numpy & non-Apple gcc?

Michael Dickens michaelld at macports.org
Wed Sep 22 19:16:26 PDT 2010


After playing around with py26-numpy a bit more as well as reading through the NumPy/SciPy MacOS X install guide < http://www.scipy.org/Installing_SciPy/Mac_OS_X >, I think I've reached a compromise that seems to meet the developers' requirements (linking fortran-created libraries with fortran, cc-created libraries with cc) while allowing for easy universal install:

use Apple's GCC for everything that requires CC, and then whatever fortran the user selects via a +gcc4x variant for linking with MacPorts' atlas if that variant is selected (or, some other fortran variant if and when other variants are created for other math ports).

* This is the recommended (by NumPy folks) way of doing this compile.

* This method allows for +universal in the usual way internally (via -arch flags), without requiring the muniversal portgroup & fixes to it & 'port' itself, so it can be used now instead of waiting for the 1.9.2 release.

* The only changes needed are to the numpy Portfile and 1 additional patch to the source.  I have attached the "svn diff" file to ticket #24942 < https://trac.macports.org/ticket/24942 >.  Please note that it creates a new patchfile in "files", as well as modifies the Portfile.

* Keeps the same default variants as current: +atlas +gcc44.

* can be build without gcc4X at all, by using -atlas variant (and thus linking with Apple's Accelerate framework, which has Apple-gcc linkage).  When -atlas is specified, none of the +gcc4X variants even show up as options :)

* easily updatable to newer gcc4X or other fortran compilers.

For those of you interested in this topic, I'm interested in hearing from you if these changes work or not.  Thanks! - MLD


More information about the macports-dev mailing list