Add cuDFInterface to work with cuDF GPU dataframes and cupy support for XArrayInterface

[philippjfr]

This PR adds a new data interface to allow HoloViews to work directly with cuDF GPU dataframes.

• values
• range
• dframe
• select
• groupby (look at optimizing)
• aggregate
• iloc
• add_dimension
• sort
• sample
• concat

Datashader support:

• Points/Scatter
• Curve/Path/Area
• QuadMesh: rectilinear and curvilinear

Other things to do:

• Implement cuDatashader support once integrated with datashader itself

One major thing to figure out is how we will set up CI tests for the GPU.

[jbednar]

Cool! Yes, it will be very important to be able to test this in CI...

[jonmmease]

Implement GPU optimized histogram operation (in this PR)

I think we could do this with cupy.histogram with something like this:

import cupy
import cudf
cdf = cudf.from_pandas(...)
ca = cupy.array(cdf["col1"].to_gpu_array(), copy=False)
hist, bin_edges = cupy.histogram(ca)

We'll probably also want to add a cupy interface for grid data eventually as well, so I'd be in favor of requiring that both cudf and cupy be installed for any GPU support.

[philippjfr]

We'll probably also want to add a cupy interface for grid data eventually as well

This PR now also supports cupy backed xarrays using a two line change.

[philippjfr]

Okay, this PR is now almost ready. I've added tests, which won't run on Travis of course but can be run locally. I'll add some additional tests for datashader integration.

There are one major class of tests that are failing which is groupby/aggregation tests. It seems like cuDF groupby(..., sort=False) results in reverse sorting which seems quite odd. I'll keep looking at that but in the worst case I'll disable those tests until the issue has been fixed upstream.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.