macOS deployment problems

[ickc]

• pixell version: 0.19.0
• Python version: 3.8.17
• Operating System: macOS 11.7.8 20G1351

What I Did

See the PR in https://github.com/conda-forge/staged-recipes/pull/22976

And the build failure in https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=746884&view=logs&j=e35d4f76-8ff2-5536-d795-df91e63eb9f7&t=fa7b4b17-b6ff-5c9c-8cfc-15f888c92310

Basically it is related to this part https://github.com/simonsobs/pixell/blob/master/setup.py#L31-L56 that implements discovery of gcc. I think it should (perhaps with helpful messages) use the gcc from the environment, not auto-discovering all available gccs and choose by itself.

[JBorrow]

Hi @ickc, I have tried to address this in #235 that now uses a unified environment for MacOS and Linux builds, and checks if the compilers have the required features. Let me know if you think this will help (my experience with conda is limited), or if we need more.

[msyriac]

@ickc does your PR get us on conda feedstock correctly?

[ickc]

It is still failing at https://github.com/conda-forge/staged-recipes/pull/22976/checks?check_run_id=19330248847

Because you are assuming gcc but

On macOS, only LLVM’s OpenMP implementation llvm-openmp is supported. This implementation is used even in Fortran code compiled using GNU’s gfortran. From https://conda-forge.org/docs/maintainer/knowledge_base.html#switching-openmp-implementation

[JBorrow]

Right... Is this because of our flags? We should be using something other than -fopenmp?

[ickc]

Unfortunately the flag -fopenmp is not the problem. CC, CXX, etc. are not problems either as you only override it when not available (but in conda-forge build stage it is already set and point to clang).

The problem is somehow it cannot locate omp.h.

I no longer have access to Intel macOS so it is very difficult to trouble shot from my side (only through CI so it is slow). Also, realistically is there anyone who is using this with Intel macOS anymore? May be we can just support conda with Linux only?

[msyriac]

I have no insight on what is happening here, but I'm fairly certain there is and will continue to be a large number of users who would prefer to install pixell on an Intel Mac with conda.

[ickc]

I opened a draft PR to drive discussion. That PR doesn't completely fix it yet: #243

Others who possess an Intel Mac can help by running conda build locally which has faster turn-around to see why fails.

My suggestion is to release the Linux build on conda-forge only for now, thoughts?

[msyriac]

That sounds good to me as a start.

[defjaf]

I agree that it probably isn't the root cause of the issue, but just to record this somewhere (I suspect it's generally known, but just in case):

To use openmp under clang (on Apple silicon or intel), you have to

1. Have openmp installed (by conda or homebrew)
2. add -Xclang -fopenmp to CPPFLAGS, add -lomp to LIBS (or equivalent)

[JBorrow]

Another complication here is that we will need to change the build tool to Meson soon. So... This may all be nullified by that change. I don't know very well how Meson works...

[ickc]

Interesting, can I ask if other solutions are explored, including poetry, CMake, etc.? And about Cython used here, is pybind11 explored? Feel free to point me to links where this is already discussed.

I am glad the build tool here will be changed.

[JBorrow]

We want the build system to be one that is used within the scientific python community, so we were looking at what numpy and scipy are transitioning to now that distutils is deprecated: https://github.com/simonsobs/pixell/issues/198

[JBorrow]

AFAIK Poetry does not support fortran extensions nicely.

[ickc]

We want the build system to be one that is used within the scientific python community, so we were looking at what numpy and scipy are transitioning to now that distutils is deprecated: #198

In the "scientific Python community", the most important build system is conda which is solving some non-trivial distribution problem for non-pure Python stack. In this sense whichever the Python build system is used, it is an intermediate step towards building the conda package.

(I have no experience on Meson and I'm not against that choice without knowing more, but I just want to make sure conda is in the picture when making that choice.)

For example, CMake is known to work well there: https://conda-forge.org/docs/maintainer/knowledge_base.html#using-cmake

On the other hand, it seems people distributing both to PyPI and conda with CMake often have 2 different built systems (setup.py magic and then plain old CMake), so it would be nice if using Meson would make that easier.

It would be interesting to see this roll out with Meson. Do you have an approximate timeline for the transition to be done?

[samueldmcdermott]

+1 for making this visible on conda, though I am now having issues on Mac with pip install pixell and a GitHub workflow for a different repo that includes python -m pip install pixell has also started failing recently. Both of these cite a ModuleNotFoundError: No module named 'numpy.distutils'

[JBorrow]

Please use python < 3.12

[JBorrow]

@samueldmcdermott you should now have no problem installing.

[samueldmcdermott]

I forgot to post here but yes I checked the other day and the new version does seem to resolve this -- thanks!