X-Mailer: QUALCOMM Windows Eudora Pro Version 4.2.2 
Date: Fri, 08 Dec 2000 10:49:19 -0600
To: Rolf Rabenseifner <Rabenseifner@XXXXXXXXXXXXXXXXXXXX>
From: William Gropp <gropp@XXXXXXXXXXX>
Subject: Re: BUG in MPI-2 Alltoallw
Cc: mpi-core@XXXXXXXXXXX,edgar.gabriel@XXXXXXXXXXXXXXXXXXXX
In-Reply-To: <200011291714.SAA00404@XXXXXXXXXXXXXXXXXXXXXXXXXX>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"; format=flowed
Sender: owner-mpi-core@XXXXXXXXXXXXX
Precedence: bulk

At 06:14 PM 11/29/2000 +0100, Rolf Rabenseifner wrote:
Hello,

a collegue has found a possible bug in the MPI-2 language bindings
of MPI_Alltoallw.

Page 165, we defined the byte-displacements as int (in C) and INTEGER
(in Fortran).

But, this should be MPI_Aint and INTEGER(INTEGER=MPI_ADDRESS_KIND)
because the displacements are byte-displacements
(and not datatype-displacements as in MPI_Alltoallv).

Pros and Cons?

Pro: One major reason for MPI_Alltoallw was, to allow different
     buffers via MPI_BOTTOM. Correct usage of MPI_BOTTOM
     requires address-length displacements.

Con: This errata would change the APIs.

I prefer the correction to the binding: MPI_Aint and MPI_ADDRESS_KIND.  Section 7.3.5 clearly states that these are byte displacements; we use Aint and ADDRESS_KIND for bytes.

Bill