Re: new filter proposal

Jean-Pierre Prost (jpprost@watson.ibm.com)
Thu, 17 Apr 1997 13:32:36 -0400

Thanks John. I think I now understands the wording of the definition
of MPI_TYPE_RESIZE.
To the miscellaneous chapter author, I would like to report a couple of
typos in the first paragraph of Page 211 ("uppoer", "datatape"), and
"datatype_resize" should be MPI_TYPE_RESIZE.
There is also a "recieve" on Page 210, line 20, and a "overiden" on Page
210, line 22.
Jean-Pierre

johnmay @ llnl.gov
04/17/97 12:47 PM

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

Jean-Pierre Prost writes:

> I am kind of confused. If MPI_TYPE_RESIZE is used for setting
> new lower and upper bounds, these bounds are not sticky. In other
> words, when replicating the new datatype, the natural lower and
> upper bounds (cf page 211 of the MPI-2 draft) are used. I do not
> see how MPI_TYPE_RESIZE could be used to define what John
> refers to as external holes. It seems to me that the tiling would not
> preserve these holes since the bounds are not sticky. Am I wrong ?

My reading of this passage is that the lb and ub set by
MPI_TYPE_RESIZE *are* signficant in the replication when
a new type is created from a resized one, but they go
away once the new type has been created. I get this from
the sentence: "If no explicit lb and ub markers are used in
datatype construction, then the lower bound and upper bound
set by datatype_resize are used to define how this datatype
is replicated when it is used to build a new datatype..."

If this new datatype is then used to create another type
(or is tiled by an MPI operation), then its natural bounds
are used.

John