Fix for function from_meshio

[keurfonluu]

Overview

• Fixed: function pyvista.from_meshio when cells data type is np.uint64.

Details

The following code raises an error:

import pygmsh
import pyvista

with pygmsh.geo.Geometry() as geom:
    geom.add_polygon(
        [
            [0.0, 0.0],
            [1.0, -0.2],
            [1.1, 1.2],
            [0.1, 0.7],
        ],
        mesh_size=0.1,
    )
    mesh = geom.generate_mesh()

mesh = pyvista.from_meshio(mesh)

This error is due to np.hstack somehow returning an array of type np.float64 when input array data type is np.uint64:

x = np.array([[0, 1], [2, 3]], dtype=np.uint64)
y = np.hstack((np.full((len(x), 1), 0), x))
print(y.dtype)

Output:

float64

[codecov]

Codecov Report

Merging #3049 (e8c65bb) into main (d3e01ee) will decrease coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main    #3049      +/-   ##
==========================================
- Coverage   94.39%   94.38%   -0.02%     
==========================================
  Files          76       76              
  Lines       16556    16566      +10     
==========================================
+ Hits        15628    15635       +7     
- Misses        928      931       +3     

[adeak]

Thanks for the PR and finding this!

This error is due to np.hstack somehow returning an array of type np.float64 when input array data type is np.uint64:

This is because of this:

>>> x = np.array([[0, 1], [2, 3]], dtype=np.uint64)

>>> x.dtype, np.full((len(x), 1), 0).dtype
(dtype('uint64'), dtype('int64'))

np.full sees native python int 0 as the fill value, and creates a numpy default int (int64 on 64-bit linux/unix, int32 roughly everywhere else). But you can't reconcile int64 and uint64 values in the same int array, because they have different ranges. In these cases NumPy switches to doubles, with the caveat that one might lose precision for very large numbers:

>>> (np.int_(1) + np.uint64(1)).dtype
dtype('float64')

I'm wondering if we should also change the hstack line, to always match the dtype of the input. That way we can avoid type promotion on the np.hstack line.

[akaszynski]

Thanks for adding in the 2D test @keurfonluu. This LGTM, and as a side benefit we've marginally improved coverage.