Emerge scipy fails, blas not found despite being installed

[andersas]

Hi

I've recently had some problems installing scipy (0.12.1 and 0.13.2-r1).

When I emerge scipy, I get the error

Code: Select all

Blas (http://www.netlib.org/blas/) libraries not found.

I tracked down the problem to be a combination of numpys distutils and the site.cfg file generated by the scipy ebuild.

In the scipy ebuild site.cfg (/var/tmp/portage/sci-libs/scipy-0.13.2-r1/work/scipy-0.13.2/site.cfg), atlas is specified twice:

Code: Select all

[blas]
include_dirs = 
library_dirs = :/usr/lib64
blas_libs = cblas,atlas,blas,atlas,
[lapack]
library_dirs = :/usr/lib64
lapack_libs = lapack,atlas,blas,atlas,cblas,atlas,

Removing the duplicates enabled a successful install.


Why does the ebuild specify atlas twice (or three times for lapack)? Does it only do this on my system?
(please somebody try emerging scipy and check /var/tmp/portage/sci-libs/scipy-0.13.2-r1/work/scipy-0.13.2/site.cfg)

The file that does the actual check is /usr/lib/python3.2/site-packages/numpy/distutils/system_info.py.
In the function _check_libs, it simply checks to see if the number of found libraries equals the number of
required libraries (if len(found_libs) == len(libs)). If a library (e.g. atlas) is specified twice, this fails.


Am I the only one who is unable to install scipy because of this?


The problem is also present in the ebuild for numpy. Here, the site.cfg file also contains duplicates,
but this does not prevent a successful compilation. It only prevents numpy from using the optimized lapack and blas routines.


Anders Søndergaard

[zhMFe3]

yet, doesn't it look like a list of (PROPERTY,VALUE) pairs? I.e.: property1,value1,property2,value2
so...

Code: Select all

[blas]
include_dirs = 
library_dirs = :/usr/lib64
blas_libs = cblas,atlas,blas,atlas,
[lapack]
library_dirs = :/usr/lib64
lapack_libs = lapack,atlas,blas,atlas,cblas,atlas,

is probably meant to be read like this:

Code: Select all

blas_libs = {cblas=atlas, blas=atlas}

like the eselect values for the same modules.

/usr/lib64 is the correct path to those libraries.

So, is there a related scipy bug report?