Fix import error on aarch64: define 'long' using ctypes.c_long

[vsevolod-misiul]

Summary

On Linux ARM64 (aarch64) platforms, installing PyOpenGL_accelerate fails with the following error:

Error compiling Cython file:
...
    self.buffer = long( buffers )
                          ^
src/vbo.pyx:191:26: undeclared name not builtin: long

This happens because the built-in long type is not defined in Python 3 on ARM64 platforms, unlike some legacy Python 2 or x86-based assumptions. This breaks Cython compilation when trying to cast buffer identifiers.

This patch imports long from ctypes.c_long, which resolves the issue on these platforms.

Changes

• Added from ctypes import c_long as long to handle missing long type on aarch64.

Platform tested

• Linux on ARM64 (Ubuntu 22.04, aarch64)

[ChrisBarker-NOAA]

I think this is only an issue with Cython 3 -- but we should be using Cython 3 by now!

Does this seem to work with Cython3 across the board?

NOTE: I'm working on a conda-forge recipe -- so I may apply this as a patch to build .... (though pinning to Cython<3 may be the more expedient route...)

[ChrisBarker-NOAA]

Got it all working on conda-forge -- but I had to add this patch -- so it would be nice to get this merged.

I do wonder how the wheels got built ????

[vsevolod-misiul]

I think this is only an issue with Cython 3 -- but we should be using Cython 3 by now!

Does this seem to work with Cython3 across the board?

NOTE: I'm working on a conda-forge recipe -- so I may apply this as a patch to build .... (though pinning to Cython<3 may be the more expedient route...)

Yes, this fix should work across the board with Cython 3. Using ctypes.c_long is the standard way to get a platform-appropriate C long type, so it should be safe on x86_64, and other platforms, preventing similar issues there as well.

[ChrisBarker-NOAA]

Indeed -- and it is working for me -- with this patch and Cython 3 -- and with Cython <3, I got segfaults.

I think -- numpy 2 changed a bit how it handles the C long types, which might explain the segfaults. I've had other issues with that).

I curse C for defining "short" and "long" way back when, rather than the modern int32 and int64, etc ... oh well.

Anyway, no need to understand all the details -- what I think we need isthis patch and also to pin Cython to >3

Maybe add that pin to this PR?

Also: I see in the MANIFEST.in file that the generated c files are still being included in the source distro -- that probably should change.

[vsevolod-misiul]

Indeed -- and it is working for me -- with this patch and Cython 3 -- and with Cython <3, I got segfaults.

I think -- numpy 2 changed a bit how it handles the C long types, which might explain the segfaults. I've had other issues with that).

I curse C for defining "short" and "long" way back when, rather than the modern int32 and int64, etc ... oh well.

Anyway, no need to understand all the details -- what I think we need isthis patch and also to pin Cython to >3

Maybe add that pin to this PR?

Also: I see in the MANIFEST.in file that the generated c files are still being included in the source distro -- that probably should change.

As you requested, I've added the version pin to require cython>=3.0.0 in pyproject.toml and just pushed the change. This should ensure robust builds going forward.

[ChrisBarker-NOAA]

Thanks! now we just need a merge ...

[vsevolod-misiul]

Thanks! now we just need a merge ...

I've been looking forward to this for a month. 😄

[ChrisBarker-NOAA]

@mcfletch: This is actually pretty bad -- pip install of accelerate on OS-X Intel fails for this reason.

There is no wheel for OS-X Intel -- so ti tries to build from source.

Also -- I'm curious how the wheels that are there were successfully built, and if they actually work (I could build with Cython < 3, but got segfaults when trying to run the tests).

BTW -- I'm sure there's a couple of us that would be glad to help with maintenance of you don't have the time.

[ChrisBarker-NOAA]

NOTE:

I've got a build for conda-forge working:

https://anaconda.org/conda-forge/pyopengl-accelerate

But I needed to apply this patch to get it to build -- see:

https://github.com/conda-forge/pyopengl-accelerate-feedstock