yes, to summarize what Hubert and Jim said and add my own comments: * accumulate_ordering and accumulate_ops should be valid for all window types (unless shared memory windows allow ONLY load-store, then we can remove them from MPI_WIN_ALLOCATE_SHARED). * same_size is primarily useful for MPI_WIN_ALLOCATE but it would be valid and mildly useful for MPI_WIN_CREATE. for example, if i want to do valid-range checking and i don't know same_size, i may need an O(N) list of (base,max_offset), whereas same_size allows me to have only O(N) for base and O(1)=1 for max_offset. * alloc_shared_noncontig only makes sense for MPI_WIN_ALLOCATE_SHARED best, jeff On Mon, Dec 3, 2012 at 8:17 AM, Hubert Ritzdorf <[email protected]> wrote:
Hi,
I think that the info argument "same_size" is defined only for MPI_Win_Allocate. It's defined in the context of this function and it's explicitly an additional info relative to the info keys of MPI_Win_create.
MPI_Win_allocate is not mentioned in the list of Page 409 Line 48 and "same_size" doesn't make sense for MPI_Win_allocate_shared.
MPI_Win_create_dynamic specifies only the info keys of MPI_Win_create (Page 412 Line 22-23) and "same_size" doesn't make sense since MPI_Win_attach is not collective.
Hubert
________________________________________ From: [email protected] [[email protected]] on behalf of Jim Dinan [[email protected]] Sent: Monday, December 03, 2012 4:07 PM To: [email protected] Subject: Re: [Mpi3-rma] missing RMA info keys
Hi Jeff,
I haven't looked at the text in a while -- alloc_shared_noncontig only makes sense in a call to MPI_Win_allocate_shared(). The other info keys were intended to be valid for all communicator creation routines. If this isn't clear, then we should definitely correct it. I created a ticket to make sure we remember to look at this:
https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/346
~Jim.
On 12/1/12 8:50 PM, Jeff Hammond wrote:
It is not clear if these info keys are predefined for MPI_WIN_ALLOCATE or MPI_WIN_CREATE_DYNAMIC.
Does the following text mean that the info keys for MPI_WIN_ALLOCATE are inclusive of those for MPI_WIN_CREATE? It really does not seem that way to me.
======================== The info argument can be used to specify hints similar to the info argument for MPI_WIN_CREATE and MPI_ALLOC_MEM. The following info key is predefined:
same_size � if set to true, then the implementation may assume that the argument size is identical on all processes. ========================
If the keys "accumulate_ordering, accumulate_ops, same_size, alloc_shared_noncontig" are not valid for MPI_WIN_ALLOCATE and/or MPI_WIN_CREATE_DYNAMIC already, they should be this task should be added to the TODO list for MPI 3.1.
Jeff
On Tue, Nov 6, 2012 at 4:41 AM, Hubert Ritzdorf <[email protected]> wrote:
Hi,
the new RMA info keys "accumulate_ordering, accumulate_ops, same_size, alloc_shared_noncontig" and the corresponding values are missing in Sections A.1.5 and A.1.6 (Pages 683 and 684).
Hubert
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
Click https://www.mailcontrol.com/sr/Bnj6dm!ZzevGX2PQPOmvUvSoGVaHHSWCqoZfLcnlBGwqN... to report this email as spam.
_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
-- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond