[Numpy-discussion] building numpy/scipy

[David Cournapeau]

Mike Landis wrote:
> Cygwin is present, so not just the dumbed down Windows CMD available.

You should not use cygwin: if you use cygwin, it will build numpy
against the cygwin python, or worse, will be very confused, because you
will mix cygwin compilers and mingw compilers.  Unless you want to build
numpy for cygwin python, you should not try anything from cygwin. Not
that it is impossible, but there are many warts to avoid, and it does
not worth the pain unless you are a numpy developer.

>
> I ran the numpy-1.2.1 superpak.  Verified that it installed (cause you
> don't get near as much output as you do from a shell prompt) by running:
>
>    "
> python -c 'import numpy; print numpy.__version__
> '
> "

See, it is easier  :)

>
> and got the numpy version number back.  Found the site.cfg that the
> installer left in the numpy package directory (it picked up on my
> ATLAS install, but not on MKL) and copied it into
> d:\temp\scipy-0.7.0b1 and ran "python setup.py install" there.

This will not work: you need blas/lapack for scipy - the site.cfg in
installed numpy refers to a blas/lapack which was used when I built the
binary installer - and it not installed (blas/lapack is statically
linked, as dynamically linked libraries is too difficult on windows).
Please install from the scipy binary installer instead:

http://sourceforge.net/project/showfiles.php?group_id=27747

Trust me, it will take you less time.

> Lots of positive looking output, but it ultimately crapped out. 
> Here's the tail end of that transcript:

Same python bug as before: the problem is that for some reason, python
checks the version of your toolchain (here binutils), and refuses to
build when the version is not the one expected. The easiest fix is to
follow Joseph suggestion. But again, you need a blas/lapack first.

David