Expose stream-ordering in column view APIs

[shrshi]

Description

Adds stream parameter to

cudf::detail::column_view_base::null_count(begin, end)
cudf::detail::column_view_base::has_nulls(begin, end)

Note: Since stream-ordered prefetching is back-logged, we defer modifying the get_data member functions to accept a stream parameter for now.

Reference:

1. [FEA] Expose public stream-ordered C++ APIs #13744
2. [Story] Enabling prefetching of unified memory #16251 (comment)

Checklist

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

[karthikeyann]

nit: Unit test updates to test default stream usage.

[davidwendt]

Some of these are for the prefetch methods and I'm not sure if they should take a stream or not.
@GregoryKimball @vyasr

[bdice]

Some of these are for the prefetch methods and I'm not sure if they should take a stream or not. @GregoryKimball @vyasr

Yes, we want prefetching to take a stream. The underlying CUDA APIs are stream-ordered. We did not do so in the original design because it would require us to add streams to get_data, which we were not yet ready for at that time. This PR is a big milestone in the streams work, as it touches the very low level column APIs that we originally could not make stream-ordered. I am curious what specific blockers we have overcome to make this possible. We do have a note here about follow-up work once stream ordered get_data methods exist:

cudf/cpp/include/cudf/utilities/prefetch.hpp

Line 100 in 4de7e7a

* removed once an method for stream-ordered data pointer access is added to

We need to add unit tests before this can be approved.

[vyasr]

CUDA's prefetching APIs are stream-ordered, and when I first implemented prefetching in libcudf I did so under the assumption that we would eventually want to make it stream-ordered. While I still believe that is the right long-term technical solution, practically speaking it would require a significant amount of work to update all of libcudf to properly use a stream-ordered version of the prefetching APIs for no real benefit since we do not currently anticipate supporting multi-stream UVM use-cases. If that changes we can certainly reprioritize, but for now I would assume that stream-ordered prefetching in libcudf is backlogged for the foreseeable future in favor of higher priority changes.

[shrshi]

/merge