Bug 72023 - [HSW fbc] cairo performance reduced over 60% by enabling FBC
Summary: [HSW fbc] cairo performance reduced over 60% by enabling FBC
Status: CLOSED INVALID
Alias: None
Product: DRI
Classification: Unclassified
Component: DRM/Intel (show other bugs)
Version: unspecified
Hardware: All Linux (All)
: medium major
Assignee: Paulo Zanoni
QA Contact: Intel GFX Bugs mailing list
URL:
Whiteboard:
Keywords:
Depends on:
Blocks: 87704
  Show dependency treegraph
 
Reported: 2013-11-26 09:33 UTC by zhoujian
Modified: 2017-07-24 22:56 UTC (History)
8 users (show)

See Also:
i915 platform: HSW
i915 features: display/FBC


Attachments

Description zhoujian 2013-11-26 09:33:21 UTC
System Environment:       
----------------------------------------------------------------------
Platform: HSW (OS:Ubuntu)
Libdrm:	(master)libdrm-2.4.49
Mesa:(master)90d185544c897ca0b4c3c677fed4290330bb0706
Xf86_video_intel:(master)2.99.906-28-g12e9
Cairo:	(master)31eff5c6eb57ad379689748fd8c60a5ffe0ba481
Kernel:	(drm-intel-nightly) 164a4cb4c1431a0689f85507868356fae24da638

Bug detailed description:
cairo-trace/firefox-planet-gnome.trace performance reduce ~35% on HSW Ubuntu.
Especially, it works well on HSW Fedora. The problem also doesn’t exist on IVB(Fedora & Ubuntu).
It’s kernel –dinq regression. By bisected, show that the first bad commit:
commit cbaef0f173c7c8bb14976f3928e5876efec444e2
Author: Ville Syrjälä <ville.syrjala@linux.intel.com>
Date:   Wed Nov 6 23:02:24 2013 +0200
    drm/i915: Set has_fbc=true for all SNB+, except VLV

Performance 
--------------------------------------------
drm-intel-fixes :   1.75
drm-intel-next-queued:2.849
drm-next:1.756 
nightlytop:2.847

Reproduce steps:
--------------------------------------------------------------------
1. xinit&
2. vblank_mode=0 CAIRO_TEST_TARGET=xlib perf/cairo-perf-trace -i 6 /home/cairo-traces/benchmark/firefox-planet-gnome.trace
Comment 1 Chris Wilson 2013-11-26 09:43:24 UTC
To confirm, can you boot with i915.i915_enable_fbc=0 and see if the performance recovers?
Comment 2 zhoujian 2013-11-26 09:52:57 UTC
(In reply to comment #1)
> To confirm, can you boot with i915.i915_enable_fbc=0 and see if the
> performance recovers?

The problem was fixed with i915.i915_enable_fbc=0.
Comment 3 Chris Wilson 2013-11-26 16:57:55 UTC
Ville, I have the funny feeling that your FBC tracking fixes are required. Of course, it still means we will have a similar impact on frontbuffer rendering, but that is par for the course.
Comment 4 Paulo Zanoni 2013-11-27 12:19:12 UTC
(In reply to comment #3)
> Ville, I have the funny feeling that your FBC tracking fixes are required.
> Of course, it still means we will have a similar impact on frontbuffer
> rendering, but that is par for the course.

I wonder if we should follow the same model as the other drivers and only allow this kind of feature when runtime PM is enabled and/or when we're on battery...
Comment 5 Daniel Vetter 2013-11-28 14:03:25 UTC
(In reply to comment #4)
> (In reply to comment #3)
> > Ville, I have the funny feeling that your FBC tracking fixes are required.
> > Of course, it still means we will have a similar impact on frontbuffer
> > rendering, but that is par for the course.
> 
> I wonder if we should follow the same model as the other drivers and only
> allow this kind of feature when runtime PM is enabled and/or when we're on
> battery...

Imo that's just chickening out of hard problems. PM features should be enabled by default, and if they're causing issues due to overhead somewhere else we need to look at that and fix it up.
Comment 6 Gordon Jin 2013-11-29 01:21:47 UTC
I'm curious why Fedora is not impacted?
Comment 7 Ville Syrjala 2013-12-04 15:03:02 UTC
I've pushed a big pile of FBC patches here:
git://gitorious.org/vsyrjala/linux.git fbc_stolen_fixes

Would be nice to know if the tracking improvements help with the performance.
Comment 8 Chris Wilson 2013-12-04 15:33:16 UTC
(In reply to comment #7)
> I've pushed a big pile of FBC patches here:
> git://gitorious.org/vsyrjala/linux.git fbc_stolen_fixes
> 
> Would be nice to know if the tracking improvements help with the performance.

Much! With fbc_stolen_fixes the impact of FBC on both blit/render frontbuffer performance is ~10%, not orders of magnitude.
Comment 9 Daniel Vetter 2013-12-04 16:02:07 UTC
Curious question: How much of those 10% can be explained with the additional memory traffic, presuming that the compressed fb is half the size of the real one in a back-of-the-envelope estimate (and that we invalidate it every frame)?
Comment 10 Daniel Vetter 2013-12-04 16:03:06 UTC
Oh and the obvious one ofc: Impact on offscreen rendering is now 0?
Comment 11 zhoujian 2013-12-06 02:05:04 UTC
(In reply to comment #7)
> I've pushed a big pile of FBC patches here:
> git://gitorious.org/vsyrjala/linux.git fbc_stolen_fixes

Hi, the kernel that from the tree ,doesn't support HSW.
IVB can start with the kernel, but i915 module can't be loaded.
Comment 12 meng 2014-01-14 08:11:40 UTC
(In reply to comment #7)
> I've pushed a big pile of FBC patches here:
> git://gitorious.org/vsyrjala/linux.git fbc_stolen_fixes
> 
> Would be nice to know if the tracking improvements help with the performance.

Hi, can you give us a patch based on kernel -nightly branch?
Comment 13 Daniel Vetter 2014-06-18 14:38:26 UTC
fbc now disabled again by default.
Comment 14 meng 2014-07-02 02:45:43 UTC
Verified it.
Comment 15 Paulo Zanoni 2014-07-02 12:37:51 UTC
Well, even though it is disabled by default, FBC is still reducing cairo performance by over 60%, isn't it? I'm not sure if closing the bug is the right approach, since the feature is still there.

Can't we test Ville's FBC tree and see if it improves the performance?
Comment 16 meng 2014-07-02 12:45:32 UTC
(In reply to comment #15)

> Can't we test Ville's FBC tree and see if it improves the performance?

 We can. Could you please us give the git repo?
Comment 17 Gordon Jin 2014-07-03 00:46:47 UTC
(In reply to comment #15)
> Well, even though it is disabled by default, FBC is still reducing cairo
> performance by over 60%, isn't it? I'm not sure if closing the bug is the
> right approach, since the feature is still there.

I agree with Paulo. Let's keep this open to track it. 
Since it's not default feature, I'm decreasing priority and not considering it as regression.
Comment 18 Jesse Barnes 2014-12-05 20:37:57 UTC
I think Paulo is working to re-enable FBC by default using the new front buffer tracking stuff; maybe that will help.
Comment 19 Jani Nikula 2015-10-23 09:39:55 UTC
Really old bug, FBC is disabled by default, and enabling taints the kernel. Closing.

I'm assuming new bugs will be filed if the same perf regression comes back if/when we start enabling FBC again.
Comment 20 Chris Wilson 2015-10-23 09:47:55 UTC
The purpose of keeping this bug open was to have a definite testcase that the new FBC code has to pass before being enabled.
Comment 21 Jani Nikula 2015-10-23 11:28:16 UTC
(In reply to Chris Wilson from comment #20)
> The purpose of keeping this bug open was to have a definite testcase that
> the new FBC code has to pass before being enabled.

I've now documented this in the relevant JIRA tasks.
Comment 22 Chris Wilson 2016-02-19 21:16:42 UTC
(In reply to Jani Nikula from comment #21)
> (In reply to Chris Wilson from comment #20)
> > The purpose of keeping this bug open was to have a definite testcase that
> > the new FBC code has to pass before being enabled.
> 
> I've now documented this in the relevant JIRA tasks.

Apparently no one reads JIRA as FBC was re-enabled without fixing the regressions.
Comment 23 Chris Wilson 2016-02-19 21:26:16 UTC
Jumping to conclusions, a secondary effect causing an annoying slowdown when fbc is enabled.


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.