[Numpy-discussion] Weird behavior of gufuncs

Nathaniel Smith njs at pobox.com
Fri Sep 6 08:09:39 EDT 2013

On Fri, Sep 6, 2013 at 12:09 AM, Jaime Fernández del Río
<jaime.frio at gmail.com> wrote:
> Hi all,
> I am seeing some very weird behavior on a gufunc I coded.
>  It has a pretty complicated signature:
> '(r,c,p),(i,j,k,n),(u,v),(d),(n,q)->(q,r,c)'
> And a single registered loop function, for types:
> uint8, uint16, uint16, uintp, uint8->uint8.
> In general it performs beautifully well, returning the expected results
> about x10 faster than a MATLAB version using a 12 core CPU (!) But today I
> found out that for certain sizes of the inputs, it doesn't even call the
> loop function and returns a zeroed output array immediately.
> I am using Python 2.7.5 and Numpy 1.7.1 under Windows 7, and have seen the
> behavior in both x64 and x86 versions. The thresholds vary, and I don't
> fully understand the logic behind it, but there is a clear threshold of
> sizes: keeping all other inputs the same, one byte more in the other
> argument and it returns all zeros, one byte less and everything works fine.
> The thresholds are different for 64 and 32 bit versions.
> I have a strong suspicion that this may be a NumPy 1.7 bug, because a
> colleague running MacOS and NumPy 1.8 compiled from the development branch
> isn't seeing any of this issues I have. It could of course be a Windows
> thing, but somehow that seems less likely, especially since there are other
> bugs in gufuncs that have been fixed in 1.8 but haven't been backported to
> 1.7. Tomorrow I am going to try and get access to a Linux box with Numpy 1.7
> to try to reproduce it there.
> Does this sound familiar to anyone? Any known (and hopefully solved) issue
> that could trigger this behavior? A search for ufunc in github's issue
> tracker didn't give me any meaningful leads...

Doesn't ring any bell directly, but as you note there are definitely
some major gufunc bug fixes in 1.8 and master, so your best bet on
tracking it down might be to (a) check whether it occurs in master,
(b) if not, bisect it.


More information about the NumPy-Discussion mailing list