MPI_Get_elements for data types used in reduction functions

[nysal]

There seems to be a difference in the way mpich and ompi behave when MPI_Get_elements is used in conjunction with datatypes used with reduction functions (MPI_DOUBLE_INT etc.). When a single count of such a datatype is received, mpich interprets the number of elements as 2 where as ompi returns 1. ompi seems to interpret the following line from the spec (chapter 4, page 114):

Note that the MPI_GET_COUNT, MPI_GET_ELEMENTS, and MPI_GET_ELEMENTS_X
return the same values when they are used with basic datatypes as long as the limits
of their respective count arguments are not exceeded.

to include the datatypes used in reduction functions. Here's the link to a test from mpich that can be used to reproduce the issue - http://git.mpich.org/mpich.git/blob_plain/HEAD:/test/mpi/datatype/get-elements-pairtype.c

[nysal]

You can change the send datatype to MPI_DOUBLE_INT, then it gives the wrong number of elements instead of MPI_UNDEFINED.

[bosilca]

This example is not correct, it assumes that MPI_DOUBLE_INT is the same as MPI_DOUBLE followed by MPI_INT. While this might be pragmatically correct, MPI_DOUBLE_INT is an MPI defined datatype and as such OMPI consider it predefined. One of the outcomes is that we cannot split the type in it's type counterparts, as in the example provided. If instead of using the predefined MPI_DOUBLE_INT you create your own type (composed of a double followed by an int) you will get the same behavior as the one you describe for MPICH.

The difference between the 2 behaviors is o line of code, line 416 in ompi/datatype/ompi_datatype_module.c.

[nysal]

ok, so there is a difference in interpretation here. Changing the test case to use:
//err = MPI_Sendrecv((const void *) &foo, 1, stype, 0, 0,
err = MPI_Sendrecv((const void *) &foo, 2, MPI_DOUBLE_INT, 0, 0,
(void *) &bar, 2, MPI_DOUBLE_INT, 0, 0, MPI_COMM_SELF, &recvstatus);

Ignore the invalid memory read, it should work though :-)
MPI_Get_elements() in mpich returns 4, where as ompi returns 2. Should we get a clarification from the forum? The spec seems to say MPI_Get_count and MPI_Get_elements return the same value for "basic" datatypes. Does "basic" include the datatypes used in reduction functions?

[bosilca]

Clearly there is a different interpretation of the standard: we consider the MPI_DOUBLE_INT as a single unitary datatype, while MPICH does not. Asking the forum might be a good idea.