Bug 66693 – [chromium] Update builder names for chromium gpu bots due to CG switchover

• Status: RESOLVED
• Resolution: FIXED
• Component: New Bugs

[James Robinson]

[chromium] Update builder names for chromium gpu bots due to CG switchover

[James Robinson]

Created attachment 104707 [details]
Patch

[Adam Barth]

Comment on attachment 104707 [details]
Patch

Why do we have chromium-gpu-cg bots?  That wasn't part of the CG migration plan.

[Dirk Pranke]

(In reply to comment #2)
> (From update of attachment 104707 [details])
> Why do we have chromium-gpu-cg bots?  That wasn't part of the CG migration plan.

The GPU tests are a separate step on the existing bots, not separate bots, and the "bot name" is formed by appending "- GPU" to the existing bot name. So, by renaming the bots and not renaming the GPU components, we stopped being able to rebaseline things.

This patch just brings everything back to consistency.

[James Robinson]

(In reply to comment #2)
> (From update of attachment 104707 [details])
> Why do we have chromium-gpu-cg bots?  That wasn't part of the CG migration plan.

We'll need to handle gpu mac baselines the same way as non-gpu mac baselines, so they definitely need to be part of the plan.

[James Robinson]

Committed r93530: <http://trac.webkit.org/changeset/93530>

[Adam Barth]

> We'll need to handle gpu mac baselines the same way as non-gpu mac baselines, so they definitely need to be part of the plan.

Is there something wrong with using the name chromium-gpu-mac-leopard rather than chromium-gpu-cg-mac-leopard?  Maybe we should update the plan?  The GPU bots seems hacked into our tools in undesirable ways.

[James Robinson]

Where do you put the pixel baselines for tests that render differently with skia+gpu than they do for cg+gpu?  I'm not really aware that we had a plan for chromium-mac-gpu, it seems like nobody thought about it when making these changes because everything broke.

[Adam Barth]

The plan doesn't ever call for having both GPU Skia and GPU CG bots running at the same time, so that problem will not occur according to the plan.

[Dirk Pranke]

(In reply to comment #8)
> The plan doesn't ever call for having both GPU Skia and GPU CG bots running at the same time, so that problem will not occur according to the plan.

In comment #3, I wrote that "The GPU tests are a separate step on the existing bots, not separate bots". As soon as you have both Skia and CG bots, you will have Skia CG GPU bots. 

It seems someone needs to update the plan.

[James Robinson]

I updated the plan document to call out adding the webkit_gpu_tests step to the skia bots as an explicit item.

[Adam Barth]

Is there some reason we represent the webkit_gpu_tests step as a separate bot in the code?  That seems like a giant hack that causes more problems then it's worth.

[James Robinson]

Actually it looks like Elliot already took care of this for the new skia mac bot, so there's no new work to do.

[Dirk Pranke]

(In reply to comment #11)
> Is there some reason we represent the webkit_gpu_tests step as a separate bot in the code?

(adding Ojan to double-check my memory here)

We did this because the TestResultsServer (aka Flakiness Dashboard) only stores results by bot name; adding an additional dimension would have required us to add a secondary key (in one form or another) in order to be able to look up things.

Arguably we could have done this by posting the bot name plus the cpu/gpu flag to the TestResultsServer, and having it do the strcat; I'm not sure if we would've run into other complications once we started down that path.

[Dirk Pranke]

I do agree that it's a hack, though.

[Adam Barth]

I guess I never fully understood why webkit_gpu_tests needs to be a separate step at all.  It's the same binary, right?  Can we just run those tests in the same step as the other tests?  That would eliminate all this special-case logic for GPU.  (Maybe we should take this discussion to a mailing list rather than a bug.)

[James Robinson]

It's the same binary but different flags passed to DRT.  One motivation for having it this way is that we want to run the canvas tests through both the HW and SW paths, since users may encounter either.  I'm not sure how to say "run this set of tests with a different set of flags than last time" without making it a different step or adding something truly bizarre to NRWT.

[Adam Barth]

I posted some of these questions to webkit-dev just now.  Sorry if that causes you to repeat some of your answers.  It's probably better to continue the discussion there.