[Numpy-discussion] Optimizing Numpy

[Pete Shinners]

i've been throwing my hand and getting more speed out of
numpy. i've birthed a little fruit from my efforts. my area
of use is specifically with 2D arrays with image info.

anyways, i've attached an 'arrayobject.c' file that is
from the 15.3 release and optimized. 

in my test case the code ran about twice the speed of
the original 15.3 release. i went and tested out other
uses and found on a 20% speedup pretty consistent.
(for example, i cranked the mandelbrot demo resolution
 to 320 x 200 and removed the 'print' command and it
 went from a runtime of 5.5 to 4.5)

i'm not sure how people 'officially' make contributions
to the code, but i hope this is easy enough to merge. i
also hope this is accepted (or at least reviewed) for
inclusion in the next release.


optimizing further...
i also plan on a few more optimizations. the least is going
to be a 'C' version of 'arrayrange' and probably 'ones'. the
current arrayrange is pretty slow (slower than the standard
python 'range' in all my tests).
the other optimization is a bit more drastic, and i'd like
to hear feedback from more 'numpy experts' before making the
change. in the file 'arraytypes.c' with all the arrays of conversion
functions, i've found that the conversion routines are a little
too 'elaborate'. these routines are only ever called from one line
and the the two "increment/skip" arguments are always hardcoded one.
there are two possible roads to speedup the conversion of array
types.
1-- optimize all the conversion routines so they aren't so generic.
this should be a pretty easy fix and should offer noticeable speed.
2-- do a better job of converting arrays. instead of creating a
whole new array of the new type and simply copying that, create
a conversion method that simply converts the data directly into
the destination array. this would mean using all those conversion
routines to their full power. this would offer more speed than
the first option, but is also a lot more work

well, what do people think? my initial thought is to make a
quick python script to take 'arraytypes.c' and convert all the
functions to be a quicker version.


numpy is amazing, and i'm glad to get a chance to better it!

-------------- next part --------------
A non-text attachment was scrubbed...
Name: arrayobject.zip
Type: application/x-zip-compressed
Size: 16499 bytes
Desc: not available
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20000708/7f3a244b/attachment.bin>