Holoviews / Datashader integration

[Huite]

Holoviews and Datashader supports TriMesh/trimesh (resp.):
https://datashader.org/user_guide/Trimesh.html

This is obviously useful, especially for larger grids.

Althought not every 2D grid is triangular, it can be easily triangulated (which is done for a number of plotting methods, since matplotlib can will only draw contours for triangles e.g.).

The following suffices for Ugrid2d topologies in the _PlotMethods accessor:

    def to_holoviews(self):
        import holoviews
        import pandas as pd

        grid = self.grid
        da = self.darray
        (_, _, triangles), index = grid.triangulation
        name = getattr(da, "name", "data")

        vertices = pd.DataFrame(
            data=grid.node_coordinates,
            columns=("x", "y"),
        )
        triangles = pd.DataFrame(
            data=triangles,
            columns=("v0", "v1", "v2"),
        )

        dim = get_ugrid_dim(grid, da)
        if dim == NODE:
            vertices[name] = da
        elif dim == FACE:
            triangles[name] = da.isel({grid.face_dimension: index})
        else:
            raise ValueError("only supports data on nodes or faces")

        return holoviews.TriMesh((triangles, vertices))

Then, to draw something:

import holoviews as hv
from datashader import rasterize

hv.extension("matplotlib")
rasterize(uda.to_holoviews()).opts(colorbar=True)

What makes it not totally trivial:

• What to do for 1D topologies? Return a network I guess? https://datashader.org/user_guide/Networks.html
• The datashader/holoviews isn't entirely obvious to me. Should we add (more) methods to let it resemble the "standard" plotting arguments (i.e. like xarray/matplotlib?)
• Should we default to bokeh/matplotlib as the extension? Not setting it results in nothing being shown.

Having to call rasterize yourself is a bit annoying. Something like .ugrid.holoviews.plot() could be an option but:

• We're stacking a lot of accessors...
• There are arguments for both the rasterize function as well as the plotting options. Should they be combined?

Secondly, there's a package called hvplot: https://hvplot.holoviz.org/
It adds an hvplot and interactive accessor to dataframes and xarray objects. This seems like it's more work, because it actually does data selection.