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