[mpich-discuss] MPI_Alltoallw bug?

[Pavan Balaji]

Wei-keng,

Yeah, I think that's a bug. The standard doesn't seem to say anything 
about it, but it's still inconvenient for users. Thanks for reporting it.

  -- Pavan

Wei-keng Liao wrote:
> In mpich2-1.0.8p1/src/mpi/coll/alltoallw.c, lines 361 - 372,
> error checking is not necessary for those sendtypes[] and recvtypes[]
> when their sendcnts[] and recvcnts[] are zeros.
> 
> I got error messages from these checking when a subset of processes
> has zero-length to send or receive, and I used MPI_DATATYPE_NULL for
> those cases.
> 
> Wei-keng

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji