very bad, and very weird, scrolling text performance xorg 1.5.3 intel 2.5.1 GM965

[Michel Dänzer]

On Mon, 2008-12-15 at 09:37 -0800, Jeffrey Baker wrote:
> On Mon, Dec 15, 2008 at 2:26 AM, Michel Dänzer
> <michel at tungstengraphics.com> wrote:
> > On Mon, 2008-12-15 at 10:57 +0100, Lukas Hejtmanek wrote:
> >> On Sun, Dec 14, 2008 at 06:48:52PM -0500, Thomas Jaeger wrote:
> >> > You really need the glyph cache in the X server to get decent text
> >> > performance out of the 2.5 intel driver.  The patches are pretty
> >> > straightforward to backport, but it is my understanding that a 1.6
> >> > server will be uploaded to jaunty soon, so you might want to wait for that.
> >>
> >> I wonder whether the Intel driver can get at least close to Nvidia driver.
> >>
> >> I ran x11perf on Nvidia system and Xserver *without* the glyph cache. Compared
> >> to these numbers, the Intel driver seems to be a bit comic with 56k glyphs/sec
> >> with EXA or 215k glyphs/sec with UXA.
> >
> > Note that the glyph cache is specific to EXA, which the nvidia driver
> > doesn't use. UXA on the other hand may have the glyph cache in its copy
> > of the EXA code.
> >
> > I'm not sure what's the bottleneck for the intel driver at this point,
> > the radeon driver easily gets 500k/s and beyond (which is still a far
> > cry from the numbers you measured, but already seems to mostly avoid it
> > being any practical limitation). Some of it may just be down to the
> > hardware (memory bandwidth, shader units etc.).
> 
> I seriously doubt it, since performance was far better last week, last
> month, and a year ago with the same hardware.  I think we can safely
> pin this current problem on the software.

The above was referring to x11perf -aa10text, not your testcase. It
sounds like you have a good start (known good and bad versions) for
tracking down the regression you're seeing.


-- 
Earthling Michel Dänzer           |          http://tungstengraphics.com
Libre software enthusiast         |          Debian, X and DRI developer