What to do about PyCXX?

[mdboom]

The Sourceforge project seems to be gone. Is it abandoned? If so, what do we do about it?

[jkseppan]

What are the alternatives? Cython, Boost.Python, Pyplusplus, Swig?

[mdboom]

I'm not a huge fan of any of the alternatives, except possibly the raw Python/C API.

A viable alternative is also to keep using our own fork of PyCXX, but make a separate release so that distro packagers don't complain.

[efiring]

In my view, the only potentially acceptable alternatives to continuing with our own PyCXX are Cython and raw. Cython works well for C, and is solidly established in our ecosystem. I know work has been done to improve C++ support, but I don't know how far it has gotten. If it is adequate for our C++ interfacing needs, then it would be my preference. Longer term, I don't like the idea of mpl needing to support a full-blown project like PyCXX, but it might be the least-bad option.
I suspect @ianthomas23 would vote for phasing it out in favor of raw APIs, correct?

[ianthomas23]

@efiring: Yes, I would vote for phasing out PyCXX. In the interests of seeing how hard it would be to support PyCXX I did attempt to add numpy arrays to it. Although I made some progress, I concluded that the code was too esoteric for occasional maintenance, i.e. I would not choose to support it.

As for replacements, I agree with the raw Python/C API and/or Cython as we concluded in the discussion on matplotlib-devel some time ago. Personally I would opt for the raw Python/C API as I am familiar and comfortable with it. I don't see the transition from PyCXX to raw Python/C API as being any big deal, as PyCXX extensions already have to use the raw Python/C API for numpy arrays anyway. Incidentally, I initially wrote my first mpl extension using the raw Python/C API but was persuaded at the time to convert it to PyCXX as that was the preferred approach!

If we want to standardise on Cython instead, I would be happy with that but as I have no experience of it I would need someone to convert one of our existing extensions from PyCXX to Cython to show me what to do, and then I would follow that.

[pelson]

I know work has been done to improve C++ support, but I don't know how far it has gotten.

From my limited experience, it seems the C++ support is very good at this point. FWIW Cartopy is making use of Cython to wrap a C library (GEOS) and some C++ code in the same module https://github.com/SciTools/cartopy/blob/master/lib/cartopy/trace.pyx.

[matthew-brett]

Cython is also being actively developed and numpy, scipy are using Cython,
so the Cython route will probably give you a better chance of help and new
developers.​

[breedlun]

Could Numba work here? (http://numba.pydata.org)

[efiring]

@stretch97: No, numba is for speeding up loops in Python that can't be vectorized efficiently with numpy. We are using CXX mainly to wrap external libraries, together with some internal C++ code. It's a completely different operation.

[mdboom]

Just to update the original issue. master is now CXX-free with the exception of _tri.cpp, which should get done, given time.

[mdboom]

Closing. CXX is gone.