[Numpy-discussion] dot and deallocation bug?
Travis Oliphant
oliphant.travis at ieee.org
Mon Mar 13 23:39:01 EST 2006
Travis Oliphant wrote:
>> Hi folks,
>>
>> I've run across an odd error in my numpy install related to numpy.dot
>> with python lists of floats.
>>
>> Now, I haven't updated from svn in a while (I'm currently using numpy
>> to do some data analysis, so I'm not changing things until that's
>> done). Perhaps someone could confirm whether this is still a bug in
>> the current version?
>>
>> Here's how I reproduce the error (OS X 10.4.5), python 2.4.2:
> This is a bug in the error-handling section of the optimized BLAS
> code. I see it too.
>
> The reason you don't see it for non floating point and/or higher than
> 2-d is that the there is an optimized dot function that is taking over
> for floating point and complex floating point code which uses the BLAS.
> So, you are exercising different sections of code. The optimized
> _dotblas.c file has been the source of quite a few bugs. The slower
> but less buggy version is always available as numpy.core.multiarray.dot
>
> Thanks for finding this one.
Update:
This should be fixed in SVN now. At least the segfault I was seeing
disappeared. There was a missing INCREF for a data-type descriptor.
This would definitely cause those kinds of errors.
-Travis
More information about the NumPy-Discussion
mailing list