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

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by Jeff Hammond 10 Jul '20

10 Jul '20

To help aid in the depreciation of mpif.h, perhaps Cray/HPE can contribute a source-to-source translation tool that can be recommended to the MPI user community. Jeff On Fri, Jul 10, 2020 at 9:54 AM Bill Long via mpiwg-fortran < mpiwg-fortran(a)lists.mpi-forum.org> wrote: > > > > On Jul 10, 2020, at 11:11 AM, Jeff Squyres (jsquyres) < > jsquyres(a)cisco.com> wrote: > > > >> > >> 2) \begin{unpopular} Why are the legacy mpi module and mpif.h still > included in the spec? These are embarrassingly obsolete. If this was > fixed, none of the above mentioned conversion routines would be needed. > \end(unpopular} > > > > > > I would love it if we could ditch -- at a minimum -- mpif.h. > > > > However, there's oodles of legacy code out there that uses it. That's > why even deprecating it gets shouted down at Forum meetings. > > We have the same discussions in the Fortran committee about deleting > things, or at least making them “obsolescent”, which seems functionally > similar to deprecated in MPI. In the end, we specify certain archaic and > now-redundant features as deleted with full knowledge that implementors > will not remove the capabilities from compilers because of old codes. > However, the Intent is to guide programmers writing new code in the > direction of using the new (hopefully better) alternatives. And, relevant > to the present discussion, deleting certain old stuff makes the standard > document cleaner. > > Fundamentally, MPI is a C library. Other languages, like C++ and Python, > don’t require separate bindings specified in the MPI spec because they can > call C library routines. The whole mpi_f08 module exercise was to > demonstrate that Fortran can now do the same. The whole idea of specifying > a separate Fortran MPI interface needs to be rethought. Perhaps something > like > > “This version of MPI specifies the contents of a vendor-required Fortran > module for MPI that allows access to the underlying C specification. The > separate language binding for Fortran has been deleted, but is still > available for historical reference in the MPI-3 and earlier specifications.” > > That should allow for a significant simplification of the MPI > specification document. > > > Cheers, > Bill > > > Bill Long > longb(a)cray.com > Principal Engineer, Fortran Technical Support & voice: 651-605-9024 > Bioinformatics Software Development fax: 651-605-9143 > Cray, a Hewlett Packard Enterprise company/ 2131 Lindau Lane/ Suite > 1000/ Bloomington, MN 55425 > > > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > https://lists.mpi-forum.org/mailman/listinfo/mpiwg-fortran > -- Jeff Hammond jeff.science(a)gmail.com http://jeffhammond.github.io/

1 0

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by Bill Long 10 Jul '20

10 Jul '20

> On Jul 10, 2020, at 11:11 AM, Jeff Squyres (jsquyres) <jsquyres(a)cisco.com> wrote: > >> >> 2) \begin{unpopular} Why are the legacy mpi module and mpif.h still included in the spec? These are embarrassingly obsolete. If this was fixed, none of the above mentioned conversion routines would be needed. \end(unpopular} > > > I would love it if we could ditch -- at a minimum -- mpif.h. > > However, there's oodles of legacy code out there that uses it. That's why even deprecating it gets shouted down at Forum meetings. We have the same discussions in the Fortran committee about deleting things, or at least making them “obsolescent”, which seems functionally similar to deprecated in MPI. In the end, we specify certain archaic and now-redundant features as deleted with full knowledge that implementors will not remove the capabilities from compilers because of old codes. However, the Intent is to guide programmers writing new code in the direction of using the new (hopefully better) alternatives. And, relevant to the present discussion, deleting certain old stuff makes the standard document cleaner. Fundamentally, MPI is a C library. Other languages, like C++ and Python, don’t require separate bindings specified in the MPI spec because they can call C library routines. The whole mpi_f08 module exercise was to demonstrate that Fortran can now do the same. The whole idea of specifying a separate Fortran MPI interface needs to be rethought. Perhaps something like “This version of MPI specifies the contents of a vendor-required Fortran module for MPI that allows access to the underlying C specification. The separate language binding for Fortran has been deleted, but is still available for historical reference in the MPI-3 and earlier specifications.” That should allow for a significant simplification of the MPI specification document. Cheers, Bill Bill Long longb(a)cray.com Principal Engineer, Fortran Technical Support & voice: 651-605-9024 Bioinformatics Software Development fax: 651-605-9143 Cray, a Hewlett Packard Enterprise company/ 2131 Lindau Lane/ Suite 1000/ Bloomington, MN 55425

1 0

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by Jeff Squyres (jsquyres) 10 Jul '20

10 Jul '20

On Jul 10, 2020, at 11:18 AM, Bill Long <longb(a)cray.com> wrote: > > In the “change” section there is this txt: > > • Within the mpi_08 Fortran module, the status was defined as TYPE(MPI_Status). Additionally, within both the mpi and the mpi_f08 modules, the constants MPI_STATUS_SIZE, MPI_SOURCE, MPI_TAG, MPI_ERROR, and TYPE(MPI_Status) are defined. New conversion routines were added: MPI_STATUS_F2F08, MPI_STATUS_F082F, MPI_Status_c2f08, and MPI_Status_f082c, In mpi.h, the new type MPI_F08_status, and the external variables MPI_F08_STATUS_IGNORE and MPI_F08_STATUSES_IGNORE were added. Good point. Just to be clear, you're referring to the changelog section in MPI-3.1, specifically bullet 30 on p802. That being said: - the changelog is non-binding ...but it does indicate our intent from that time - the changelog text states that the mpi module has TYPE(MPI_Status) -- but it does not say it was added to mpif.h > 1) Why would the F08 status be defined different from the C definition? (If that were the case, conversions between f08 and C would be irrelevant). I remember that there was a lot of discussion about this at the time, which is what resulted in Figure 17.1. I know there were discussions about making the F08 and C statuses the same, but for some reason we chose not to mandate it. Perhaps we wanted to allow implementations to do whatever they wanted...? (e.g., allow Status_c2f08 be a no-op if the implementation wanted to, but not mandate it) > 2) \begin{unpopular} Why are the legacy mpi module and mpif.h still included in the spec? These are embarrassingly obsolete. If this was fixed, none of the above mentioned conversion routines would be needed. \end(unpopular} I would love it if we could ditch -- at a minimum -- mpif.h. However, there's oodles of legacy code out there that uses it. That's why even deprecating it gets shouted down at Forum meetings. -- Jeff Squyres jsquyres(a)cisco.com

1 0

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by Bill Long 10 Jul '20

10 Jul '20

In the “change” section there is this txt: • Within the mpi_08 Fortran module, the status was defined as TYPE(MPI_Status). Additionally, within both the mpi and the mpi_f08 modules, the constants MPI_STATUS_SIZE, MPI_SOURCE, MPI_TAG, MPI_ERROR, and TYPE(MPI_Status) are defined. New conversion routines were added: MPI_STATUS_F2F08, MPI_STATUS_F082F, MPI_Status_c2f08, and MPI_Status_f082c, In mpi.h, the new type MPI_F08_status, and the external variables MPI_F08_STATUS_IGNORE and MPI_F08_STATUSES_IGNORE were added. 1) Why would the F08 status be defined different from the C definition? (If that were the case, conversions between f08 and C would be irrelevant). 2) \begin{unpopular} Why are the legacy mpi module and mpif.h still included in the spec? These are embarrassingly obsolete. If this was fixed, none of the above mentioned conversion routines would be needed. \end(unpopular} Cheers, Bill > On Jul 9, 2020, at 7:54 PM, Jeff Squyres (jsquyres) via mpiwg-fortran <mpiwg-fortran(a)lists.mpi-forum.org> wrote: > > MPI-3.1 Figure 17.1 (in section 17.2.5, starting on page 656) defines the following routines: > > 1. MPI_Status_c2f() (from MPI-2) > 2. MPI_Status_f2c() (from MPI-2) > 3. MPI_Status_c2f08() (new in MPI-3) > 4. MPI_Status_f082c() (new in MPI-3) > 5. MPI_Status_f2f08() (new in MPI-3) > 6. MPI_Status_f082f() (new in MPI-3) > > All of the functions are fairly straightforward to implement in C. > > But per the bottom part of Figure 17.1, two of the routines are supposed to be implemented in Fortran: #5 and #6. > > Does this mean that TYPE(MPI_Status) needs to be defined in mpif.h and the mpi module? > > I'm not sure how an MPI application would be able to invoke functions #5 and #6 without a definition of TYPE(MPI_Status) -- or even have a variable of that type that they need to convert to INTEGER(MPI_STATUS_SIZE). > > -- > Jeff Squyres > jsquyres(a)cisco.com > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > https://lists.mpi-forum.org/mailman/listinfo/mpiwg-fortran Bill Long longb(a)cray.com Principal Engineer, Fortran Technical Support & voice: 651-605-9024 Bioinformatics Software Development fax: 651-605-9143 Cray, a Hewlett Packard Enterprise company/ 2131 Lindau Lane/ Suite 1000/ Bloomington, MN 55425

1 0

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by Jeff Squyres (jsquyres) 10 Jul '20

10 Jul '20

On Jul 10, 2020, at 9:46 AM, William Gropp <wgropp(a)illinois.edu> wrote: > > Granted, an error, not a clarification. Does any implementation do this now (define Type_status in mpif.h/mpi module)? If not, it is easier to consider this an error. Open MPI does not -- this issue came up a few months ago in the Open MPI project and just re-surfaced again. It led me to a deeper read of MPI-3.1 about the definition of these functions. I don't immediately see it in MPICH, either. -- Jeff Squyres jsquyres(a)cisco.com

1 0

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by William Gropp 10 Jul '20

10 Jul '20

Granted, an error, not a clarification. Does any implementation do this now (define Type_status in mpif.h/mpi module)? If not, it is easier to consider this an error. Bill William Gropp Director and Chief Scientist, NCSA Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign > On Jul 10, 2020, at 8:39 AM, Jeff Squyres (jsquyres) <jsquyres(a)cisco.com> wrote: > > Gah. Closer reading shows that MPI-3.1 page 658 actually has mpif.h/mpi module bindings for MPI_Status_f2f08 and _f082f that include TYPE(MPI_Status) dummy parameters. > > So the standard text actually strongly infers that TYPE(MPI_Status) is supposed to be defined in mpif.h and the mpi module. > > Granted, a user *could* do something like this, but I can't believe that this is what we intended users to do: > > ----- > ! This is super ugly IMNSHO > use :: mpi_f08, only : MPI_Status > implicit none > include 'mpif.h' > ----- > > Do we really want to change the spec to state that the _f2f08 and _f082f subroutines are only available in the mpi_f08 module? > > I think that that would be a *change*, not a *clarification*. > > That being said, we might be able to squint at this and call it an "error" in MPI-3.1. > > But *that* being said, it's *not* actually a technical error. It's just that we seemed to have missed the implication back then that TYPE(MPI_Status) would need to be defined in mpif.h and the mpi module. That's not a technical *problem* -- it's just a little weird. > > > > >> On Jul 10, 2020, at 8:49 AM, William Gropp <wgropp(a)illinois.edu> wrote: >> >> That certainly was not my understanding - I expected that these would be defined only for F08, allowing code in F08 to move status data to/from the integer status array for the prior Fortran bindings. I am in favor of clarifying the standard to make this clear. >> >> Bill >> >> William Gropp >> Director and Chief Scientist, NCSA >> Thomas M. Siebel Chair in Computer Science >> University of Illinois Urbana-Champaign >> >> >> >> >> >> >>> On Jul 9, 2020, at 7:54 PM, Jeff Squyres (jsquyres) via mpiwg-fortran <mpiwg-fortran(a)lists.mpi-forum.org> wrote: >>> >>> MPI-3.1 Figure 17.1 (in section 17.2.5, starting on page 656) defines the following routines: >>> >>> 1. MPI_Status_c2f() (from MPI-2) >>> 2. MPI_Status_f2c() (from MPI-2) >>> 3. MPI_Status_c2f08() (new in MPI-3) >>> 4. MPI_Status_f082c() (new in MPI-3) >>> 5. MPI_Status_f2f08() (new in MPI-3) >>> 6. MPI_Status_f082f() (new in MPI-3) >>> >>> All of the functions are fairly straightforward to implement in C. >>> >>> But per the bottom part of Figure 17.1, two of the routines are supposed to be implemented in Fortran: #5 and #6. >>> >>> Does this mean that TYPE(MPI_Status) needs to be defined in mpif.h and the mpi module? >>> >>> I'm not sure how an MPI application would be able to invoke functions #5 and #6 without a definition of TYPE(MPI_Status) -- or even have a variable of that type that they need to convert to INTEGER(MPI_STATUS_SIZE). >>> >>> -- >>> Jeff Squyres >>> jsquyres(a)cisco.com >>> >>> _______________________________________________ >>> mpiwg-fortran mailing list >>> mpiwg-fortran(a)lists.mpi-forum.org >>> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-fortran >> > > > -- > Jeff Squyres > jsquyres(a)cisco.com >

1 0

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by Jeff Squyres (jsquyres) 10 Jul '20

10 Jul '20

Gah. Closer reading shows that MPI-3.1 page 658 actually has mpif.h/mpi module bindings for MPI_Status_f2f08 and _f082f that include TYPE(MPI_Status) dummy parameters. So the standard text actually strongly infers that TYPE(MPI_Status) is supposed to be defined in mpif.h and the mpi module. Granted, a user *could* do something like this, but I can't believe that this is what we intended users to do: ----- ! This is super ugly IMNSHO use :: mpi_f08, only : MPI_Status implicit none include 'mpif.h' ----- Do we really want to change the spec to state that the _f2f08 and _f082f subroutines are only available in the mpi_f08 module? I think that that would be a *change*, not a *clarification*. That being said, we might be able to squint at this and call it an "error" in MPI-3.1. But *that* being said, it's *not* actually a technical error. It's just that we seemed to have missed the implication back then that TYPE(MPI_Status) would need to be defined in mpif.h and the mpi module. That's not a technical *problem* -- it's just a little weird. > On Jul 10, 2020, at 8:49 AM, William Gropp <wgropp(a)illinois.edu> wrote: > > That certainly was not my understanding - I expected that these would be defined only for F08, allowing code in F08 to move status data to/from the integer status array for the prior Fortran bindings. I am in favor of clarifying the standard to make this clear. > > Bill > > William Gropp > Director and Chief Scientist, NCSA > Thomas M. Siebel Chair in Computer Science > University of Illinois Urbana-Champaign > > > > > > >> On Jul 9, 2020, at 7:54 PM, Jeff Squyres (jsquyres) via mpiwg-fortran <mpiwg-fortran(a)lists.mpi-forum.org> wrote: >> >> MPI-3.1 Figure 17.1 (in section 17.2.5, starting on page 656) defines the following routines: >> >> 1. MPI_Status_c2f() (from MPI-2) >> 2. MPI_Status_f2c() (from MPI-2) >> 3. MPI_Status_c2f08() (new in MPI-3) >> 4. MPI_Status_f082c() (new in MPI-3) >> 5. MPI_Status_f2f08() (new in MPI-3) >> 6. MPI_Status_f082f() (new in MPI-3) >> >> All of the functions are fairly straightforward to implement in C. >> >> But per the bottom part of Figure 17.1, two of the routines are supposed to be implemented in Fortran: #5 and #6. >> >> Does this mean that TYPE(MPI_Status) needs to be defined in mpif.h and the mpi module? >> >> I'm not sure how an MPI application would be able to invoke functions #5 and #6 without a definition of TYPE(MPI_Status) -- or even have a variable of that type that they need to convert to INTEGER(MPI_STATUS_SIZE). >> >> -- >> Jeff Squyres >> jsquyres(a)cisco.com >> >> _______________________________________________ >> mpiwg-fortran mailing list >> mpiwg-fortran(a)lists.mpi-forum.org >> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-fortran > -- Jeff Squyres jsquyres(a)cisco.com

1 0

Re: [MPIWG Fortran] Question about MPI_Status_f2f08() and _f082f()
by William Gropp 10 Jul '20

10 Jul '20

That certainly was not my understanding - I expected that these would be defined only for F08, allowing code in F08 to move status data to/from the integer status array for the prior Fortran bindings. I am in favor of clarifying the standard to make this clear. Bill William Gropp Director and Chief Scientist, NCSA Thomas M. Siebel Chair in Computer Science University of Illinois Urbana-Champaign > On Jul 9, 2020, at 7:54 PM, Jeff Squyres (jsquyres) via mpiwg-fortran <mpiwg-fortran(a)lists.mpi-forum.org> wrote: > > MPI-3.1 Figure 17.1 (in section 17.2.5, starting on page 656) defines the following routines: > > 1. MPI_Status_c2f() (from MPI-2) > 2. MPI_Status_f2c() (from MPI-2) > 3. MPI_Status_c2f08() (new in MPI-3) > 4. MPI_Status_f082c() (new in MPI-3) > 5. MPI_Status_f2f08() (new in MPI-3) > 6. MPI_Status_f082f() (new in MPI-3) > > All of the functions are fairly straightforward to implement in C. > > But per the bottom part of Figure 17.1, two of the routines are supposed to be implemented in Fortran: #5 and #6. > > Does this mean that TYPE(MPI_Status) needs to be defined in mpif.h and the mpi module? > > I'm not sure how an MPI application would be able to invoke functions #5 and #6 without a definition of TYPE(MPI_Status) -- or even have a variable of that type that they need to convert to INTEGER(MPI_STATUS_SIZE). > > -- > Jeff Squyres > jsquyres(a)cisco.com > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > https://lists.mpi-forum.org/mailman/listinfo/mpiwg-fortran

1 0

Question about MPI_Status_f2f08() and _f082f()
by Jeff Squyres (jsquyres) 09 Jul '20

09 Jul '20

MPI-3.1 Figure 17.1 (in section 17.2.5, starting on page 656) defines the following routines: 1. MPI_Status_c2f() (from MPI-2) 2. MPI_Status_f2c() (from MPI-2) 3. MPI_Status_c2f08() (new in MPI-3) 4. MPI_Status_f082c() (new in MPI-3) 5. MPI_Status_f2f08() (new in MPI-3) 6. MPI_Status_f082f() (new in MPI-3) All of the functions are fairly straightforward to implement in C. But per the bottom part of Figure 17.1, two of the routines are supposed to be implemented in Fortran: #5 and #6. Does this mean that TYPE(MPI_Status) needs to be defined in mpif.h and the mpi module? I'm not sure how an MPI application would be able to invoke functions #5 and #6 without a definition of TYPE(MPI_Status) -- or even have a variable of that type that they need to convert to INTEGER(MPI_STATUS_SIZE). -- Jeff Squyres jsquyres(a)cisco.com

1 0

Re: [MPIWG Fortran] DRAFT: Chapter Committees ACTION NEEDED
by Jeff Squyres (jsquyres) 17 Dec '19

17 Dec '19

Oops! I meant to send this to the large count WG, not the fortran WG. Sorry for the noise, folks... > On Dec 13, 2019, at 12:33 PM, Jeff Squyres (jsquyres) via mpiwg-fortran <mpiwg-fortran(a)lists.mpi-forum.org> wrote: > > This is a DRAFT of what I'd like to send to the Forum to get all Chapter Committees to review the document-wide stuff that we need reviewed. Please send comments / suggestions to me. Thanks! > > ----------------------------------------- > > We need Chapter Committees to review their chapters in 2 sets of upcoming document-wide changes that are slated to be merged "soon". > > After you complete a review, please put an "X" in the corresponding row/column for your chapter and the specified PR in this Google spreadsheet. This will let us track what has been done / what still needs to be done: > > https://docs.google.com/spreadsheets/d/1qPAREYgeuSCMtU-EQniDilrHIYb8W0-f6Qz… > > ==> FIRST SET > > Bill Gropp (the MPI Standard document Editor) has two PRs with "ticket 0" changes. However, the changes are pervasive and touch a LOT of the document. We therefore need Chapter Committees to review their chapters in the following two PRs: > > 1a. "Ticket 0 changes for MPI-4.x" > https://github.com/mpi-forum/mpi-standard/pull/145 > 1b. "Uniform style for Fortran code" > https://github.com/mpi-forum/mpi-standard/pull/151 > > **** BILL: what deadline do you want to set? > > ==> SECOND SET > > The Pythonization group has Pythonized the bindings in the entire document. We've completed a first correctness pass over the work, but we need more eyes to compare the end result of our work to (at least) the MPI-3.1 document -- or, better yet, to the head of the mpi-4.x branch. > > See this wiki page for instructions: https://github.com/mpiwg-large-count/large-count-issues/wiki/pythonizing-in… > > Also please see the Pythonization slides presented at the Forum this week: https://docs.google.com/presentation/d/1QZyh6TbjDMu7mEcNqBFfe9fFOltaCtxgZ2Z… > > The Pythonization group needs your reviews BY *** 21 JAN 2020 *** to give us time to finalize everything before the T-2 week deadline for the Portland Forum meeting. > > NOTE: At the ABQ Forum meeting, I asked you to commit/PR your approvals.json file that resulted from running the tool. This is no longer necessary -- please just mark the google spreadsheet as described above. > > -- > Jeff Squyres > jsquyres(a)cisco.com > > _______________________________________________ > mpiwg-fortran mailing list > mpiwg-fortran(a)lists.mpi-forum.org > https://lists.mpi-forum.org/mailman/listinfo/mpiwg-fortran -- Jeff Squyres jsquyres(a)cisco.com

1 0