id	summary	reporter	owner	description	type	status	priority	milestone	component	version	resolution	keywords	cc
380	Operations on uninitialized data in NETCDF_GET_FVAR_1D	m.hadfield	arango	"I have a limited-area ocean modelling application in which the usual variables are being read from a boundary file. On one of the platforms I run this on (Cray T3E), it fails in '''netcdf_get_fvar_1d''' with a message about an invalid floating-point value. The failure occurs somewhere in the following section of code:

{{{
        DO i=1,Asize(1)
          IF (ABS(A(i)).ge.ABS(Aspval)) THEN
            A(i)=0.0_r8
          ELSE
            A(i)=Afactor*A(i)+Aoffset
          END IF
        END DO
}}}

I set up a toy example with Lm=78, Mm=87 and N=25, serial with 1 tile in each direction. The failure occurred when reading vbar data from the southern boundary. '''netcdf_get_fvar_1d''' was being called by '''get_ngfld''' with the optional argument total set to 80 (the number of v points on the southern boundary, as expected) and being passed a 1D array much larger than this: I think the size was 2025 or 2225. In the above loop, Asize(1) was set to the size of the output array and the failure occurred on the 81st point, which had a value of NaN.

As I said, this failure occurs on only one platform, but processing uninitialized values is always dangerous and, in this case, inefficient: why process 2000-odd values when you only want 80 of them?

The fix I propose is to set Asize to the size of the data returned from the netCDF file, if that differs from the size of the output array. So the following

{{{
        Asize(1)=UBOUND(A, DIM=1)
}}}
becomes
{{{
      IF (PRESENT(start).and.PRESENT(total)) THEN
        Asize=total
      ELSE
        Asize(1)=UBOUND(A, DIM=1)
      END IF
}}}

NB: I test for the presence of both start and total because that is the test done when '''nf90_get_var''' is called.

I see no reason not to apply the same logic to all the '''netcdf_get_$var_$d''' routines, so I have done so in my copy of the source code, which is here:

https://www.myroms.org/projects/omlab/browser/branches/hadfield/trunk/ROMS/Modules/mod_netcdf.F

PS: I don't know why '''get_ngfld''' is creating such a large array to collect 1D boundary data, but I presume the same array is being used for other purposes.


"	bug	closed	major	Release ROMS/TOMS 3.3	Nonlinear	3.3	Fixed