standard variants for py26-numpy

[vincent habchi]

Le 14 févr. 2011 à 20:50, Olaf Foellinger a écrit :

> Today I've noticed during an upgrade that macports tries to install
> atlas. That's because it's a standard dependency of py26-numpy.
> py26-numpy also adds gcc44 this way which is quite a lot for gnucash
> users. I would recommend to omit atlas from the default dependencies of
> p25-numpy. What do you think? Should I enter a ticket for this?	

Numpy, gewissermaßen :), is a computing package: scipy uses numpy. As such, it strives to work as fast as possible (nobody wants to wait ten hours the result of a matrix inversion knowing that it can be done in ten seconds when the algorithm is optimized). ATLAS takes hours to compile on 32-bit machines because it *really tries hard* to find the most efficient algorithm for each BLAS routine. Besides, ATLAS can use the latest GCC version, whereas the Accelerate framework embedded in MacOS is probably compiled with the same compiler as the one furnished with Xcode. It would be worth benchmarking code linked with an up to date ATLAS and the Apple Accelerate framework, and see which wins.

IMHO, the problem is not in numpy depending on Atlas, which is correct ; it is in other packages depending on numpy for tasks that may involve very little of it (e.g. using numpy as a convenient means to get array functions). Gimp relies ultimately on numpy for matrix convolutions in image processing, which can take quite a lot of time when the size of the image is large.

FWIW
Viele Grüße auch aus Paris unter Wolken,
Vincent