To: Hubert Ritzdorf <ritzdorf@XXXXXXXXXXXXXXXXXXXXXXXXXXX>
cc: mpi-core@XXXXXXXXXXX
Subject: Re: Question to MPI/IO
Date: Tue, 02 Feb 1999 10:55:12 -0800
From: Bill Nitzberg <nitzberg@XXXXXXXXXXXX>
Sender: owner-mpi-core@XXXXXXXXXXX

I agree with Jean-Pierre and John, version (i) is correct.

- bill

Bill Nitzberg, PhD			Work: nitzberg@XXXXXXXXXXXX
MRJ Technology Solutions		Tel: (650) 604-4513
NASA Ames Research Center, M/S 258-6	FAX: (650) 966-8669
Moffett Field, CA 94035-1000		Personal: nitzberg@XXXXXXXXXX

Jean-Pierre Prost replied:
> Version (i) is the version intended by the standard, which says:
> "The function must begin copying at the location in userbuf specified
> by position into the (tiled) datatype."
> Note that your note was confusing, because the descriptions of the two
> alternatives before the code and within the code are opposite. The
> right description is the one given within the code. Instead of saying
> "the user must find...", I would rather say "the user defined conversion
> routine must determine..."

John May <johnmay@XXXXXXXX> replied:> I agree with Jean-Pierre that version (i) is correct.  Version
> (ii) would not be as simple as it looks because the user might
> call MPI_File_write with complex derived data type, and the
> first call to the conversion function might return after 
> traversing only part of the type.  Having the MPI implementation
> figure out the right buffer location for the second call would
> be pretty complicated, but the user function can just cache
> this information somewhere when it finishes the first call and
> then pick it up when it starts the second call.
> 
> Since you're working on the datarep converters, you might be
> interested in a test code we posted on the web at
> http://www.llnl.gov/sccd/lc/piop/ .  Look for the links at
> the end of the MPI-IO section.

Hubert Ritzdorf <ritzdorf@XXXXXXXXXXXXXXXXXXXXXXXXXXX> wrote:> 
> I am implementing MPI for NEC SX computers and I have a question to 
> Datarep Conversion functions (chapter 9.5.3 of the MPI-2 document)
> 
> (*)  If MPI cannot allocate a (file) buffer large enough to hold all the
>      the data to be converted, the standard says:
> 
>      For the first call, MPI will call the (conversion) function with  
>      ``position'' set to zero. ...
>      Then in subsequent calls to the conversion function,
>      MPI will increment the value in ``position'' by the ``count''
>      of items converted in the previous call.
> 
>      My question is:
>      Is the user buffer ``buf'' incremented correspondingly
>      in the second call (and further calls) of the conversion function
>      (cf. Version i below) or 
>      must the user find the correct position in the user buffer ``buf''
>      within the conversion function (cf. Version ii below)?
> 
>      i.e.
> 
>         /* first call */
>            position = 0;
>            write_fn (buf, datatype, count, filebuf, position, extra_state);
> 
>         /* write filebuf to the file */
> 
>            write (.., filebuf, count*datarep_size);
> 
> 
> 
>         /* (Version i)
>             second call, the user must find the correct position in buf */
> 
>            position = count;
>            write_fn (buf, datatype, count, filebuf, position, extra_state);
> 
> 
>      or
> 
>         /* (Version ii)
>            second call, MPI transfers the new user buffer location */
> 
>            position = count;
>            next_buf = (char *) buf + count * native_size;
>            write_fn (next_buf, datatype, count, filebuf, position, 
> extra_state);
>         /* write filebuf to the file */
> 
>            write (.., filebuf, count*datarep_size);
>