Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Already on GitHub? Sign in to your account

compiling and installing scipy in AIX with xlf (Trac #1299) #1825

scipy-gitbot opened this Issue Apr 25, 2013 · 11 comments


None yet
1 participant

Original ticket http://projects.scipy.org/scipy/ticket/1299 on 2010-10-12 by trac user davide, assigned to unknown.

compiling and installing scipy in AIX with xlf does not work out of the box. Here is a quick and (very) dirty way to get it installed, although not correctly with everything one might want.
See also http://projects.scipy.org/numpy/ticket/1582

Attachment added by trac user davide on 2010-10-12: README.txt

Attachment added by trac user davide on 2010-10-12: scipy-patch.txt

Attachment added by trac user davide on 2010-10-12: numpy-patch.txt

@rgommers wrote on 2010-11-28

The NPY__INTP_FMT part of the numpy patch was probably fixed by r6646.

Most of the rest may be useful for others on AIX to get things to work, but can't be applied to trunk.

trac user davide wrote on 2011-02-04

Sure, I didn't submit this patch with the intent to be used for trunk as is. But I believe there is more you can (and should!) apply to trunk. Especially the following two things (from scipy-patch.txt). It shouldn't be too hard. Do you think it's more appropriate that I submit individual patches for each of them with more details?

  •    extra_compile_args=['-Wall'],
    +# extra_compile_args=['-Wall'], WRONG for IBM's xl

as much as I love gcc, it's a fact of the life that it's not the only compiler on the earth. And many (if not all) of the non-gcc compilers do not understand -Wall option and fail in the weirdest possible way when they see that option. So IMHO the installer should use -Wall only when it's sure gcc is going to run

  1. there are four occurrences of something like the following:
  •    X=X+PI+MAX((0.0972+0.0679_N-0.000354_N**2)/L, 0d0)
  •    X=X+PI+MAX(DBLE((0.0972+0.0679_N-0.000354_N**2)/L), 0d0)

I'm not 100% sure about what the Fortran standards say, but I don't think being more explicit like my patch "recommends" would hurt anyone.

Thanks and Regards,
Davide Del Vento,
NCAR Computational & Information Services Laboratory

@rgommers wrote on 2011-02-06

Your point 1, the -Wall compile argument, was already fixed in 308535a. I indeed don't think there's any harm in the changes of your point 2. I can send a pull request on github for that, so someone who knows more about Fortran can double check.

@rgommers wrote on 2011-02-13

Will send request after scipy actually moves to github (~10 days).

All that is left from the numpy/scipy patches is related to NO_ATLAS_INFO. Do you think there's something that can be improved in numpy.distutils related to this?

trac user davide wrote on 2011-02-18

Yes, I believe that there must be a way to enable/disable NO_ATLAS_INFO from the command line when installing. Of course my patch is not doing that, but just a quick and dirty delete of the "offending" lines, so you don't want to check it in as is.

If somebody could take the time to implement pass and parse a cli argument to enable/disable NO_ATLAS_INFO, it would be great. Otherwise affected people can always manually remove those like like I did (and hopefully they'll find this ticket helping them with it).

@pv wrote on 2011-07-09

The change to specfun is in 0547eea

trac user carletes wrote on 2011-12-06

Very helpful info here, Davide: After following your instructions I have now Scipy 0.10.0 on AIX.

Carlos Valiente carlos.valiente@ecmwf.int

trac user davide wrote on 2011-12-29

In case somebody has to install matplotlib on AIX with xlc/xlf, it has similar issues. I described how to solve those here:


Davide Del Vento,
NCAR Computational & Information Services Laboratory
Consulting Services Software Engineer
SEA http://sea.ucar.edu/

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment