API: Support Matrix Multiplication Ops (@)

[sinhrks]

Python 3.5 is being released on Sep.13. Maybe we can wait for v0.17?

• http://legacy.python.org/dev/peps/pep-0465/
• numpy fix: Quick and dirty matmul numpy/numpy#5878

[jreback]

we need Travis to support 3.5 before can do this
iirc for 3.4 this support took some time

[shoyer]

We can certainly test @ by calling __matmul__ explicitly.

But we certainly should wait until numpy 1.10 is out -- until then we don't even have a np.matmul implementation to use internally.

[charris]

Note that the current numpy implementation does not support object arrays.

[bnaul]

I was a bit surprised to find that this isn't already supported since I've gotten so used to @ in numpy. This case seems a bit simpler since there's no analog of scalars; does that mean that just setting __matmul__ = dot is sufficient or am I missing some other subtlety?

[shoyer]

does that mean that just setting matmul = dot is sufficient or am I missing some other subtlety?

There is a small amount of subtlety in returning NotImplemented when the other argument is the wrong type (standard for binary operators), but yes this should be pretty straightforward.

[charris]

Note that @ and dot broadcast differently. I don't know if that means anything for pandas ...

[shoyer]

In pandas we only have 1D and 2D data structures (ignoring Panel, which is deprecated), so the broadcasting behavior is irrelevant.