Reduce OMP syncs
Open
enhancement
- Dominant language
- C++
- Stars
- 119
- Forks
- 31
- PR merge metrics
- No merged PRs in 30d
Description
Currently, there is an OMP sync after exterior blocks are computed, then the pack and halo exchange is started (with OMP syncs in the packing), then interior blocks, then final sync. Only the final sync is truly needed. Getting rid of the others would require all loops to be in one parallel section.
Contributor guide
Assessment
This issue has not been assessed yet.