Is there a reason you do not want to use mpi_reduce with a user-defined reduction? Perhaps there is something that can be improved about that feature set. Jeff On Thu, Aug 9, 2018 at 5:14 PM Jiancang Zhuang <[email protected]> wrote:

Thanks a lot.

In fact, I have tried MPI_REAL16. The same thing happened. mpi_gather does not work for it either. I have decided to rewrite a new subroutine to replace mpi_reduce.

Best regards,

On Fri, Aug 10, 2018 at 12:38 AM Jeff Hammond <[email protected]> wrote:

...

The specific issue here is that MPI_LONG_DOUBLE does not map to REAL*16. You should try MPI_REAL16. I suspect it generates an error telling you that it is not supported.

MPI_LONG_DOUBLE corresponds to C "long double", which is an awful type that means a bunch of different things depending on your hardware and compiler. If you are using x86 Linux, it should be the x87 80-bit type. See https://en.wikipedia.org/wiki/Long_double for details.

Jeff

On Thu, Aug 9, 2018 at 8:33 AM, Jeff Hammond <[email protected]> wrote:

...

There is a challenge here because there is no standard type in C corresponding to REAL*16. Either the reduction operation needs to be written in Fortran or MPICH needs to figure out the compiler-dependent equivalent of REAL*16 that works in C. While GCC __float128 and Intel _Quad might be equivalent, this is not a rigorous assumption.

I recommend that you write your own user-defined reduction for REAL*16 with the reduction operation callback in Fortran.

Jeff

On Thu, Aug 9, 2018 at 2:31 AM, Jiancang Zhuang <[email protected]> wrote:

...

I have found that the fortran version mpi_reduce does not work for real*18. This can be shown by the following program. I have not test the C version of mpi_reduce.

c-----------------------Fortran code begins ----------------------------- implicit real*8 (a-h, o-z) include 'mpif.h' real*16 h1, h

call mpi_init(ierr) call mpi_comm_size(mpi_comm_world, nprocs, ierr) call mpi_comm_rank(mpi_comm_world, myrank, ierr)

h1 = (myrank+4) *2.00000000000000 write(*,'(''before reduce --'', i4,2f12.8)')myrank, h1,h

call mpi_reduce(h1,h,1,mpi_long_double,mpi_sum,0, & mpi_comm_world,ierr) write(*,'(''after reduce --'',i4,2f12.8)')myrank, h1,h

call mpi_bcast(h,1,mpi_long_double,0, & mpi_comm_world,ierr)

write(*,'(''bcastvalue -- '',i4,2f12.8)')myrank, h1,h

call mpi_finalize(ierr) end

c-----------------------Fortran code begins -----------------------------

$ mpif77 a.f -o a.out $ mpirun -np 3 ./a.out before reduce -- 1 10.00000000 0.00000000 after reduce -- 1 10.00000000 0.00000000 before reduce -- 2 12.00000000 0.00000000 before reduce -- 0 8.00000000 0.00000000 after reduce -- 2 12.00000000 0.00000000 after reduce -- 0 8.00000000 8.00000000 bcastvalue -- 0 8.00000000 8.00000000 bcastvalue -- 1 10.00000000 8.00000000 bcastvalue -- 2 12.00000000 8.00000000

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

-- Jeff Hammond [email protected] http://jeffhammond.github.io/

-- Jeff Hammond [email protected] http://jeffhammond.github.io/ _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

Address: 10-3 Midori-cho, Tachikawa, Tokyo 190-8562, Japan Phone: +81-50-5533-8532 Fax: +81-42-526-4335 email: [email protected] homepage: http://bemlar.ism.ac.jp/zhuang

---------------------------------------------------------------------------------------- _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

-- Jeff Hammond [email protected] http://jeffhammond.github.io/ _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss