Relaxed stride checking in numpy 1.12 causes bottleneck unit test failures

[fmaussion]

Hi, this is what reproduces the problem on my machine (linux mint):

In [1]: import numpy as np

In [2]: import bottleneck as bn

In [3]: print('numpy:', np.__version__)
numpy: 1.12.0

In [4]: print('bottleneck:', bn.__version__)
bottleneck: 1.2.0

In [5]: a = np.ones((3, 2))

In [6]: np.sum(a[:, [1]]), bn.nansum(a[:, [1]])
Out[6]: (3.0, 1.0)

In [7]: np.sum(a[:, [1]].copy()), bn.nansum(a[:, [1]].copy())
Out[7]: (3.0, 3.0)

This has been hitting the xarray test suite on debian.

Thanks!

[kwgoodman]

It works fine with numpy 1.11 which is what bottleneck 1.2 supports. So my guess is that you are using numpy 1.12. My second guess is that the unit test failures with numpy 1.12 is due to the change in numpy's relaxed stride checking. My third guess is that I don't understand what relaxed stride checking is. (That last one is not a guess.)

I've added your example as a unit test.

[kwgoodman]

OK, I made the fix in master.

[fmaussion]

That was fast! Thanks a lot.

[kwgoodman]

Well, you rang the fire alarm. Thanks for reporting and thanks for a simple example.

[shoyer]

Consider issuing a bug fix release including this fix? I don't think it break any NumPy 1.11 users...

[kwgoodman]

A bug in numpy 1.12.0 prevented me from making a bug fix release. But it should be possible now that 1.12.1 is out.

Ugh. The current blocker is here: #166.

[kwgoodman]

See #168