982338 - (osx-tiling) Use tiled thebes layers on OSX

Status: RESOLVED FIXED

(Core :: Graphics: Layers, defect)

Description

[Matt Woodrow (:mattwoodrow)]

OSX content is one of (if not the) last place(s) we are using deprecated textures.

We should switch to using tiling assuming we can get it to perform as well, so we can delete a whole bunch of code.

https://tbpl.mozilla.org/?tree=Try&rev=26fed6aa3b9a

Comment 1

[Andreas Gal :gal]

\o/

Comment 2

[Matt Woodrow (:mattwoodrow)]

Attachment: Enable tiling for OSX 10.8

http://compare-talos.mattn.ca/?oldRevs=d697d649c765,8da1906596d6,1138e8f07c48,47c9418fbc28,2a15dc07ddaa&newRev=c35a86bbdf70&server=graphs.mozilla.org&submit=true

Comment 3

[Bas Schouten (:bas.schouten)]

Comment on attachment 8481196 [details] [diff] [review]
Enable tiling for OSX 10.8

Review of attachment 8481196 [details] [diff] [review]:
-----------------------------------------------------------------

Please add a pref to force tiling on 10.6.

Comment 4

[Matt Woodrow (:mattwoodrow)]

Attachment: Enable tiling for OSX 10.8 v2

Comment 5

[Matt Woodrow (:mattwoodrow)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/7de1b4a42f1c

Comment 6

[Matt Woodrow (:mattwoodrow)]

Adding Avih and Joel to the CC list since I expect some talos changes with this push.

Looking at compare talos it looks like we have a TART and tp5_scroll win, and a tsvg_opacity and CART regression. The improvements are both bigger than the regressions.

Tscrollx is also showing as a slight regression, but graphserver ignores the slowest test (of only 6!), which is the one that improved significantly. With a true average of the six we'd be winning here too.

I think this is a pretty reasonable trade-off especially since getting tiling is important to the graphics team in terms of cleaning up old code and making it easier for us to progress to fully async scrolling.

Comment 7

[Carsten Book [:Tomcat]]

sorry had to backout this change for crashes like https://tbpl.mozilla.org/php/getParsedLog.php?id=47156878&tree=Mozilla-Inbound on OS X due the backout of other changes

Comment 8

[Matt Woodrow (:mattwoodrow)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/4df1bd30364d

Comment 9

[Phil Ringnalda (:philor)]

And out again in https://hg.mozilla.org/integration/mozilla-inbound/rev/21b1152c66cc for build bustage.

Comment 10

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

it appears this patch gives us a 9.48% win on 10.6 for svgx, we will see when this sticks if we still retain that win :)

Comment 11

[Matt Woodrow (:mattwoodrow)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/4effa50d8934

Comment 12

[Matt Woodrow (:mattwoodrow)]

Landing this for 10.6 as well, since the performance difference ended up not being much.

Looks like we're expecting a big for svgx, tscrollx (though won't show up in graphserver results again), CART and Tresize. tp5_scroll, svg_opacity and tp5_rss_main look to be worse.

Comment 13

[Matt Woodrow (:mattwoodrow)]

And disabled again for 10.6 since it's having intermittent reftest failures that look like some sort of driver bug.

 https://hg.mozilla.org/integration/mozilla-inbound/rev/ab5cfc402f77

Comment 14

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/4effa50d8934
https://hg.mozilla.org/mozilla-central/rev/ab5cfc402f77

Comment 15

[Milan Sreckovic [:milan] (needinfo for best results)]

Comment on attachment 8482031 [details] [diff] [review]
Enable tiling for OSX 10.8 v2

Review of attachment 8482031 [details] [diff] [review]:
-----------------------------------------------------------------

Should we stop using gfxPrefs::LayersTilesEnabled() everywhere, and instead use gfxPlatform::UseTiling() or risk having some of the code think we're in tiling, and some that we're not?  There are a couple of more places that still use gfxPrefs::LayersTilesEnabled() directly.

Comment 16

[Milan Sreckovic [:milan] (needinfo for best results)]

Matt, question for you in comment 15.

Comment 17

[Matt Woodrow (:mattwoodrow)]

(In reply to Milan Sreckovic [:milan] from comment #15)
> Comment on attachment 8482031 [details] [diff] [review]
> Enable tiling for OSX 10.8 v2
> 
> Review of attachment 8482031 [details] [diff] [review]:
> -----------------------------------------------------------------
> 
> Should we stop using gfxPrefs::LayersTilesEnabled() everywhere, and instead
> use gfxPlatform::UseTiling() or risk having some of the code think we're in
> tiling, and some that we're not?  There are a couple of more places that
> still use gfxPrefs::LayersTilesEnabled() directly.

Probably, yes. I don't think any of the other callers actually affect desktop so we're ok for now, but it would be nice to clean up.

Comment 18

[Jonathan Watt [:jwatt]]

This caused a huge regression in bug 982338.

Comment 19

[Jonathan Watt [:jwatt]]

And bug 1077842.

Comment 20

[Jonathan Watt [:jwatt]]

And bug 554004.