1837027 - Need Mesa Lavapipe 22.1.2 or later to run WebGPU CTS on Linux

Status: RESOLVED FIXED

(Testing :: General, task, P2)

Description

[Jim Blandy :jimb]

We need to run the WebGPU conformance test suite in CI on Linux for Mozilla Central. This requires a conformant Vulkan implementation. Linux Mesa lavapipe 22.1.2 was certified as a conformant implementation in 2022-07, so that should be good enough. Lavapipe is a CPU-based renderer, so it does not require a machine with a GPU. It seems that Ubuntu 22.04 LTS only has Mesa 22.0.1, which is not good enough, but Jammy Updates has Mesa 22.2.5.

Comment 1

[Joel Maher ( :jmaher ) (UTC -8)]

Currently we are building an ubuntu 22.04 VM running Wayland (not X11). Will WebGPU/Mesa work on that, or does it need X11?

From my little digging on Google searches, it seems that it doesn't matter if Ubuntu is running X11 or Wayland.

:jimb, can you confirm?

Comment 2

[Erich Gubler [:ErichDonGubler]]

:jmaher: Answering for Jim here, we believe that this should work. The best way to figure this out is going to be with testing, though. Is there a way to access an interactive environment with the current WIP of the Ubuntu 22.04 environment?

Comment 3

[Joel Maher ( :jmaher ) (UTC -8)]

I am looking into options with the current Ubuntu 22.04 Wayland- I just got back from a long PTO, so I am not sure who is on PTO, etc. - I assume later this week or first thing next week I can have an answer.

Comment 4

[Erich Gubler [:ErichDonGubler]]

Mostly for :jimb: I found this upstream issue in wgpu noting that lavapipe seems to stall on certain hardware profiles: https://github.com/gfx-rs/wgpu/issues/1974. It's hard to tell if this might affect us until we know more about the hardware profile of CI runners, though.

Comment 5

[Andrew Erickson [:aerickson]]

The current Wayland 22.04 image has version 22.2.5-0ubuntu0.1~22.04.1 of mesa-vulkan-drivers. There is a minor update available (details in Bug 1841867).

Have we tried running the tests on the workers that use the image (https://firefox-ci-tc.services.mozilla.com/worker-manager/gecko-t%2Ft-linux-vm-2204-wayland)?

Comment 6

[Erich Gubler [:ErichDonGubler]]

:aerickson: I can answer that. Short answer: nope! Do you have some instructions to help us test consuming this image? I'm unfamiliar with plumbing specific images to Taskcluster jobs, but I'd likely be the one to take it on. 😅

Comment 7

[Joel Maher ( :jmaher ) (UTC -8)]

./mach try fuzzy -q 'test linux webgpu' --worker-override="t-linux-large-gcp=gecko-t/t-linux-vm-2204-wayland"

Comment 8

[Erich Gubler [:ErichDonGubler]]

Just tried before and after Try builds. Both are broken. 🙁 Looks like the worker type is different (t-linux-large-gcp vs t-linux-xlarge-gcp) and the image didn't get picked up in these test runs, so I'll try pushing again with an additional --worker-override="t-linux-xlarge-gcp=gecko-t/t-linux-vm-2204-wayland".

EDIT: That didn't work either. 😩 Gonna reach out to Andrew and Joel directly, see if I can get some instruction on how to get this consumed properly for a test run.

Comment 9

[Joel Maher ( :jmaher ) (UTC -8)]

the override isn't working, is there going to be a new push with the new --worker-override... ?

Comment 10

[Erich Gubler [:ErichDonGubler]]

:jmaher: I'm not sure who you're addressing (does support for --worker-override=… need to be added still?), but attempting to override xlarge workers on Linux didn't work either.

Comment 11

[Andrew Erickson [:aerickson]]

I think you want just t-linux-large and t-linux-xlarge like:

--worker-override="t-linux-xlarge=gecko-t/t-linux-vm-2204-wayland"

See https://mozilla-hub.atlassian.net/wiki/spaces/ROPS/pages/318996714/Using+mach+try+fuzzy+--worker-override - it's confusing. I couldn't find an alias t-linux-xlarge-gcp in taskcluster/ci/config.yml.

Comment 12

[Erich Gubler [:ErichDonGubler]]

:aerickson: Aha, that seems to change something, but now all non-build jobs are ending in exception with no logs. 😩

Comment 13

[Andrew Erickson [:aerickson]]

The original workers are docker-workers (that use a different payoload). The new workers are generic-workers and the worker-override option doesn't seem to change the payload to match what worker expects. I think some taskgraph hacking is required to get the jobs to run on the instances.

Comment 14

[Erich Gubler [:ErichDonGubler]]

:aerickson and :jmaher have been extremely helpful in actually getting a working Try build that exercises this environment, and it seems that we are successfully running most WebGPU tests. I believe that the spirit of this ticket has been resolved, and that we just need to adjust test expectation metadata with the WIP patch stack I have against bug 1836805.

Thanks for your help, everyone! 👏🏻🙏🏻