Re: tiling issues

Jean-Pierre Prost (jpprost@watson.ibm.com)
Wed, 16 Apr 1997 14:01:58 -0400

No, the extent (both the one defined by MPI-1 and the true extent defined
by MPI-2)
in the case of {(int, 4)} is 4.
Typically, such a type map is obtained by calling (in C):
bklen = 1;
disp = 4;
MPI_Type_indexed(1, &bklen, &disp, MPI_INT, &newtype);
This is very different from {(lb, 0), (int, 4)}, which can be obtained by
calling (in C):
bklens[0] = 1;
bklens[1] = 1;
disps[0] = 0;
disps[1] = 4;
types[0] = MPI_LB;
types[1] = MPI_INT;
MPI_Type_struct(2, bklens, disps, types, &newtype)

In the second case, the extent is 8, and the true extent is 4.

Jean-Pierre

linda @ k2.llnl.gov
04/16/97 01:14 PM

To: mpi-io @ mcs.anl.gov
cc: (bcc: Jean-Pierre Prost/Watson/IBM Research)
Subject: Re: tiling issues

I agree that the contiguous type repeated bufcount times is determined
from the extent of the buftype, but in your example I believe the
extent is 4+sizeof(int)--that is, the extent includes the size of the
hole in the datatype. On the other hand, the size of the datatype does
not include the hole and would be just the sizeof(int). I am not
convinced, therefore, that there is any difference in tiling using
buftypes and tiling using filetypes.

Linda