refactor: Refactor/cleanup partitions #3609
Conversation
rrsettgast
requested review from
CusiniM,
wrtobin and
untereiner
as code owners
rrsettgast
changed the title
rrsettgast
added
ci: run CUDA builds
| { | ||
| secondNeighborRanks.erase( neighbor.neighborRank() ); | ||
| } | ||
|
|
||
| for( integer const neighborRank : secondNeighborRanks ) | ||
| { | ||
| m_neighbors.emplace_back( neighborRank ); | ||
| neighbors.emplace_back( neighborRank ); |
There was a problem hiding this comment.
We used to have two versions of m_neighbors: one in DomainPartition, one in SpatialPartition.
Despite sharing the same name, I have the impression that:
- the
SpatialPartitionversion used to be a "strict" list of neighbors. - the
DomainPartitionversion also appended the neighbors of neighbors.
After this refactoring, I understand only one version will remain. It will be stored in SpatialPartition, and will now include the neighbors of neighbors.
If that's correct, this might conflict notably with the Particles machinery (e.g. SpatialPartition::repartitionMasterParticles) which was relying so far on the "strict" definition of neighbors.
Best case scenario: the MPM will simply iterate over more neighbors for nothing.
NB: the CI tests for MPM may still run fine as they seem to use a 1x1x1 or 2x2x1 partition... meaning they won't generate neighbors of neighbors anyway.
This PR seeks to remove a redundant definition of
DomainPartition::m_neighborsand use the one that exists inPartitionBase