Re: [Devel] Benchmarking FreeType

[David Turner]

Salut,

Vincent Caron a écrit :
> 
> On Fri, 2002-06-07 at 09:39, David Turner wrote:
> > I'd also like to include "ftbench.c" as part of the "ft2demos" package,
> > however, it will need a few small cleanups since its source code probably
> > only compiles with GCC due to statements like:
> >
> >    {
> >      int   array1[ face->num_glyphs ];   // dynamic stack allocation !!
> >      int   array2[ face->num_glyphs ];
> >
> >      ....
> >    }
> >
> > Could you polish it a bit before we include it ??
> 
> Done, still available here : http://zerodeux.net/ft/
> 
> - the stack allocation code has been replaced by a good old calloc().
> Anyway I was fetching the whole charmap at every bench(), yek. This kind
> of code works well with gcc or MSVC, but it's effectively a C extension
> (non-POSIX).
> 
> - the 'bench time' constant is available from the command line
> 
> - the tests can be selected from the command line
> 
> - added the legal ft2demo blurb. I'm hereby giving this code and its
> copyright to the FreeType project, if it has to be said clearly.
>

Thanks a lot, it is now in "ft2demos/src/ftbench" and should compile
on all platforms supported by FreeType. There are probably a couple
compiler warnings still, but I'll get rid of them very soon..

Since I was optimizing the cache manager when you first posted about
your benchmark program, I had the pleasure to use it to refine the
code in efficient ways. This results in a 20% to 50% speed-up in
all cache hits, depending on your compiler, CPU and operating
system.

Note that all caches are now significantly fasters. For example, the
CMap cache is now a _lot_ faster than calling FT_Get_Char_Index
directly while preserving "abstractness"...

Here's the output for Arial.ttf on a 233 MHz Mobile Pentium:

  Load                           : 428.819 us/op
  Load + Get_Glyph               : 471.258 us/op
  Load + Get_Glyph + Get_CBox    : 475.309 us/op
  Get_Char_Index                 :   1.882 us/op
  CMap cache (1st run)           :  -0.000 us/op
  CMap cache                     :   0.899 us/op
  Outline cache (1st run)        : 455.247 us/op
  Outline cache                  :   2.189 us/op
  Bitmap cache (1st run)         : 564.043 us/op
  Bitmap cache                   :   1.708 us/op
  SBit cache (1st run)           : 556.327 us/op
  SBit cache                     :   1.369 us/op


Just another good reason to use the FT2 cache in your applications
and libraries, instead of rewriting your own :-)


Thanks,

- David Turner
- The FreeType Project  (www.freetype.org)


PS:

For the technically savvy, these improvements were obtained by
applying three important techniques:

  - inlining a few common routine calls and optimizing
    the result (i.e. rewriting ftc_node_mru_up which called
    ftc_node_mru_unlink and ftc_node_mru_link)

  - performing template instantiation (possible in C with
    a few pre-processor tricks. see the file src/cache/ftccache.i)
    to speed-up cache hits.
 

  - finally, changing the internal hash table computation. The
    previous algorithm was based on the algorithm used by the GLib.
    I replaced it with a straight-forward implementation of
    linear hashing, which is significantly faster while preventing
    any "chokes" when resizing the buckets array.

    for more details on this algorithm, see:

         http://citeseer.nj.nec.com/34453.html