Use CuPy array in pip_bitmap_column_to_binary_array

[isVoid]

Description

The performance regression in #1413 is due to numba's DeviceNDArray
is slow in slicing. Recent cudf's DataFrame construction has simplified the construction and delegated construction
to similar logic that handles cuda_array_interface. Since the construction involves slicing the array, we need
this operation to be fast. In that sense, we should cast the use of DeviceNDArray to cupy array to support fast
slicing.

closes #1413

Checklist

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

[trxcllnt]

Looks like the test needs to be updated as well:

    def test_pip_bitmap_column_to_binary_array():
        col = cudf.Series([0b00000000, 0b00001101, 0b00000011, 0b00001001])._column
        got = pip_bitmap_column_to_binary_array(col, width=4)
        expected = np.array(
            [[0, 0, 0, 0], [1, 1, 0, 1], [0, 0, 1, 1], [1, 0, 0, 1]], dtype="int8"
        )
>       np.testing.assert_array_equal(got.copy_to_host(), expected)
E       AttributeError: 'ndarray' object has no attribute 'copy_to_host'

[mroeschke]

If this is 2D, you could use asfortranarray so when sliced into 1D columns the memory is contiguous. (I'll also ensure this on the cudf side)

[isVoid]

asfortranarray will reorder the data to F-contiguous with new allocation and increases peak memory usage. It would be nice if we don't have to assume memory order for return values? Cupy support fast indexing for any memory order anyway.

[jakirkham]

If we would like to change the order of the array, this could be done by passing the order flag to reshape

That said, given this wasn't done before with the DeviceNDArray (and may involve other subtle changes in associated code), maybe this should be converted to a new issue/PR?

[isVoid]

maybe this should be converted to a new issue/PR?

Yes - again, unless explicitly required by downstream libraries, I would rather cuspatial not having to dictate a memory order here but just stick to the defaults. Switching to cupy gives the flexibility not having to be explicit on memory orders, and cudf is adding back support to different memory order arrays. This relieves the responsibility from cuspatial.

[jakirkham]

Have retitled the PR to reflect the current state. Hope that is ok 🙂

[harrism]

/merge