Re: testing pango with 770

[Tapani Pälli <tapani palli nokia com>]

ext Federico Mena Quintero wrote:

>On Wed, 2005-12-21 at 15:15 +0200, Tapani P�i wrote:
>
>  
>
>>Text 'abcdefghijklmnopqrstuvwxyz' is rendered to 20 rows 100 times using
>>'sans serif-18' font.
>>(Each test is renderering 52000 glyphs.)
>>    
>>
>
>This test could be more thorough.  You aren't testing word breaking or
>the non-latin shapers, for example.
>
>  
>
Yes, I will think of real-life usecases also. But writing lines of text
is pretty much simplest usecase there is ... I wonder why the difference
is that big, I will try to find that out.

>Also, your test needs to run for a longer time.  Otherwise, you'll get a
>lot of variation among runs.  Make it so that each run takes a few
>minutes.
>
>  
>
This is true, I will run this again with longer time. For gtk I think it
will speed up but xlib test should stay pretty neutral.

>You mentioned that you were using gettimeofday().  Can you please test
>the "pango-profile" module from CVS?  It has a slightly more
>sophisticated timer, using times().  I've found it to be quite reliable,
>and it gives reproducible results.
>
>  
>
I will try pango-profile, thanks!

>>* What is causing slowness?
>>    
>>
>
>Ask a profiler :)
>
>For your test, I'd like to know the CPU usage for both Pango and the X
>server.  On my machine, a "render a lot of text" benchmark spends 50% of
>the time in the X server (xorg without the MMX compositing stuff), about
>30% in Pango, and the rest in other processes.  Sysprof can tell you
>that very nicely.
>
>  
>

Yep the whole pipeline has to be inspected, the biggest bottleneck can
be elsewhere (whether the actual blitting can be optimized more that's
another question). but we haven't actually had any changes with our X
for awhile and I was using same X build for all tests.

>If you use the benchmark in pango-profile, beware that it just tests the
>Pango machinery, not the rendering machinery.  You may want to stick a
>gdk_draw_layout() in the code.
>
>  
>
Ok, this is good to know.

>Does Sysprof run on the 770 yet?
>  
>
Nope :-/

// Tapani