[SciPy-User] test issues with 0.10
Mon Nov 14 14:56:36 CST 2011
On Mon, Nov 14, 2011 at 9:17 PM, Paul Anton Letnes <
> On 14. nov. 2011, at 20:48, Ralf Gommers wrote:
> > On Mon, Nov 14, 2011 at 3:20 PM, Paul Anton Letnes <
> firstname.lastname@example.org> wrote:
> > Hi, when trying to build scipy 0.10 with python 2.7 (from homebrew/
> python.org), numpy 1.6.1 (built manually) on Mac OS X 10.7.2 and with gcc
> 4.6.0 (built from source), scipy.test() breaks. Any ideas?
> > Was everything including python itself built with gcc 4.6.0?
> No; I'm sorry about the screwup on version numbers. Here are the correct
> i-courant ~/src/scipy-0.10.0 % python
> Python 2.7.2 (default, Oct 9 2011, 18:03:13)
> [GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
> Type "help", "copyright", "credits" or "license" for more information.
> so python was built with Apple GCC 4.2.1. The same goes for numpy. The
> reason is that numpy and scipy setup tries to send the -faltivec flag to
> the compiler, which only Apple's gcc accepts. Same gcc version with scipy.
> The gfortran version, however, is 4.6.0. Apple's gcc installation does not
> include gfortran, so that's why I installed a separate gfortran.
That's not the correct gfortran. Through homebrew you should be able to get
the right one, or download it directly: (
http://r.research.att.com/gfortran-lion-5666-3.pkg). It is version 4.2.4,
but may report itself as 4.2.1.
The -faltivec thing was a numpy distutils issue that is fixed in numpy
master. It shouldn't have given you any problems, it just caused you to
miss out on SSE optimizations.
> (As a side question, what's a reliable way of getting all this (recent
> gfortran, numpy, scipy) installed? Homebrew is not very helpful with
> alternative gcc installations, neither is Apple…)
On OS X 10.7 this is still a bit painful due to the switch to llvm-gcc as
default compiler. But if you get the right gfortran, linked also from
http://scipy.org/Installing_SciPy/Mac_OS_X, and set gcc as default compiler
you should be good to go. Of course there are binary installers too that
work on 10.7 if you don't want to fight with compilers. That's just a
matter of downloading Python from python.org and numpy/scipy from
> > What was the last version of scipy that did work for you?
> 0.10.b2 was working fine, although not all tests passed. I figured it
> would be ironed out before release. (I did not use any of the problematic
> sub-modules anyway.)
> With the same compilers, that's odd. No changes to fftpack code or tests
went in since beta 2.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the SciPy-User