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

Re: [MPIWG Fortran] BIND(C) attribute of MPI callback functions
by N.M. Maclaren 02 Jul '14

02 Jul '14

On Jul 2 2014, Bill Long wrote: >On Jul 2, 2014, at 2:11 AM, Rolf Rabenseifner <rabenseifner(a)hlrs.de> wrote: >> >> The reason for removal of BIND(C) was simple: >> Four callbacks contain LOGICAL arguments >> and we learnt that LOGICAL and BIND(C) does not work together. I considered an extension, mentioned it, and we discussed it. The problem is virtually insoluble, because the semantic models for Fortran LOGICAL and C truth values are so different, and any attempt to make them interoperate would be a truly ghastly mess. There are outstanding issues that could be changed, but this is not one. Yes, I know that Fortran 2008 makes LOGICAL(C_BOOL) interoperate with _Bool, but I regard that as being even more of a mistake than specifying that INTEGER(C_SIZE_T) interoperates with size_t. Don't go there; REALLY don't go there :-( Regards, Nick Maclaren.

1 0

Re: [MPIWG Fortran] BIND(C) attribute of MPI callback functions
by Bill Long 02 Jul '14

02 Jul '14

On Jul 2, 2014, at 2:11 AM, Rolf Rabenseifner <rabenseifner(a)hlrs.de> wrote: > Bill, > > The reason for removal of BIND(C) was simple: > Four callbacks contain LOGICAL arguments > and we learnt that LOGICAL and BIND(C) does not work together. > Therefore the errata is correct. Agreed. > > I hope you find a solution for storing the callback pointer > in a C structure and to bring it back to Fortran. > I expect that calling the Fortran callbacks > in a portable way may require Fortran code > that calls the callback. For compilers that support the TS, there should be no problem. As noted by Junchao, Cray and Gfortran did not have a problem. The IBM compiler appears to have overlooked the TS change. A bug report to IBM should resolve the issue. The fix is trivial - just disabling an error message. Cheers, Bill > > Best regards > Rolf > > > ----- Original Message ----- >> From: "Bill Long" <longb(a)cray.com> >> To: "MPI-WG Fortran working group" <mpiwg-fortran(a)lists.mpi-forum.org> >> Sent: Tuesday, July 1, 2014 9:06:05 PM >> Subject: Re: [MPIWG Fortran] BIND(C) attribute of MPI callback functions >> >> >> On Jul 1, 2014, at 2:00 PM, Junchao Zhang <jczhang(a)mcs.anl.gov> >> wrote: >> >>> Hello, >>> Errata for MPI-3.0 removed the BIND(C) attribute of MPI callback >>> functions. But in an implementation, to pass the Fortran >>> procedure to the backend C, I need to get its C function pointer >>> with c_funloc, as the following code shows. When I compile it >>> with IBM XLF 15.1, it complains with such an error "1516-203 (S) >>> The argument to the intrinsic procedure C_FUNLOC must be a >>> procedure with the BIND(C) attribute." In my experiments, gcc >>> and Cray ftn are fine with it. Looking at >>> https://gcc.gnu.org/onlinedocs/gfortran/C_005fFUNLOC.html, I >>> find c_funloc indeed requires an interoperable function >>> argument. Any comments? >> >> This requirement on the argument of C_FUNLOC was removed by TS 29113, >> subclause 8.1 “Removed restrictions on ISO_C_BINDING module >> procedures”. >> >> Cheers, >> Bill >> >> >> >>> >>> subroutine MPI_Comm_create_errhandler_f08(comm_errhandler_fn, >>> errhandler, ierror) >>> use, intrinsic :: iso_c_binding, only : c_funloc, c_int, >>> c_funptr >>> use :: mpi_f08, only : MPI_Errhandler, >>> MPI_Comm_errhandler_function >>> use :: mpi_c_interface, only : c_Errhandler, >>> MPIR_Comm_create_errhandler_c >>> >>> implicit none >>> >>> procedure(MPI_Comm_errhandler_function) :: comm_errhandler_fn >>> type(MPI_Errhandler), intent(out) :: errhandler >>> integer, optional, intent(out) :: ierror >>> >>> type(c_funptr) :: comm_errhandler_fn_c >>> >>> comm_errhandler_fn_c = c_funloc(comm_errhandler_fn) !! xlf >>> issues an error >>> ierror_c = MPIR_Comm_create_errhandler_c(comm_errhandler_fn_c, >>> errhandler%MPI_VAL) >>> ... >>> end subroutine MPI_Comm_create_errhandler_f08 >>> >>> --Junchao Zhang >>> _______________________________________________ >>> mpiwg-fortran mailing list >>> mpiwg-fortran(a)lists.mpi-forum.org >>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran >> >> Bill Long >> longb(a)cray.com >> Fortran Technical Suport & voice: >> 651-605-9024 >> Bioinformatics Software Development fax: >> 651-605-9142 >> Cray Inc./ Cray Plaza, Suite 210/ 380 Jackson St./ St. Paul, MN 55101 >> >> >> _______________________________________________ >> mpiwg-fortran mailing list >> mpiwg-fortran(a)lists.mpi-forum.org >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran >> > > -- > Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner(a)hlrs.de > 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) > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran Bill Long longb(a)cray.com Fortran Technical Suport & voice: 651-605-9024 Bioinformatics Software Development fax: 651-605-9142 Cray Inc./ Cray Plaza, Suite 210/ 380 Jackson St./ St. Paul, MN 55101

1 0

Re: [MPIWG Fortran] BIND(C) attribute of MPI callback functions
by Rolf Rabenseifner 02 Jul '14

02 Jul '14

Bill, The reason for removal of BIND(C) was simple: Four callbacks contain LOGICAL arguments and we learnt that LOGICAL and BIND(C) does not work together. Therefore the errata is correct. I hope you find a solution for storing the callback pointer in a C structure and to bring it back to Fortran. I expect that calling the Fortran callbacks in a portable way may require Fortran code that calls the callback. Best regards Rolf ----- Original Message ----- > From: "Bill Long" <longb(a)cray.com> > To: "MPI-WG Fortran working group" <mpiwg-fortran(a)lists.mpi-forum.org> > Sent: Tuesday, July 1, 2014 9:06:05 PM > Subject: Re: [MPIWG Fortran] BIND(C) attribute of MPI callback functions > > > On Jul 1, 2014, at 2:00 PM, Junchao Zhang <jczhang(a)mcs.anl.gov> > wrote: > > > Hello, > > Errata for MPI-3.0 removed the BIND(C) attribute of MPI callback > > functions. But in an implementation, to pass the Fortran > > procedure to the backend C, I need to get its C function pointer > > with c_funloc, as the following code shows. When I compile it > > with IBM XLF 15.1, it complains with such an error "1516-203 (S) > > The argument to the intrinsic procedure C_FUNLOC must be a > > procedure with the BIND(C) attribute." In my experiments, gcc > > and Cray ftn are fine with it. Looking at > > https://gcc.gnu.org/onlinedocs/gfortran/C_005fFUNLOC.html, I > > find c_funloc indeed requires an interoperable function > > argument. Any comments? > > This requirement on the argument of C_FUNLOC was removed by TS 29113, > subclause 8.1 “Removed restrictions on ISO_C_BINDING module > procedures”. > > Cheers, > Bill > > > > > > > subroutine MPI_Comm_create_errhandler_f08(comm_errhandler_fn, > > errhandler, ierror) > > use, intrinsic :: iso_c_binding, only : c_funloc, c_int, > > c_funptr > > use :: mpi_f08, only : MPI_Errhandler, > > MPI_Comm_errhandler_function > > use :: mpi_c_interface, only : c_Errhandler, > > MPIR_Comm_create_errhandler_c > > > > implicit none > > > > procedure(MPI_Comm_errhandler_function) :: comm_errhandler_fn > > type(MPI_Errhandler), intent(out) :: errhandler > > integer, optional, intent(out) :: ierror > > > > type(c_funptr) :: comm_errhandler_fn_c > > > > comm_errhandler_fn_c = c_funloc(comm_errhandler_fn) !! xlf > > issues an error > > ierror_c = MPIR_Comm_create_errhandler_c(comm_errhandler_fn_c, > > errhandler%MPI_VAL) > > ... > > end subroutine MPI_Comm_create_errhandler_f08 > > > > --Junchao Zhang > > _______________________________________________ > > mpiwg-fortran mailing list > > mpiwg-fortran(a)lists.mpi-forum.org > > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran > > Bill Long > longb(a)cray.com > Fortran Technical Suport & voice: > 651-605-9024 > Bioinformatics Software Development fax: > 651-605-9142 > Cray Inc./ Cray Plaza, Suite 210/ 380 Jackson St./ St. Paul, MN 55101 > > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran > -- Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner(a)hlrs.de 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)

1 0

Re: [MPIWG Fortran] BIND(C) attribute of MPI callback functions
by Bill Long 01 Jul '14

01 Jul '14

On Jul 1, 2014, at 2:00 PM, Junchao Zhang <jczhang(a)mcs.anl.gov> wrote: > Hello, > Errata for MPI-3.0 removed the BIND(C) attribute of MPI callback functions. But in an implementation, to pass the Fortran procedure to the backend C, I need to get its C function pointer with c_funloc, as the following code shows. When I compile it with IBM XLF 15.1, it complains with such an error "1516-203 (S) The argument to the intrinsic procedure C_FUNLOC must be a procedure with the BIND(C) attribute." In my experiments, gcc and Cray ftn are fine with it. Looking at https://gcc.gnu.org/onlinedocs/gfortran/C_005fFUNLOC.html, I find c_funloc indeed requires an interoperable function argument. Any comments? This requirement on the argument of C_FUNLOC was removed by TS 29113, subclause 8.1 “Removed restrictions on ISO_C_BINDING module procedures”. Cheers, Bill > > subroutine MPI_Comm_create_errhandler_f08(comm_errhandler_fn, errhandler, ierror) > use, intrinsic :: iso_c_binding, only : c_funloc, c_int, c_funptr > use :: mpi_f08, only : MPI_Errhandler, MPI_Comm_errhandler_function > use :: mpi_c_interface, only : c_Errhandler, MPIR_Comm_create_errhandler_c > > implicit none > > procedure(MPI_Comm_errhandler_function) :: comm_errhandler_fn > type(MPI_Errhandler), intent(out) :: errhandler > integer, optional, intent(out) :: ierror > > type(c_funptr) :: comm_errhandler_fn_c > > comm_errhandler_fn_c = c_funloc(comm_errhandler_fn) !! xlf issues an error > ierror_c = MPIR_Comm_create_errhandler_c(comm_errhandler_fn_c, errhandler%MPI_VAL) > ... > end subroutine MPI_Comm_create_errhandler_f08 > > --Junchao Zhang > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran Bill Long longb(a)cray.com Fortran Technical Suport & voice: 651-605-9024 Bioinformatics Software Development fax: 651-605-9142 Cray Inc./ Cray Plaza, Suite 210/ 380 Jackson St./ St. Paul, MN 55101

1 0

BIND(C) attribute of MPI callback functions
by Junchao Zhang 01 Jul '14

01 Jul '14

Hello, Errata for MPI-3.0 removed the BIND(C) attribute of MPI callback functions. But in an implementation, to pass the Fortran procedure to the backend C, I need to get its C function pointer with c_funloc, as the following code shows. When I compile it with IBM XLF 15.1, it complains with such an error "1516-203 (S) The argument to the intrinsic procedure C_FUNLOC must be a procedure with the BIND(C) attribute." In my experiments, gcc and Cray ftn are fine with it. Looking at https://gcc.gnu.org/onlinedocs/gfortran/C_005fFUNLOC.html, I find c_funloc indeed requires an interoperable function argument. Any comments? subroutine MPI_Comm_create_errhandler_f08(comm_errhandler_fn, errhandler, ierror) use, intrinsic :: iso_c_binding, only : c_funloc, c_int, c_funptr use :: mpi_f08, only : MPI_Errhandler, MPI_Comm_errhandler_function use :: mpi_c_interface, only : c_Errhandler, MPIR_Comm_create_errhandler_c implicit none procedure(MPI_Comm_errhandler_function) :: comm_errhandler_fn type(MPI_Errhandler), intent(out) :: errhandler integer, optional, intent(out) :: ierror type(c_funptr) :: comm_errhandler_fn_c comm_errhandler_fn_c = c_funloc(comm_errhandler_fn) !! xlf issues an error ierror_c = MPIR_Comm_create_errhandler_c(comm_errhandler_fn_c, errhandler%MPI_VAL) ... end subroutine MPI_Comm_create_errhandler_f08 --Junchao Zhang

1 0

Re: [MPIWG Fortran] errata
by Jeff Squyres (jsquyres) 17 Jun '14

17 Jun '14

Looks like Rolf already handled this on the ticket. On Jun 17, 2014, at 7:44 AM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > See https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/432 > > Jeff > > -- > Jeff Hammond > jeff.science(a)gmail.com > http://jeffhammond.github.io/ > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran -- Jeff Squyres jsquyres(a)cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/

1 0

errata
by Jeff Hammond 17 Jun '14

17 Jun '14

See https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/432 Jeff -- Jeff Hammond jeff.science(a)gmail.com http://jeffhammond.github.io/

1 0

Re: [MPIWG Fortran] MPI-3.0 Fortran errata latex integration: Reviews requested
by Jeff Squyres (jsquyres) 16 Jun '14

16 Jun '14

I’ve received one review so far (from Hubert; he found a mistake). Can others please review this text to ensure I did the merge properly? It’s a LOT of text to merge — I would really appreciate some additional eyeballs on it to ensure it was done properly. Thanks! On May 14, 2014, at 12:54 PM, Jeff Squyres (jsquyres) <jsquyres(a)cisco.com> wrote: > I have integrated all the MPI-3.0 Fortran-related errata as of this morning into the MPI-3.1 tree latex. It is not committed in the SVN MPI-3.1 tree because I'd like to get reviews first. > > Per Forum direction, I did *not* use the latex change macros. Instead, we're relying on diffs to review the changes. > > Once I get 2-3 reviews to ensure that I have applied the errata properly, I'll commit all of it to the official MPI-3.1 SVN tree. > > ----- > > Here's how I've made the changes available for review: > > 0. All the files listed below are available here: > > http://www.open-mpi.org/~jsquyres/mpi-forum/ > > 1. A PDF of the entire standard is there; all Fortran-related errata have been applied. You should proof this document. To know exactly what to proof in this PDF, however, see the follow documents... > > 2. I also put my copy of errata-30.pdf there, including BRIGHT RED annotations indicating which errata items I incorporated into the PDF from #1. > > 3. I also posted a tarball of a git tree where I did my work (NOTE: because this git tree includes the entire latex source, it is behind a trivial password. Email me to get it). There are 36 commits; each commit is the application of an individual errata item. You can use normal git commands to view the diffs. For example: > > tar xf mpi-3.1-tree-with-fortran-errata-applied-git.tar.bz2 > cd mpi-3.1 > git log > # ...see all my individual commits... > git show c126721 > # ...see the trivial addition of Type(MPI_Message) into the annex... > > 4. Or, if you don't like/know git, the individual patch files (numbered in order) are also available. For example: > > See a list of all the individual patches at > http://www.open-mpi.org/~jsquyres/mpi-forum/individual-patches/ > > See an individual patch: > wget http://www.open-mpi.org/~jsquyres/mpi-forum/individual-patches/0033-Add-TYP… > > -- > Jeff Squyres > jsquyres(a)cisco.com > For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/ > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran -- Jeff Squyres jsquyres(a)cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/

1 0

Re: [MPIWG Fortran] MPI_SIZEOF: Errata required (was: Another MPI_SIZEOF question)
by N.M. Maclaren 15 Jun '14

15 Jun '14

On Jun 14 2014, William Gropp wrote: > > My feeling on this is (a) there have already been too many oversights > because of the tremendous scope of the changes and (b) the mpif.h > interface, in practice, should and will be used mostly for legacy > applications - thus either defining it as similar to the original "basic" > support, with any new constants, will satisfy the users without the risk > of finding another MPI_SIZEOF issue. At this point I have no confidence > that there isn't something else lurking in the wings, and that argues for > as conservative an approach as possible. In my view, that is the only reasonable approach. Sane people use it only in 'legacy' code or because the code is extremely defensive. In neither case are major new facilities relevant. Regards, Nick Maclaren. Regards, Nick Maclaren.

1 0

Re: [MPIWG Fortran] MPI_SIZEOF: Errata required (was: Another MPI_SIZEOF question)
by William Gropp 14 Jun '14

14 Jun '14

My feeling on this is (a) there have already been too many oversights because of the tremendous scope of the changes and (b) the mpif.h interface, in practice, should and will be used mostly for legacy applications - thus either defining it as similar to the original "basic" support, with any new constants, will satisfy the users without the risk of finding another MPI_SIZEOF issue. At this point I have no confidence that there isn't something else lurking in the wings, and that argues for as conservative an approach as possible. Bill William Gropp Director, Parallel Computing Institute Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign On Jun 14, 2014, at 12:49 PM, Rolf Rabenseifner wrote: > Yes, the "Extended Support" listed of about 5 additional routines, > but as far as I understand, MPI_SIZEOF is the only one > that needs an INTERFACE description in mpif.h. > > All other routines are without any implementation-problems. > > Therefore, it would be enough to allow in the special case > that MODULE (and INTERFACE) are not supported by the target compiler, > that > - the Fortran Support requires only mpif.h (same as in MPI-2.0), > and > - MPI_SIZEOF is not provided. > > Additionally, we can restrict the MPI_SIZEOF support only 23 dimensions > - in mpif.h > - or generally if TS29113 or other non-Standard methods > are not available, > supposed that "DOUBLE PRECISION" is the longest > Fortran type Name: > > 123456789012345678901234567890123456789012345678901234567890123456789012 > SUBROUTINE MPI_SIZEOF(x,n) > DOUBLE PRECISION x(1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,*) > ! 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 > What did mpich and OpenMPI in the past with MPI_SIZEOF? > > Rolf > > ----- Original Message ----- >> From: "William Gropp" <wgropp(a)illinois.edu> >> To: "MPI-WG Fortran working group" <mpiwg-fortran(a)lists.mpi-forum.org> >> Sent: Saturday, June 14, 2014 6:36:35 AM >> Subject: Re: [MPIWG Fortran] MPI_SIZEOF: Errata required (was: Another MPI_SIZEOF question) >> >> >> Not quite true. See my email, which gives specific locations in the >> standard for the relevant text. Note also that MPI_SIZEOF is *not* >> the only function that is not included in the "basic" Fortran >> support - the routines to create a type based on precision and range >> are also not required for basic support. >> >> >> I think there are two options for the so-called Fortran 77 support. >> Continue to support the "basic" support, as defined in MPI-2, or >> delete it entirely. Anything else is likely to leave errors and >> mistakes. And while I agree with Jeff that its not good practice to >> remove something like F77 support at a point release, I believe that >> the language in the standard removed Fortran 77 with 3.0. We're >> just now recognizing that. >> >> >> Bill >> >> >> >> >> >> >> >> William Gropp >> Director, Parallel Computing Institute Thomas M. Siebel Chair in >> Computer Science >> >> >> University of Illinois Urbana-Champaign >> >> >> >> >> >> >> >> On Jun 13, 2014, at 10:32 AM, Rolf Rabenseifner wrote: >> >> >> 2) The requirement of having this function in mpif.h was changed: >> >> MPI-2.0 through 2.2 requiered MPI_SIZEOF in mpif.h >> "unless the target compiler does not support modules >> or KIND-parametrized types." >> >> This restriction was removed. >> >> >> _______________________________________________ >> mpiwg-fortran mailing list >> mpiwg-fortran(a)lists.mpi-forum.org >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran > > -- > Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner(a)hlrs.de > 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) > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran

1 0