Avoid patching sort algorithms from CCCL#18364
Avoid patching sort algorithms from CCCL#18364rapids-bot[bot] merged 3 commits intorapidsai:branch-25.06from
Conversation
CCCL provides a cusom `CCCL_AVOID_SORT_UNROLL` escape hatch that will selectively avoid unrolling the problematic loops
github-actions
Bot
added
libcudf
|
See CCCL PR |
davidwendt
added
improvement
vyasr
left a comment
vyasr
left a comment
There was a problem hiding this comment.
Just to make sure that I understand, we're adding a nonfunction preprocessor definition now and marking the patch as no longer necessary as of 3.0 so that when we drop the patch in 3.0 the definition will start having the same effect as our patch, right? i.e. adding this definition right now is a no-op?
That is my understanding as well but will wait for @miscco to verify. |
|
Yes, that's correct. I asked @miscco if we might be able to get a backport of this to 2.8.x so that our 25.06 release can be patch-free (maybe!): NVIDIA/cccl#4254 (comment) |
|
That would be a huge step forward yeah. |
There was a problem hiding this comment.
I need this for #18235 to be patch-free. A backport to CCCL branch/2.8.x was made here: NVIDIA/cccl#4387
This is safe to merge now.
|
/merge |
|
I am also merging this into #18235 for testing. |
4 participants
CCCL provides a custom
CCCL_AVOID_SORT_UNROLLescape hatch that will selectively avoid unrolling the problematic loops