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
Non-blocking atomic fetch uses fi_fetch_atomic with DELIVERY_COMPLETE flag. This ensures that the value is fetched before returning. Should it use the FI_INJECT without any completion semantics?
Also, for the same reason, does the blocking atomic fetch require a call to get_wait?
The text was updated successfully, but these errors were encountered:
@wrrobin I am looking at shmem_transport_fetch_atomic_nbi function in transport_ofi.h and I see that in the fi_fetch_atomicmsg call the argument passed is FI_INJECT. Has these bug already been fixed then or am I missing something?
So OFI uses that line of code I described so I think that one checks out. UCX is calling which I also assume in non-blocking. Portals has this comment "transport_fetch_atomic already buffers the source argument" and then calls a regular fetch atomic so I am assuming the comment is accurate. The onnly place I don't know if we are calling something that is blocking is in shmem_shr_transport_fetch_atomic, I don't see any non-blocking stuff there and in the code you reference here: #1062
Non-blocking atomic fetch uses fi_fetch_atomic with DELIVERY_COMPLETE flag. This ensures that the value is fetched before returning. Should it use the FI_INJECT without any completion semantics?
Also, for the same reason, does the blocking atomic fetch require a call to
get_wait
?The text was updated successfully, but these errors were encountered: