Using parallel NetCDF to save a distributed 3D complex array

问题

I have an MPI-based program written in Fortran which produces a 3D array of complex data at each node (sections of a 2D time-series). I would like to use parallel I/O to write these arrays to a single file which can be relatively easily opened in python for further analysis/visualization. Ideally I would like the solution to be memory efficient (i.e. avoid the creation of intermediate temporary arrays).

Using NetCDF, I have managed to adapt a subroutine which achieves this for a 3D array of real numbers. However, I've hit a stumbling block when it comes to complex arrays.

In the following code I have attempted to extend the subroutine from reals to complex numbers by creating a compound datatype consisting of two reals and assuming that the real and imaginary components of the Fortran complex datatype are stored contiguously in the 1st dimension of the 3D array.

 module IO

    use NetCDF
    use, intrinsic :: iso_fortran_env, only: dp => real64

    implicit none

    contains

        subroutine output_3D(dataname, starts, ends, global_data_dims, &
            local_data, MPI_communicator)

        character(len=*), intent(in) :: dataname
        integer, dimension(3), intent(in) :: starts
        integer, dimension(3), intent(in) :: ends
        integer, dimension(3), intent(in) :: global_data_dims

        complex(dp), intent(in) :: local_data(   1:(ends(1) - starts(1)+ 1), &
                                                1:(ends(2) - starts(2) + 1), &
                                                1:(ends(3) - starts(3) + 1))

        integer, dimension(3) :: expanded_starts

        integer, intent(in) :: MPI_communicator

        integer :: ncid, varid, dimid(3)
        integer :: counts(3)

        integer :: typeid

        expanded_starts(1) = (starts(1))* 2 + 1
        expanded_starts = starts(2) 
        expanded_starts(3) = starts(3)

        call check(nf90_create( trim(dataname)//'.cdf', &
                                IOR(NF90_NETCDF4, NF90_MPIIO), &
                                ncid, &
                                comm = MPI_communicator, &
                                info = MPI_INFO_NULL))

        call check(nf90_def_dim(ncid, "x", global_data_dims(1), dimid(1)))
        call check(nf90_def_dim(ncid, "y", global_data_dims(2) * 2, dimid(2)))
        call check(nf90_def_dim(ncid, "z", global_data_dims(3), dimid(3)))

        ! define a complex data type consisting of two real(8)
        call check(nf90_def_compound(ncid, 16, "COMPLEX", typeid))
        call check(nf90_insert_compound(ncid, typeid, "REAL", 0, NF90_DOUBLE))
        call check(nf90_insert_compound(ncid, typeid, "IMAG", 8, NF90_DOUBLE))

        ! define a 3D variable of type "complex"
        call check(nf90_def_var(ncid, dataname, typeid, dimid, varid))

        ! exit define mode
        call check(nf90_enddef(ncid))

        ! Now in NETCDF data mode

        ! set to use MPI/PnetCDF collective I/O
        call check(nf90_var_par_access(ncid, varid, NF90_COLLECTIVE))

        counts(1) = (ends(1) - starts(1) + 1) * 2
        counts(2) = (ends(2) - starts(2) + 1)
        counts(3) = (ends(3) - starts(3) + 1)

        call check(nf90_put_var(ncid, &
                                varid, &
                                local_data, &
                                start = expanded_starts,&
                                count = counts))

        ! close the file
        call check(nf90_close(ncid))

        return

    end subroutine output_3D

    subroutine check(status)

        integer, intent ( in) :: status

        if(status /= nf90_noerr) then
            print *, trim(nf90_strerror(status))
            stop 2
        end if

    end subroutine check

end module IO

program test_write

    use IO
    use MPI

    complex(dp) :: data(2,2,3)

    integer :: flock
    integer :: rank
    integer :: ierr

    integer :: i, j, k

    call MPI_init(ierr)
    call MPI_comm_size(MPI_comm_world, flock, ierr)
    call MPI_comm_rank(MPI_comm_world, rank, ierr)

    do k = 1, 3
        do j = 1, 2
            do i = 1, 2
                data(i,j,k) = cmplx(i, j, 8)
            enddo
        enddo
    enddo

    if (rank == 0) then

        call output_3D_hdf5('out', [1,1,1], [2,2,3], [2,2,6], &
                data, MPI_comm_world)

    else

        call output_3D_hdf5('out', [1,1,4], [2,2,6], [2,2,6],  &
                data, MPI_comm_world)

    endif

    call MPI_finalize(ierr)

end program test_write

The above code results in an "There is no specific function for nf90_put_var" error on compilation. This indicates that the function is not happy with the data type of the input array, so clearly there is something I'm missing regarding the usage of compound data types.

EDIT: One simple workaround is to assign the complex array to a real pointer as described in this post. The array can then be reshaped/recast using numpy to arrive at the complex array in python. It's a bit clunky, and somewhat dissatisfying - but is probably good enough for my purposes for now.

回答1:

This is only a partial answer for reasons you will see below - but it is too long for a comment. Hopefully I will be able to find the missing info and "upgrade" it, but this is what I have so far.

If you look in the NetCDF4 documentation under the "Compound type introduction" at https://www.unidata.ucar.edu/software/netcdf/fortran/docs/f90-user-defined-data-types.html#f90-compound-types-introduction you will see:

To write data in a compound type, first use nf90_def_compound to create the type, multiple calls to nf90_insert_compound to add to the compound type, and then write data with the appropriate nf90_put_var1, nf90_put_vara, nf90_put_vars, or nf90_put_varm call.

Note it doesn't mention nf90_put_var at all, but 4 different functions. This makes some kind of sense, nf90_put_var is presumably nicely overloaded to deal with all intrinsic types NetCDF support (and it is utterly crap it doesn't support complex), so for non-intrinsic type presumably there is some C like interface with something like a void *, and I'm guessing it is this that the four functions mentioned above implement.

So far so good, you should call one of nf90_put_var1, nf90_put_vara, nf90_put_vars, or nf90_put_varm rather than nf90_put_var. Now the bad news - I can't find any documentation for these 4 functions. The equivalent C functions are at https://www.unidata.ucar.edu/software/netcdf/docs/group__variables.html so you might be able to work out what is required from there, but it's not very good - I'd at the very least put in a bug report to Unidata, but that said for me lack of intrinsic support for complex is enough to make me look elsewhere for my I/O solution ...

While I am here you really shouldn't use explicit numbers for kinds for variables, I can show you compilers where complex(8) will fail to compile. Instead use Selected_real_kind or similar, or use the constant in the intrinsic module iso_fortran_env, or possibly those in iso_c_binding, and the fact that the kind of a complex number is the same as the kind of the reals that compose it.

来源：https://stackoverflow.com/questions/57130311/using-parallel-netcdf-to-save-a-distributed-3d-complex-array