The Cost Of ATI Kernel Mode-Setting On Fedora 12
Phoronix: The Cost Of ATI Kernel Mode-Setting On Fedora 12
One of the articles on Phoronix last week was entitled Intel Linux Graphics Shine With Fedora 12, which showed off the nice state of Intel graphics on this latest Red Hat release when it came to kernel mode-setting and its 3D stack with it working well "out of the box" and offering some nice performance gains over the earlier Fedora 10 and Fedora 11 releases. While the Intel stack may be improved in Constantine, the ATI support has taken a hit, as users were quick to point out in response to last week's article. In particular, when using the ATI kernel mode-setting driver in Fedora 12 (which is the default for pre-R600 hardware), there is a large performance discrepancy compared to using the traditional user-space mode-setting for ATI Radeon hardware. Today we are looking at what exactly the performance cost is for using ATI KMS in this new release.
Hmm yet another fact finding failure, we do ship r600 with kms by default.
3D isn't but thats easy to get working.
also F12 has GL tearing avoidance enabled, so you'll get vsynced back/front copies instead of tearing like you get with UMS.
Also color tiling isn't enabled by default under KMS which is probably something we should fix.
I'd be really interested in a follow up one or two tests to see if colortiling actually could help on the r500, I just forgot to enable it for F12 final.
Last edited by airlied; 11-25-2009 at 06:23 AM.
Ok good, thought I was the only one that noticed KMS was on by default...
Originally Posted by airlied
ufaogros: they never said it was news. They said they were benchmarking it.
Phoronix: looking whether you get the graphical boot or progress bar is not the perfect test for KMS, as it also depends on Plymouth being able to find the graphical splash. If KMS is enabled but Plymouth has some problem displaying the graphical splash, it'll fall back to the progress bar but *at native resolution*. The resolution's the key - if you see boot at native resolution, either graphical splash or progress bar, KMS is enabled. If you see boot as the progress bar at console resolution (720x400, the resolution used for POST and DOS and stuff like that), KMS is disabled.
You told me that only
yum install mesa-dri-drivers-experimental
was needed for r600 3d, that's not that complicated. But for testing selinux sucked more when you wanted to try the driver with Quake Live.
Am I the only one who noticed the first graph maximum to be dangerously close to 33%, that is 1 out of the three cores? I don't know what you used for decoding, but libav is single threaded for the most part... To me it looks like with KMS it played at a slower speed and stopped playing to early after the time the movie should have finished.
And of course I wonder why the performance for UT2004 does increase with the resolution with UMS.
Well H264 with 33% is very low bitrate, i would not even call that a needed test. Use something that really hits 100%.
This isn't news. I have filled bugs against this issue (https://bugzilla.redhat.com/show_bug.cgi?id=533244 and https://bugzilla.redhat.com/show_bug.cgi?id=466695, to mention some). I always start with no modeset because the performance hit is absurd.
You may want to take a look here: http://bugs.freedesktop.org/show_bug.cgi?id=22743
Originally Posted by ufaogros
This was a darkplaces bug, not a driver bug.