ENH: Added submodule numpy.random_intel

[oleksandr-pavlyk]

numpy.random_intel is MKL-based alternative for numpy.random, first
featured in the Intel Distribution for Python.

It implements all distributions of numpy.random as well as
numpy.random_intel.multinormal_cholesky, which allows for fast
sampling from multinormal distribution when the Cholesky factor
is known ahead of time.

Module numpy.random_intel implements sampling from all distributions
in a vectorized fashion, just like MKL does, which allows MKL to use
vectorized instructions for efficiency.

That means that instead of mtrand's rk_normal which produces a single
variate sampled from normal distribution, mklrand offers vrk_normal_vec,
which produces a vector of variates of requested size.

EDIT: link to mailing list design discussion on this feature: https://mail.scipy.org/pipermail/numpy-discussion/2016-July/075822.html

[oleksandr-pavlyk]

@njsmith @rkern

[charris]

Take a look at numpy/core/include/numpy/npy_3kcompat.h. Given the current organization, we might want to move that or this work. Not sure what the best organozation will be.

[charris]

Also note that PyCapsule is available in Python 2.7 and we no longer support earlier versions. In fact, PyCapsule has been used in numpy/random/mtrand/mtrand.pyx since 1.11.x was branched. As long as it is private, it no longer matters.

[oleksandr-pavlyk]

I copied mtrand/Python.pxi over to `mklrand/Python.pxi'. Thanks!

[charris]

What I notice at first glance are long lines (we like < 80 char/line) and not much documentation. There is doc/C_STYLE_GUIDE.rst.txt for numpy C code, and PEP8 for Python. I'm not sure of the best way to document without too much repetition of current random function documentation. I'm also not sure of the best way will be to handle two distinct modules covering the same material, or whether you should put this work up on PyPI as a separately loadable module -- which was pretty much my first thought.

It would be good to bring this up on the numpy mailing list for discussion: numpy-discussion@scipy.org. @rkern, @njsmith Thoughts?

[rgommers]

I've edited the PR description to link to the mailing list discussion in July on this.

As far as I can judge from a browse of this PR, this implementation is still standalone and doesn't reflect the design @rkern said he preferred right?

[charris]

On Mon, Oct 24, 2016 at 4:03 PM, Oleksandr Pavlyk notifications@github.com
wrote:

@njsmith https://github.com/njsmith @rkern https://github.com/rkern

Hmm, I don't see any replies so far. Other projects that might be
interested are scikit-learn and stats-models.

Chuck

[njsmith]

@oleksandr-pavlyk: So you know how this works... usually we use pull request comments to talk about details of code, and for higher-level discussions that affect public API (like: "should numpy have a random_intel module at all?") we switch to the mailing list. That discussion is happening there now, and you probably want to read and participate :-). The mailing list thread starts here: https://mail.scipy.org/pipermail/numpy-discussion/2016-October/076144.html

[rkern]

Hmm, I don't see any replies so far. Other projects that might be interested are scikit-learn and stats-models.

I've made my thoughts clear on the mailing list, in this incarnation and the previous thread. This should be in a separate package.

[charris]

@oleksandr-pavlyk Sounds like a separate package it is, so closing this. Thanks for your work on this.