FIX: Skip any nan coordinates in quadmesh

[greglucas]

Even though the coordinates of quadmesh are meant to be finite, a transform may make them nonfinite. This adds a check to avoid the nan quads in the quadmesh rendering routine.

This fixes an issue in Cartopy+mplcairo
xref: SciTools/cartopy#2305

It would be nice if the backend spec (essentially the docstrings of the base methods in RendererBase) specified clearly which methods need to support nans, and for which inputs.

I agree, and perhaps even try to add some tests on that assertion, not just docstrings if we could. Do you want me to try and implement a test here, or look into adding something upstream in matplotlib?

[greglucas]

Here is a quick MPL-only example demonstrating the bad quads get propagated to lower values of x instead of just the quad with the bad coordinate.

Without this patch, and with (fully filled color).

import matplotlib.pyplot as plt
import matplotlib
import numpy as np

matplotlib.use("module://mplcairo.macosx")

fig, ax = plt.subplots()

x = np.linspace(-20, 20)
y = np.linspace(-20, 20)
xx, yy = np.meshgrid(x, y)

mesh = ax.pcolormesh(xx, yy, xx[:-1, :-1], shading='flat', cmap='viridis')
mesh._coordinates[30, 30, 0] = np.nan
plt.show()

[anntzer]

Thanks for the patch.

Do you want me to try and implement a test here, or look into adding something upstream in matplotlib?

The test should go to matplotlib: mplcairo picks up the matplotlib test suite and runs it while replacing the agg backend by itself (see run-mpl-test-suite.py).