SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillswebgpu-agents-quality-validator

webgpu-agents-quality-validator

Agents & models

View repository

Updated 10-15 22:18

About this skill

WebGPU Quality Validator

Reviews generated WebGPU host code and WGSL shaders for correctness, then routes each defect to the sibling skill that holds the authoritative fix. This is a REVIEW skill: it checks and routes, it does NOT re-teach the APIs.

Quick Reference

Review category What it catches Routes to
Initialization Missing adapter null-check, missing device.lost handler webgpu-core-architecture, webgpu-errors-device-loss
Features and limits Optional feature in requiredFeatures without detection webgpu-core-limits-features, webgpu-core-cross-browser
Buffers Illegal usage-flag combinations, missing usage flags webgpu-syntax-buffers
Alignment Size not mult-4, dynamic offset not 256, bytesPerRow not 256 webgpu-core-memory-model, webgpu-impl-buffer-upload
Textures and samplers Format/usage mismatch, sampler binding-type mismatch webgpu-syntax-textures, webgpu-wgsl-textures
Bind groups @binding/@group mismatch, layout mismatch, mixed auto/explicit webgpu-syntax-bind-groups
Pipelines Target format mismatch, sync pipeline build per frame webgpu-syntax-render-pipeline, webgpu-core-pipeline-architecture
Render passes Attachment count/format mismatch, sampleCount mismatch webgpu-impl-render-targets, webgpu-impl-multipass
WGSL correctness Recursion, missing default:, uniform struct layout webgpu-wgsl-syntax, webgpu-wgsl-memory-layout, webgpu-wgsl-uniformity
Compute correctness Missing @workgroup_size, missing barrier webgpu-wgsl-compute-shaders, webgpu-syntax-compute-pipeline
Error handling No error scopes, missing labels webgpu-errors-validation, webgpu-errors-debugging
Performance Per-frame pipeline or bind-group rebuild, render-loop stall webgpu-impl-performance, webgpu-impl-async-patterns

Review Checklist

Run the checklist in this fixed order. Each line is a binary pass/fail check. Full checklist text and the issue-to-skill routing map are in references/methods.md.

1. Initialization

  • ALWAYS verify the code null-checks the adapter: requestAdapter() can resolve to null, which is NOT a rejection. Flag any code that calls adapter.requestDevice() without a preceding if (!adapter) guard.
  • ALWAYS verify a device.lost handler is attached before any GPU resource is created. Flag any WebGPU app with no device.lost.then(...).
  • ALWAYS verify device.queue is read as a property. Flag device.queue() (call form).

2. Features and limits

  • ALWAYS verify every entry in requiredFeatures is gated by adapter.features.has(...). Flag a hard-coded feature string in requiredFeatures.
  • ALWAYS verify requiredLimits requests only what the app needs. Flag values copied from adapter.limits wholesale.

3. Buffers

  • ALWAYS verify usage flags are legal: MAP_READ combines only with COPY_DST, MAP_WRITE combines only with COPY_SRC. Flag MAP_READ | STORAGE and similar.
  • ALWAYS verify the usage flag set matches how the buffer is used (a vertex buffer needs VERTEX, an indirect buffer needs INDIRECT).

4. Alignment

  • ALWAYS verify buffer size is a multiple of 4 when mappedAtCreation: true.
  • ALWAYS verify writeBuffer bufferOffset and size are multiples of 4.
  • ALWAYS verify dynamic bind-group offsets are multiples of 256.
  • ALWAYS verify bytesPerRow in buffer-texture copies is a multiple of 256.

5. Textures and samplers

  • ALWAYS verify texture usage includes RENDER_ATTACHMENT for render targets, TEXTURE_BINDING for sampled textures, STORAGE_BINDING for storage textures.
  • ALWAYS verify a comparison sampler binds to a "comparison" sampler layout entry and a filtering sampler binds to a "filtering" entry.
  • ALWAYS verify the canvas format comes from getPreferredCanvasFormat(), not a hard-coded "bgra8unorm".

6. Bind groups and layouts

  • ALWAYS verify each WGSL @group(g) @binding(b) has a matching GPUBindGroupLayoutEntry with the same binding and a compatible resource type.
  • ALWAYS verify visibility covers every shader stage that reads the binding.
  • ALWAYS verify a bind group built from pipeline.getBindGroupLayout(i) is used only with that pipeline. Flag a shared bind group built from an "auto" layout.

7. Pipelines

  • ALWAYS verify each fragment.targets[i].format equals the matching render-pass colorAttachments[i] texture format.
  • ALWAYS verify pipelines are built once at load time, never inside the frame loop.

8. Render passes

  • ALWAYS verify colorAttachments count and formats match the pipeline fragment.targets.
  • ALWAYS verify pipeline multisample.count equals the attachment texture sampleCount and that any resolveTarget is single-sampled.
  • ALWAYS verify every beginRenderPass/beginComputePass is closed with .end() before encoder.finish().

9. WGSL correctness

  • ALWAYS verify the static call graph is acyclic. Flag any recursive WGSL function.
  • ALWAYS verify every switch has a default: clause.
  • ALWAYS verify a uniform struct accounts for vec3 16-byte alignment and 16-byte array element stride; the host Float32Array offsets MUST match the WGSL layout.
  • ALWAYS verify integer inter-stage variables carry @interpolate(flat).
  • ALWAYS verify textureSample, dpdx, and barriers run in uniform control flow.

10. Compute correctness

  • ALWAYS verify every @compute entry point declares @workgroup_size(...).
  • ALWAYS verify a workgroupBarrier() separates a write to var<workgroup> from a later read by other invocations.

11. Error handling and debugging

  • ALWAYS verify suspect resource creation is bracketed by pushErrorScope/popErrorScope.
  • ALWAYS verify every buffer, texture, pipeline, bind group, and shader module has a label.
  • ALWAYS verify getCompilationInfo() is checked after createShaderModule.

12. Performance

  • ALWAYS verify the render loop does not await mapAsync or await onSubmittedWorkDone.
  • ALWAYS verify render bundles, when used, are built once and replayed, never rebuilt per frame.

Decision Tree

Use this tree to route a confirmed defect to the authoritative skill.

Issue found in review
├─ API name / descriptor field / enum / WGSL builtin not in the 1.0 spec surface?
│   └─ HALLUCINATED API -> see references/methods.md "API surface cross-check"
├─ Initialization defect
│   ├─ adapter not null-checked / queue called -> webgpu-core-architecture
│   └─ no device.lost handler / unsafe recovery -> webgpu-errors-device-loss
├─ Feature or limit defect -> webgpu-core-limits-features, webgpu-core-cross-browser
├─ Buffer usage / mapping defect -> webgpu-syntax-buffers
├─ Alignment defect (4 / 8 / 256) -> webgpu-core-memory-model, webgpu-impl-buffer-upload
├─ Texture / sampler defect -> webgpu-syntax-textures
├─ Bind-group / layout / @binding mismatch -> webgpu-syntax-bind-groups
├─ Pipeline format / build defect -> webgpu-syntax-render-pipeline,
│                                    webgpu-core-pipeline-architecture
├─ Render-pass attachment / MSAA defect -> webgpu-impl-render-targets, webgpu-impl-multipass
├─ WGSL language defect
│   ├─ types / control flow / recursion -> webgpu-wgsl-syntax
│   ├─ struct layout / alignment / address space -> webgpu-wgsl-memory-layout
│   ├─ texture handle / sampler type -> webgpu-wgsl-textures
│   └─ uniform control flow / diagnostics -> webgpu-wgsl-uniformity
├─ Shader-stage defect -> webgpu-wgsl-vertex-shaders, webgpu-wgsl-fragment-shaders,
│                         webgpu-wgsl-compute-shaders
├─ Error-handling / labeling defect -> webgpu-errors-validation, webgpu-errors-debugging
└─ Performance defect -> webgpu-impl-performance, webgpu-impl-async-patterns

Core Patterns

  • ALWAYS cross-check every API name, descriptor field, enum value, and WGSL builtin against the WebGPU 1.0-stable surface. Anything not in the spec is a hallucinated API and a hard failure. references/methods.md lists the legal surface for cross-checking.
  • ALWAYS treat a missing device.lost handler as a defect, not a style issue. Without it the app silently produces a dead device after a GPU process crash.
  • ALWAYS report the FIRST defect in the contagious chain. WebGPU errors propagate: an invalid descriptor poisons every dependent object, so later messages are downstream noise. Fix the root cause.
  • NEVER pass code with a positional mismatch between WGSL @binding/@group and the GPUBindGroupLayout entries. WebGPU binds by index, not by name; a silent off-by-one reads the wrong resource.
  • NEVER accept a uniform struct whose host-side Float32Array offsets ignore the WGSL vec3 16-byte alignment. A vec3f member lands at the next 16-byte offset.
  • ALWAYS verify alignment numbers literally: size mult-4, mapAsync offset mult-8, dynamic offset mult-256, bytesPerRow mult-256. These are the largest source of WebGPU validation errors.

Common Anti-Patterns

Top 3 inline. The full 18-entry catalog with detection cue and fix is in references/anti-patterns.md.

  • Hallucinated or invented API. Detection cue: an API name, descriptor field, enum value, or WGSL builtin not in the 1.0-stable spec (for example device.createBufferSync, GPUBuffer.update, passEncoder.bindPipeline, textureSampleArray). Fix: replace with the real API or remove the call; cross-check against references/methods.md.
  • Missing device-loss handling. Detection cue: a WebGPU app with no device.lost.then(...) attached before resource creation. Fix: route to webgpu-errors-device-loss for the recovery pattern.
  • Alignment error. Detection cue: a buffer size, writeBuffer offset, dynamic bind-group offset, or bytesPerRow that is not the required multiple (4 / 8 / 256). Fix: route to webgpu-core-memory-model.

Critical Warnings

  • NEVER approve WebGPU code that calls an API absent from the WebGPU 1.0-stable spec.
  • NEVER approve a WebGPU app that omits the adapter null-check or the device.lost handler.
  • NEVER approve a buffer whose usage flags are an illegal combination, for example MAP_READ with STORAGE or UNIFORM.
  • NEVER approve a uniform struct whose host offsets ignore the WGSL 16-byte alignment rules.
  • NEVER approve a @compute entry point without @workgroup_size, or a switch without default:; both are WGSL shader-creation errors.
  • NEVER report a downstream contagious error as the root cause; trace back to the first invalid object.

Reference Files

  • references/methods.md: the full category-by-category review checklist, the WebGPU 1.0-stable API surface for hallucination cross-check, and the issue-to-skill routing map.
  • references/examples.md: before/after pairs of flawed WebGPU and WGSL code with the corrected version.
  • references/anti-patterns.md: the consolidated 18-entry WebGPU anti-pattern catalog, each with a detection cue and a fix.