[FEA] Add new parameter to cudf::gather to enable behavior of negative indices beget null elements

[shwina]

This is the second of a series of issues as part of a larger join refactor.

It would be great if cudf::gather exposed a way to get null when negative indices are specified in the gather map.

cudf::gather({1, 2, 3, 4, 5, 6}, {1, 3, -1}) == {2, 4, null}

The join API is going to be refactored to return gather map(s) instead of the actual result of a join. This is so callers have the freedom to construct the output table in whatever way they like. Currently, the libcudf join code and join API are both complicated because it tries to construct an output table that meets Pandas' expectations. As a consequence, the Cython and Python code is also complicated.

This feature will enable the libcudf join API to return a gather map with negative values to indicate non-matches in outer joins.

cc: @jrhemstad

[jrhemstad]

negative indices

While the original intention is for negative indices, I think we can expand it just to say that any out of bounds index yields a null.

[shwina]

cc: @brandon-b-miller as well

[mythrocks]

It would be great if cudf::gather exposed a way to get null when negative indices are specified in the gather map.

I didn't realize that the following code wasn't the prescribed way to achieve that effect:

    auto output_table = detail::gather(table_view{{input}},
                                       output->view(),
                                       detail::out_of_bounds_policy::IGNORE,
                                       detail::negative_index_policy::NOT_ALLOWED,
                                       mr,
                                       stream);

This is from rolling.cu.

[shwina]

I think that code does do that, but this is not exposed in the public API

[mythrocks]

Ah, so. Gotcha.

[github-actions]

This issue has been marked rotten due to no recent activity in the past 90d. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[shwina]

This is now enabled with the NULLIFY out of bounds policy introduced in #6523. Closing.