February 2008 - mpiwg-fortran - mailman3.cels.anl.gov

Re: [MPI3 Fortran] Start discussing new Fortran binding
by Craig Rasmussen 19 Feb '08

19 Feb '08

Silly me, after a couple of brain cycles I figured out how to use ints. I'll put some information up on the wiki about it. How about a conference call on Wednesday or Thursday morning say 11:00 EST? Although if another time works out better, I'll pretty free all week. Craig On Feb 15, 2008, at 1:56 PM, Craig E. Rasmussen wrote: >> > It would be easier to use ints (as we did in the paper). I just > wasn't > sure how to implement everything as ints using OpenMPI. Perhaps > you could > remind me how to do it. Ints would be preferable as I've already > run into > trouble with the current implementation attempt (see attached > files) as > the pointers are not compile time constants (and it seems that > MPI_Recv is > messing around with my datatype pointer). > > Cheers, > Craig > > <mpi.f90><ring.f90><print_ptr.c>

1 0

Re: [MPI3 Fortran] Start discussing new Fortran binding
by Jeff Squyres 18 Feb '08

18 Feb '08

On Feb 15, 2008, at 3:56 PM, Craig E. Rasmussen wrote: >> If I grok this, it looks like the Fortran binding for MPI_COMM_WORLD >> will always be a C pointer. How would that translate automatically >> back to the C bindings if the back-end C type is an integer? I >> thought the goal was to make the C and Fortran types be the same (or >> automatically translatable by the compiler). >> >> (note that OMPI uses pointers in C, so ompi_mpi_comm_world is a >> pointer, not an integer. But you could s/ompi_mpi_comm_world/ >> mpich_mpi_comm_world/ and have a valid-ish example) > > It would be easier to use ints (as we did in the paper). I just > wasn't > sure how to implement everything as ints using OpenMPI. Perhaps you > could > remind me how to do it. No idea. :-) > Ints would be preferable as I've already run into > trouble with the current implementation attempt (see attached files) > as > the pointers are not compile time constants (and it seems that > MPI_Recv is > messing around with my datatype pointer). Here's what I see in your MPI_Recv: function MPI_Recv(buf, count, datatype, source, tag, comm, status) & BIND(C, name="MPI_Recv") result(err) use, intrinsic :: ISO_C_BINDING import MPI_HANDLE_KIND implicit none type(C_PTR), value, intent(in) :: buf integer(C_INT), value, intent(in) :: count type(C_PTR), value, intent(in) :: datatype integer(C_INT), value, intent(in) :: source integer(C_INT), value, intent(in) :: tag type(C_PTR), value, intent(in) :: comm type(C_PTR), intent(out) :: status integer(C_INT) :: err end function MPI_Recv So I think we need to leave it up to the implementation as to what the type of the MPI handles are (datatype and comm in this case). There's two cases that I'm aware of: - the handle is an int in C - the handle is a pointer in C Not being a fortran guy, here's a few options that jump to mind (some/ all of these make be senseless in Fortran): 1. Can there be a fortran integer equivalent than can map to either of those, given an appropriate KIND? 2. Is there a way to use a preprocessor to replace something like this integer(MPI_HANDLE_TYPE), value, intent(in) :: datatype with either C_INT or C_PTR, depending on what the implementation requires? 3. Does F03 have the equivalent to C's "typedef" such that we could create handle types that map to integer(C_INT) or type(C_PTR) (and the user doesn't need to know)? Remember that MPI only has to be source code compatible, not binary compatible. Sidenote: I don't know why MPI_Recv would modify the value of MPI_INT in F03 -- it doesn't on the C side. -- Jeff Squyres Cisco Systems

1 0

Re: [MPI3 Fortran] Start discussing new Fortran binding
by Craig E. Rasmussen 15 Feb '08

15 Feb '08

> On Feb 14, 2008, at 9:03 PM, Craig E. Rasmussen wrote: > >>>> >>>> ! We need to figure out the type of MPI handles. >>>> ! Currently I'm leaning toward TYPE(C_PTR) rather than >>>> ! the INTEGER(MPI_HANDLE_KIND) Jeff and I used in our paper. >>> >>> Will this be able to handle MPI implementations that use integers as >>> MPI handles in C? >> >> So are we talking about global constants like MPI_COMM_WORLD? If so >> it >> looks like we can use the same global integers that are used in C. > > That's the point, though -- some MPI implementations (e.g., Open MPI) > use pointers for MPI handles in C, while others (e.g., MPICH) use C > int's. The standard doesn't specify what the type has to be, so > implementations do what they please. > >> We >> would just initialize MPI_COMM_WORLD in Fortran MPI_Init to the >> address of >> the global C integer. If you can grok the Fortran, I've included >> the code >> snippets: >> >> ! the global integer in OMPI >> integer, bind(C) :: ompi_mpi_comm_world >> >> ! the declaration of the handle in Fortran >> type(C_PTR) :: MPI_COMM_WORLD >> >> ! the initialization in MPI_Init() >> MPI_COMM_WORLD = C_LOC(ompi_mpi_comm_world) > > > If I grok this, it looks like the Fortran binding for MPI_COMM_WORLD > will always be a C pointer. How would that translate automatically > back to the C bindings if the back-end C type is an integer? I > thought the goal was to make the C and Fortran types be the same (or > automatically translatable by the compiler). > > (note that OMPI uses pointers in C, so ompi_mpi_comm_world is a > pointer, not an integer. But you could s/ompi_mpi_comm_world/ > mpich_mpi_comm_world/ and have a valid-ish example) > It would be easier to use ints (as we did in the paper). I just wasn't sure how to implement everything as ints using OpenMPI. Perhaps you could remind me how to do it. Ints would be preferable as I've already run into trouble with the current implementation attempt (see attached files) as the pointers are not compile time constants (and it seems that MPI_Recv is messing around with my datatype pointer). Cheers, Craig

1 0

Re: [MPI3 Fortran] Start discussing new Fortran binding
by Jeff Squyres 14 Feb '08

14 Feb '08

On Feb 14, 2008, at 9:03 PM, Craig E. Rasmussen wrote: >>> >>> ! We need to figure out the type of MPI handles. >>> ! Currently I'm leaning toward TYPE(C_PTR) rather than >>> ! the INTEGER(MPI_HANDLE_KIND) Jeff and I used in our paper. >> >> Will this be able to handle MPI implementations that use integers as >> MPI handles in C? > > So are we talking about global constants like MPI_COMM_WORLD? If so > it > looks like we can use the same global integers that are used in C. That's the point, though -- some MPI implementations (e.g., Open MPI) use pointers for MPI handles in C, while others (e.g., MPICH) use C int's. The standard doesn't specify what the type has to be, so implementations do what they please. > We > would just initialize MPI_COMM_WORLD in Fortran MPI_Init to the > address of > the global C integer. If you can grok the Fortran, I've included > the code > snippets: > > ! the global integer in OMPI > integer, bind(C) :: ompi_mpi_comm_world > > ! the declaration of the handle in Fortran > type(C_PTR) :: MPI_COMM_WORLD > > ! the initialization in MPI_Init() > MPI_COMM_WORLD = C_LOC(ompi_mpi_comm_world) If I grok this, it looks like the Fortran binding for MPI_COMM_WORLD will always be a C pointer. How would that translate automatically back to the C bindings if the back-end C type is an integer? I thought the goal was to make the C and Fortran types be the same (or automatically translatable by the compiler). (note that OMPI uses pointers in C, so ompi_mpi_comm_world is a pointer, not an integer. But you could s/ompi_mpi_comm_world/ mpich_mpi_comm_world/ and have a valid-ish example) -- Jeff Squyres Cisco Systems

1 0

Re: [MPI3 Fortran] Start discussing new Fortran binding
by Craig E. Rasmussen 14 Feb '08

14 Feb '08

> On Feb 13, 2008, at 5:07 PM, Craig E. Rasmussen wrote: > >> ! I think we should provide better interfaces to command line >> arguments. >> ! It can be done through GET_COMMAND and GET_COMMAND_ARGUMENT. >> ! MPI_Init should be able to pass back something like modified C >> argc and >> argv arguments. >> >> err = MPI_Init() > > What do you mean? Pretty much what you said (see below). > > I think that if an MPI implementation needs the argv, it can have its > own implementation of MPI_Init in Fortran that calls GET_COMMAND and > GET_COMMAND_ARGUMENT; the argc/argv arguments to MPI_INIT in C are > allowed be NULL,NULL. Right. I'm thinking that we can wrap MPI_Init so that the functionality is to pass back essentially the same argc, argv list that C users get. > >> ! We need to figure out the type of MPI handles. >> ! Currently I'm leaning toward TYPE(C_PTR) rather than >> ! the INTEGER(MPI_HANDLE_KIND) Jeff and I used in our paper. > > Will this be able to handle MPI implementations that use integers as > MPI handles in C? So are we talking about global constants like MPI_COMM_WORLD? If so it looks like we can use the same global integers that are used in C. We would just initialize MPI_COMM_WORLD in Fortran MPI_Init to the address of the global C integer. If you can grok the Fortran, I've included the code snippets: ! the global integer in OMPI integer, bind(C) :: ompi_mpi_comm_world ! the declaration of the handle in Fortran type(C_PTR) :: MPI_COMM_WORLD ! the initialization in MPI_Init() MPI_COMM_WORLD = C_LOC(ompi_mpi_comm_world) Cheers, Craig

1 0

Re: [MPI3 Fortran] Start discussing new Fortran binding
by Craig E. Rasmussen 14 Feb '08

14 Feb '08

> On Feb 13, 2008, at 5:07 PM, Craig E. Rasmussen wrote: > >> ! I think we should provide better interfaces to command line >> arguments. >> ! It can be done through GET_COMMAND and GET_COMMAND_ARGUMENT. >> ! MPI_Init should be able to pass back something like modified C >> argc and >> argv arguments. >> >> err = MPI_Init() > > What do you mean? Pretty much what you do (see below). > > I think that if an MPI implementation needs the argv, it can have its > own implementation of MPI_Init in Fortran that calls GET_COMMAND and > GET_COMMAND_ARGUMENT; the argc/argv arguments to MPI_INIT in C are > allowed be NULL,NULL. Right. I'm thinking that we can wrap MPI_Init so that the functionality is to pass back essentially the argc, argv list that C users get. > >> ! We need to figure out the type of MPI handles. >> ! Currently I'm leaning toward TYPE(C_PTR) rather than >> ! the INTEGER(MPI_HANDLE_KIND) Jeff and I used in our paper. > > Will this be able to handle MPI implementations that use integers as > MPI handles in C? So are we talking about global constants like MPI_COMM_WORLD? If so it looks like we can use the same global integers that are used in C. We would just initialize MPI_COMM_WORLD in Fortran MPI_Init to the address of the global C integer. If you can grok the Fortran, I've included the code snippets: ! the global integer in OMPI integer, bind(C) :: ompi_mpi_comm_world ! the declaration of the handle in Fortran type(C_PTR) :: MPI_COMM_WORLD ! the initialization in MPI_Init() MPI_COMM_WORLD = C_LOC(ompi_mpi_comm_world) Cheers, Craig

1 0

Re: [MPI3 Fortran] Start discussing new Fortran binding
by Jeff Squyres 13 Feb '08

13 Feb '08

On Feb 13, 2008, at 5:07 PM, Craig E. Rasmussen wrote: > ! I think we should provide better interfaces to command line > arguments. > ! It can be done through GET_COMMAND and GET_COMMAND_ARGUMENT. > ! MPI_Init should be able to pass back something like modified C > argc and > argv arguments. > > err = MPI_Init() What do you mean? I think that if an MPI implementation needs the argv, it can have its own implementation of MPI_Init in Fortran that calls GET_COMMAND and GET_COMMAND_ARGUMENT; the argc/argv arguments to MPI_INIT in C are allowed be NULL,NULL. > ! We need to figure out the type of MPI handles. > ! Currently I'm leaning toward TYPE(C_PTR) rather than > ! the INTEGER(MPI_HANDLE_KIND) Jeff and I used in our paper. Will this be able to handle MPI implementations that use integers as MPI handles in C? -- Jeff Squyres Cisco Systems

1 0

[MPI3 Fortran] Start discussing new Fortran binding
by Craig E. Rasmussen 13 Feb '08

13 Feb '08

I propose that we have a phone call next week sometime to begin to discuss potential new Fortran bindings. I have put some ideas on the wiki as a starting point for discussion (text copied below). I'll add more in the next couple of days. Cheers, Craig ------------------ ! import C interop and MPI interfaces use, intrinsic :: ISO_C_BINDING use :: MPI_C_BINDING ! this needs to be given a name integer(C_INT) :: rank, size, err ! this module needs to be given a size (C_INT here) ! I think we should provide better interfaces to command line arguments. ! It can be done through GET_COMMAND and GET_COMMAND_ARGUMENT. ! MPI_Init should be able to pass back something like modified C argc and argv arguments. err = MPI_Init() ! We need to figure out the type of MPI handles. ! Currently I'm leaning toward TYPE(C_PTR) rather than ! the INTEGER(MPI_HANDLE_KIND) Jeff and I used in our paper. err = MPI_Comm_rank(MPI_COMM_WORLD, rank) err = MPI_Comm_size(MPI_COMM_WORLD, size) err = MPI_Finalize()

1 0