Investigate wait needed after async_buffer host launched copies
Open
- Dominant language
- C++
- Stars
- 2.5k
- Forks
- 486
- Avg merge
- 2d 6h
- Merged PRs (30d)
- 295
Description
In #3460, we had to add extra synchronization between copying data using a stream callback and copying from the destination to the device using `cudaMemcpyAsync`. This synchronization should not be necessary, but we could not figure it out in reasonable time.
_Originally posted by @bernhardmgruber in https://github.com/NVIDIA/cccl/pull/3460#discussion_r1955852408_
Contributor guide
Assessment
This issue has not been assessed yet.