Just to wrap this up for the web archives... George and I chatted on the phone; but before we talked, the user had already changed their code to "include "mpif.h'" and declared "problem solved!".
I'm still curious as to what the exact problem was, because as far as I can tell, the OMPI code is correct. So if this issue ever comes up again and someone stumbles across this thread via googling, feel free to ping me and we can dig more. On Feb 16, 2011, at 4:39 PM, Jeff Squyres wrote: > Were you able to check that their code matches the type signature that Open > MPI is exporting? Specifically, double check that the types of the > parameters that the code is passing as arguments to MPI_TYPE_CREATE_SUBARRAY > match the types in the signature that I sent in my original reply. I > verified that OMPI v1.4.x's signature matches what is listed in MPI-2.2 p87. > > Absent any other information, I am guessing that his code is incorrect. > > I'm hazarding a guess here: it's possible that this code works with other > MPI's because they don't offer an F90 binding for MPI_TYPE_CREATE_SUBARRAY. > I.e., your user can call this function from an F90 MPI code, but it just > "falls through" (with no compiler checking) to that MPI's underlying F77 > function. > > Open MPI actually implements the F90 compiler checking before letting the > implementation "fall through" to the underlying F77 function -- many other > MPI's do not. ...or at least, many other MPI implementations *did* not; but > I admit that I haven't checked in a while. Note, too, that it's perfectly > allowable for MPI implementations to *not* do the compiler checking because > the F90 bindings are classified as "optional" in the MPI specification. > > Finally, I note that we've had a few F90 bindings fixes in between 1.4.x and > 1.5.x, but only in the following functions: > > MPI_COMM_CREATE_KEYVAL > MPI_GREQUEST_START > MPI_OP_CREATE > MPI_REGISTER_DATAREP > MPI_TYPE_CREATE_KEYVAL > MPI_WIN_CREATE_KEYVAL > > I'm actually a little surprised that we (I) forgot to back-port these fixes > to the 1.4.x series; I'll do so now... > > > > > On Feb 16, 2011, at 1:16 PM, Petit, George (Cont, ARL/CISD) wrote: > >> Jeff, >> >> The problem occurs in both OpenMPI 1.3 and 1.4. The user is also reporting >> that he's experiencing this problem on other systems as well. >> >> George Petit >> Lockheed Martin >> ARL DSRC Customer Support Team Lead >> george.pe...@us.army.mil >> 410-278-9283 >> >> >> From: users-boun...@open-mpi.org on behalf of Jeff Squyres >> Sent: Wed 2/16/2011 10:40 AM >> To: Open MPI Users >> Subject: Re: [OMPI users] MPI_TYPE_CREATE_SUBARRAY not in mpif90 library >> >> What version of Open MPI are you using? >> >> I just checked a trivial f90 MPI program and MPI_TYPE_CREATE_SUBARRAY seemed >> to work ok. >> >> I am not a fortran expert, but when I have seen this type of error message >> before, it *usually* meant that there was a type mismatch between the types >> passed to the function and the function signature available in the f90 >> module. I.e., the compiler found a function matching the *name*, but wasn't >> able to find a match for the *type signature*. >> >> Here's Open MPI's declaration of MPI_TYPE_CREATE_SUBARRAY: >> >> ----- >> interface MPI_Type_create_subarray >> subroutine MPI_Type_create_subarray(ndims, size_array, subsize_array, >> start_array, order, & >> oldtype, newtype, ierr) >> integer, intent(in) :: ndims >> integer, dimension(*), intent(in) :: size_array >> integer, dimension(*), intent(in) :: subsize_array >> integer, dimension(*), intent(in) :: start_array >> integer, intent(in) :: order >> integer, intent(in) :: oldtype >> integer, intent(out) :: newtype >> integer, intent(out) :: ierr >> end subroutine MPI_Type_create_subarray >> end interface >> ----- >> >> One thing you might want to double check is that OMPI was compiled with >> MPI_REAL8 support (since MPI_REAL8 was passed to the subroutine). Run this: >> >> ----- >> shell$ ompi_info --all | grep real8 >> Fort have real8: yes >> Fort real8 size: 8 >> Fort real8 align: 8 >> shell$ >> ----- >> >> Finally, note that Open MPI doesn't have many of the F90 symbols that you >> might expect in libmpi_f90.so because most of our F90 support is "direct >> pass through" to the F77 subroutines. Specifically, our F90 mpi module >> simply provides compile-time checking for expected signatures, but the >> subroutine call is actually automatically linked against the corresponding >> symbol in libmpi_f77.so. Only a few subroutines -- i.e., the ones that have >> multiple different type signatures -- actually appear in libmpi_f90.so. >> >> Meaning: MPI_TYPE_CREATE_SUBARRAY only has one type signature. So it's >> normal / expected that you don't see /mpi_type_subarray/i in libmpi_f90.so. >> >> >> On Feb 15, 2011, at 1:18 PM, Petit, George (Cont, ARL/CISD) wrote: >> >>> A user posted the following ticket to our site. Should the referenced MPI >>> routine be included in the mpif90 library? If not, why? And is there a >>> way to get it added. >>> >>> ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------- >>> I receive the following error when I try to compile my code with ifort with >>> the >>> openmpi library: >>> >>> grid.f(38): error #6285: There is no matching specific subroutine for this >>> generic >>> subroutine call. [MPI_TYPE_CREATE_SUBARRAY] >>> call >>> mpi_type_create_subarray(3,isizes,isubs,istarts,MPI_ORDER_FORTRAN,MPI_REAL8,DTghost(1),ierr) >>> -------------^ >>> >>> It looks like the 'mpi_type_create_subarray' routine is missing from the >>> openmpi >>> library libmpi_f90.so. The routine is in libmpi_f77.so. Is it possible to >>> add it >>> to the f90 library? The fortran 90 versions of mpi for sgi_mpt, xt-mpt, >>> ibm-mpi and >>> intelmpi all include this routine. Thanks. -Rob >>> >>> George Petit >>> Lockheed Martin >>> ARL DSRC Customer Support Team Lead >>> george.pe...@us.army.mil >>> 410-278-9283 >>> >>> >>> _______________________________________________ >>> users mailing list >>> us...@open-mpi.org >>> http://www.open-mpi.org/mailman/listinfo.cgi/users >> >> >> -- >> Jeff Squyres >> jsquy...@cisco.com >> For corporate legal information go to: >> http://www.cisco.com/web/about/doing_business/legal/cri/ >> >> >> _______________________________________________ >> users mailing list >> us...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/users >> >> _______________________________________________ >> users mailing list >> us...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/users > > > -- > Jeff Squyres > jsquy...@cisco.com > For corporate legal information go to: > http://www.cisco.com/web/about/doing_business/legal/cri/ > > > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
