[SciPy-dev] segmentation with scipy+icc during test

Matthieu Brucher matthieu.brucher@gmail....
Fri Dec 5 15:29:43 CST 2008


> no. i modified the numpy distutils a little bit so it would accept the
> -Bstatic and -Bdynamic switches, for everything else i used standard
> numpy buld procedure (i also change intelccompiler.py a bit so numpy
> would use icpc as cxx compiler)

I think a lot of people would love to what you did ;)

>> Did you do the same for scipy ?
> i could be completely wrong, but as far as i can tell, scipy reuses the
> numpy distutils and site.cfg, and looking at the build logs, the same
> options were used.

In this case, yes, it should ;)

>> Because the first time I ran
>> into this issue was with the FFT functions.
> when i say numpy works, i mean the unittests don't fail. (but i have set
> the fftw3 intel interfaces in the numpy site.cfg. i haven't bothered yet
> to check what scipy uses for fft, as i first want to reproduce the
> segfault outside the test environment.)
>
>> And the guy is me.
>>
> ah. well, thanks for contacting intel for their advice. it saved me some
> work ;)

No problem, I wanted to have an answer as well.

>>  it's scipy that causes the segfault (although
>> > scipy uses the same build options as numpy. (i'm not saying that the
>> > segfault has nothing to do with using icc/ifort/icpc, just that it's not
>> > this specific problem).
>>
>> Another solution:
>> - make a simple Python script with the test inside
> i tried but failed. and i don't want to figure out the whole nose test
> environment just to do this ;)

Then this is strange. I thought nose would help us in that matter.
I don't have icc and mkl on my own box now, so I cannot try myself. If
you give us the diff for numpy, we may eventually help you.

>> - call gdb python
>> - call r script.py
>>
>> and tell us where it crashes ;)
>>
>> > i actually just found out that while running the scipy tests, gcc is
>> > called. i always assumed that scipy takes whatever numpy uses. (it
>> > seemed to work pretty well). maybe the tests don't?
>>
>> It may be weave that calls gcc.
>>
> but this should be 100% harmless? (i'm not familiar with the inner
> working of scipy)

Completely. Weave is just a way to compile on the fly C++ code, and it
has no impact on fftpack.

Matthie
-- 
Information System Engineer, Ph.D.
Website: http://matthieu-brucher.developpez.com/
Blogs: http://matt.eifelle.com and http://blog.developpez.com/?blog=92
LinkedIn: http://www.linkedin.com/in/matthieubrucher


More information about the Scipy-dev mailing list