Refactor cibuildwheel setup

[mwtoews]

There are a few aims in this PR:

• Move cibuildwheel data from GitHub Actions variables to pyproject.toml
• Build each platform, but allow to work with any Python 3 version, since this is a Pure Python project
• Use tox to run pytest for all Python versions
• For Linux, remove duplicated rtree/lib; close Linux binary wheels have two copies of libspatialindex #270
• Modify finder to look for auditwheel-built shared library (using importlib.metadata.files)

The repair_wheel.py script was modified from cmake-python-distributions. Admittedly, there are several things that I don't understand, but have left alone. It uses wheel unpack and wheel pack to modify the binary wheels.

[adamjstewart]

After this is merged we should consider cutting a 1.0.2 release so that we'll have wheels that can be installed on Python 3.12.

[mwtoews]

The failures with manylinux2014_i686.manylinux_2_17_i686 are puzzling, since there were no issues with the most-recent build. And for some reason the platform tags (separated by .) are flipped from manylinux_2_17_i686.manylinux2014_i686. Does the order of these matter? I'm able to reproduce these locally too, so they can be investigated further.

The macos wheels are also a bit odd. The previous build had two platform tags macosx_10_9_x86_64 and macosx_11_0_arm64 while the new has only one macosx_12_0_arm64.macosx_12_0_universal2.macosx_12_0_x86_64. Does this one wheel work for older platforms? I'm unable to do much testing with these.

[mwtoews]

Oh wait, the previous builds worked for manylinux*i686 since these tests were skipped with 6763788 . The skip could be re-instated, but I don't understand the root cause of the failures yet, they could be a 32-bit/64-bit thing.

Things are running oddly smooth for Windows, however! Additional platforms could be considered.

[mwtoews]

I've enabled cp312 testing to see what happens. Currently testing is only possible on Linux, but all of the platforms show a consistent error. See #277 for this issue.

Output of error

  platform linux -- Python 3.12.0rc1, pytest-7.4.0, pluggy-1.2.0
  cachedir: .tox/py312/.pytest_cache
  rootdir: /project
  collected 38 items
  tests/test_index.py ........x............F...............                [ 97%]
  tests/test_tpr.py .                                                      [100%]
  =================================== FAILURES ===================================
  _____________________ IndexSerialization.test_interleaving _____________________
  self = <tests.test_index.IndexSerialization testMethod=test_interleaving>
      def test_interleaving(self) -> None:
          """Streaming against a persisted index without interleaving"""
      
          def data_gen(
              interleaved: bool = True,
          ) -> Iterator[Tuple[int, Tuple[float, float, float, float], int]]:
              for i, (minx, miny, maxx, maxy) in enumerate(self.boxes15):
                  if interleaved:
                      yield (i, (minx, miny, maxx, maxy), 42)
                  else:
                      yield (i, (minx, maxx, miny, maxy), 42)
      
          p = index.Property()
          tname = tempfile.mktemp()
          idx = index.Index(
              tname, data_gen(interleaved=False), properties=p, interleaved=False
          )
          hits1 = sorted(list(idx.intersection((0, 60, 0, 60))))
          self.assertTrue(len(hits1), 10)
          self.assertEqual(hits1, [0, 4, 16, 27, 35, 40, 47, 50, 76, 80])
      
  >       leaves = idx.leaves()
  /project/tests/test_index.py:475: 
  _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
  self = rtree.index.Index(bounds=[-186.673789279, 184.761387556, -96.7177218184, 96.6043699778], size=100)
      def leaves(self):
          leaf_node_count = ctypes.c_uint32()
          p_leafsizes = ctypes.pointer(ctypes.c_uint32())
          p_leafids = ctypes.pointer(ctypes.c_int64())
          pp_childids = ctypes.pointer(ctypes.pointer(ctypes.c_int64()))
      
          pp_mins = ctypes.pointer(ctypes.pointer(ctypes.c_double()))
          pp_maxs = ctypes.pointer(ctypes.pointer(ctypes.c_double()))
          dimension = ctypes.c_uint32(0)
      
          core.rt.Index_GetLeaves(
              self.handle,
              ctypes.byref(leaf_node_count),
              ctypes.byref(p_leafsizes),
              ctypes.byref(p_leafids),
              ctypes.byref(pp_childids),
              ctypes.byref(pp_mins),
              ctypes.byref(pp_maxs),
              ctypes.byref(dimension),
          )
      
          output = []
      
          count = leaf_node_count.value
          sizes = ctypes.cast(p_leafsizes, ctypes.POINTER(ctypes.c_uint32 * count))
          ids = ctypes.cast(p_leafids, ctypes.POINTER(ctypes.c_int64 * count))
          child = ctypes.cast(
              pp_childids, ctypes.POINTER(ctypes.POINTER(ctypes.c_int64) * count)
          )
          mins = ctypes.cast(
              pp_mins, ctypes.POINTER(ctypes.POINTER(ctypes.c_double) * count)
          )
          maxs = ctypes.cast(
              pp_maxs, ctypes.POINTER(ctypes.POINTER(ctypes.c_double) * count)
          )
          for i in range(count):
              p_child_ids = child.contents[i]
      
              id = ids.contents[i]
              size = sizes.contents[i]
              child_ids_array = ctypes.cast(
                  p_child_ids, ctypes.POINTER(ctypes.c_int64 * size)
              )
      
              child_ids = []
              for j in range(size):
  >               child_ids.append(child_ids_array.contents[j])
  E               ValueError: NULL pointer access
  /project/rtree/index.py:1311: ValueError
  =============================== warnings summary ===============================
  tests/test_index.py::IndexSerialization::test_interleaving
    /project/rtree/index.py:283: DeprecationWarning: index.get_size() is deprecated, use len(index) instead
      warnings.warn(
  -- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
  =========================== short test summary info ============================
  FAILED tests/test_index.py::IndexSerialization::test_interleaving - ValueErro...
  ============== 1 failed, 36 passed, 1 xfailed, 1 warning in 0.62s ==============

[mwtoews]

Finally ready for review. A few notes:

• Firstly, I'm newish to cibuildwheel, so review by anyone with more experience would be appreciated. Would appreciate @EwoutH to take a look too.
• There are many fewer built distribution files, e.g. linux went from 24 files 28.7 MB to 4 files 4.24 MB. This is a good thing.
• macOS now has a platform tag: macosx_10_9_universal2.macosx_10_9_x86_64.macosx_11_0_arm64.macosx_11_0_universal2 which is better? Anyone with M1/M2 chip want to test this one?
• There is still an issue with cp312, these errors are shown in the workflow, but the tox config ignore_outcome = True allows the wheel builds to succeed
• This is still an issue with the i686 tests, but these don't seem to be major. The tests could be re-written to be more robust for 32-bit and 64-bits

[adamjstewart]

Anyone with M1/M2 chip want to test this one?

Me! Where can I download the wheel?

[adamjstewart]

Would it be possible to have separate wheels for x86_64 and arm64 on macOS? We might see better performance if binaries aren't running through Rosetta.

[mwtoews]

Go to the "Summary" of any build (e.g. here), scroll to the bottom and find the "Artifacts", e.g. macos-latest-whl.

As for separate macos builds, sure I can take another look at the repair wheel script. As mentioned previously, I don't really understand much about the macos stuff, so advice would be appreciated.

[adamjstewart]

I was able to import rtree. Anything else I should test?

I don't know much about cibuildwheel but I know some stuff about macOS. I'm going to guess that the macOS platform tag args are an artifact of when macOS arm64 first came out and no one had runners to properly build it so they instead built a single universal wheel. Are there any other repos that do something similar to this (single wheel for all Python versions) without universal wheels? I feel like there should be an easier way to do this than having a repair script... Maybe reach out to the cibuildwheel devs?

@pradyunsg @rgommers may also have ideas on how to create platform-specific Python-version-independent wheels with cibuildwheel.

[mwtoews]

Note the previous macos artifacts have macosx_10_9_x86_64 and macosx_11_0_arm64 platform tags, but they are actually both x86_64 binaries. Testing again with CMAKE_OSX_ARCHITECTURES from ARCHFLAGS variable...

[mwtoews]

@adamjstewart mind testing the latest artifacts again? There are now two macOS wheels, each with only one architecture. Importing rtree is the first smoke test. But I suppose a performance benchmark would be more useful. There is no easy benchmark, unfortunately. The closest is benchmarks.py.off by @hobu

[adamjstewart]

New artifacts work great:

Stream load:
14350.47 usec/pass

One-at-a-time load:
58568.11 usec/pass


30000 points
Query box:  (1240000, 1010000, 1400000, 1390000)


Brute Force:
53 hits
1657.80 usec/pass

Memory-based Rtree Intersection:
53 hits
620.89 usec/pass

Disk-based Rtree Intersection:
53 hits
608.25 usec/pass

Disk-based Rtree Intersection without Item() wrapper (objects='raw'):
53 raw hits
47.25 usec/pass

[adamjstewart]

Once this is merged we should think about cutting a 1.0.2 release so we have wheels for Python 3.12

[mwtoews]

Thanks all for testing and feedback! Last few things to mention.

1. The magic that rewrites the wheels is wheel tags, using --python-tag=py3 --abi-tag=none options. Additionally the unpack and pack steps are not really needed for Windows and macOS, but was there as a workaround for a (now closed) issue wheel tags adds directories to RECORD causing issues with pip uninstall pypa/wheel#558
2. The new wheels with "none" ABI tag work with PyPy, which supports ctypes. I could have added pp{37,38,39} to tox.ini, perhaps later.
3. Tests against each Python version are automated using tox. Tests are only run if a binary version of numpy is available, otherwise a long wait is needed to build numpy from source. This is done via tox option:
   install_command = python -I -m pip install --only-binary=:all: {opts} {packages}
4. Tox tests always pass due to options ignore_outcome = True (and possibly ignore_errors = True). This means that manual inspection may be required, rather than just looking at a green checkmark.