Addressing align/barrier clarifications #14
Conversation
|
@agrippa @minsii @swpoole @jamesaross @tonycurtis Can you please review? |
|
I agree with the sentiment in the original issue: users shouldn't be relying on collective memory allocation/deallocation for quieting memory operations. In that case, I think that allocating routines (calloc, malloc, realloc) should have a shmem_sync_all at exit. De-allocating routines (free, realloc) should have a shmem_barrier_all at entry? However, if there are hypothetical programs out that there that do rely on that property (intentionally or otherwise) this is a change that breaks those programs because they change OpenSHMEM semantics. So I think we need to discuss whether this should be handled in the committee. |
|
Just making a note here. The current semantics are a shmem_barrier_all on exit for allocating routines and a shmem_barrier_all on entry for de-allocating routines. I also think we might need changes to the new shmem_malloc_with_hints? |
|
Will address this as a part of OpenSHMEM 1.6 |
Clarify element sizes for collectives
Convert \hyperref on non-identifiers to \ref
Specify deprecated p2p types for `shmem_wait_until` and `shmem_test`
Addressing openshmem-org#180