SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillswebgpu-impl-buffer-upload

webgpu-impl-buffer-upload

Data & analytics

View repository

Updated 10-15 22:18

About this skill

WebGPU Buffer Upload and Readback

Moving data CPU to GPU (upload) and GPU to CPU (readback) in WebGPU 1.0-stable (Chrome 113+, Safari 26+, Firefox 141+), with the alignment rules that cause the most validation errors.

Quick Reference

Upload path When to use Needs COPY_DST API
queue.writeBuffer Safe default for any CPU to GPU upload yes device.queue.writeBuffer(buf, offset, data)
mappedAtCreation: true Initial vertex/index/uniform data, set once no createBuffer({ mappedAtCreation: true })
Staging-buffer ring (2-3 buffers) Large uploads issued every frame yes (on destination) mapAsync + copyBufferToBuffer
Readback path When to use API
Buffer readback Read a storage/compute-written buffer copyBufferToBuffer into a COPY_DST | MAP_READ buffer
Texture readback Read pixels / download a texture copyTextureToBuffer with 256-byte bytesPerRow padding

Alignment numbers in one place: writeBuffer bufferOffset and size multiples of 4; mappedAtCreation size a multiple of 4; mapAsync/getMappedRange offset a multiple of 8 and size a multiple of 4; bytesPerRow a multiple of 256. The per-rule rationale lives in webgpu-core-memory-model.

Decision Tree

Choosing an upload path

Uploading data to a buffer?
├─ Is this the buffer's initial content, written once at creation?
│   └─ YES → mappedAtCreation: true
│            (no COPY_DST needed, no extra copy, size MUST be multiple of 4)
├─ Is this a large buffer rewritten EVERY frame and writeBuffer profiles as a bottleneck?
│   └─ YES → staging-buffer ring (2-3 buffers, see references/methods.md)
└─ Anything else (per-frame uniforms, occasional updates, unknown)
    └─ queue.writeBuffer  ← ALWAYS the default

NEVER reach for a staging ring before writeBuffer is measured as the bottleneck. The ring adds buffer pool management and async bookkeeping; writeBuffer lets the user agent pick the optimal internal path with zero extra code.

GPU to CPU readback steps

1. Create a destination buffer: usage = COPY_DST | MAP_READ
   (MAP_READ ALWAYS combines only with COPY_DST, never STORAGE/UNIFORM/VERTEX)
2. Encode the copy INTO that buffer:
   - source is a buffer  → encoder.copyBufferToBuffer(...)
   - source is a texture → encoder.copyTextureToBuffer(...) with padded bytesPerRow
3. queue.submit([encoder.finish()])
4. await readbackBuffer.mapAsync(GPUMapMode.READ)   ← waits for the GPU
5. const range = readbackBuffer.getMappedRange()    ← ArrayBuffer view
6. Copy bytes OUT now: new Uint8Array(range).slice() or .set() into your own array
7. readbackBuffer.unmap()   ← range is detached after this; copy BEFORE unmap

Step 4 is what guarantees the GPU finished. The mapAsync promise resolves only after all prior submitted work touching the buffer completes, so readback never sees stale data when the copy and the mapAsync are on the same buffer.

Core Patterns

Pattern 1: writeBuffer is the default upload

ALWAYS use queue.writeBuffer for routine CPU to GPU uploads. The destination buffer MUST have COPY_DST.

const buf = device.createBuffer({
  size: data.byteLength,                       // multiple of 4
  usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,
});
device.queue.writeBuffer(buf, 0, data);        // bufferOffset 0, size = data.byteLength

NEVER call writeBuffer on a buffer that is currently mapped or destroyed; it fails validation. bufferOffset and the data size MUST both be multiples of 4.

Pattern 2: mappedAtCreation for initial data

ALWAYS use mappedAtCreation: true for vertex/index/uniform data written once. It needs NO COPY_DST flag and avoids an internal copy.

const vbuf = device.createBuffer({
  size: vertices.byteLength,                   // MUST be a multiple of 4
  usage: GPUBufferUsage.VERTEX,                // no COPY_DST required
  mappedAtCreation: true,
});
new Float32Array(vbuf.getMappedRange()).set(vertices);
vbuf.unmap();                                  // buffer is now GPU-owned

NEVER use the mapped range after unmap(); the ArrayBuffer is detached and any view over it throws TypeError.

Pattern 3: storage buffers cannot be mapped directly

A STORAGE buffer (and any GPU-written buffer not declared MAP_READ) CANNOT be mapped. ALWAYS copy it into a separate COPY_DST | MAP_READ staging buffer first.

const readback = device.createBuffer({
  size: storageBuffer.size,
  usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ,
});
const enc = device.createCommandEncoder();
enc.copyBufferToBuffer(storageBuffer, 0, readback, 0, storageBuffer.size);
device.queue.submit([enc.finish()]);
await readback.mapAsync(GPUMapMode.READ);
const out = new Float32Array(readback.getMappedRange()).slice();  // copy OUT
readback.unmap();

MAP_READ is ONLY valid combined with COPY_DST. NEVER add STORAGE, UNIFORM, or VERTEX to a MAP_READ buffer; that combination fails validation.

Pattern 4: bytesPerRow MUST be a multiple of 256

For copyTextureToBuffer, bytesPerRow MUST round up to the next multiple of 256. ALWAYS size the destination buffer as paddedBytesPerRow * height and strip the padding on the CPU after mapping.

const unpaddedBytesPerRow = width * 4;                        // RGBA8 = 4 bytes/texel
const paddedBytesPerRow = Math.ceil(unpaddedBytesPerRow / 256) * 256;
const readback = device.createBuffer({
  size: paddedBytesPerRow * height,                           // padded, not width*height*4
  usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ,
});
enc.copyTextureToBuffer(
  { texture },
  { buffer: readback, bytesPerRow: paddedBytesPerRow, rowsPerImage: height },
  { width, height },
);

After mapping, copy each logical row out of the padded buffer (full strip code in references/examples.md). NEVER size the buffer as width * height * 4 with bytesPerRow = width * 4 unless width * 4 is already a multiple of 256.

Pattern 5: sequence readback so it does not read stale data

ALWAYS encode the copy and call mapAsync on the SAME buffer in the SAME flow. The mapAsync promise resolves only after prior submitted work finishes.

device.queue.submit([enc.finish()]);
await readback.mapAsync(GPUMapMode.READ);   // resolves after the copy completes

When you must confirm an unrelated submission finished before reading a different buffer, ALWAYS await device.queue.onSubmittedWorkDone(). NEVER assume data is ready right after submit() returns; submit() only schedules the work.

Pattern 6: never re-map a pending or mapped buffer

NEVER call mapAsync on a buffer whose mapState is "pending" or "mapped"; it fails with "buffer is already mapped". ALWAYS unmap() first, or sequence map/unmap strictly. A staging ring (see references/methods.md) keeps each buffer single-mapped by rotating across 2-3 buffers.

Common Anti-Patterns

  1. Unpadded readback buffer. Sizing a copyTextureToBuffer destination as width * height * 4 with bytesPerRow = width * 4. When width * 4 is not a multiple of 256 the copy fails validation. Pad bytesPerRow to 256 and size the buffer as paddedBytesPerRow * height.
  2. Mapping a storage buffer directly. Storage buffers cannot be mapped; MAP_READ only combines with COPY_DST. Copy into a separate staging buffer.
  3. Reading before the GPU finished. Calling getMappedRange without awaiting mapAsync (or before its promise resolves) reads stale or undefined data. The mapAsync promise is the synchronization point.

Critical Warnings

  • NEVER combine MAP_READ with anything except COPY_DST, or MAP_WRITE with anything except COPY_SRC. Other combinations fail validation.
  • NEVER use a getMappedRange() ArrayBuffer after unmap(); it is detached. ALWAYS copy bytes out with .slice() or .set() before unmap().
  • NEVER pass a bytesPerRow that is not a multiple of 256 to copyTextureToBuffer or copyBufferToTexture.
  • NEVER call writeBuffer on a mapped or destroyed buffer.
  • NEVER build a staging ring when writeBuffer is not measured as a bottleneck; it is added complexity with no benefit for non-per-frame uploads.

Reference Files

  • references/methods.md : writeBuffer, mappedAtCreation, the staging-buffer ring, and the full readback workflow with bytesPerRow padding.
  • references/examples.md : verified working code for each upload path and a complete GPU to CPU texture readback with 256-byte padding stripping.
  • references/anti-patterns.md : upload and readback mistakes with WHY-it-fails explanations.

Related skills: webgpu-syntax-buffers (buffer creation and usage flags), webgpu-core-memory-model (the alignment numbers in detail), webgpu-impl-async-patterns (mapAsync, onSubmittedWorkDone, frame loop), webgpu-syntax-command-encoder (copyBufferToBuffer, copyTextureToBuffer).