Use meson to compile the fortran modules.

kalekundert commented 1 year ago

This PR moves to the meson build system for compiling the fortran extension modules. I mostly followed the instructions on this page, with some inspiration taken from the scikit-image project. This would fix #22.

Here is an example of the install process:

$ cd /path/to/py3nj
$ python -m venv venv
$ . venv/bin/activate
$ pip install --upgrade pip
$ python --version
Python 3.10.0

$ pip --version
pip 23.2.1 from /home/kale/research/software/forks/py3nj/venv/lib/python3.10/site-packages/pip (python 3.10)

$ pip install .
Processing /home/kale/research/software/forks/py3nj
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
Collecting numpy (from py3nj==0.1.2)
  Obtaining dependency information for numpy from https://files.pythonhosted.org/packages/71/3c/3b1981c6a1986adc9ee7db760c0c34ea5b14ac3da9ecfcf1ea2a4ec6c398/numpy-1.25.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata
  Using cached numpy-1.25.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (5.6 kB)
Using cached numpy-1.25.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (18.2 MB)
Building wheels for collected packages: py3nj
  Building wheel for py3nj (pyproject.toml) ... done
  Created wheel for py3nj: filename=py3nj-0.1.2-cp310-cp310-linux_x86_64.whl size=114741 sha256=c4414c46343473a4579435f21a4c264772303c30b6b4f379f4d5c32b3f5dba5b
  Stored in directory: /tmp/pip-ephem-wheel-cache-jyt0ibue/wheels/54/46/c8/698fe6f7850e3962dd32c95d6a4166d337c037a9bcbba56d9b
Successfully built py3nj
Installing collected packages: numpy, py3nj
Successfully installed numpy-1.25.2 py3nj-0.1.2

$ pip install pytest
$ python -m pytest
============================= test session starts ==============================
platform linux -- Python 3.10.0, pytest-7.4.0, pluggy-1.2.0
rootdir: /home/kale/research/software/forks/py3nj
collected 22 items

tests/test_3j.py ............                                            [ 54%]
tests/test_6j.py .......                                                 [ 86%]
tests/test_9j.py ..                                                      [ 95%]
tests/test_cg.py .                                                       [100%]

============================== 22 passed in 1.36s ==============================

Some caveats:

While I think I have a good understanding of python packaging in general, I don't know anything about meson. So I'm not sure I did the meson-related things in the best way possible.
This branch won't work yet for python==3.12.0rc1, because pip install numpy doesn't work yet in that environment: numpy/numpy#23808. The issue is that all the numpy distributions uploaded to PyPI still depend on distutils. Presumably this will be fixed soon, but for now I did all my testing with python==3.10.
Editable installs don't work. The mesonpy build backend is supposed to setup some clever system where recompiles extension modules on the fly when you import them, as described here. But for some reason it fails to find the numpy headers when it does this. I couldn't figure out why. It's possible that this is a bug in mesonpy, since the non-editable install does work. For troubleshooting this, it's probably useful to set the $MESONPY_EDITABLE_VERBOSE environment variable.

fujiisoup commented 1 year ago

This looks good to me. I enforced pip=23.2.1 explicitly in CI, and it works perfect. If you think it is ready, I will merge it.

kalekundert commented 1 year ago

I think it's ready to merge. Thanks for getting this done so fast!

fujiisoup commented 1 year ago

Merged. Thank you for your contribution!

fujiisoup commented 1 year ago

Now the new version is on Pypi. Thank you again for your contribution.

fujiisoup / py3nj

Use meson to compile the fortran modules. #24