[MITgcm-devel] SX8 changes in diagstats_calc.F

[Jean-Michel Campin]

Hi Martin,

The new secondary test "advect_xz.nlfs" seems to have also some problems
on sx8 in diagstats_calc.
I don't know which diagnostic is the problem; the new ones are:
 a) implicit vertical advection flux "ADVr_TH" & "ADVr_SLT" 
and 
 b) "gSinAB"
If you can take a look at this when you find time.

Thanks,
Jean-Michel

On Wed, Nov 16, 2011 at 10:40:59AM +0100, Martin Losch wrote:
> Hi Jean-Michel,
> 
> thanks for catching this. The problem is that "frcArr" seems to have some garbage on the overlaps. In the TARGET_NEC_SX version I use array assignments that include the overlaps, whereas in the original version, the loops range from 1,iRun/jRun. Improving vectorization requires that the loops go  over the entire array (that way the compiler can combine two loops into one), so I'd like to have this behavior also in (the TARGET_NEC_SX version of) diagstats_calc.F
> 
> The diagnostics that cause this problem are SI_Qice1/2, because in thsice_diagnostics_state.F the "frcArr" tmpFld is not initialized in the overlaps. I suggest fixing the loop ranges (and thus properly initialising tmpFild) in thsice_diagnostics_state.F (and likewise iceVol in thsice_advdiff.F, the associated diagnostics SI_Adv01/02 are not tested anywhere).
> 
> What do you think?
> 
> Martin
> 
> On Nov 16, 2011, at 1:43 AM, Jean-Michel Campin wrote:
> 
> > Hi Martin,
> > 
> > I noticed that 2 experiments (global_ocean.cs32x15.icedyn
> > & global_ocean.cs32x15.thsice) fail in your latest testreport from sx8,
> > and looks like from the error message that it has something to do
> > with diagstats_calc.F (the one you recently modified).
> > Could you check this ?
> > 
> > Thanks,
> > Jean-Michel
> > 
> > _______________________________________________
> > MITgcm-devel mailing list
> > MITgcm-devel at mitgcm.org
> > http://mitgcm.org/mailman/listinfo/mitgcm-devel
> 
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel