Re: [MPI3 Fortran] [Mpi-comments] MPI 3.0: Fortran 2008 interface - issue with the LOGICAL kind
On 3/21/13 2:44 AM, Rolf Rabenseifner wrote:
Thank you for the answers. Four comments:
1. About the linker names and intenal call sequence:
I think that Bill's proposal kills the Fortran Profiling Interface. If the user defines a subroutine MPI_Test (....) do something call PMPI _Test() do something end subroutine MPI_Test
Yes, this is true. Having these two implementation options A. Internal real MPI C calls with A1. standardized Fortran linker names, or with A2. non-standardized Fortran names, i.e. within a Fortran module. B. Standardized Fortran linker names. would kill the upper Fortran example as soon as an implementation chooses A2.
Bill, do you have an idea how to solve the A2 problem. Is it still possible to provide the standardized names when you somehow inline the wrappers?
Yes, this is possible. In fact, I think it will be required because of the requirement to still support mpif.h. Users who use only mpif.h would not have access to the module routines. Therefore, it would be necessary for the implementation to take a copy of the wrappers in the mpi module (the new descriptor versions are not available with mpif.h if I understand the rules) and make a standalone file that is compiled separately. This would be done mechanically to avoid maintaining this file separately. The stand-alone file would result in external linker names like mpi_send_ for the generic Linux case. The compiler would likely not inline these, though a particularly aggressive one might, unless told otherwise. As far as the PMPI versions, one option would be to just replicate the wrapper routines in the standalone file made for the mpif.h users, but with the routine name changed to add a P at the beginning. Again, this would be a mechanical change. Another option would be to insert an ENTRY statement in the files as a way to give the routine a second name (saves space in the library, but ENTRY is considered obsolete). So I'd suggest the duplicate wrappers with the P-changed name. These routines are all short, so library size seems like a minor issue. Since you cannot have two routines with external name pmpi_send_, I assume that the pmpi_xxx Fortran versions do not exist for the new descriptor argument variants.
Then, we would have only A1 which allows that a pure PMPI Fortran user can freely choose to intercept the Fortran or the C interface.
2. About MPI_Send_desc:
Yes, Bill, you are absolutely right. I thought that the MPI_Isend_desc internally calls MPI_Isend, but this would be an implementation that internally copies the non-contiguous Fortran application buffer into a contiguous scratch buffer and then calls MPI_Isend (the scratch buffer is released latest in MPI_Wait or MPI_Test). Optimized solutions would do flattening on the fly of the combination of the description of the Fortran buffer combined with the description in the datatype handle, and this double-flattening within the MPI_Isend_desc.
Right. I had assumed that the MPI_Send_cdesc would do something like if (buf is contiguous) then extract the buf address from the descriptor and use it to call MPI_Send. The other arguments are identical and just pass through. else work through the descriptor, calling low level transport routines on sub-blocks of the buffer. We should at least write the rules such that an implementation like this is allowed. Making a copy of the whole buffer is potentially dangerous of you are low on available memory and the buffer is large. Cheers, Bill
Therefore special naming for option A must be defined.
This is only for the MPI routine groups MPI_SEND and MPI_ISEND (see MPI-3.0, page 607 and ticket #364)
3. Routine groups:
All what we do should be done per routine group, i.e., mpi.h tells by macros how it is done within the specific routine groups.
4. My apologies to Hubert.
The list should have been
I need at least answers for - mpich (Bill Long), - OpenMPI (Jeff or Craig), - NEC MPI (Hubert), - other, if there are other independent mpi_f08 implementations, - some tools people.
Best regards Rolf
----- Original Message -----
From: "Bill Long" <[email protected]> To: "Jeff Squyres (jsquyres)" <[email protected]> Cc: "Martin Schulz" <[email protected]>, "MPI-3 Fortran working group" <[email protected]> Sent: Wednesday, March 20, 2013 9:40:08 PM Subject: Re: [MPI3 Fortran] [Mpi-comments] MPI 3.0: Fortran 2008 interface - issue with the LOGICAL kind My impression is that Rolf's proposal was to offer two options:
a) no bind(c) on the Fortran wrappers, no requirements on the Fortran external names, but the wrappers are REQUIRED to call the C library routines.
b) specific external names are required for the Fortran-called routines, but they do not have to be wrappers. (Still a bit unclear to me in this case where bind(c) might be required).
On 3/20/13 2:58 PM, Jeff Squyres (jsquyres) wrote:
Not only that, I think that there are some cases where tools *need* to intercept the Fortran call, not the C call. In OMPI, for example, we have a few corner cases where the Fortran MPI subroutine does *not* call the corresponding C MPI API function. Instead, it calls a back-end OMPI routine.
True for implementations using option (b), but not for the (a) case.
One obvious case that jumps to mind is anything involving a callback.
Looking at the callback examples on page 268, I have some questions:
1) The BIND(C) and an argument of type LOGICAL are likely incompatible, so either the BIND(C) or the LOGICAL needs to go.
2) The external names for the C and Fortran versions are different. I assume this does not matter assuming that the function that is calling these has only an address, and not a name, to call. But how does the function know the calling signature for the callback routine with only a procedure pointer to work with?
If I were writing a code with a callback like this, I would want to assume the C prototype even of the code was in Fortran. I assume the existence of the separate (and apparently flawed) Fortran interface is for legacy code. But there should not be legacy mpi_f08 code. A bit confusing.
If you have integer handles in C, they might be able to get away with not swizzling/de-swizzling the arguments to the callback to the proper language. But this is not always the case.
On Mar 20, 2013, at 12:29 PM, Hubert Ritzdorf <[email protected]> wrote:
Hi,
I think that Bill's proposal kills the Fortran Profiling Interface.
If the user defines a
subroutine MPI_Test (....)
do something
call PMPI _Test()
do something
end subroutine MPI_Test
I was under the impression that users would not directly call PMPI routines. Those were for the tools people. Why would a normal user want to do something like the above?
then this function will never be executed because the Fortrran compiler may have already inlined the MPI_Test subroutine in the original code or the module version of MPI_Test is used.
In the scope where this routine is called, the Fortran wrapper would have to be inaccessible. That would be done by NOT having a USE of the module in this scope. The external name of the module procedure would be mangled to include the name of the module, whereas a stand-alone routine like this would have an external name like mpi_test_ . The linker would see these as different routines.
Cheers, Bill
Hubert ________________________________________ From: [email protected] [[email protected]] on behalf of Rolf Rabenseifner [[email protected]] Sent: Wednesday, March 20, 2013 6:09 PM To: [email protected]; MPI-3 Fortran working group Cc: Martin Schulz Subject: Re: [MPI3 Fortran] [Mpi-comments] MPI 3.0: Fortran 2008 interface - issue with the LOGICAL kind
This mail contains a proposal how we can solve the current Fortran problems and possibly making the tools people also happy.
I need at least answers for - mpich (Bill Long), - OpenMPI (Jeff or Craig), - some tools people.
Let me summarize what I've learnt so far for mpi_f08:
- #364 about LOGICAL correction is a must. - MPI-3.0 allows BIND(C) and non-BIND(C), with the new restrictions in #364 - BIND(C) needs TR29113, therefore any implementation now should use non-BIND(C) for all non-TR29113 compilers. - Most (or all) try to use the non-BIND(C) interface. - Best way to do this are thin wrappers in Fortran that internally call - the oficial MPI-3.0 C binding (here, the tools people have a problem because they expect to wrap the Fortran interface and may therefore double-wrap), or - or an internal MPI-3.0 C routine (which is half-perfect for the tools folk, because they have to handle the implementation-dependant Fortran names with underscores). - And at least Bill wants to include the Fortran MPI library routines itselves in the mpi_f08 module. This is in conflict with the current MPI-3.0 linker naming rules.
Implications and proposal: - We should rewrite the linker names section MPI-3.0 17.1.5 with following goals: - Allow additionally to define that the PMPI is provided only on the MPI C interface and the Fortran MPI routines internally call the C interface; again, a macro definition in mpi.h tells whether this is chosen. Question: Would this cause a problem when linking a C or Fortran user application with some own MPI_Test etc. profiling routines that internally call PMPI_Test? - If Fortran profiling is done through the MPI C binding then there are no rules about the Fortran linker names needed, and therefore, one may put the thin Fortran wrappers into a Fortran module. - We can define BIND(C) for MPI routine groups as a future option that is currently not used. This would reduce the current effort for the tools people. Is there anybody who wants to implement one of the Fortran routine groups with BIND(C)?
I would expect, that this is the best to solve the problem that #364 prohibits BIND(C) for some routines and implementors want to do the same method for all routines.
Don't forget that BIND(C) and TR29113 is still important for implementing these thin wrappers internally.
Without using BIND(C) on the official mpi_f08 and mpi module interface, some features of TR29113 are not used: - the extension for interoperabple CHARACTER(LEN=*), - the extension for OPTIONAL dummy arguments.
Other TR29113 features are heavily used: - The DIMENSION(..), TYPE(*) is needed for some internal C-written MPI_Isend_desc. - The ASYNCHRONOUS usage for communication buffers.
Should we go in this direction?
Best regards Rolf
----- Original Message -----
From: "Bill Long" <[email protected]> To: "Jeff Squyres (jsquyres)" <[email protected]> Cc: "Martin Schulz" <[email protected]>, "MPI-3 Fortran working group" <[email protected]> Sent: Wednesday, March 20, 2013 4:01:52 PM Subject: Re: [MPI3 Fortran] [Mpi-comments] MPI 3.0: Fortran 2008 interface - issue with the LOGICAL kind On 3/20/13 8:26 AM, Jeff Squyres (jsquyres) wrote:
On Mar 20, 2013, at 6:25 AM, Bill Long <[email protected]> wrote:
interface !--> MPI_Test
! int MPI_Test(MPI_Request *request, int *flag, MPI_Status *status);
Function MPI_Test_C( request, flag, status) & BIND(C, name="MPI_Test") RESULT (res) import :: C_request, c_int, MPI_Status_C integer(C_request) :: request
This is not correct. You cannot assume that MPI handles are integers (they're pointers in Open MPI).
This is correct for mpich. OpenMPI would evidently have a different version. But the concept is still the same - make a correct interface to the C library routine, however it is defined by the implementation, and then call that from a thin Fortran wrapper that takes care of the issues related to LOGICAL, the use of default INTEGER, and OPTIONAL arguments.
Cheers, Bill
-- Bill Long [email protected] Fortran Technical Support & voice: 651-605-9024 Bioinformatics Software Development fax: 651-605-9142 Cray Inc./Cray Plaza, Suite 210/380 Jackson St./St. Paul, MN 55101
_______________________________________________ mpi3-fortran mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-fortran
-- Dr. Rolf Rabenseifner . . . . . . . . . .. email [email protected] High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530 University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832 Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307) _______________________________________________ mpi3-fortran mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-fortran
Click https://www.mailcontrol.com/sr/oUBxarVXlArGX2PQPOmvUrwgXREt4LZXdolWasoQmWDbX... to report this email as spam.
_______________________________________________ mpi3-fortran mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-fortran
-- Bill Long [email protected] Fortran Technical Support & voice: 651-605-9024 Bioinformatics Software Development fax: 651-605-9142 Cray Inc./Cray Plaza, Suite 210/380 Jackson St./St. Paul, MN 55101
_______________________________________________ mpi3-fortran mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-fortran
-- Bill Long [email protected] Fortran Technical Support & voice: 651-605-9024 Bioinformatics Software Development fax: 651-605-9142 Cray Inc./Cray Plaza, Suite 210/380 Jackson St./St. Paul, MN 55101
participants (1)
-
Bill Long