Bug 54582 - [regression] 33% performance regression on GLBenchmark2.1-Egypt from ETC1 support
Summary: [regression] 33% performance regression on GLBenchmark2.1-Egypt from ETC1 sup...
Status: CLOSED FIXED
Alias: None
Product: Mesa
Classification: Unclassified
Component: Drivers/DRI/i965 (show other bugs)
Version: 9.0
Hardware: x86 (IA32) Linux (All)
: high major
Assignee: Eric Anholt
QA Contact:
URL:
Whiteboard:
Keywords:
Depends on:
Blocks:
 
Reported: 2012-09-06 06:17 UTC by libo
Modified: 2013-07-04 00:24 UTC (History)
3 users (show)

See Also:
i915 platform:
i915 features:

Attachments

Description libo 2012-09-06 06:17:27 UTC 
System Environment:
--------------------------
Libdrm:	(master)libdrm-2.4.39-2-g92fd0ce4f659d7b0680543e9e5b96a3c7737a5f3
Mesa:	(9.0)284fe97515998178962b5abb4ed0b828b3497c77
Xserver:(server-1.12-branch)xorg-server-1.12.4
Cairo:	(master)c0b1b178184ccfba0fe60c2a49901cadef9df36b
Libva:	(master)44b03ffcdb814392a9acf025615f99a06fed63d4
Kernel:	(drm-intel-fixes) 28dcc2d60cb570d9f549c329b2f51400553412a1

Bug detailed description:
---------------------------
 There are 20.22% regression on GLBench2.1-Egypt caused by Mesa. 
 Bad commit : 284fe97515998178962b5abb4ed0b828b3497c77   52.26 fps      
 Good commit: 59d247c829cce0fe14957ff32221b329fde6a918   78.20 fps

 I have bisected it and result as follows:
 The first bad commit could be any of:
8dc074cd92dfc2f80f6c69f80f0d169ad5bb10aa     (bad commit)
a5a34b153d494ad4374e7b2c8ea13b1073a887e2     (build failed)
 
Reproduce steps:
----------------------------
1.xinit&
2.gnome-session
3.vblank_mode=0 ./GLBenchmark2  -data ../../data/  -w 1920 -h 1080 -t 2000
Comment 1 Gordon Jin 2012-09-12 03:00:01 UTC 
Ian/Ken, can you reproduce this?
Comment 2 Eric Anholt 2012-10-16 00:29:48 UTC 
Those two commits are right next to each other, and are the enabling of ETC1 support on non-ETC1-supporting hw:

commit a5a34b153d494ad4374e7b2c8ea13b1073a887e2
Author: Chad Versace <chad.versace@linux.intel.com>
Date:   Mon Jun 18 18:25:25 2012 -0700

    intel: Enable GL_OES_compressed_ETC1_RGB8_texture
Comment 3 Ian Romanick 2012-10-16 07:02:29 UTC 
We couldn't reproduce this.  We got identical performance in the ETC1 path and the S3TC path.
Comment 4 Gordon Jin 2012-12-06 03:49:57 UTC 
Ian, any idea to diagnose this issue?
Comment 5 Eric Anholt 2012-12-18 21:36:40 UTC 
Patches on the list.  Apparently they were testing with a GLB2.5 build using the GLB2.1 data, which makes different choices for texture formats (pretty much invalidating the point of having compatibility with the old datasets).
Comment 6 meng 2012-12-19 03:17:50 UTC 
(In reply to comment #5)

With the path "i965: Fix validation of ETC miptrees", GLBenchmark2.1 Egypt performance improve ~100%(70fps → 148fps).

Environment:
----------------------
Platform:IVB
mesa:(master)dc613f11ddf1f3a6e10c2d99830fb1a84fdb55b2
kernel:(drm-intel-fixes)b45305fce5bb1abec263fcff9d81ebecd6306ede
case:GLBenchmark2.1 GLB2_EgyptTestStandard(2000)
Comment 7 Eric Anholt 2012-12-23 00:18:14 UTC 
commit 46386816a77c6e5751a0685776c0a6320f46f8fe
Author: Eric Anholt <eric@anholt.net>
Date:   Tue Dec 18 12:03:49 2012 -0800

    i965: Fix validation of ETC miptrees.
Comment 8 libo 2012-12-24 06:03:08 UTC 
Please merge it into 9.0 branch because the issue still exists on 9.0 branch.
(In reply to comment #7)
> commit 46386816a77c6e5751a0685776c0a6320f46f8fe
> Author: Eric Anholt <eric@anholt.net>
> Date:   Tue Dec 18 12:03:49 2012 -0800
> 
>     i965: Fix validation of ETC miptrees.
Comment 9 Gordon Jin 2013-04-02 05:55:31 UTC 
Eric, do you have plan to cherry-pick?
Comment 10 Gordon Jin 2013-07-02 05:59:01 UTC 
Eric/Ian, please decide if you want to cherry-pick to 9.0.
Comment 11 Ian Romanick 2013-07-03 20:28:17 UTC 
I went ahead and picked it to 9.0 branch (ba73bea), but I doubt anyone will care.  To the best of my knowledge, no distro is still shipping 9.0.
Comment 12 Gordon Jin 2013-07-04 00:24:24 UTC 
I don't care about 9.0 either. I just wanted to close this one.

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.