Nvfortran Direct Access I/O issue

Hi everyone,

I’m dealing with I/O operations in a MPI Fortran code compiled with mpif90 under nvhpc-23.1. I’m trying to read a binary file in this way:

open(UNIT=fh, file=file_name,FORM='UNFORMATTED',RECL=reclen, &

do k = MINk(Zone)-mnk, MAXk(Zone)+mxk
  do j = MINj(Zone)-mnj, MAXj(Zone)+mxj
    do i = MINi(Zone)-mni, MAXi(Zone)+mxi
     displ = (k - K_IND_MIN(Zone) + GHOSTS)*(range_i*range_j)+ &
             (j - J_IND_MIN(Zone) + GHOSTS)*(range_i)+ &
             (i - I_IND_MIN(Zone) + GHOSTS) + 1
     READ(fh,rec=displ) buf
    end do
  end do
end do

The code uses OpenACC for GPU offloading but this section should be executed by the CPUs. I do not have problems in compilation but during execution the following issue arise:

FIO-F-219/unformatted read/unit=23/attempt to read/write past end of record.
File name = 'DNS_110.ZONE0',    unformatted, direct access   record = 1
In source file ../SHARED_FILES/nuovo_read_field.f90, at line number 80

This file (DNS_110.ZONE0), was written by the same code compiled with mpif90 shipped under openmpi_intel17-3.1.2 in the following way:

open(UNIT=fh, file=file_name,FORM='UNFORMATTED',RECL=reclen * range_i,  &

do k= MINk(Zone)-mnk, MAXk(Zone)+mxk
  do j= MINj(Zone)-mnj , MAXj(Zone)+mxj
     do i= MINi(Zone)-mni , MAXi(Zone)+mxi

        ....Some calculations...

        displ_f = (k - K_IND_MIN(Zone) + n_ghosts)*(range_j)+ (j - J_IND_MIN(Zone) + n_ghosts) + 1
        write(fh, rec=displ_f, err=100, iostat=stato) buf

     end do
  end do
end do

For the sake of clarity, this code works 100% compiled with openmpi-intel17 but it does not compiled with nvhpc, so the issue. Am I doing something wrong dealing with this type of operation? Should I include some flag during compilation? Thanks everybody in advance for the support!


If I remember correctly, Ifort uses a recl in units of 4-byte words while we as well as gfortran use 1-byte. Try setting the environment variable “FORTRANOPT=vaxio” to use the old VAX/IO.

Alternatively, rebuild the ifort code with the “-assume byterecl” so they output using 1-byte words.