fix: ensure render pipelines have at least 1 target

[ErichDonGubler]

Connections

None of note.

Description

Important validation is performed on render pipeline configuration (like multisampling) against render targets. However, when no targets are specified, WGPU still compiles pipelines. It misses not just this validation¹, but also the case mandated by WebGPU spec., section 10.3.1. (Render Pipeline Creation), which states in the subsection titled validating GPURenderPipelineDescriptor(descriptor, layout, device):

There must exist at least one attachment, either:

• A non-null value in descriptor.fragment.targets, or

• A descriptor.depthStencil.

Even if the entire space of possible render targets would have rejected the pipeline's configuration, wgpu-core still asks backends to compile pipelines using potentially invalid configuration when no targets are specified. These backends can then misbehave in very different ways, which I've tested with erichdongubler-mozilla#82 for multisampling specifically; examples:

1. No issues are observed on MacOS CI. Perhaps because it supports the specific offending configurations I've tested?
2. DX12 rejects the graphics pipeline creation on API call, resulting in an internal error.
3. Vulkan on Linux simply segfaults. 😬

While there are valid use cases for executing render pipelines with no render targets (DX12 and Vulkan support this, for instance), the WebGPU standard currently does not enable them.

Testing

Tests have been added in this PR; see the changes in tests/ for more details. As described above, these tests were cherry-picked separately and tested in erichdongubler-mozilla#82; this PR definitely fixes them.

Checklist

• Run cargo fmt.
• Run cargo clippy. If applicable, add:
  • --target wasm32-unknown-unknown
  • --target wasm32-unknown-emscripten
• Run cargo xtask test to run tests.
• Add change to CHANGELOG.md. See simple instructions inside file.

Footnotes

1. While not germane to this bug, specifying a render target is the only way to get a usable render pipeline, so functioning applications are unlikely to do this in shipped code. ↩

[ErichDonGubler]

Oof, it appears this change has caught at least one place the CTS is doing something that seems non-standard in webgpu:api,operation,labels:wrappers_do_not_share_labels:* in a CI run I did this to test in Firefox:

• Source: https://github.com/gpuweb/cts/blob/007e1015875a0a91794edbc14feb3ad5fe60f90b/src/webgpu/api/operation/labels.spec.ts#L264-L270

• Sample Firefox CI output; since this disappears eventually, I'll inline the output below:

   [task 2024-05-18T17:04:21.442Z] 17:04:21     INFO - TEST-UNEXPECTED-FAIL | /_mozilla/webgpu/cts/webgpu/api/operation/labels/cts.https.html?q=webgpu:api,operation,labels:wrappers_do_not_share_labels:* | : - assert_unreached: 
   [task 2024-05-18T17:04:21.442Z] 17:04:21     INFO -   - EXCEPTION: Error: Unexpected validation error occurred: At least one color attachment or depth-stencil attachment was expected, but no render target for the pipeline was specified.
   [task 2024-05-18T17:04:21.442Z] 17:04:21     INFO -     TestFailedButDeviceReusable@https://web-platform.test:8443/_mozilla/webgpu/webgpu/util/device_pool.js:22:1
   [task 2024-05-18T17:04:21.442Z] 17:04:21     INFO -     attemptEndTestScope@https://web-platform.test:8443/_mozilla/webgpu/webgpu/util/device_pool.js:409:13
   [task 2024-05-18T17:04:21.442Z] 17:04:21     INFO -     
   [task 2024-05-18T17:04:21.442Z] 17:04:21     INFO -  Reached unreachable code
   [task 2024-05-18T17:04:21.442Z] 17:04:21     INFO - wpt_fn@https://web-platform.test:8443/_mozilla/webgpu/common/runtime/wpt.js:81:25
   [task 2024-05-18T17:04:21.444Z] 17:04:21     INFO - TEST-OK | /_mozilla/webgpu/cts/webgpu/api/operation/labels/cts.https.html?q=webgpu:api,operation,labels:wrappers_do_not_share_labels:* | took 1255ms
  

[ErichDonGubler]

CTS fix for webgpu:api,operation,labels:wrappers_do_not_share_labels:*: gpuweb/cts#3756

[sagudev]

I think there are more such occurrences in cts: https://github.com/sagudev/servo/actions/runs/9602065915/job/26483847356,

but on my repro even Edge https://sagudev.github.io/briefcase/WGPURenderPipeline does not trigger error, so I thought I interpreted spec wrong.

EDIT: Opened gpuweb/cts#3806