Jeff,
Thanks for the detailed response. It is about how I figured. I guess
what we need to do now is fiugure out where MPI_OFFSET_KIND is defined
in ROMIO. I did not find it in mpiof.h. Perhaps Rajeev can help us
with that one.
Dave.
Jeff Squyres wrote:
>Rajeev is exactly right -- we're using a third party package for
>MPI-2 IO functionality named ROMIO. It has its own functions, types,
>and header files. Hence, if anything is going to define the
>OFFSET_KIND type, it will be ROMIO.
>
>As for the mpiof.h issue: in LAM, as you saw, mpi.h includes mpio.h,
>so the user don't have to do anything. In mpif.h, we did *not*
>include mpiof.h because some fortran compilers don't accept the
>"include" statement. So we left it up to the Fortran users to
>manually include both mpif.h and mpiof.h (if you want to use ROMIO).
>
>Rajeev is also correct that this technically violates the standard --
>the standard says that all you should have to do is include mpif.h
>and that should have everything. But for similar integration
>reasons, ROMIO implements different types and function names (the
>MPIO_Request type in C, and MPIO_TEST and MPIO_WAIT in C/Fortran).
>Hence, users wanting to use MPI-2 IO functionality already have to
>violate the standard by using these non-standard names.
>
>We have a measley section on ROMIO in the User's Guide -- I should
>really expand it to include the above explanation, as well as the
>fact that you need to include mpiof.h in Fortran.
>
>All this being said, we fully integrated ROMIO in Open MPI -- you
>never need to include mpio.h or mpiof.h, you don't have to use
>MPIO_Request's, and you don't have to use MPIO_Test or MPIO_Wait.
>You use regular MPI_Request's in C, and can mix I/O, point-to-point,
>and generalized requests in the array versions of MPI_TEST and
>MPI_WAIT (in both C and Fortran) -- there is no need to use MPIO_TEST
>or MPIO_WAIT.
>
>I'm somewhat resistant to fix it so that you don't have to include
>mpiof.h in LAM for a few reasons:
>
>1. We are putting 99% of our effort into Open MPI these days, and the
>ROMIO integration is Much Better there (see above).
>
>2. The ROMIO integration that we did with Open MPI is at a much
>deeper/more fundamental level than we did with LAM; it is not
>possible to port that integration back to LAM.
>
>3. Even if we fix this one thing (no need to include mpiof.h) for our
>Last Big LAM Release (7.1.2), there will still be lots of old
>installations of LAM out there where you do need to include mpiof.h.
>
>
>
>On Jan 27, 2006, at 4:53 PM, David Cronk wrote:
>
>
>
>>I am hoping one of the LAM developers will provide some input.
>>That is
>>why I waited to respond. I looked through the header files for LAM.
>>MPI_Offset is defined in the C header file. However, I could not find
>>MPI_OFFSET_KIND in any of the Fortran header files.
>>
>>I have used LAM for MPI I/O in C programs in the past and it has
>>worked.
>> I have never tried it with Fortran programs, though the fact that I
>>don't see MPI_OFFSET_KIND being defined suggests to me that it will
>>not
>>work with Fortran.
>>
>>As far as what to include. As was mentioned earlier, it SHOULD not be
>>necessary to include both. LAM's mpi.h includes mpio.h for you so
>>with
>>C programs including just mpi.h is sufficient. This does not
>>appear to
>>hold for the Fortran headers, though someone from the LAM development
>>team may correct me there.
>>
>>So, it looks like the question for the LAM development team is,
>>where is
>>MPI_OFFSET_KIND defined?
>>
>>Dave.
>>
>>Barry A. Croker wrote:
>>
>>
>>>David, others,
>>>
>>>This might be an indication of a larger misunderstanding, that
>>>maybe you
>>>can help with. The documentation that I have is not entirely clear
>>>about some of the included header files needed for MPI I/O.
>>>
>>>For "normal" MPI programs I "INCLUDE mpif.h". When I started writing
>>>the MPI I/O routines on my local machine (using LAM-MPI) I found
>>>that I
>>>also had to "INCLUDE mpiof.h" for everything to be recognized.
>>>However,
>>>even with both of these includeded I still get "Entity
>>>mpi_offset_kind
>>>has undefined type", so I didn't use it (bad call I guess). When I
>>>ported our code to other platforms, I had to modify these include
>>>statements. For example, the MPICH-2 distribution gave me redundant
>>>declaration errors when I included both, but no complaints when I
>>>only
>>>"INCLUDE mpif.h". Likewise, the IBM SP4 had similar behavior.
>>>
>>>Is there something incorrect with my installation of MPI? Should
>>>I need
>>>both headers, or only one? The odd thing is the only time that
>>>MPI I/O
>>>appears to work correct is on my local machine, when I include
>>>both files.
>>>
>>>Thanks for everyone's help and patience with this issue.
>>>
>>>David Cronk wrote:
>>>
>>>
>>>
>>>>Barry,
>>>>
>>>>
>>>>try INTEGER(KIND=MPI_OFFSET_KIND) offset, disp
>>>>
>>>>
>>>>I am guessing your INTEGER is 32 bits while MPI requires 64 bits for
>>>>offset and displacement (files are often larger than 2 gig).
>>>>
>>>>
>>>>Dave.
>>>>
>>>>
>>>
>>>
>>>
>>>---------------------------------------------------------------------
>>>---
>>>
>>>_______________________________________________
>>>This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>>>
>>>
>>--
>>Dr. David Cronk, Ph.D. phone: (865) 974-3735
>>Research Director fax: (865) 974-8296
>>Innovative Computing Lab http://www.cs.utk.edu/
>>~cronk
>>University of Tennessee, Knoxville
>>_______________________________________________
>>This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>>
>>
>
>
>
>
|
