[cairo] Text performance on windows?

[Vladimir Vukicevic]

On Sep 10, 2008, at 5:52 PM, Damian Frank wrote:

> On Tue, Sep 9, 2008 at 2:39 AM, Vladimir Vukicevic  
> <vladimir at pobox.com> wrote:
>> Is the code for this something that you can share?  (I can write my  
>> own, but
>> want to avoid redoing the work if at all possible.)  I'm guessing  
>> it does
>> something like:
>>
>> do_cairo_setup();
>>
>> for (int i = 0; i < bignum; i++) {
>>  cairo_show_glyphs (cr, glyphs, nglyphs);
>> }
>>
>> And times the loop?  (Want to make sure that you're not using  
>> show_text or
>> similar.)  And for the pure GDI case, you're just calling  
>> ExtTextOutW,
>> ideally with ETO_GLYPH_INDEX?
>
> Unfortunately I can't share, both because of licensing and because of
> the excessive complexity of the software.  However, it essentially
> does that, except it's time-based.  The other important difference is
> that we *are* using show_text for testing purposes.  Here are some of
> the timings breaking the overhead down a little bit.  cairo_show_text
> represents time spent in the whole routine, and c_gs_* breaks it down
> into those two routines alone.

Yep, no worries, just wanted to make sure we'd be testing the same  
thing.  cairo_show_text is part of the "toy" text api, and is really  
just there for prototyping and testing; it's not really a production- 
quality API.  So as your timings show, it looks like on average half  
of the time is spent inside text_to_glyphs, which should go away if  
you use show_glyphs directly.  The perf with a dib32 and an image  
surface looks to be the close, so I'm guessing we're taking a fallback  
path there -- I'll definitely take a look there, as we may be able to  
avoid that.  I'm surprised by DDBs being slower than DIB24, though.

> As for our GDI comparison, yes, we have a text layout engine of our
> own and we do the rendering with ETO_GLYPH_INDEX.

Ok; I would suggest keeping that piece and just handing the glyphs to  
cairo render via show_glyphs -- you won't be able to replace your  
entire layout engine with cairo.

I'll try to create a small testcase for this in the next few days and  
send it to the list and we can compare numbers.

     - Vlad

>
>
> Some timings using QueryPerformanceCounter:
> ddb:
>         ExtTextOutW: last:      8 avg:    8.5  min:      6  max:    
> 2163
>     cairo_show_text: last:    969 avg: 1012.6  min:    833  max:   
> 86489
> c_gs_text_to_glyphs: last:    404 avg:  431.7  min:    385  max:   
> 85762
> c_gs_show_tx_glyphs: last:    540 avg:  555.0  min:    420  max:   
> 58285
>
> dib24:
>         ExtTextOutW: last:    222 avg:  247.7  min:    198  max:   
> 15359
>     cairo_show_text: last:    737 avg:  786.3  min:    709  max:   
> 15064
> c_gs_text_to_glyphs: last:    394 avg:  427.7  min:    384  max:   
> 11327
> c_gs_show_tx_glyphs: last:    314 avg:  332.8  min:    297  max:    
> 9396
>
> dib32:
>     cairo_show_text: last:   1203 avg: 1294.8  min:   1167  max:    
> 9301
> c_gs_text_to_glyphs: last:    397 avg:  430.3  min:    384  max:    
> 5215
> c_gs_show_tx_glyphs: last:    782 avg:  839.5  min:    754  max:    
> 7481
>
> image:
>     cairo_show_text: last:   1073 avg: 1145.6  min:   1051  max:   
> 51262
> c_gs_text_to_glyphs: last:    387 avg:  406.0  min:    383  max:    
> 3776
> c_gs_show_tx_glyphs: last:    662 avg:  715.1  min:    640  max:   
> 50815
>
>
> Thanks for your help.  If you have any trouble reproducing the
> performance, I'm willing to try to reproduce what we do in a simpler C
> program.
>
> Damian