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

Re: [MPIWG Fortran] MPI_SIZEOF: redux
by Jeff Squyres (jsquyres) 09 Jun '14

09 Jun '14

Yep -- looks like it was quoted as being part of F90 incorrectly. So -- ya: storage_size is F2008 and later. But I still think it would be ok to deprecate MPI_SIZEOF, since the language now supports this. And as Bill Long pointed out, having MPI deprecate MPI_SIZEOF might actually pressure compiler vendors to implement storage_size(). And I think we'd give plenty of warning to compiler vendors (i.e., this would be MPI-4). And also have no plan to actually remove MPI_SIZEOF. On Jun 6, 2014, at 10:29 AM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > Maybe GCC docs are wrong? > > http://gcc.gnu.org/onlinedocs/gfortran/STORAGE_005fSIZE.html > > 8.238 STORAGE_SIZE — Storage size in bits > > > Description: > Returns the storage size of argument A in bits. > Standard: > Fortran 2008 and later > Class: > Inquiry function > Syntax: > RESULT = STORAGE_SIZE(A [, KIND]) > Arguments: > > A Shall be a scalar or array of any type. > KIND (Optional) shall be a scalar integer constant expression. > > > Return Value: > The result is a scalar integer with the kind type parameter specified by KIND (or default integer type if KIND is missing). The result value is the size expressed in bits for an element of an array that has the dynamic type and type parameters of A. > See also: > C_SIZEOF, SIZEOF > > Sent from my iPhone > > On Jun 6, 2014, at 7:16 AM, "Jeff Squyres (jsquyres)" <jsquyres(a)cisco.com> wrote: > >> I'm still not sure why you're associating MPI_SIZEOF, F08, and storage_size together...? >> >> According to earlier in the thread, storage_size has been around since F90. I.e., 20+ years. >> >> Bill also mentioned that even though some compilers don't yet have it, a) it's easy to implement, and b) giving 2+ years warning to compilers that MPI intends to deprecate MPI_SIZEOF might well give compilers the nudge they need to go implement it. >> >> >> >> >> On Jun 6, 2014, at 9:00 AM, Jeff Hammond <jeff.science(a)gmail.com> wrote: >> >>> The majority of Fortran users in HPC are ignorant of F08 (beyond trivial notions, eg existence) and will still be by the time of MPI-4. I think deprecating this function must be coincident with deprecating the pre-F08 bindings since it is ootebtionskly required for pre-F08 MPI usage. >>> >>> If we are serious about forcing everyone to use F08 (which I think is a terrible idea), then let's at least do it consistently. >>> >>> Jeff >>> >>> Sent from my iPhone >>> >>>> On Jun 6, 2014, at 5:15 AM, Bill Long <longb(a)cray.com> wrote: >>>> >>>> >>>>> On Jun 6, 2014, at 6:31 AM, Jeff Squyres (jsquyres) <jsquyres(a)cisco.com> wrote: >>>>> >>>>> Would it still be ok to deprecate MPI_SIZEOF in MPI-4 (which is at least ~2 years away)? We don't have to have any plan for removing it. >>>> >>>> I think that would be reasonable. In two years time more vendors will have implemented storage_size. And if it were well known that MPI wants people to migrate to this function, vendors would be motivated to add support sooner. The function itself is pretty trivial to implement as inline code by the compiler. It’s just a matter of vendor priorities. >>>> >>>> Cheers, >>>> Bill >>>> >>>> >>>> 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 >>> _______________________________________________ >>> 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/ >> >> _______________________________________________ >> mpiwg-fortran mailing list >> mpiwg-fortran(a)lists.mpi-forum.org >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran > _______________________________________________ > 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] Another MPI_SIZEOF question
by William Gropp 06 Jun '14

06 Jun '14

There's no (small) errata. The Forum decided that MPI would no longer support Fortran 77 (and this was pretty clear from the discussions). In MPI-2, MPI_SIZEOF was part of the extended support; the basic support did not include it and thus mpif.h could be implemented with Fortran 77 (see page 291, 10.2.4 in MPI-2.0). This is no longer true; see 17.1.4 (mpif.h) which says the second set of function prototypes must be supported for all MPI functions, and page 621, where MPI_SIZEOF has a second set of function prototypes. There are other references to overloading in mpif.h; see, for example, page 339, which would require an MPI implementation to provide different mpif.h files depending on which level of Fortran support is provided by the compiler (which might depend on a compiler switch!). Bill William Gropp Director, Parallel Computing Institute Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign On Jun 6, 2014, at 6:27 PM, Jeff Squyres (jsquyres) wrote: > Do we need to add some errata text about mpif.h and MPI-sizeof, particularly wrt older fortran compilers? > > Sent from my phone. No type good. > >> On Jun 6, 2014, at 9:32 AM, "William Gropp" <wgropp(a)illinois.edu> wrote: >> >> Yes, despite the new claims in MPI-3, the mpif.h file for MPICH has been required to be backward compatible with Fortran 77, since there are still Fortran 77 codes. MPI-3 essentially mandated that Fortran 77 compilers be unable to compile standard conforming MPI codes, and the MPI implementations that choose to permit existing codes to be compiled decided to violate the standard in this place. >> >> Note that storage_size appeared in Fortran 2008, not earlier versions of Fortran 9x, which is why we added MPI_SIZEOF. >> >> Having said that, yes, I think that deprecating but not yet removing MPI_SIZEOF makes sense. >> >> Bill >> >> William Gropp >> Director, Parallel Computing Institute >> Thomas M. Siebel Chair in Computer Science >> University of Illinois Urbana-Champaign >> >> >> >> >> >>> On Jun 6, 2014, at 9:11 AM, Jeff Squyres (jsquyres) wrote: >>> >>> It strikes me that the current MPI_SIZEOF text does not limit this function to the mpi or mpi_f08 modules, meaning that it needs to be supported/available in mpif.h as well. >>> >>> Since SIZEOF can only be implemented via overloaded procedures, this means that implementations have to put an INTERFACE block with the overloaded procedures for MPI_SIZEOF in mpif.h. Yowza! This goes against a long-standing tradition of not including procedure prototypes in mpif.h. >>> >>> If I'm interpreting this properly, it looks like neither Open MPI nor MPICH do this correctly. >>> >>> Am I correct in this deduction? >>> >>> Random side note: this makes me wonder how much MPI_SIZEOF is really used... >>> >>> -- >>> 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 >> >> _______________________________________________ >> mpiwg-fortran mailing list >> mpiwg-fortran(a)lists.mpi-forum.org >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran

1 0

Re: [MPIWG Fortran] Another MPI_SIZEOF question
by Jeff Squyres (jsquyres) 06 Jun '14

06 Jun '14

Do we need to add some errata text about mpif.h and MPI-sizeof, particularly wrt older fortran compilers? Sent from my phone. No type good. > On Jun 6, 2014, at 9:32 AM, "William Gropp" <wgropp(a)illinois.edu> wrote: > > Yes, despite the new claims in MPI-3, the mpif.h file for MPICH has been required to be backward compatible with Fortran 77, since there are still Fortran 77 codes. MPI-3 essentially mandated that Fortran 77 compilers be unable to compile standard conforming MPI codes, and the MPI implementations that choose to permit existing codes to be compiled decided to violate the standard in this place. > > Note that storage_size appeared in Fortran 2008, not earlier versions of Fortran 9x, which is why we added MPI_SIZEOF. > > Having said that, yes, I think that deprecating but not yet removing MPI_SIZEOF makes sense. > > Bill > > William Gropp > Director, Parallel Computing Institute > Thomas M. Siebel Chair in Computer Science > University of Illinois Urbana-Champaign > > > > > >> On Jun 6, 2014, at 9:11 AM, Jeff Squyres (jsquyres) wrote: >> >> It strikes me that the current MPI_SIZEOF text does not limit this function to the mpi or mpi_f08 modules, meaning that it needs to be supported/available in mpif.h as well. >> >> Since SIZEOF can only be implemented via overloaded procedures, this means that implementations have to put an INTERFACE block with the overloaded procedures for MPI_SIZEOF in mpif.h. Yowza! This goes against a long-standing tradition of not including procedure prototypes in mpif.h. >> >> If I'm interpreting this properly, it looks like neither Open MPI nor MPICH do this correctly. >> >> Am I correct in this deduction? >> >> Random side note: this makes me wonder how much MPI_SIZEOF is really used... >> >> -- >> 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 > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran

1 0

Re: [MPIWG Fortran] MPI_SIZEOF: redux
by Jeff Hammond 06 Jun '14

06 Jun '14

Maybe GCC docs are wrong? http://gcc.gnu.org/onlinedocs/gfortran/STORAGE_005fSIZE.html 8.238 STORAGE_SIZE — Storage size in bits Description: Returns the storage size of argument A in bits. Standard: Fortran 2008 and later Class: Inquiry function Syntax: RESULT = STORAGE_SIZE(A [, KIND]) Arguments: A Shall be a scalar or array of any type. KIND (Optional) shall be a scalar integer constant expression. Return Value: The result is a scalar integer with the kind type parameter specified by KIND (or default integer type if KIND is missing). The result value is the size expressed in bits for an element of an array that has the dynamic type and type parameters of A. See also: C_SIZEOF, SIZEOF Sent from my iPhone > On Jun 6, 2014, at 7:16 AM, "Jeff Squyres (jsquyres)" <jsquyres(a)cisco.com> wrote: > > I'm still not sure why you're associating MPI_SIZEOF, F08, and storage_size together...? > > According to earlier in the thread, storage_size has been around since F90. I.e., 20+ years. > > Bill also mentioned that even though some compilers don't yet have it, a) it's easy to implement, and b) giving 2+ years warning to compilers that MPI intends to deprecate MPI_SIZEOF might well give compilers the nudge they need to go implement it. > > > > >> On Jun 6, 2014, at 9:00 AM, Jeff Hammond <jeff.science(a)gmail.com> wrote: >> >> The majority of Fortran users in HPC are ignorant of F08 (beyond trivial notions, eg existence) and will still be by the time of MPI-4. I think deprecating this function must be coincident with deprecating the pre-F08 bindings since it is ootebtionskly required for pre-F08 MPI usage. >> >> If we are serious about forcing everyone to use F08 (which I think is a terrible idea), then let's at least do it consistently. >> >> Jeff >> >> Sent from my iPhone >> >>> On Jun 6, 2014, at 5:15 AM, Bill Long <longb(a)cray.com> wrote: >>> >>> >>>> On Jun 6, 2014, at 6:31 AM, Jeff Squyres (jsquyres) <jsquyres(a)cisco.com> wrote: >>>> >>>> Would it still be ok to deprecate MPI_SIZEOF in MPI-4 (which is at least ~2 years away)? We don't have to have any plan for removing it. >>> >>> I think that would be reasonable. In two years time more vendors will have implemented storage_size. And if it were well known that MPI wants people to migrate to this function, vendors would be motivated to add support sooner. The function itself is pretty trivial to implement as inline code by the compiler. It’s just a matter of vendor priorities. >>> >>> Cheers, >>> Bill >>> >>> >>> 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 >> _______________________________________________ >> 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/ > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-fortran

1 0

Re: [MPIWG Fortran] Another MPI_SIZEOF question
by William Gropp 06 Jun '14

06 Jun '14

Yes, despite the new claims in MPI-3, the mpif.h file for MPICH has been required to be backward compatible with Fortran 77, since there are still Fortran 77 codes. MPI-3 essentially mandated that Fortran 77 compilers be unable to compile standard conforming MPI codes, and the MPI implementations that choose to permit existing codes to be compiled decided to violate the standard in this place. Note that storage_size appeared in Fortran 2008, not earlier versions of Fortran 9x, which is why we added MPI_SIZEOF. Having said that, yes, I think that deprecating but not yet removing MPI_SIZEOF makes sense. Bill William Gropp Director, Parallel Computing Institute Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign On Jun 6, 2014, at 9:11 AM, Jeff Squyres (jsquyres) wrote: > It strikes me that the current MPI_SIZEOF text does not limit this function to the mpi or mpi_f08 modules, meaning that it needs to be supported/available in mpif.h as well. > > Since SIZEOF can only be implemented via overloaded procedures, this means that implementations have to put an INTERFACE block with the overloaded procedures for MPI_SIZEOF in mpif.h. Yowza! This goes against a long-standing tradition of not including procedure prototypes in mpif.h. > > If I'm interpreting this properly, it looks like neither Open MPI nor MPICH do this correctly. > > Am I correct in this deduction? > > Random side note: this makes me wonder how much MPI_SIZEOF is really used... > > -- > 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

1 0

Re: [MPIWG Fortran] Another MPI_SIZEOF question
by Jeff Hammond 06 Jun '14

06 Jun '14

I want to be absolutely sure that "INTERFACE block with the overloaded procedures for MPI_SIZEOF in mpif.h" does not break user experience for those using legacy Fortran. Can someone give me a prototype implementation to test? Best, Jeff Sent from my iPhone > On Jun 6, 2014, at 7:11 AM, "Jeff Squyres (jsquyres)" <jsquyres(a)cisco.com> wrote: > > INTERFACE block with the overloaded procedures for MPI_SIZEOF in mpif.h

1 0

Re: [MPIWG Fortran] Another MPI_SIZEOF question
by Junchao Zhang 06 Jun '14

06 Jun '14

Let me paste an email I got a few days ago from a mailing list of mpich :) ---------------- Hi Pavan, did you ever implement mpi_sizeof in fortran 77 bindings? I get it undefined: $ ifort test.f -lfmpich -lmpich /tmp/ifort1AJSk8.o: In function `MAIN__': test.f:(.text+0x5b): undefined reference to `mpi_sizeof_' Best regards, Michael. --Junchao Zhang On Fri, Jun 6, 2014 at 9:11 AM, Jeff Squyres (jsquyres) <jsquyres(a)cisco.com> wrote: > It strikes me that the current MPI_SIZEOF text does not limit this > function to the mpi or mpi_f08 modules, meaning that it needs to be > supported/available in mpif.h as well. > > Since SIZEOF can only be implemented via overloaded procedures, this means > that implementations have to put an INTERFACE block with the overloaded > procedures for MPI_SIZEOF in mpif.h. Yowza! This goes against a > long-standing tradition of not including procedure prototypes in mpif.h. > > If I'm interpreting this properly, it looks like neither Open MPI nor > MPICH do this correctly. > > Am I correct in this deduction? > > Random side note: this makes me wonder how much MPI_SIZEOF is really > used... > > -- > 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 >

1 0

Re: [MPIWG Fortran] MPI_SIZEOF: redux
by Jeff Squyres (jsquyres) 06 Jun '14

06 Jun '14

I'm still not sure why you're associating MPI_SIZEOF, F08, and storage_size together...? According to earlier in the thread, storage_size has been around since F90. I.e., 20+ years. Bill also mentioned that even though some compilers don't yet have it, a) it's easy to implement, and b) giving 2+ years warning to compilers that MPI intends to deprecate MPI_SIZEOF might well give compilers the nudge they need to go implement it. On Jun 6, 2014, at 9:00 AM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > The majority of Fortran users in HPC are ignorant of F08 (beyond trivial notions, eg existence) and will still be by the time of MPI-4. I think deprecating this function must be coincident with deprecating the pre-F08 bindings since it is ootebtionskly required for pre-F08 MPI usage. > > If we are serious about forcing everyone to use F08 (which I think is a terrible idea), then let's at least do it consistently. > > Jeff > > Sent from my iPhone > >> On Jun 6, 2014, at 5:15 AM, Bill Long <longb(a)cray.com> wrote: >> >> >>> On Jun 6, 2014, at 6:31 AM, Jeff Squyres (jsquyres) <jsquyres(a)cisco.com> wrote: >>> >>> Would it still be ok to deprecate MPI_SIZEOF in MPI-4 (which is at least ~2 years away)? We don't have to have any plan for removing it. >> >> I think that would be reasonable. In two years time more vendors will have implemented storage_size. And if it were well known that MPI wants people to migrate to this function, vendors would be motivated to add support sooner. The function itself is pretty trivial to implement as inline code by the compiler. It’s just a matter of vendor priorities. >> >> Cheers, >> Bill >> >> >> 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 > _______________________________________________ > 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

Another MPI_SIZEOF question
by Jeff Squyres (jsquyres) 06 Jun '14

06 Jun '14

It strikes me that the current MPI_SIZEOF text does not limit this function to the mpi or mpi_f08 modules, meaning that it needs to be supported/available in mpif.h as well. Since SIZEOF can only be implemented via overloaded procedures, this means that implementations have to put an INTERFACE block with the overloaded procedures for MPI_SIZEOF in mpif.h. Yowza! This goes against a long-standing tradition of not including procedure prototypes in mpif.h. If I'm interpreting this properly, it looks like neither Open MPI nor MPICH do this correctly. Am I correct in this deduction? Random side note: this makes me wonder how much MPI_SIZEOF is really used... -- 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: redux
by Jeff Hammond 06 Jun '14

06 Jun '14

ootebtionskly -> ostensibly Lack-of-obvious-autocorrect fail. Sent from my iPhone > On Jun 6, 2014, at 7:00 AM, Jeff Hammond <jeff.science(a)gmail.com> wrote: > > ootebtionskly

1 0