[arch-general] scipy package - is there anything important that stops the update to 0.9?

Andrzej Giniewicz gginiu at gmail.com
Fri Apr 22 13:59:26 EDT 2011


Hi,

> The main issue now is linking to lapack, which fails. There is a
> symlink hack (you can refer to the python3- PKGBUILD in AUR), but we
> do not know why and how (this works). Without that hack, SciPy should
> not be choosy about linking to static or dynamic lapack libs, but in
> this case it builds only if a static lapack is present. The reason for
> the stall in releasing an update is basically waiting to sort out this
> behaviour. Otherwise, it's good to go.

I think I know what this is about, i.e. iirc it was bug in Numpy. In
Python 3.2 because of PEP 3149 (
http://www.python.org/dev/peps/pep-3149/ ) all .so files are prefixed
with SOABI name, i.e. cpython-32 in that case. It was made so compiled
files by multiple python versions could coexist. Python itself uses
sysconfig to determine the name:

>>> sysconfig.get_config_var('SO')
'.cpython-32mu.so'
>>> sysconfig.get_config_var('SOABI')
'cpython-32mu'

the problem is that it's use is aimed for internal python use, not
external libraries. Numpy used sysconfig.get_config_var('SO') - the
internel python extension name, for external libraries. This is fixed
for Numpy 1.6, when they instead check if there is SOABI and if yes,
they remove it from SO to get previous behaviour. This is actually
affecting numpy and scipy and probably others, because that use
distutils of numpy. The patch is attached to ticket:

http://projects.scipy.org/numpy/ticket/1749

and was accepted upstream/merged and it seems to work.

2011/4/22 Ángel Velásquez <angvp at archlinux.org>:
> Maybe you should try to build it by yourself first, before asking if
> there is anything important that is stoping the release...

tried and it worked, because I'm running prerelease of numpy 1.6 -
sorry, I probably should downgrade numpy to 1.5.1 as in official
package and test then, my bad, did not see that coming at first,
before I started to dig, but I've seen no issues with scipy 0.9 or
git, that's why I asked.

> As Ray explained to you, to build this version, for now it have to be
> hacked horribly, and isn't being possible, plus, I had on vacations,
> and the unique person who helped me to see this behaviour by far have
> been Ray, nobody from the community, not you btw who next time before
> ask this kind of stuff you should contact first the maintainer.

Sorry again. Though I better ask on list/start discussion when
multiple people see that flow and can add to finally find solution,
that's why lists are for, to talk, not? Especially for people who
prefer lists before forums like myself, I like to keep the copy in my
mailbox. I looked trough list and found nothing, checked forum and
found one issue marked as solved which I wasn't able to reproduce - so
I asked if there is something I don't know about and maybe help...
Anyway next time I will contact peer to peer if this is preferred way,
no problem, you have my word.

Cheers,
Andrzej.


More information about the arch-general mailing list