[cairo] Re: Cairo gears performance benchmarks

[David Reveman]

On Mon, 2004-04-19 at 13:38 -0600, Soorya Kuloor wrote: 
> Hi,
> 
> I compiled the latest cairo CVS. I then ran the cairogears program
> from David's download site (http://www.cs.umu.se/~c99drn/). Here are
> some results I thought people would be interested in.
> 
> Hardware:
> ---------
> AMD Athlon64 3200
> 1.5GB RAM
> Graphics Chipset: ATI R350 NH (R9800)
> 
> Software:
> ---------
> Fedora Core 1 (32 bit version)
> XFree: ATI's proprietary drivers downloaded from their website
> Display resolution: 1280x1024
> 
> The numbers below are in frames per second as reported by the
> test. For full-screen numbers I just maximized the demo window.
> 
> Test   backend       normal-size    full-screen
> ----   -------       -----------    -----------
> 
> TRAP   image           9.2             2.2
>        xrender        11.8             2.9
>        glx           150.2           135.2
> 
> GRAD   image           8.4             2.0
>        xrender         9.2             2.4
>        glx           125.4           103.0
> 
> COMP   image           5.0             1.1
>        xrender         6.4             1.4
>        glx          2087.0           686.8
> 
> TEXT   image          18.4             1.4
>        xrender        22.6             1.9
>        glx            61.8            49.4
> 
> SHADOW image           2.1             0.35
>        xrender         2.62            0.5
>        glx           133.0           130.0
> 
> The performance of the OpenGL back-end is really impressive. Obviously,
> the text caching problem that David mentioned slows down the OpenGL
> back-end considerably.

Hi Soorya

I haven't been able to run any tests on radeon 9800 cards, so these
results are very interesting, thanks!

The results are similar to those that we've been getting with nvidia's
geforce FX cards.

The text caching problem isn't actually present in the above TEXT test.
This test uses cairo_text_path, which works very well. However,
decomposing and tessellating text seems to be very CPU intensive and
that's why the OpenGL backend is not as superior in this test.

The SHADOW test is actually not really fare to the OpenGL backend. The
OpenGL backend support CAIRO_FILTER_GAUSSIAN with your radeon card.
Hence, filter each shadow through a 3x3 convolution filter for every
frame. Without convolution filtering the OpenGL backend would probably
perform about 4 times faster.

If you like to, you could get my updated cairogears program with the
bitmap text test and also present results from this test. Both with and
without my new glyph cache patch. That would be interesting.

- David

-- 
David Reveman <c99drn at cs.umu.se>