1101635 - Enable tiling on 10.6

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Markus Stange [:mstange]]

When tiling for OS X landed in bug 982338, there were some very strange reftest failures on 10.6, so tiling was disabled on 10.6 in bug 982338 comment 13.

There also were a few minor performance regressions, but according to bug 982338 comment 12 "the performance difference ended up not being much".

Yesterday I started a try push to see which reftests are failing:
https://treeherder.mozilla.org/ui/#/jobs?repo=try&revision=1d6dfd6fe0c9

The result was that the intermittent garbage is gone, the only remaining failures are fuzzable gradient differences (compare bug 1059622).

I've also started two try pushes so that we can look at the performance difference again:
Baseline: https://treeherder.mozilla.org/ui/#/jobs?repo=try&revision=90ddac8c04a0
Enable: https://treeherder.mozilla.org/ui/#/jobs?repo=try&revision=513fe0805cb9

Comment 1

[Markus Stange [:mstange]]

Attachment: adjust fuzzing

Comment 2

[Markus Stange [:mstange]]

Attachment: enable tiling

This just backs out https://hg.mozilla.org/integration/mozilla-inbound/rev/ab5cfc402f77 and http://hg.mozilla.org/mozilla-central/rev/862c00a40fb1 .

Comment 3

[Markus Stange [:mstange]]

Here's the Talos comparison:
http://compare-talos.mattn.ca/?oldRevs=90ddac8c04a0&newRev=513fe0805cb9&server=graphs.mozilla.org&submit=true

The only really worrying change is the tp5o_scroll regression by 16.95%.

tcanvasmark and tsvgx get faster.

Comment 4

[Jeff Muizelaar [:jrmuizel]]

We should probably look at this a little bit.

Comment 5

[Markus Stange [:mstange]]

New pushes with more up-to-date m-c base revision:
Baseline:    https://treeherder.mozilla.org/ui/#/jobs?repo=try&revision=796baeda6da8
With tiling: https://treeherder.mozilla.org/ui/#/jobs?repo=try&revision=e1de81e48b5e

Comment 6

[Markus Stange [:mstange]]

The tp5o_scroll regression is the same with the up-to-date m-c base.
http://compare-talos.mattn.ca/?oldRevs=90ddac8c04a0&newRev=513fe0805cb9&server=graphs.mozilla.org&submit=true

Comment 7

[Markus Stange [:mstange]]

Er, that's because I was looking at the wrong comparison. But it's still similarly bad:
http://compare-talos.mattn.ca/?oldRevs=796baeda6da8&newRev=e1de81e48b5e&server=graphs.mozilla.org&submit=true

Comment 8

[Markus Stange [:mstange]]

The tp5o_scroll regression was exaggerated because tp5o_scroll was measuring the wrong thing (see bug 1112339).
Here's a new comparison with a fixed tp5o_scroll test:
http://compare-talos.mattn.ca/?oldRevs=b897e85ee379&newRev=a962f05dca55&server=graphs.mozilla.org&submit=true

Comment 9

[Markus Stange [:mstange]]

I've started another try run for reftests:
https://treeherder.mozilla.org/ui/#/jobs?repo=try&revision=182a78da47a1

Comment 10

[Markus Stange [:mstange]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/c49b6abe67e5
https://hg.mozilla.org/integration/mozilla-inbound/rev/e7b95eb855a2

Comment 11

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/c49b6abe67e5
https://hg.mozilla.org/mozilla-central/rev/e7b95eb855a2