If we thought it was a mistake in MPI-2 to disallow it, wouldn't we have tried to relax it in MPI-3? Brian had commented before that we might have been confused when defining same_op_no_op and really meant same_op_no_op_replace, which is a perfectly reasonable default semantic. I would argue that we should admit the latter mistake and fix this in MPI-3.1. I know of no implementation that need to change because of this. same_op_no_op_replace might not be completely SHMEM-compliant but it almost certainly meets the needs of reasonable SHMEM programs. Right now, the incompatibility of REPLACE and <any reduce op> puts me in a very bad place for essentially all PGAS models. Jeff On Thu, Mar 13, 2014 at 2:40 PM, Balaji, Pavan <[email protected]> wrote:
On Mar 13, 2014, at 2:00 PM, Jeff Hammond <[email protected]> wrote:
Do you think a user can determine from MPI-3 that it is an erroneous program to have accumulate to same loc with diff ops using _reasonable_ effort?
I’m not arguing as to whether it is convenient for the user. I’m just stating that the standard disallows it. IIRC, it was the explicit intention of the WG to disallow it, because we thought MPI-2 disallowed it as well (which was a mistake).
— Pavan
_______________________________________________ mpiwg-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
-- Jeff Hammond [email protected]