Bug 23084

Summary:	[G43] OpenGL 5-10% performance regression with QGears2 test
Product:	xorg	Reporter:	Bryce Harrington <bryce>
Component:	Driver/intel	Assignee:	Eric Anholt <eric>
Status:	RESOLVED INVALID	QA Contact:	Xorg Project Team <xorg-team>
Severity:	major
Priority:	medium	CC:	Michael
Version:	7.4 (2008.09)	Keywords:	NEEDINFO
Hardware:	All
OS:	Linux (All)
Whiteboard:
i915 platform:		i915 features:

Description Bryce Harrington 2009-08-01 15:23:05 UTC

Forwarding this issue reported in Phoronix by Michael Larabel:
http://www.phoronix.com/scan.php?page=article&item=intel_q309_flakes&num=3

[Problem]
The OpenGL-backed QGears2 rendering test Michael ran on his hardware is showing a very large regression (over 50% performance regression).

Several other of the QGears2 tests show slight regressions but this one is the worst.  I've not heard many other complaints about OpenGL performance with the 2.8.0 driver, but not many people provide the quantitative test results that Michael does, so this would be further investigation.

Comment 1 Bryce Harrington 2009-08-01 15:24:05 UTC

Michael, when you subscribe to this bug, can you please also attach your Xorg.0.log with ModeDebug enabled, and the output of dmesg.  Maybe also glxinfo output.

Also, we have a newer mesa snapshot in the xorg-edgers PPA.  Would you mind re-running this test with that installed to see if the newer mesa code will improve this?

Comment 2 Michael Larabel 2009-08-01 18:37:59 UTC

I'll try to get the relevant logs shortly. I am in the process of relocating and most of my office is already packed up, including those systems, but will see if I can unpack the G43 one tomorrow and capture the verbose logs. Due to that, testing out the PPA probably won't until my test farm is completely back up in a week and a half.

Comment 3 Eric Anholt 2009-08-03 19:03:51 UTC

Is qgears2 -gl supposed to represent actual QT workloads?  I strongly suspect it doesn't, given that the QT GL backend appears to do no antialiasing at all (surely people don't tolerate that?), and even if people do, I don't think that stroking and filling large non-pixel-aligned paths is common like that.

qt probably needs a real performance tracing system, like cairo-perf-trace

So, my question is: is this really a high priority?

Comment 4 Gordon Jin 2009-08-07 00:31:47 UTC

Similar to bug#23083, I'm sharing my test result which replicates Michael's test configurations (G43, with x86-64 build). It shows 2D is improving but 3D does have regression (though not so much as Michael's).

                               Ubuntu 9.04    Ubuntu 9.10 Alpha3
QGears2 XRender - Gears:         24.96             40.63
QGears2 XRender - Image Scaling: 63.00             90.80 
QGears2 OpenGL - Gears:          97.10             61.84

Comment 5 Eric Anholt 2009-08-10 08:37:35 UTC

Since 9.04 doesn't sync to vblank, did you disable vblank syncing in your tests on 9.10?

Comment 6 Eric Anholt 2009-10-19 13:48:56 UTC

Feedback timeout, bad testcase, and the problem may well have been in the testing procedure.

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.