted command, so the two timelines can no longer overlap and frame rate collapses. Fix: use a rotating staging buffer (Pattern 4) and read results one or two frames late.

  2. Calling mapAsync on an already-mapped or pending buffer. WebGPU rejects the second call with a "buffer is already mapped" validation error, because a buffer can only be re-mapped after unmap() returns it to "unmapped". Fix: guard with buffer.mapState === "unmapped" and strictly pair every mapAsync with unmap.

  3. Using a detached ArrayBuffer after unmap(). unmap() detaches the buffer returned by getMappedRange(); any typed view over it then throws TypeError. Fix: copy the data out (.slice(0)) before unmap().

Critical Warnings

  • NEVER await mapAsync or onSubmittedWorkDone inside a requestAnimationFrame callback. It collapses frame pipelining.
  • NEVER call mapAsync unless buffer.mapState === "unmapped".
  • NEVER read or write a getMappedRange() ArrayBuffer after unmap(); it is detached.
  • NEVER reuse a GPUCommandEncoder or GPUCommandBuffer across frames; both are single-use. Create a fresh encoder every frame.
  • NEVER recreate pipelines per frame; pre-create them with the async variants at load time. See webgpu-impl-performance.
  • ALWAYS keep mapAsync offset a multiple of 8 and size a multiple of 4.
  • ALWAYS feature-gate mapSync; it is experimental and absent on 1.0-stable.

Reference Files

  • references/methods.md : mapAsync and mapState lifecycle, onSubmittedWorkDone, createRenderPipelineAsync, and the frame-loop anatomy.
  • references/examples.md : working verified code for a requestAnimationFrame render loop, async pipeline loading, and non-blocking readback.
  • references/anti-patterns.md : async mistakes with WHY-it-fails explanations.

Related skills: webgpu-syntax-buffers (buffer creation and usage flags), webgpu-impl-buffer-upload (CPU-to-GPU upload strategies), webgpu-impl-performance (pipeline reuse, staging rings, bundle caching), webgpu-core-pipeline-architecture (pipeline objects and the timeline model).