[Numpy-discussion] Vectorized ufuncs with OS X frameworks
Francesc Altet
faltet at carabos.com
Wed Mar 14 08:47:36 EDT 2007
Hi Eric,
A Divendres 09 Març 2007 15:32, Eric Brown escrigué:
> Hi All,
>
> I have a set of large arrays to which I have to do exp functions
> over and over again. I am wondering if there is any benefit to
> teaching numpy how to use the vForce frameworks to do this function.
>
> Does anyone have experience hacking the source to get numpy to use
> system libraries to do vectorized versions of sqrt, exp, log,
> addition, etc.? I am also interested in doing this for other
> operating systems.
Maybe it is not what you want, but perhaps it is worth for you to have a look
at numexpr [1] which has a kernel for vectorizing the sort of computations
you are interested in (sqrt, exp, log, pow, trigonometric, addition...). It
does the vectorization by extracting blocks of the numpy arrays and passing
them to the computing kernel (which are macros in C for achieving maximum
performance).
Also, it should be possible to adapt the kernel of numexpr for VForce,
although I'm not sure about how vForce is going to achieve more efficiency
than numexpr, as IMO the cost for these kind of computations is mostly
dominated by fetch/store memory cells rather than doing actual computations
(although I might be wrong).
[1] http://projects.scipy.org/scipy/scipy/browser/trunk/Lib/sandbox/numexpr/
Cheers,
--
>0,0< Francesc Altet http://www.carabos.com/
V V Cárabos Coop. V. Enjoy Data
"-"
More information about the NumPy-Discussion
mailing list