BLD: add ppc64le, aarch64, s390x builds

[mattip]

Using the recent updates to multibuild, this add multilinux2014-compatible tarballs for aarch64, ppc64le, and s390x. appveyor builds are disabled, let's see if travis likes this first

[tylerjereddy]

This will be pretty cool when it is ready to go! Should remove the manual building on the other platforms, etc.

[mattip]

rebased, squashed, and added more set -v to see why the builds are failing

[mattip]

Removed the WIP label. The changes can be summarized as:

• update the multibuild repo
• update the OpenBLAS repo, even though the build checks out its own repo, it is confusing to me to have that dissonance. If reviewers request, I can revert this change
• tweak some build scripts, especially the duplicate get_distutils_platform function xref The version of get_distutils_platform here differs from the one in multibuild gfortran-install#2
• add more architectures to the travis builds

I think this is ready for review.

[mattip]

It seems strange that the other architectures all build in <10 minutes, and the x86 ones take over 30 minutes. Are there some optimizations or flags missing?

[matthew-brett]

Is it just the fact that there are many more kernels for Intel? Are the other architectures also multi-kernel? Or are they being built for only one CPU architecture?

[mattip]

Ahh, right. I guess there is no log of which variant kernels are being compiled.

[mattip]

I added the _64 variants on ppc64le and s390x, for some reason the aarch64 one segfaults so I removed it for now. TBD in a future PR, maybe we need to compare with the upstream CI configuration, but this can get us started with tarballs on the rackspace instance

[tylerjereddy]

Keep in mind that the tarballs will be built for v0.3.8-dev instead of the v0.3.7 stable release used in i.e., NumPy CI. I don't think it is that big of a deal, but Pauli did recently ask me to separately build the _64 ones for v0.3.7 for a better testing match.

There is a way to do that, but I'd prefer not to unless it is really necessary, since I made a bit of a mess of it yesterday with the hash change for some v0.3.7 tarballs (they can easily get regenerated with a new hash if not careful).

[tylerjereddy]

The NumPy CI would just need to be adjusted to check for the correct/bumped OpenBLAS version for those new tarballs vs. stable release tarballs in some matrix entries.

[tylerjereddy]

Then when 0.3.8 stable comes out everything can sync up again.

[mattip]

waiting on MacPython/gfortran-install#2

[mattip]

seems to work: the aarch64 run built openblas-v0.3.5-605-gc815b8fb-manylinux2014_aarch64.tar.gz

[mattip]

@tylerjereddy does this look ok?

[tylerjereddy]

Ok, this looks like a reasonable expansion as the ecosystem prepares for supporting more platforms. CI is green and the merge into master should not be disruptive to downstream projects immediately since we are pointing to a 0.3.8.dev build of OpenBLAS instead of the latest stable release (so not many projects should suffer from tarball hash regenerations, etc.).

Thanks Matti---let's give this a try & hopefully we can discover any issues early on before wheels have to be distributed for those platforms, etc.