Description

Contributes to #16483 by adding fast paths to DataFrame.to_cupy (which is called when DataFrame.values is called). The PR follows up #18450 to add cython bindings for cudf::table_to_array to pylibcudf and plumbs those changes through cudf classic.

I benchmarked the fast (True) and slow (False) when the dataframe has 1, 6, 20, and 100 columns. The fast paths use cudf::table_to_array if the number of columns is greater than 1 and cp.asarray directly if the dataframe has only one column. The slow path uses a raw python loop + assignment to create the cupy array.

I used the median because the CUDA overhead of calling cudf::table_to_array is large (so there are outliers in the times). Here is a profile of calling to_cupy twice for both the slow and fast paths.

In the first calls, the fast path takes 7.3 ms vs 4.8 ms for the slow path. The first call to cudf::table_to_array is the bottleneck. But if you compare the second calls, the fast path is much faster (79 us vs 2.3ms)

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.