Re: [pygccxml-development] Any recent commits that may affect performance

[Roman Y. <rom...@gm...>]

On 2/19/07, Allen Bierbaum <al...@vr...> wrote:
>
> > I thought about it and I don't know. May be you are doing something
> wrong.
>
> That is always possible although nothing has really changed in the
> script in the time since it used to run fast.
>
> > Python-Ogre project is not a smallest one, but it takes only 5 minutes(
> without
> > cache and files ) to generate code. During code generation they also
> parse
> > Ogre source tree to extract documentation. Can you show your script?
>
> The generation script is here:
>
> https://realityforge.vrsource.org/trac/pyopensg/browser/trunk/src/gen_bindings.py
>
> I have added some additional timing analysis that shows the vast
> majority (>90%)of the time being spent in two places.
>
> 1) line 587, the call to the module builder initialization where all the
> code is parsed
> 2) line 1040, the call to build_code_creator where all the code creators
> are built up and everything is setup to run the code generation.
>
> Neither of these is a surprise since this is where py++ does all of it's
> work.
>
> I am still collecting more information that should be able to help track
> down why the full_name and get_name methods are being called so many
> millions of times.  (calling full_name 50 million times for a single run
> just seems excessive.)  I am using stack monitoring technique I used
> when I helped out with the performance last time.  I have added code to
> a local version of py++ that is not only counting the number of times
> full_name is called but also keeps a dictionary that contains a list of
> every call path (stack) into the code and how many times full_name has
> been called along each path.  This should be able to show us if there
> are some code paths that end up making the majority of the calls and
> could be optimized.
>
> Collecting this information can take a long time though.  I have already
> had the instrumented code running for about 20 hours and I expect it
> will run for another 24 hours before it is complete.  Once the run is
> done I will post the resulting information so everyone can take a look
> and present any ideas they have for helping out performance.
>
> -Allen
>
>
Grate. We make Py++ faster. Thank you for all your work.

-- 
Roman Yakovenko
C++ Python language binding
http://www.language-binding.net/