The standard is already clear - commit is only required when the datatype is used for communication. It is unnecessary for the standard to list all cases where commit is not required - it is enough to list where commit *is* required, which it does. From a practical standpoint, the purpose of commit is to give the MPI implementation the opportunity to build specialized datatype pack/unpack code at runtime. This is irrelevant for the accessor functions. For example, the size of the datatype is independent of any optimizations for data communication, which is what the commit step is for. This is a clear error in MPICH; it is unnecessary to bother the Forum with this. Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign On Dec 27, 2012, at 1:08 PM, Lisandro Dalcin wrote:
On 27 December 2012 16:01, Pavan Balaji <[email protected]> wrote:
On 12/27/2012 12:56 PM US Central Time, Jeff Hammond wrote:
I don't disagree with your interpretation, but I am always hesitant to assert my understanding of the text in the absence of input from those who were present when it was written.
I'll file an MPI Forum ticket and see what happens.
This does look like something that's ill-defined in the standard. A ticket to address in MPI-3.1 is a good idea.
Lisandro: till this is fixed in the standard, the workaround will be to commit all datatypes.
So, in MPICH 3.x, given an uncommited datatype, MPI_Type_size() will succeed, but MPI_Type_size_x() will fail ?
-- Lisandro Dalcin --------------- CIMEC (INTEC/CONICET-UNL) Predio CONICET-Santa Fe Colectora RN 168 Km 472, Paraje El Pozo 3000 Santa Fe, Argentina Tel: +54-342-4511594 (ext 1011) Tel/Fax: +54-342-4511169