[C++][GPU] Writing sliced GPU record batches to IPC

[zeroshade]

Describe the enhancement requested

Currently, it isn't possible to write GPU-memory backed record batches to IPC if they are sliced. When Arrays are sliced, the null-count is set to a sentinel to indicate "UNKNOWN" rather than recompute it immediately. We need to add some interfaces to the GPU / MemoryManager / Device libraries to perform the bit counting and copying of bitmaps to handle sliced arrays properly and write them to IPC when they are using device memory.

Component(s)

C++, GPU

[jorisvandenbossche]

I think it's not only the null count that is the issue? At least for variable size binary or list data types, the IPC writing will also convert the offsets to be zero-based in case of sliced arrays. That's also something that will need custom handling for non-CPU data I think.

[zeroshade]

Good point, that's also an issue that needs to be handled

[pitrou]

The offsets conversion is tracked in #46747