Latest numpy 1.21.0 release causes "Unknown ufunc mode" for set_precision

Seen on a PR, but I think it’s not related to the changes in that PR but to the bump of the numpy version, which just had a 1.21.0 release (https://numpy.org/doc/stable/release/1.21.0-notes.html#compatibility-notes):

 =================================== FAILURES ===================================
______________________________ test_get_precision ______________________________

    @pytest.mark.skipif(pygeos.geos_version < (3, 6, 0), reason="GEOS < 3.6")
    def test_get_precision():
        geometries = all_types + (point_z, empty_point, empty_line_string, empty_polygon)
        # default is 0
        actual = pygeos.get_precision(geometries).tolist()
        assert actual == [0] * len(geometries)
    
>       geometry = pygeos.set_precision(geometries, 1)

pygeos/test/test_geometry.py:430: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
pygeos/decorators.py:70: in wrapped
    return func(*args, **kwargs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

geometry = (<pygeos.Geometry POINT (2 3)>, <pygeos.Geometry LINESTRING (0 0, 1 0, 1 1)>, <pygeos.Geometry LINEARRING (0 0, 1 0, 1...0, 2 0, 2 2, 0 2, 0 0))>, <pygeos.Geometry MULTIPOINT (0 0, 1 2)>, <pygeos.Geometry MULTILINESTRING ((0 0, 1 2))>, ...)
grid_size = 1, preserve_topology = False, kwargs = {}

    @requires_geos("3.6.0")
    @multithreading_enabled
    def set_precision(geometry, grid_size, preserve_topology=False, **kwargs):
        """Returns geometry with the precision set to a precision grid size.
    
        By default, geometries use double precision coordinates (grid_size = 0).
    
        Coordinates will be rounded if a precision grid is less precise than the
        input geometry. Duplicated vertices will be dropped from lines and
        polygons for grid sizes greater than 0. Line and polygon geometries may
        collapse to empty geometries if all vertices are closer together than
        grid_size. Z values, if present, will not be modified.
    
        Note: subsequent operations will always be performed in the precision of
        the geometry with higher precision (smaller "grid_size"). That same
        precision will be attached to the operation outputs.
    
        Also note: input geometries should be geometrically valid; unexpected
        results may occur if input geometries are not.
    
        Returns None if geometry is None.
    
        Parameters
        ----------
        geometry : Geometry or array_like
        grid_size : float
            Precision grid size. If 0, will use double precision (will not modify
            geometry if precision grid size was not previously set). If this
            value is more precise than input geometry, the input geometry will
            not be modified.
        preserve_topology : bool, default False
            If True, will attempt to preserve the topology of a geometry after
            rounding coordinates.
        **kwargs
            For other keyword-only arguments, see the
            `NumPy ufunc docs <https://numpy.org/doc/stable/reference/ufuncs.html#ufuncs-kwargs>`_.
    
        See also
        --------
        get_precision
    
        Examples
        --------
        >>> set_precision(Geometry("POINT (0.9 0.9)"), 1.0)
        <pygeos.Geometry POINT (1 1)>
        >>> set_precision(Geometry("POINT (0.9 0.9 0.9)"), 1.0)
        <pygeos.Geometry POINT Z (1 1 0.9)>
        >>> set_precision(Geometry("LINESTRING (0 0, 0 0.1, 0 1, 1 1)"), 1.0)
        <pygeos.Geometry LINESTRING (0 0, 0 1, 1 1)>
        >>> set_precision(None, 1.0) is None
        True
        """
    
>       return lib.set_precision(geometry, grid_size, preserve_topology, **kwargs)
E       NotImplementedError: Unknown ufunc mode with args[0]=0x7f91f8161790, args[N]=0x1, steps[0]=8, steps[N]=0, dimensions[0]=13.

pygeos/geometry.py:737: NotImplementedError

[jorisvandenbossche]

OK, so actually for the failing test case, it doesn’t seem something changed with the latest numpy version in the steps. So because of the out-of-bounds index bug, we probably just got undefined behaviour. This is fixed with #367.

---

If I add a print statement in the CHECK_NO_INPLACE_OUTPUT macro, we get the following for an older and latest version of numpy. After correcting the index, for set_precision with an array as input and output, we get the same steps (it’s for the ufunc that receives a scalar as input that the steps of the output changed from 0 to 8, but that doesn’t matter for us, it’s also a single element output)

In [1]: from pygeos.test.common import all_types, point_z, empty_point, empty_line_string, empty_polygon
ufunc mode with args[0]=0x55ae3750ad00, args[N]=0x55ae3750ab30, steps[0]=0, steps[N]=0, dimensions[0]=1.

In [2]: geometries = all_types + (point_z, empty_point, empty_line_string, empty_polygon)

In [3]: import pygeos

In [4]: pygeos.set_precision(geometries, 1)
ufunc mode with args[0]=0x55ae375283a0, args[N]=0x55ae3753bce0, steps[0]=8, steps[N]=8, dimensions[0]=13.
Out[4]: 
array([<pygeos.Geometry POINT (2 3)>, ...

In [5]: np.__version__
Out[5]: '1.18.4'

In [1]: from pygeos.test.common import all_types, point_z, empty_point, empty_line_string, empty_polygon
ufunc mode with args[0]=0x55d879227920, args[N]=0x55d879270b50, steps[0]=0, steps[N]=8, dimensions[0]=1.

In [2]: geometries = all_types + (point_z, empty_point, empty_line_string, empty_polygon)

In [3]: import pygeos

In [4]: pygeos.set_precision(geometries, 1)
ufunc mode with args[0]=0x55d878e9c860, args[N]=0x55d878df19c0, steps[0]=8, steps[N]=8, dimensions[0]=13.
Out[4]: 
array([<pygeos.Geometry POINT (2 3)>, ...

In [5]: np.__version__
Out[5]: '1.21.0'

[seberg]

OK, let me know if you find anything, it probably is related to the contiguity flags not matching somehow. The problem with continuity may just be that the scalar is both Fortran- and C-order contiguous, and the other argument(s) are only C- or only F-order contiguous…

I expect our change is good, but I may be able to undo it in the release branch if it gives you a lot of headaches.