Re: [Matplotlib-users] mpl-1.2.1: Speedup code by removing .startswith() calls and some for loops

[Martin M. <mmo...@fo...>]

Hi Ben,
  thank your for your comments. OK, here are revised patches. I see a hot spot
in artist.py where the getattr() calls are too expensive. Actually, those under
the callable() path.


 793 class ArtistInspector:

 817     def get_aliases(self):
 818         """
 819         Get a dict mapping *fullname* -> *alias* for each *alias* in
 820         the :class:`~matplotlib.artist.ArtistInspector`.
 821 
 822         Eg., for lines::
 823 
 824           {'markerfacecolor': 'mfc',
 825            'linewidth'      : 'lw',
 826           }
 827 
 828         """
 829         names = [name for name in dir(self.o) if
 830                  (name[:4] in ['set_', 'get_'])
 831                  and callable(getattr(self.o, name))]
 832         aliases = {}
 833         for name in names:
 834             func = getattr(self.o, name)
 835             if not self.is_alias(func):
 836                 continue
 837             docstring = func.__doc__
 838             fullname = docstring[10:]
 839             aliases.setdefault(fullname[4:], {})[name[4:]] = None
 840         return aliases


Another hot spot is setp() artist.py, actually its get_aliases() on line 817,
which again leads to getattr() and callable(). The problem is there are millions of
their calls.


So, once again, my all my patches attached (figure.py.patch artist.py.patch have changed).
Martin



Benjamin Root wrote:
> 
> On Fri, Aug 9, 2013 at 9:04 AM, Martin Mokrejs <mmo...@fo... <mailto:mmo...@fo...>> wrote:
> 
>     Hi Phil,
> 
>     Phil Elson wrote:
>     > Hi Martin,
>     >
>     > Thanks for this - we are really interested in speeding up the scatter and barchart plotting with large data sets. In fact, we've done some work (https://github.com/matplotlib/matplotlib/pull/2156) recently to make the situation better.
>     >
>     > I'd really like to review these changes (against matplotlib master), and the best possible solution to doing this is if you were to submit a pull request. If the changes you have made are logically seperable, then I'd encourage you to make a few PRs, but otherwise, a single PR with all of these changes would be great.
> 
>     I went through the changes there and they just cope with other pieces of matplotlib.
>     My changes are general python improvements moving away from str.startswith()
>     and using generators instead of for loops. Just apply the patches yourself and see.
>     ;)
> 
>     >
>     > Would you mind turning these patches into PR(s)? (https://github.com/matplotlib/matplotlib/compare/)
> 
>     Um, I don't know what to do on that page, sorry. I don't see how to upload my patch file or patched file
>     to be compared with master. :(
> 
>     >
>     > Thanks!
> 
>     I am sorry but I just don't have time to fiddle with github. It is just awkward. I even failed to download
>     diffs of the changes from https://github.com/matplotlib/matplotlib/pull/2156/commits.
> 
>     I rather continue studying runsnake output. ;-)
> 
>     Martin
> 
> 
> A snippet from one of you patches:
> 
>          dsu = []
>  
> -        for a in self.patches:
> -            dsu.append( (a.get_zorder(), a, a.draw, [renderer]))
> +        [ dsu.append( (x.get_zorder(), x, x.draw, [renderer])) for x in self.patches ]
> 
> Yes, we certainly should use list-comprehensions here, but you are using it incorrectly.  It should be:
> 
> dsu = [(x.get_zorder(), x, x.draw, [renderer])) for x in self.patches ]
> 
> And then, further down, do the following:
> 
> dsu.extend((x.get_zorder(), x, x.draw, [renderer])) for x in self.lines)
> 
> Note the generator form of the comprehension as opposed to the list comprehension form. List comprehensions should *always* be assigned to something. List comprehensions should only be for replacing the for-append idiom in python.
> 
> Thank you though for pointing out parts of the code that can benefit from revisions. I certainly hope you can get this put together as a pull request on github so we can work to make this patch better!
> 
> Ben Root
> 
> P.S. - I <3 runsnakerun!