[Intel-gfx] Reasons for no ring-buffer on 2.6.27?

[Clemens Eisserer]

Thanks a lot for your investigation.

Unfourtunatly this doesn't fix the performance regression I see.
I opened bug https://bugs.freedesktop.org/show_bug.cgi?id=18075 with a
benchmark as well as the sysprof profile.

Thanks, Clemens

2008/10/15 Eric Anholt <eric at anholt.net>:
> On Wed, 2008-10-15 at 16:48 +0200, Clemens Eisserer wrote:
>> Hi,
>>
>> I've played a bit with latest intel/xserver git and noticed very poor
>> performance on Fedora-10-beta.
>> When I execute exactly the same x/driver binaries on a self-compiled
>> 2.6.20.1 on Fedora8 performance was as expected (well, actually a bit
>> worse than I hoped for ;) )
>>
>> What I noticed was that according to sysprof a lot of time was spent
>> inside flush-functions like intel-batch-flush (called by
>> i830exaDoneSolid ~70% and i830sync ~25%), and that some driver
>> initialization messages differ in Xorg.log.
>> Any idea why a ring-buffer is initialized and used with 2.6.20.1 and
>> it works so well for my testcase, but not with 2.6.27?
>
> Batchbuffers are dispatched through the ringbuffer.  The ring is used in
> either case (it's the only way to program acceleration).
>
> This is probably the EXA bug introduced with UXA that keithp said he was
> going to fix the other day, but I don't see in the tree yet.  It's there
> now.
>
> commit 6707371176147340fabc9ab6f1e3d6d5ac980662
> Author: Eric Anholt <eric at anholt.net>
> Date:   Wed Oct 15 08:12:11 2008 -0700
>
>    Remove gratuitous flushing in EXA after solid operations.
>
>    This snuck in with the UXA rename commit.
>
> Thanks for the sysprof results, it easily reminded me of the correct issue.
>
> --
> Eric Anholt
> eric at anholt.net                         eric.anholt at intel.com
>
>
>