[Numpy-discussion] numpy failure under Windows Vista 64 bit
Thu May 21 00:11:49 CDT 2009
Klaus Nökel wrote:
>> Klaus Noekel wrote:
>>> I doubt that the DLL was not physically present and rather suspect a
>>> dependency on some other DLL that was missing. The INSTALL.TXT
>>> unfortunately was not helpful. Can anybody please explain what other
>>> dependencies exist? Anything else I need to install?
>> This exact problem is specific to IDLE - I don't know what triggers it.
>> Today, the best solution for a 64 bits numpy on windows is to built it
>> yourself with MS compilers - the distributed one is built with mingw
>> compilers, and there still seems to be some stability problems with
>> those. Unfortunately, as the mingw debugger does not work either on 64
>> bits archs, finding the problem is quite hard.
> I don't believe that the problem is specific to IDLE. Python also crashes when I put nothing but "import numpy" in a file and execute it with python.exe.
That's not the same problem - in one case, you have a dll not found, and
in another case, a crash. I am sorry I can't tell more, but I have no
idea about what's going on: sometimes, it works, sometimes, it does not.
When it works, it runs the full test, and when it does not, it crashes
at import - but before even initializing the first numpy extension ! The
crash always happen in some conditions, and seldom in others (executing
in a cmd shell vs being executed by nosetests, for example). The problem
is difficult to track without a debugger, I am afraid (mingw compilers
do not seem to generate debugging symbols usable by MS debugger).
> Regarding the note on building numpy myself: the discussion in this forum scared me a little, because of the challenge to build LAPACK with a compatible Fortran compiler etc. That and the fact that I do not have MSVC 2008 (only 2005) keeps me from trying it. Any chance that a MS-based installer will materialize soon?
I don't intend on doing one myself, no. Note that you don't need
blas/lapack to build numpy - it is required for scipy. That's why I am
interested in making numpy work with the mingw toolchain: once it works
reliably, it will give scipy as well. Actually, I managed to build scipy
for windows 64, but as for numpy, it sometimes crash.
> Or are there any mingw-specific runtime libraries that I need to install so that the mingw-based numpy works?
No, there should not be anything else to install. There is a bug
somewhere, which needs to be found.
More information about the Numpy-discussion