Efficient Laplacian calculations

[nickeopti]

The current implementation of trimesh.smoothing.laplacian_calculation relies on list comprehensions. The case of equal_weight=True, at least, can be computed directly using sparse matrix operations. This PR does that, which turns out to be nearly 20 times faster.

I have been unable to figure out a way of computing the equal_weight=False case using sparse matrix operations, so I have let the current implementation of that be. This implies quite a performance difference between the two flag values. This is noted in the docstring, which is now also more descriptive.

Correctness tests are added, which include a verbatim copy of the original implementation to test against.

Performance characteristics:

In [1]: import trimesh

In [2]: from tests import test_smoothing

In [3]: m = trimesh.creation.icosphere(8, 10)

In [4]: m.vertices.shape, m.faces.shape
Out[4]: ((655362, 3), (1310720, 3))

In [5]: %timeit trimesh.smoothing.laplacian_calculation(m, equal_weight=True)
44.4 ms ± 783 μs per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [6]: %timeit test_smoothing.explicit_laplacian_calculation(m, equal_weight=True)
790 ms ± 9.89 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [7]: %timeit trimesh.smoothing.laplacian_calculation(m, equal_weight=False)
2.87 s ± 31.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [8]: %timeit test_smoothing.explicit_laplacian_calculation(m, equal_weight=False)
2.91 s ± 34.8 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

where test_smoothing.explicit_laplacian_calculation is the original implementation.

[mikedh]

Looks like a huge improvement, thanks for the PR and great test! It seems like since equal_weight=False is not called by the internal uses of the function it's pretty reasonable for the flag to have a large performance penalty, since people should know what they're asking for.

Don't worry about the CI failing, it's just some minor housekeeping.