It seems that I didn't understand MPI_BOTTOM completely, so the error checks needed to be different than described below, but the high-level issues are the same. Interested parties may consult http://trac.mpich.org/projects/mpich/ticket/1863 for details and further discussion. Jeff On Wed, May 29, 2013 at 7:41 AM, Jeff Hammond <[email protected]> wrote:

I see Jim used MPIR_ERRTEST_ARGNULL in (r)get_accumulate, which makes sense (I am working on verifying that they should be in the others as well). However, these calls are not guarded by "if (count)", which means that one presumably (I have not written a test for it but it seems obvious) cannot use addr=NULL with count=0. It seems to me that 0-byte transfers are find with NULL buffers (this was recently verified for MPI_Ssend because it's a common way to implement p2p synchronization).

It seems prudent to change something like this:

MPIR_ERRTEST_COUNT(origin_count, mpi_errno); MPIR_ERRTEST_DATATYPE(origin_datatype, "origin_datatype", mpi_errno); MPIR_ERRTEST_ARGNULL(origin_addr, "origin_addr", mpi_errno);

To this:

MPIR_ERRTEST_COUNT(origin_count, mpi_errno); MPIR_ERRTEST_DATATYPE(origin_datatype, "origin_datatype", mpi_errno); if (origin_count) MPIR_ERRTEST_ARGNULL(origin_addr, "origin_addr", mpi_errno);

Then I got to thinking, what if someone - almost certainly not intentionally but because of a general implementation - uses MPI_Type_vector to create a zero-length vector type. I don't see why one can't do a count>0 transfer with with a NULL buffer if the datatype is empty.

I wrote the following test of a 0-length vector and see that it works just fine with bcast.

#include <stdio.h> #include <stdlib.h>

#include <mpi.h>

int main(int argc, char ** argv) { int rank, nproc; MPI_Datatype emptyvec; MPI_Aint lb, ext;

MPI_Init(&argc, &argv);

MPI_Comm_rank(MPI_COMM_WORLD, &rank); MPI_Comm_size(MPI_COMM_WORLD, &nproc);

if (rank == 0) printf("Starting MPI datatype test with %d processes\n", nproc);

MPI_Type_vector(0,0,0,MPI_DOUBLE,&emptyvec); MPI_Type_commit(&emptyvec); MPI_Type_get_extent(emptyvec, &lb, &ext); MPI_Type_size(emptyvec, &size); if (rank == 0) printf("lb = %zu ext = %zu size = %d \n", lb, ext, size);

MPI_Bcast(NULL, 0, emptyvec, 0, MPI_COMM_WORLD); MPI_Bcast(NULL, 1, emptyvec, 0, MPI_COMM_WORLD);

MPI_Type_free(&emptyvec);

if (rank == 0) printf("Test complete: PASS.\n");

MPI_Finalize();

return 0; }

This suggests that it should be valid to use addr=NULL and count>0 as long as the extent of the type is 0, hence the aforementioned code should really be:

int origin_datatype_size; ... MPIR_ERRTEST_COUNT(origin_count, mpi_errno); MPIR_ERRTEST_DATATYPE(origin_datatype, "origin_datatype", mpi_errno); MPID_Datatype_get_size_macro(origin_datatype, origin_datatype_size); if (origin_count && origin_datatype_size) MPIR_ERRTEST_ARGNULL(origin_addr, "origin_addr", mpi_errno);

Do people agree with this reasoning?

Thanks,

Jeff

-- 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 ALCF docs: http://www.alcf.anl.gov/user-guides

-- 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 ALCF docs: http://www.alcf.anl.gov/user-guides