[MITgcm-devel] [MITgcm-cvs] MITgcm/pkg/seaice CVS Commit

[Jean-Michel Campin]

Hi Martin,

I am going to put back 2 old output_adm.* (among the 4 I undated yesterday)
and update a 3rd one, all in lab_sea/results:
The reason is that the recip_nDim is computed in single precision,
and there is no reason to do so (going to change this).
with gfortran on my laptop:
> recip_nDim old,new=  0.14285714924335480       0.14285714285714285

It's only used for the SEAICE_MULTICATEGORY code, which is tested 
only in the lab_sea adjoint test (why ? it's a mystery).
Once I fixed the precision calculation of recip_nDim,
the 2 non-evp lab_sea ad test pass (when compared to old output)
and just the evp ad test fails (but we know this one is ultra sensitive).
Here is the comparaison using old lab_sea/results/output_adm*.txt :
Y Y Y Y 16>13< 4 pass  lab_sea
Y Y Y Y 11> 7< 2 FAIL  lab_sea.evp
Y Y Y Y 16>16<16 pass  lab_sea.noseaice
Y Y Y Y 16>13< 4 pass  lab_sea.noseaicedyn

Cheers,
Jean-Michel

On Wed, Feb 15, 2012 at 11:02:27AM +0100, Martin Losch wrote:
> Hi Jean-Michel,
> 
> thanks for covering my back. 
> 
> I apologize: I have to admit that I only checked the forward runs assuming that the truncation issues would be the same for the adjoint runs. Gael pointed me to the one forward run that lost some precision (global_ocean.cs32x12.seaice), but that was too late for me to do anything else (I had to go home).
> 
> These divisions within loops have been bugging me for a while now. I am glad they are gone (except for recip_HO=1./HO), but I am sorry that this caused so many differences.
> 
> Martin
> 
> On Feb 15, 2012, at 3:06 AM, Jean-Michel Campin wrote:
> 
> > Martin,
> > 
> > The 1rst Pb should be fixed. 
> > And regarding AD tests, the ones that fail are:
> > Y Y Y Y  8>11< 2 pass  global_ocean.cs32x15.seaice_dynmix
> > Y Y Y Y 12> 7< 4 FAIL  lab_sea
> > Y Y Y Y  9> 5< 2 FAIL  lab_sea.evp
> > Y Y Y Y 12> 7< 4 FAIL  lab_sea.noseaicedyn
> > I am going to update the corresponding output_adm.*txt,
> > since it's likely a machine truncation problem.
> > And if not, we can put back the old results.
> > 
> > Jean-Michel
> > 
> > On Tue, Feb 14, 2012 at 07:19:55PM -0500, Jean-Michel Campin wrote:
> >> Hi Martin,
> >> 
> >> I am getting a floating point exception when trying to run
> >> 1D_ocean_ice_column (fwd), and also global_ocean.cs32x15.seaice (fwd)
> >> from the line you added:
> >>>     recip_HCUT = 1./HCUT
> >> because HCUT is zero.
> >> 
> >> Also, it looks like the changes you made change the way
> >> machine truncate the number so that a couple of adjoint verification
> >> experiment are affected and don't pass anymore (13 digits with
> >> gfortran on baudelaire). I am now checking this directly on baudelaire
> >> (was trying before on my laptop and could be due to something else).
> >> 
> >> Cheers,
> >> Jean-Michel
> >> 
> >> On Tue, Feb 14, 2012 at 09:52:09AM -0500, Martin Losch wrote:
> >>> Update of /u/gcmpack/MITgcm/pkg/seaice
> >>> In directory forge:/tmp/cvs-serv1373/pkg/seaice
> >>> 
> >>> Modified Files:
> >>> 	seaice_solve4temp.F 
> >>> Log Message:
> >>> remove one divisions by HCUT within a loop, without changing the results
> >>> 
> >>> 
> >>> _______________________________________________
> >>> MITgcm-cvs mailing list
> >>> MITgcm-cvs at mitgcm.org
> >>> http://mitgcm.org/mailman/listinfo/mitgcm-cvs
> >> 
> >> _______________________________________________
> >> 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
> 
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel