[mpich-devel] binding a new MPI type with fortran

[Rajeev Thakur]

Another reminder that you should call it MPIX_Pattern instead of MPI_Pattern.

Rajeev

On Mar 29, 2013, at 3:11 PM, Tatiana V. Martsinkevich wrote:

> Hello, 
> 
> Thanks again for your help Dave.
> 
> So we went the other way around and changed the type of MPI_Pattern from void* to integer and now fortran interfaces are being generated correctly. However, now the question is how to register the new type MPI_Pattern so that we could use it in fortran programs. This also doesn't seem to be trivial...
> 
> In buildface there are this kind of code sections: 
> 
>     # Datatypes
>     # These are determined and set at configure time
>     foreach $key (COMPLEX, DOUBLE_COMPLEX, LOGICAL, REAL, DOUBLE_PRECISION, INTEGER, '2INTEGER', '2DOUBLE_PRECISION', '2REAL', CHARACTER) {
>     print MPIFFD "       INTEGER MPI_$key\n";
>     print MPIFFD "       PARAMETER (MPI_$key=\@MPI_$key\@)\n";
>     }
> 
> this looks like where I should start digging but apparently, simply adding PATTERN to the key set doesn't work:
> 
>        Included at src/binding/f90/mpi_constants.f90:6:
> 
>        PARAMETER (MPI_PATTERN=@MPI_PATTERN@)
>                               1
>        Error: Expected expression at (1) in PARAMETER statement
> 
> Also, could someone please explain what does this notion - @MPI_COMPLEX@ - in mpi.h.in mean
> 
> /* Fortran types */
> #define MPI_COMPLEX           ((MPI_Datatype)@MPI_COMPLEX@)
> 
> 
> regards,
> Tatiana
> 
> 
> 2013/3/28 Dave Goodell <goodell at mcs.anl.gov>
> You almost certainly will also need some modifications in "src/binding/f77/buildiface".  Whenever I need to modify this script I have to stare at it for a while before I'm sure where the modifications need to go.
> 
> I don't think we have any existing types which map to pointers under the hood.  So there isn't a good example for you to use here.  The "MPI_Fint *" changes in particular do not seem correct to me in general.  They *might* work in a pinch on a platform with where INTEGER and (void*) are the same size.  They might also work on little endian platforms like x86/x86_64, but it will be purely by luck if they do.
> 
> You may be better writing your own Fortran bindings by hand in the short term.
> 
> Bill may have something to add on this subject, since he's the primary maintainer for these scripts.
> 
> -Dave
> 
> On Mar 28, 2013, at 3:04 PM CDT, Tatiana V. Martsinkevich <tanya.manekineko at gmail.com> wrote:
> 
> > Thanks for reply.
> >
> > Yes, unfortunately we need fortran interfaces because most of the
> > applications that we use for testing are written in fortran.
> > Will it be enough if I just add a conversion rule to %parmc2f in
> > ./src/binding/f90/buildface? Something like:
> >
> > 'MPI_Pattern' => 'INTEGER(KIND=MPI_ADDRESS_KIND)',
> > 'MPI_Pattern*'  => 'INTEGER(KIND=MPI_ADDRESS_KIND)',
> >
> > And for f77 it will be:
> >
> > 'MPI_Pattern' => 'MPI_Fint *'
> > 'MPI_Pattern*'  => 'MPI_Fint *'
> >
> >
> >> Also, as a side note, please use "MPIX_" for nonstandard MPI extensions.  It makes nonstandard interfaces very obvious to users and will help prevent potential issues with standardization in the future.
> >
> > Duly noted.
> >
> > Tatiana
> >
> > 2013/3/28 Dave Goodell <goodell at mcs.anl.gov>:
> >> On Mar 28, 2013, at 2:16 PM CDT, Tatiana V. Martsinkevich <tanya.manekineko at gmail.com> wrote:
> >>
> >>> We are implementing some fault tolerance related functionality in
> >>> MPICH-3.0.2. We define a new MPI type in
> >>> ~/mpich-3.0.2/src/include/mpi.h.in as:
> >>>
> >>> typedef void* MPI_Pattern;
> >>>
> >>> then in the same file we define:
> >>>
> >>> int MPI_Pattern_declare(char* name, MPI_Pattern* handler);
> >>>
> >>> However, when I run autogen.sh and fortran binding script executes it
> >>> gives out:
> >>>
> >>> Pattern_declare: No parm type for MPI_Pattern* (MPI_Pattern*)
> >>> Use of uninitialized value $fparm in pattern match (m//) at
> >>> ./buildiface line 1289, <FD> line 1710.
> >>
> >> Do you want Fortran interfaces at this time or would you be just as happy to have C-only bindings right now?  The former will require hacking on one or more of the "buildiface" scripts in the "src/binding/LANGUAGE" directories.  The latter can be accomplished by moving your prototypes and types near the end of mpi.h.in, near the prototypes for "MPIX_Mutex_unlock" and "PMPIX_Mutex_unlock".
> >>
> >> Also, as a side note, please use "MPIX_" for nonstandard MPI extensions.  It makes nonstandard interfaces very obvious to users and will help prevent potential issues with standardization in the future.
> >>
> >> -Dave
> >>
> 
>