meeting notes from mingwpy progress call

[Ralf Gommers]

Hi all,

Here are some brief minutes from a call we had last night.

Attendance: Carl Kleffner, Matthew Brett, Olivier Grisel, Ralf Gommers


Technical discussions:
- Carl tried to build Scipy against Matthew's numpy ATLAS wheel; ran into an issue with static gcc symbols that shouldn't be exported by the ATLAS dll. Carl will send def file to Matthew to resolve it.
- Carl made some changes to MingwPy to map extended double to double always. Played with a flag to configure this behavior, but this can/will be left for later. This is a technical choice that should be documented on the MingwPy website.

- Discussed runtime switching of BLAS/LAPACK libraries via a dll with special name (like numpy-blas.dll) that uses ctypes to pick ATLAS, OpenBLAS, etc. Could be an interesting idea, but for later. Put this kind of thing on an idea list / issue on the Github repo. (done today at https://github.com/mingwpy/mingwpy-wheel-builder/issues/3)

State of repo:
- Carl to check in latest changes

- Olivier will then rebuild the toolchain, check it works and "officially confirm" that phase 1 milestone is complete.

- Some cleanup/rebase of commits would be nice, to make the set of patches on top of upstream mingw-w64 as clean as possible. It may be possible to push some build options upstream, Carl will look at that.

Next steps:

- numpy.distutils patch

- Setting up CI for Numpy + MingwPy on Appveyor. Olivier can help with Appveyor, Ralf to look at $$ to be able to do parallel test runs on Appveyor.

- Scipy CI a bit more difficult because no Appveyor. Matthew to look into access to Windows CI machine from Berkeley.

Discussion on putting up PyPi wheels:

- Important to note that this is an orthogonal issue to MingwPy. Carl will continue as planned based on OpenBLAS.

- Numpy ATLAS wheels are ready, could go up tomorrow.
- Matthew to email the Numpy list with the proposal.
- Make clear that these are slow-ish but reliable, and meant as a stopgap until we have something more performant. Whether those use BLIS or OpenBLAS is not that relevant for the decision to put ATLAS based wheels up now.
- No Scipy wheels for now.

Cheers,
Ralf