629568 - scrolling content with a plugin is slow (on Ubuntu w/ NVIDIA & layers acceleration force-enabled)

Status: NEW

Description

[Daniel Holbert [:dholbert]]

On my Linux desktop, my soon-to-be-attached testcase scrolls *much* slower when I force-enable layers acceleration.

STEPS TO REPRODUCE:
 1. Add the following bookmarklet to bookmarks toolbar
   (minor tweak on roc's from bug 579260)...

javascript:void((function(){var%20i=0;var%20start=Date.now();function%20f(){if(++i==50){alert((Date.now()-start)+"ms");return;}document.getElementsByTagName('iframe')[0].contentWindow.scrollTo(0,i*5);setTimeout(f,10);}f();})())

 2. Open testcase; wait for it to start autoplaying.
 3. Click bookmarklet. (let it scroll & alert() its results)

With layers disabled, I get results around 500ms-950ms
With layers enabled, I get results around 55000ms
(That's a 50x - 100x increase.)

Some of this may be explained by X.org inefficiencies and/or silly GL drivers, but I'm hoping there's something we can do on our side to ameliorate this...

Comment 1

[Daniel Holbert [:dholbert]]

Created attachment 507732 [details]
(helper file -- iframe contents)

Comment 2

[Daniel Holbert [:dholbert]]

Created attachment 507733 [details]
testcase 1

Comment 3

[Daniel Holbert [:dholbert]]

Mozilla/5.0 (X11; Linux x86_64; rv:2.0b11pre) Gecko/20110127 Firefox/4.0b11pre
OS: Ubuntu 10.10 x86_64 (w/ compiz enabled, as it is by default)
Graphics card: GeForce 9800 GT
NVIDIA proprietary driver version (from nvidia-settings): 260.19.06

Comment 4

[Daniel Holbert [:dholbert]]

FWIW: on my laptop (with a setup as described in comment 3 except that the card is an NVIDIA NVS 3100M), I immediately hit the crash from bug 629265 upon loading the testcase.  No crash on my desktop, though - just slooow scrolling.

Comment 5

[Daniel Holbert [:dholbert]]

On my mac mini running OS X 10.6 (with layers accel defaulted to being on in nightlies), there doesn't seem to be a significant difference for this bug's testcase, for layers acceleration being enabled vs. disabled -- both seem to yield results ranging between 1600ms-2800ms.

So this bug's testcase appears to hit a Linux-specific issue.

Comment 6

[Constantine]

For me it is the same. But when I opened Bug 655017 it was ~4x faster.