1848857 - (webgpu-commenc-valid-time) `GPUCommandEncoder` commands generate validation error on call, instead of deferring to `GPUCommandEncoder.finish`

Status: NEW

(Core :: Graphics: WebGPU, defect, P2)

Description

[Erich Gubler [:ErichDonGubler]]

According to the current draft of WebGPU v1 spec., Firefox's implementation generates validation errors at the wrong time for GPUCommandEncoder commands:

Currently, Firefox implementation emits validation errors when a command that invalidates the command encoder is issued. However, the current v1 spec. draft states that an invalid command should only invalidate the command buffer. In the draft, the following language is used for validation checks with each command:

If any of the following conditions are unsatisfied, make this invalid and stop. …

GPUCommandEncoder.finish()'s entry in the spec. states that validation error should only be generated once is called:

Device timeline finish steps:

1. Let validationSucceeded be true if all of the following requirements are met, and false otherwise.

   • this must be valid.
   • …

2. …

3. If validationSucceeded is false, then:

   1. Generate a validation error.

   2. Return a new invalid GPUCommandBuffer.

…

Comment 1

[Erich Gubler [:ErichDonGubler]]

Steps to reproduce

1. Apply (maybe WIP) patch(es) currently attached to bug 1848866.

2. With a fully built checkout of Firefox Nightly, run the following test command:

   ./mach wpt '/_mozilla/webgpu/chunked/16/cts.https.html?q=webgpu:api,validation,encoding,cmds,copyTextureToTexture:copy_ranges:*'
   

Comment 2

[Erich Gubler [:ErichDonGubler]]

:jgilbert has noted that this may not actually be desirable spec. behavior, and is lobbying to change it to specify an eager generation of a GPUValidationError (our current behavior) and invalidating the command encoder, instead of only invalidating: https://github.com/gpuweb/gpuweb/issues/4269

Comment 3

[BugBot [:suhaib / :marco/ :calixte]]

:ErichDonGubler, if you think that's a regression, could you try to find a regression range using for example mozregression?

Comment 4

[Erich Gubler [:ErichDonGubler]]

BugBot: AFAIK, this behavior has always been part of wgpu-core. I don't know when the spec. may or may not have embraced it, but I don't think it's useful to figure it out, TBH.

Comment 5

[Erich Gubler [:ErichDonGubler]]

Assigning to :jgilbert for now, since she'll be bringing this up the W3C spec. meeting tomorrow.

Comment 6

[Erich Gubler [:ErichDonGubler]]

Decision from today's WebGPU team sync: We'll be using this issue to become spec-compliant, regardless of the outcome of gpuweb#4269. If and when a change to the spec. happens because of our advocacy for a better API, we will create a separate follow-up issue for it.

Un-assigning from :jgilbert, since we have decoupled the original reason for assigning this bug to her from blocking this issue.