You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Description: Once a team is created, the user can create a context with the team via SHMEM_TEAM_CREATE_CTX. The context means a set of local communication resource, similar to an MPI VCI. The context will then be used with any RMA or atomics (AMO) operation. When a team-context is used, the dest PE of each RMA or AMO operation needs to be read as team-relative PE, similar to the win-relative target_rank in MPI RMA.
Suggested implementation
# updates the comm based implementation as posted in #36
# For each team, creates `num_contexts` (>=0) number of wins at team creation time.
# It is safe to call win_create here as team creation is also collective.
SHMEM_TEAM_SPLIT_STRIDED
SHMEM_TEAM_SPLIT_2D
SHMEM_TEAM_CREATE_CTX -> return previously created win (local op)
SHMEM_CTX_GET_TEAM-> return associated team object (local op)
TODO
Implement SHMEM_TEAM_CREATE_CTX. Can skip optimization with SHMEM_CTX_SERIALIZED | SHMEM_CTX_PRIVATE | SHMEM_CTX_NOSTORE for now.
Update remote PE in all RMA and AMO operations by using team-relative PE.
The text was updated successfully, but these errors were encountered:
minsii
changed the title
Team: Adapt RMA and AMO with teams
Team: Adapt RMA and AMO with team/context
Feb 26, 2021
minsii
changed the title
Team: Adapt RMA and AMO with team/context
Context: Adapt RMA and AMO with team/context
Feb 26, 2021
Description: Once a
team
is created, the user can create acontext
with the team viaSHMEM_TEAM_CREATE_CTX
. Thecontext
means a set of local communication resource, similar to anMPI VCI
. The context will then be used with any RMA or atomics (AMO) operation. When a team-context is used, thedest PE
of each RMA or AMO operation needs to be read as team-relative PE, similar to the win-relativetarget_rank
in MPI RMA.Suggested implementation
TODO
SHMEM_TEAM_CREATE_CTX
. Can skip optimization withSHMEM_CTX_SERIALIZED | SHMEM_CTX_PRIVATE | SHMEM_CTX_NOSTORE
for now.The text was updated successfully, but these errors were encountered: