[Bug 51787] New: performance regression with llvm shader compiler in ut2004

[Bug 51787] New: performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

             Bug #: 51787
           Summary: performance regression with llvm shader compiler in
                    ut2004
    Classification: Unclassified
           Product: Mesa
           Version: git
          Platform: Other
        OS/Version: All
            Status: NEW
          Severity: normal
          Priority: medium
         Component: Drivers/Gallium/r600
        AssignedTo: dri-devel@lists.freedesktop.org
        ReportedBy: aaalmosss@gmail.com


Whenever some new geometry appears (even for weapon projectiles), the framerate
drops drastically for a short time. This is a frequent event, so the game
becomes unplayable overall.

There is also a big preformance drop with the old compiler, when several
players get gibbed at once (happens frequently on as-convoy), but with the llvm
compiler this is much worse.

I'm not an expert on this, but it seems as if a new vertex shader would be
compiled for each vertex array submitted, and the llvm compiler is an order of
magnitude slower.

-- 
Configure bugmail: https://bugs.freedesktop.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.

[Bug 51787] performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

--- Comment #1 from Michel Dänzer <michel@daenzer.net> 2012-07-06 08:25:03 PDT ---
FWIW, I suspect it's more likely (re)compiling fragment shaders than vertex
shaders. Possibly related to constantly changing fixed-function emulation
shaders, UT2004 seems to burn through those quickly (this also caused slowdowns
with fglrx's shader cache when the game was new). Just an idea, maybe this is
different.

-- 
Configure bugmail: https://bugs.freedesktop.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

[Bug 51787] performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

--- Comment #2 from almos <aaalmosss@gmail.com> 2012-08-02 22:04:09 UTC ---
It might not help in fixing this, but I found that the framerate is much more
consistent if I load all cpu cores with something while playing ut2004.
Normally, the framerate counter is green, but flashes into yellow and purple,
which makes movements very choppy. When something is running in the background,
the framerate counter stays green, and all movements are smooth.

With the llvm compiler the game is unplayable either way.

-- 
Configure bugmail: https://bugs.freedesktop.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.

[Bug 51787] performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

--- Comment #3 from Andy Furniss <lists@andyfurniss.entadsl.com> 2012-08-02 22:32:43 UTC ---
(In reply to comment #2)
> It might not help in fixing this, but I found that the framerate is much more
> consistent if I load all cpu cores with something while playing ut2004.
> Normally, the framerate counter is green, but flashes into yellow and purple,
> which makes movements very choppy. When something is running in the background,
> the framerate counter stays green, and all movements are smooth.

Maybe cpufreq is causing this here's what I get on the demo on an HD4890 +
4x3.4GHz phenom II.

R600_LLVM=0 ut2004
as-convoy?spectatoronly=1?numbots=8?quickstart=1?attractcam=1 -benchmark
-seconds=120 -nosound

cpufreq ondemand - 

30.027578 / 76.497787 / 162.457611 fps

cpufreq set to performance

37.627441 / 93.523949 / 197.051376 fps

> 
> With the llvm compiler the game is unplayable either way.

I see the stalling in the demo benchmark.

2.030273 / 70.589706 / 196.047073 fps

-- 
Configure bugmail: https://bugs.freedesktop.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.

[Bug 51787] performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

[-- Attachment #1.1: Type: text/plain, Size: 309 bytes --]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

--- Comment #4 from Tobias Jakobi <liquid.acid@gmx.net> ---
This is not related to cpufreq. I've fixed my governor to the highest frequency
and still experience these massive lags.

-- 
You are receiving this mail because:
You are the assignee for the bug.

[-- Attachment #1.2: Type: text/html, Size: 1101 bytes --]

[-- Attachment #2: Type: text/plain, Size: 159 bytes --]

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

[Bug 51787] performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

[-- Attachment #1.1: Type: text/plain, Size: 493 bytes --]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

--- Comment #5 from almos <aaalmosss@gmail.com> ---
(In reply to comment #4)
> This is not related to cpufreq. I've fixed my governor to the highest
> frequency and still experience these massive lags.

Since I cerated a startup script for ut2004 that does 'set all cores to
performance ; ut2004 ; set all cores to ondemand' the game runs smoothly
without any lag.

-- 
You are receiving this mail because:
You are the assignee for the bug.

[-- Attachment #1.2: Type: text/html, Size: 1350 bytes --]

[-- Attachment #2: Type: text/plain, Size: 159 bytes --]

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

[Bug 51787] performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

[-- Attachment #1.1: Type: text/plain, Size: 1206 bytes --]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

--- Comment #6 from almos <aaalmosss@gmail.com> ---
Now I took a new look at the issue with GALLIUM_HUD=fps, and there is
definitely a serious problem with cpufreq. I have a few scripts to switch
cpufreq governor between performance and ondemand (Phenom II x4), and to switch
between high and low gpu power profiles (HD6850). When ondemand is selected, it
seriously limits performance, and causes noticeable variance in fps. Some
examples:

Psychonauts, main menu (Raz standing on top of the giant brain)
ondemand/low 20 fps
ondemand/high 40-50 fps
performance/low 29 fps
performance/high 104 fps

SW:KoToR, the main character standing alone on Dantuine, looking at the Ebon
Hawk
ondemand/low 20-30 fps
ondemand/high 20-30 fps
performance/low 55-60 fps
performance/high 68 fps

The difference between ondemand and performance makes a huge difference in
native games as well. Quake Wars, Team Fortress 2, and UT2004 are all
borderline playable with ondemand, but run perfectly with performance. It's not
just a drop in fps, but a very noticeable laggyness, that kills the games.

-- 
You are receiving this mail because:
You are the assignee for the bug.

[-- Attachment #1.2: Type: text/html, Size: 1990 bytes --]

[-- Attachment #2: Type: text/plain, Size: 159 bytes --]

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/dri-devel

[Bug 51787] performance regression with llvm shader compiler in ut2004

[bugzilla-daemon]

[-- Attachment #1.1: Type: text/plain, Size: 840 bytes --]

https://bugs.freedesktop.org/show_bug.cgi?id=51787

GitLab Migration User <gitlab-migration@fdo.invalid> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |MOVED
             Status|NEW                         |RESOLVED

--- Comment #7 from GitLab Migration User <gitlab-migration@fdo.invalid> ---
-- GitLab Migration Automatic Message --

This bug has been migrated to freedesktop.org's GitLab instance and has been
closed from further activity.

You can subscribe and participate further through the new bug through this link
to our GitLab instance: https://gitlab.freedesktop.org/mesa/mesa/issues/413.

-- 
You are receiving this mail because:
You are the assignee for the bug.

[-- Attachment #1.2: Type: text/html, Size: 2465 bytes --]

[-- Attachment #2: Type: text/plain, Size: 159 bytes --]

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel