Unable to add _FillValue to INT64 variable

[adamcpovey]

Using version 4.4.4 for the Fortran library and 4.4.1 of the C library compiled with either ifort 15.0.2 or gfortran 4.7.4, the following program:

program test
   use netcdf

   implicit none

   character(len=30) :: path="trash.nc"
   integer :: ierr, ncid, dim_var(1), varid
   integer, parameter :: k = selected_int_kind(18)
   integer, parameter :: len = 3
   integer, parameter :: s(1) = [1]
   integer, parameter :: c(1) = [len]
   integer(kind=k) :: data(l)

   ! Make data
   do ierr=1,len
      data(ierr) = ierr
   end do

   ! Make simple netcdf file
   ierr = nf90_create(path, IOR(NF90_NETCDF4, NF90_CLOBBER), ncid)
   ierr = nf90_def_dim(ncid, "len", len, dim_var(1))

   ! Make variable
   ierr = nf90_def_var(ncid, "test", NF90_INT64, dim_var, varid)

   ! This will succeed, but make a LONG variable rather than INT64
   ierr = nf90_put_att(ncid, varid, 'scale_factor', int(2, kind=k))

   ! This will fail
   ierr = nf90_put_att(ncid, varid, '_FillValue', int(1, kind=k))
   if (ierr /= NF90_NOERR) then
      write(*,*) trim(nf90_strerror(ierr))
      stop 5
   end if

   ierr = nf90_enddef(ncid)

   ! Write data
   ierr = nf90_put_var(ncid, varid, data, s, c, s)
   ierr = nf90_close(ncid)

end program test

exists prematurely returning,

NetCDF: Not a valid data type or _FillValue type mismatch STOP 5

The program runs as expected if k=2 and the variable is defined for NF90_SHORT or k=4 and NF90_INT.

My best guess is that here nf90_get_att_one_EightByteInt() casts values to an int before passing to the C code. If that's intended as a feature, could the documentation be updated to indicate that INT64 isn't intended for general use?

[WardF]

Thank you; I believe it should be available to use, although I am not a Fortran expert. I will investigate this and see what we can do to fix it. If an obvious fix suggests itself to you, please feel free to submit a pull request as well! Thanks, I'll address this for the next release one way or the other.

[DennisHeimbigner]

I did a quick look at the code and it looks to me like fixing this is going be a prolonged effort. As near as I can tell, this code was written before int64 was available, so it assumes that all integers are either 4 or 8 bytes (ala cray). The notion of having 4-byte and 8-byte integers mixed was not considered. I hope this problem is isolated to the attribute code, but not having looked, I bet it is a problem for int64 typed variables as well. How urgent is this?

[adamcpovey]

I wouldn't say it's massively urgent as I've added the necessary variable to my outputs, but until I can add attributes it won't be CF-compliant, which will be a problem in a few months when the compliance review happens.

[DennisHeimbigner]

Ok. I will put it on my stack.

[wkliao]

I believe this affects not only fill values but also variable and attribute values. Given both NetCDF-4 and CDF5 formats support NC_INT64 data type, it is expected netcdf-fortran include F77 functions like nf_put_vara_int8, nf_get_vara_int8, etc. so F90 functions can also call them accordingly. Will this issue be addressed in the next release?