[SciPy-Dev] test_qz_double_sort seg fault - upgrade mingw for binaries?

Ralf Gommers ralf.gommers at gmail.com
Thu Aug 9 17:54:58 EDT 2012


On Wed, Aug 8, 2012 at 11:28 PM, Ralf Gommers <ralf.gommers at gmail.com>wrote:

>
>
> On Wed, Aug 8, 2012 at 9:13 PM, Skipper Seabold <jsseabold at gmail.com>wrote:
>
>> Hi,
>>
>> There is a failure that's holding up the release reported here.
>>
>> projects.scipy.org/scipy/ticket/1717
>>
>> AFAIK, this test seg faults only on the official SSE2/SSE3 windows 32-bit
>> binaries that are built using mingw 3.14 with gcc 3.4. It's only the double
>> precision test that has a callback function that break.
>>
>> https://github.com/numpy/vendor
>>
>> The suspicion is that it's a compiler issue since the tests pass on other
>> architectures and with msvc/mkl on windows as reported by Christoph. I
>> didn't see anything in the LAPACK change sets to suggest it's an issue with
>> the ATLAS/LAPACK, though I suppose this is still possible.
>>
>> How difficult would it be for someone to upgrade mingw/gcc and see if the
>> seg fault goes away? I could try though I'm not set up for this and am not
>> that familiar with cygwin/mingw.
>>
>> Are there any other mingw issues on windows to be aware of that prevents
>> using a newer compiler?
>>
>
> We can't use gcc 4.x for binaries we want to distribute, see
> http://thread.gmane.org/gmane.comp.python.numeric.general/46536/focus=47813
>
>
>>
>> Any other way to resolve this?
>>
>
> Disabling either the test or the sort keyword of qz() until it's resolved
> is one possibility.
>

Skipper has updated the test so it doesn't crash:
https://github.com/scipy/scipy/pull/285. There may still be a problem for
some combinations of inputs and sort functions, but I think this shouldn't
hold up the 0.11.0 release any longer. So unless someone has a better idea
or a fix, I intend to merge that PR and prepare 0.11.0 RC2 in the next few
days.

Ralf
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scipy-dev/attachments/20120809/3043e191/attachment.html>


More information about the SciPy-Dev mailing list