1864698 - Remove MacOS from WebGPU blocklist

Status: RESOLVED FIXED

(Core :: Graphics: WebGPU, defect, P1)

Description

[Erich Gubler [:ErichDonGubler]]

From WebGPU team chat on 2023-11-14:

ErichDonGubler: [In an experiment I did recently (try:a0e6d9c553c1),] MacOS tests ran chunked tests in a single chunk in <50m … . That (a) doesn't sound like we're doing much there, eek things broken, and (b) makes our current chunk configuration feel really wasteful. Observe that webgpu9 runs all webgpu/cts/chunked tests, and webgpu10 runs all webgpu/cts/webgpu tests.

:jimb: Yeah, and every test seems to take a consistent ~600ms. Very suspicious.

:ErichDonGubler: I'd wager that's how long it takes each test to get to trying to use navigator.gpu, or something close and early in the common API sequence to do things? 😅

:jimb: Yeah, probably.
:jimb: I'll bet they're all hitting the blocklist

Comment 1

[Jim Blandy :jimb]

If it is indeed the blocklist, then we can employ the gfx.webgpu.ignore-blocklist pref.

Comment 2

[Brad Werth [:bradwerth]]

(In reply to Jim Blandy :jimb from comment #1)

If it is indeed the blocklist, then we can employ the gfx.webgpu.ignore-blocklist pref.

We've already set this pref for webgpu mochitests.

Comment 3

[Erich Gubler [:ErichDonGubler]]

Attachment: Bug 1864698: test(webgpu): set `gfx.webgpu.ignore-blocklist` to `true` for CTS testing on Mac r=#webgpu-reviewers!

Comment 4

[Erich Gubler [:ErichDonGubler]]

Attachment: Bug 1864698: feat(webgpu): make `gfx.webgpu.ignore-blocklist` pref. `mirror: always` r=#webgpu-reviewers!

Comment 5

[Erich Gubler [:ErichDonGubler]]

Attachment: Bug 1864698: test(webgpu): merge exps. for observed CTS intermittents r=#webgpu-reviewers!

Comment 6

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/42778d7f458c
test(webgpu): merge exps. for observed CTS intermittents r=webgpu-reviewers,nical DONTBUILD

Comment 7

[Erich Gubler [:ErichDonGubler]]

Status update: quoting this comment in D193612:

Hmmm…the latest Try push I've submitted with D193682 included (try:440344eacf9b) has not worked as I'd hoped. The decision to initialize WebGPU does not seem like something we can easily live-update with current code. Trying to change this seems to entail a bunch of design decisions that are a distraction from the core of this problem, which is that we expect hardware that supports MacOS 10.15 to be compatible with WebGPU, and not blocklisted[^1].

I'd prefer to explore other options like fixing our blocklist before going further down this hacky path. Abandoning for now.

---

[^1]: Examples: How do we transition currently open tabs when this preference changes?

:jimb, I think we simply need to focus on fixing the blocklist, but this is not something I have experience with. NI'ing you for help here, since I think it shouldn't be a huge lift.

Will un-assign this bug from myself once the test expectations patch I landed in autoland makes it to central.

Comment 8

[Natalia Csoregi [:nataliaCs]]

https://hg.mozilla.org/mozilla-central/rev/42778d7f458c

Comment 9

[Jim Blandy :jimb]

I don't know much about the blocklist code either, unfortunately. Happy to pair on this if you like.

Comment 10

[Erich Gubler [:ErichDonGubler]]

Currently blocked on gfx-rs/wgpu#2432. Without it, we shouldn't consider exposing any users to the unsafety with vertex attribute data. We have a blocklist entry that we want corresponding to it at widget/cocoa/GfxInfo.mm:478-482:

    // wgpu doesn't safely support OOB behavior on Metal yet.
    IMPLEMENT_MAC_DRIVER_BLOCKLIST(
        OperatingSystem::OSX, DeviceFamily::All, nsIGfxInfo::FEATURE_WEBGPU,
        nsIGfxInfo::FEATURE_BLOCKED_DEVICE,
        "FEATURE_FAILURE_MAC_WGPU_NO_METAL_BOUNDS_CHECKS");

Comment 11

[Erich Gubler [:ErichDonGubler]]

Attachment: WIP: Bug 1864698: test(webgpu): prep. refac. for second approx. of CTS test splitting r=#webgpu-reviewers!

Comment 12

[Erich Gubler [:ErichDonGubler]]

Attachment: WIP: Bug 1864698: test(webgpu): make CTS case splitting logic determine WPT timeout len. r=#webgpu-reviewers!

Depends on D198120

Comment 13

[Erich Gubler [:ErichDonGubler]]

Attachment: WIP: Bug 1864698: test(webgpu): make `…:device_lost:destroy` CTS cases use short WPT timeout r=#webgpu-reviewers!

Depends on D198121

Comment 14

[Erich Gubler [:ErichDonGubler]]

Attachment: WIP: Bug 1864698: test(webgpu): adjust vendored CTS file perms. to match upstream for Unix platforms r=#webgpu-reviewers!

Depends on D198122

Comment 15

[Erich Gubler [:ErichDonGubler]]

D'oh, I erroneously filed the above patches against this bug, instead of bug 1873696. Sorry for the noise!

Comment 16

[Phabricator Automation]

Comment on attachment 9371923 [details]
WIP: Bug 1864698: test(webgpu): prep. refac. for second approx. of CTS test splitting r=#webgpu-reviewers!

Revision D198120 was moved to bug 1873696. Setting attachment 9371923 [details] to obsolete.

Comment 17

[Phabricator Automation]

Comment on attachment 9371924 [details]
WIP: Bug 1864698: test(webgpu): make CTS case splitting logic determine WPT timeout len. r=#webgpu-reviewers!

Revision D198121 was moved to bug 1873696. Setting attachment 9371924 [details] to obsolete.

Comment 18

[Phabricator Automation]

Comment on attachment 9371925 [details]
WIP: Bug 1864698: test(webgpu): make …:device_lost:destroy CTS cases use short WPT timeout r=#webgpu-reviewers!

Revision D198122 was moved to bug 1873696. Setting attachment 9371925 [details] to obsolete.

Comment 19

[Phabricator Automation]

Comment on attachment 9371926 [details]
WIP: Bug 1864698: test(webgpu): adjust vendored CTS file perms. to match upstream for Unix platforms r=#webgpu-reviewers!

Revision D198123 was moved to bug 1873696. Setting attachment 9371926 [details] to obsolete.

Comment 20

[Erich Gubler [:ErichDonGubler]]

Attachment: Bug 1864698: feat(webgpu): re-enable MacOS support r=#webgpu-reviewers!

TODO: Validate that this is actually safe to do, per the comment being
removed. @ErichDonGubler has asked here:
https://matrix.to/#/!AQDMTeppDxxlzBSBzM:mozilla.org/$JO9ErjRjlLAGKGHAZOJz8K0Md3fB1_63z3Yhd5pLQqo?via=mozilla.org

Comment 21

[Bob Hood [:bhood]]

WebGPU is unreleased tech. Dropping Severity to S3 to remove from org tracking.

Comment 22

[Isaac Marovitz]

Since https://github.com/gfx-rs/wgpu/issues/2432 is now resolved and the changes pulled into Firefox can macOS be un-blacklisted?

Comment 23

[Jim Blandy :jimb]

I just verified that we are requesting Naga-injected bounds checking on other indexing operations, so yeah, indexed draw calls were the last step. I'm not aware of any reason to keep this blocklisted in nightly.

Comment 24

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/bb79dcf0336e
feat(webgpu): re-enable MacOS support r=webgpu-reviewers,teoxoy

Comment 25

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/e5b11a4833d1
Check properly for timeout multipler in mozharness. CLOSED TREE

Comment 26

[pstanciu]

https://hg.mozilla.org/mozilla-central/rev/bb79dcf0336e
https://hg.mozilla.org/mozilla-central/rev/e5b11a4833d1

Comment 27

[Erich Gubler [:ErichDonGubler]]

I'd forgotten to remove the CI configuration that worked around the thing this bug fixes. Catching up with that in bug 1909265.