Re: [AMBER] mpi-allreduce in AMBER11

From: Jason Swails <>
Date: Fri, 16 Mar 2012 09:38:40 -0400

On Fri, Mar 16, 2012 at 9:23 AM, Carlos Sosa <> wrote:

> Hello,
> AMBER11 uses MPI_ALLREDUCE. Is it use with 32 bit words or 64? I am
> currently checking the code as well.

I'm not sure I understand the question. Every MPI call takes an
MPI_Datatype (e.g. MPI_DOUBLE_PRECISION, MPI_REAL, MPI_INT for Fortran or
MPI_DOUBLE, MPI_LONG_INT, etc. for C/C++). The size of the MPI_Datatype is
what MPI uses to assign variables properly in all function calls that pass
information (including the allreduce). The way it's implemented should be
platform-independent, so it should work with both 32-bit and 64-bit words.

The aim for sander was to use a preprocessor define, MPI_AMBER_REAL or
something, that could be toggled to single or double precision, much like
the preprocessor define _REAL_ can switch between real and double
precision. MPI_DOUBLE_PRECISION is hard-coded in too many places for the
single-precision build to work anymore, not that it is recommended, anyway.


> Thanks
> --
> Carlos P Sosa, Ph.D.
> Computational Chemistry Consultant and Adjunct Assistant Professor
> *Biomedical Informatics and Computational Biology (BICB) *
> Minnesota Supercomputing Institute for Advanced Computational Research,
> University of Minnesota, Walters Library # 509, 117 Pleasant Street,
> Minneapolis, MN 55455.
> _______________________________________________
> AMBER mailing list

Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
AMBER mailing list
Received on Fri Mar 16 2012 - 07:00:03 PDT
Custom Search