In order to test scenarios where non-contiguous split would be more appropriate than cudf::contiguous_split (e.g. potentially for large tables where a contiguous split would force large allocations per split), I would like to add JNI bindings to cudf::split.
I have this coded up, but I do need to add a test. I should have a PR today or tomorrow at the latest.
Proposing splitAndCopy at the Table level, which means we would copy the table views returned by cudf::split to make them owning tables.
This one slipped due to other things taking a lot longer. At this stage, I'll target 23.08 early for a patch but I am looking to experiment with it as a performance enhancement in the plugin as well.
In order to test scenarios where non-contiguous split would be more appropriate than
cudf::contiguous_split
(e.g. potentially for large tables where a contiguous split would force large allocations per split), I would like to add JNI bindings tocudf::split
.I have this coded up, but I do need to add a test. I should have a PR today or tomorrow at the latest.
Proposing
splitAndCopy
at theTable
level, which means we would copy the table views returned bycudf::split
to make them owning tables.