[Numpy-discussion] Re: Summer of Code 2006
oliphant.travis at ieee.org
Sat Apr 15 10:51:18 CDT 2006
Albert Strasheim wrote:
> Hello all
> Let me start by saying that the build system works fine for what I think is
> the default case, i.e. building NumPy on Linux with preinstalled LAPACK and
> BLAS. However, as soon as you vary any of those parameters, things get
It also builds fine with mingw and pre-installed ATLAS (I do it all the
time). It also builds fine with no-installed ATLAS (or LAPACK or BLAS)
with mingw32 and Linux. It also builds on Mac OS X. It also builds on
Solaris, AIX, and Cygwin. Work also went in recently to make sure it
builds with a Visual Studio Compiler (the one Tim Hochberg was using...)
So, I think it's a bit unfair to say that varying from only a Linux
build causes "things to get interesting". Definitely there are
configurations that can require a specialized site.cfg file and it can
be difficult if you build with a compiler that was not used to build
Python itself. But, it's not a one-platform build system. I just
want that to be clear.
Documentation on the site.cfg file could be more prominent, of course,
and this was aided recently by the addition of an example file to the
The expert on the build system is Pearu Peterson. He has been very
responsive to suggested fixes and problems that people have
experienced. Robert Kern, David Cooke, and I also have some
familiarity with the build system enough to assist from time to time.
All help is greatly appreciated, however, as I know you can come up with
configurations that do cause things to "get interesting." The more
configurations that we get tested and working, the better off we will
be. The more people who understand the build system well enough to
help fix it, the better off we'll be as well. So, I definitely don't
want to discourage any ideas you have on improving the build system.
Thanks for being willing to dive in and help.
> I've spent the past couple of days trying to build NumPy on Windows with
> ATLAS and CLAPACK with MinGW and Visual Studio .NET 2003 and VS 8. I don't
> know if it's just me, but this seems to be very hard. This could probably be
> partly attributed to the build systems of these libraries and to the lack of
> documentation, but I've also run into problems with NumPy build scripts.
> For example, the inclusion of the gcc library in the list of libraries when
> building Fortran code with MinGW causes the build to break. Also, building
> FLAPACK from source causes the build to fail (too many open files).
> While these errors on their own aren't particularly serious, I think it
> would be helpful to set up an automated system to check that builds of the
> various configurations NumPy supports can actually be done. There are
> probably a few million ways to build NumPy, but it would be nice if we could
> make sure that the N most common configurations always work, and provide
> documentation for "trying this at home."
> I also think it would be useful to set up a system that performs regular
> builds of the latest revision from the SVN repository. I think anyone
> attempting this is going to run into a few issues with the build scripts,
> especially when trying to build on multiple platforms.
> Things I would like to get right, which I think are much harder than they
> need to be (feel free to disagree):
> - Windows builds in general
> - Visual Studio .NET 2003 builds
> - Visual C++ Toolkit 2003 builds
> - Visual Studio 2005 builds
> - Builds with ATLAS and CLAPACK
> The reason I'm interested in the Microsoft compilers is that they have many
> features to help us make sure that the code is correct, both at compile time
> and at run time.
> Any comments? Anybody building on Windows that finds the process to be
> completely painless?
> This SF.Net email is sponsored by xPML, a groundbreaking scripting language
> that extends applications into web and mobile media. Attend the live webcast
> and join the prime developer group breaking into this new coding territory!
> Numpy-discussion mailing list
> Numpy-discussion at lists.sourceforge.net
More information about the Numpy-discussion