Re: [fbsd] Re: DRM update for testing

On Tue, 2005-12-20 at 09:06 +0100, Jeremie Le Hen wrote:
> Hi, Eric,
> 
> > glxgears is not a benchmark.
> 
> What benchmark do you advice ?

Find an app whose performance you care about.  Something that you would
run regularly, and probably where the fps with hardware is approximately
<= the refresh rate.  For me that's been quake3, other developers ut2kX
or doom3 or their own custom apps.

Why is glxgears performance measurement pointless?

1) In software, many people see it going faster than the refresh rate.
So why do you care if you can go 10x the refresh rate or only 1.5x?
You're only displaying so much.
2) It's flat or smooth-shaded triangles.  Does anybody use these?  Well,
some GL screensavers do, but those in particular I find work fine in
software as well.
3) It's display lists of triangles.  Most software where performance
matters is going to be using some sort of vertex arrays of triangles
(any game, basically) or lines (CAD, scientific apps).  Those are going
to be totally different rendering paths.
4) relative significance of swapping.  In very few apps do you see
swapping back to front (or perhaps waiting for pending swaps) being a
major participant in the total time.

Notably, 4) has led to bad optimizations like page flipping.  It's a 20%
or so performance improvement (over an existing 1000+ fps) in tests I've
seen for glxgears, but I've only been able to measure .4% to 5%
improvements for quake3 under rather ideal conditions.  On the other
hand, page flipping increases the complexity of your X driver code (it
took a long time to get it to even render right), increases overhead of
2d operations to perform the extra accounting and copying, and is likely
nonconformant for X drawing on top of GL drawing, which some apps will
do as far as I've heard.  That's what optimizing for bad benchmarks gets
you.

Oh, and for those using software, in software you're probably going to
be measuring your CPU-framebuffer write speeds more than the-rest-of-GL
speeds, anyway.  Those rates are rather different between cards
(notably, Intel stuff seems to do quite well with operations on
framebuffer, perhaps due to its UMAness.  I need to actually quantify it
and see if I'm right), and it's unimportant for hardware acceleration.

This is why I say that posting glxgears fps on mailing lists or IRC is
just noise.  That is, unless all you care about is the screensaver with
the gears.

-- 
Eric Anholt                                     eta_at_lclark.edu
http://people.freebsd.org/~anholt/              anholt_at_FreeBSD.org