It looks like we're sending work to the GPU that takes >7 seconds. This causes a timeout and the driver gets reset.

It seems like it might have to do with our initial allocations.

Disabling picture caching seems to improve things substantially

It looks like others see this as well:
https://www.reddit.com/r/firefox/comments/g9u0r0/call_for_webrender_testers_on_intel_sandybridge/foy3sq9/

I'm able to load Firefox captures using wrench without issue, so there's definitely something weird going on.

(In reply to Jeff Muizelaar [:jrmuizel] from comment #0)

initial allocations.

Is this a regression from bug 1598380 comment 8? (bug 1631353 comment 11)

Could be, I'll try reverting it.

Further investigation has discovered that disabling dual source blending avoids the hangs.

We still have the mystery of why dual source blending seems to work in other situations (wrench/picture-caching disabled) without problem.

So we discovered some more yesterday. It looks like the combination of dual source blending and bug 1580063 caused this.
That change added scissored clears and our theory is that the scissored clear ends up going through the regular 3d path and if the previous draw used dual source blending that would cause a TDR.

I was able to get an apitrace recording of the problem and replaying it also causes TDR, I confirmed that it was a call to ClearView() that causes the TDR and that the blend state is set to use dual source blending prior to calling ClearView().

I tried resetting the blend state to the default before doing the ClearView() but that didn't seem to help. As Lee suggested, it could well be that the blend state doesn't end up set on the GPU unless you actually do a draw with it.

ANGLE has an alternative code path that avoids the use of ClearView(), forcing that path avoids the problem.

https://hg.mozilla.org/mozilla-central/rev/302c30b598b0

https://hg.mozilla.org/mozilla-central/rev/754436d8b3a2

The patch landed in nightly and beta is affected.
:jrmuizel, is this bug important enough to require an uplift?
If not please set status_beta to wontfix.
If yes, don't forget to request an uplift for the patches in the regression caused by this fix.

For more information, please visit auto_nag documentation.

Nope. We're not turning on gen6 in beta