On 2/7/07, Jonathan Griffitts <griffitts_lists@comcast.net> wrote:
In message <45C510E8.3060602@comcast.net>, John Hassler
< hasslerjc@comcast.net> wrote
>Jonathan Griffitts wrote:
>  Hi!
>  I have installed SciPy 0.5.2 for Python 2.5 on several Windows
>  computers, using the precompiled binary scipy-0.5.2.win32-py2.5.exe.
>  This works fine on some, but on others it crashes Python with an
>  Illegal
>  Instruction exception. The crash is easy to find by running the
>  test()
>  suite or by attempting to make any use of scipy.integrate.quad.
>  Digging into it, I see that the exception comes from _quadpack.pyd,
>  and
>  it dies at a MOVSD instruction. I believe MOVSD is an SSE2
>  instruction
>  that is only implemented on the more recent CPUs from both Intel and
>  AMD.
>
>It's a problem with the Athlon. It also happened with a previous
>version of SciPy this summer.

This is not just an Athlon issue.  It also fails on a Pentium 3
processor.

To reiterate, this binary (scipy-0.5.2.win32-py2.5) uses SSE2
instructions, which are available only on the newer CPUs from both AMD
and Intel.

Thanks for the info. We linked the 0.5.0 and (I think) 0.4.9 Win32 binaries to the ATLAS library labelled "ATLAS-P2" that doesn't require SSE2 instructions. Sorry that we've reverted to the SSE2 builds in the the last release or two. Rebuilding takes time, and it may be more productive for us to wait for the next release, when we'll try to get it right. Meanwhile I've changed the release notes on the SourceForge page to reflect the SSE2 requirement. I'll also explain the situation on the SciPy.org Download page (as soon as I can log in ;) ...

-- Ed