Newsgroups: comp.parallel.mpi From: "Timothy H. Kaiser" Reply-To: tkaiser@arc.unm.edu Subject: Re: MPI_REAL and autodouble Organization: Albuquerque Resource Center, HPCERC Date: Wed, 26 Feb 1997 19:48:57 -0700 Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Message-ID: <3314F5C1.67C4@arc.unm.edu> Jeff Nucciarone asks If I compile using an autodoubling compiler option, thus converting arrays a and b from real*4 to real*8, will tasks 0 and 1 send and receive the proper length of data or does the MPI_REAL datatype need to be rewritten as MPI_DOUBLE? If so I suppose we could then use a preprocessor definition to substitute the proper datatype. Don't count on this working. Why? The libraries are normally compiled by them self with MPI_REAL and MPI_Double as constants. What can you do? There is a very good fix to this problem. Define your own MPI data type such as MY_MPI_REAL and use it instead of MPI_REAL and MPI_Double. My_MPI_Real can be set to either 8 or 16 bytes at run time. I'm on a short time clock so I will not go into this here today. Also, if you use Fortan 90 there is a way to define your own Fortran data types which can be globally changed between 8 and 16 bytes by changing a single line. If your are interested I can send you the code to do both of these but not this second. Tim Kaiser