[Wheel-builders] BLAS/LAPACK for manylinux1 wheels

[Matthew Brett]

Hi,

On Mon, Mar 7, 2016 at 4:34 AM, Olivier Grisel <olivier.grisel at ensta.org> wrote:
> Hi Matthew,
>
> It is my understanding that you have built manylinux1 wheels for the
> latest numpy and scipy releases at:
>
> https://nipy.bic.berkeley.edu/manylinux

I have.  I was testing out the manylinux build system, trying to get a
set of wheels we can use on travis-ci for different versions of
Ubuntu.

> Those builds embed openblas but I don't see how to find the version
> number. Other downstream projects such as scikit-learn might need to
> run on the same openblas version.
>
> How did you build those wheels, do you have travis config somewhere or
> do you use docker locally?

I am using docker locally for now.  My build scripts are here:

https://github.com/matthew-brett/manylinux-builds

> Do you plan to update the travis config script for the matrix entry
> that generate wheels to generate manylinux1 wheels instead?

You mean, within the travis-ci docker container, open up another
docker with the manylinux image?

> https://github.com/numpy/numpy/blob/master/.travis.yml#L62
>
> At the moment those dev wheels are uploaded to:
>
> http://travis-dev-wheels.scipy.org/
>
> The goal is to make it possible for downstream projects to test
> against numpy master on travis but they need to be careful to install
> atlas as done in numpy's .travis. Using manylinux1 wheels would make
> it easier to setup CI for dowstream project without having to worry
> about atlas.

Yes, true.

> Do you plan to generate  a clib_openblas wheel in the longer term?
> This would reduce the size of scipy, pandas and scikit-learn
> manylinux1 wheels and increase installation speed for users of the
> scipy stack.

There's the main sticking point(s).

First - should we be using openblas for the manylinux wheels?  We know
of at least one bug that is not yet fixed in openblas master, and our
vague suspicion is that there may be quite a few more:

https://github.com/xianyi/OpenBLAS/issues/783

As you know, for Windows, we dropped back to ATLAS for now :

https://github.com/numpy/numpy/issues/5479
https://mail.scipy.org/pipermail/numpy-discussion/2016-March/075125.html

Second - how should we go about naming / building / distributing our
external (non-Python) dependencies?

Cheers,

Matthew