[Intel-gfx] FireFox performance regressions XAA -> EXA -> UXA

[Chris Wilson]

On Tue, 26 Jan 2010 07:30:25 -0500, Clemens Eisserer <linuxhippy at gmail.com> wrote:
> > Whether the driver uses xaa, exa or uxa should be mostly invisible to
> > the client though?  It doesn't affect whether the server supports the
> > render extension, just how it's handled internally, so shouldn't change
> > how cairo does stuff.  Or am I missing something?
> 
> Yes, you're right. With XAA most RENDER requests simply end up
> somewhere in pixman.
> Therefore, firefox/cairo should behave the same, sending exactly the
> same requests/commands (minus possible workarrounds) to the X-Server
> in the XAA/EXA/UXA case.

What is the reported RENDER version for the XAA server (and EXA, UXA...)?
For I thought EXA marked the beginning of support for RENDER on -intel.
And no, I expect the xtraces between XAA/EXA/UXA to be different due to
the varying levels of RENDER support and X.org version. We haven't even
stated what version(s) of cairo we are comparing either.

This is all incidental anyway, we know the CPU is generally faster than
our driver for the typical mix of drawing operations performed by
applications. (As a counter point, we can also show applications where the
driver is much faster than the CPU.)  All I'm trying to ascertain is
whether trender is a good benchmark of why so many people complain that
firefox feels slow when scrolling. To which my initial response is that
the system profiles look very similar to cairo-trace and the visible
trender behaviour does not seem to suggest any new information.
-ickle

-- 
Chris Wilson, Intel Open Source Technology Centre