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

[Martin Losch]

Great, then I’ll go ahead now and check the updated output in.

Martin

> On 21 Jan 2016, at 16:21, Jean-Michel Campin <jmc at ocean.mit.edu> wrote:
> 
> Hi Martin,
> 
> Just one comment regarding forward grad of exp. global_ocean.cs32x15.seaice_dynmix :
> Even before you made changes with recip_deltaT in seaice_lsr.F,
> other than gfortran compilers were not giving very accurate fwd grad:
> 1) pgf77, mpi, -ieee:
> tr_acesgrid-adm_20160120_0/summary.txt:    Y Y Y Y  8>10< 2 pass  global_ocean.cs32x15.seaice_dynmix
> 2) pgf77, mpi, -fast:
> tr_acesgrid-adm_20160119_0/summary.txt:    Y Y Y Y  9> 3< 2 FAIL  global_ocean.cs32x15.seaice_dynmix
> 3) open64, no mpi, -devel:
> tr_acesgrid-tuv_20160119_0/summary.txt:    Y Y Y Y  8>11< 2 pass  global_ocean.cs32x15.seaice_dynmix
> 4) open64, mpi, -devel:
> tr_engaging-o64Adm_20160120_0/summary.txt: Y Y Y Y  8>11< 2 pass  global_ocean.cs32x15.seaice_dynmix
> 5) open64, mpi, -fast:
> tr_engaging-o64Adm_20160119_0/summary.txt: Y Y Y Y  8>11< 2 pass  global_ocean.cs32x15.seaice_dynmix
> 
> It's not impossible that, in some cases, using recip_deltaT will reduce differences between compilers.
> 
> Cheers,
> Jean-Michel
> 
> On Thu, Jan 21, 2016 at 07:27:25AM +0100, Martin Losch wrote:
>> OK,
>> 
>> I am waiting for the tests to complete in the afternoon an then do the update of adjoint and tlm results
>> 
>> Martin
>> 
>>> On 20 Jan 2016, at 18:12, Jean-Michel Campin <jmc at ocean.mit.edu> wrote:
>>> 
>>> Hi Martin,
>>> 
>>> I would suggest to go for 2 or 3, so that we get updated output for 
>>> all the experiments we understand why the results did change.
>>> Regarding global_ocean.cs32x15.seaice_dynmix, it's up to you:
>>> you could keep the old (and good) reference output or update it with
>>> a new and less good output.
>>> 
>>> Cheers,
>>> Jean-Michel
>>> 
>>> On Wed, Jan 20, 2016 at 05:00:35PM +0100, Martin Losch wrote:
>>>> Hi,
>>>> 
>>>> this recent change in seaice_lsr.F (replace 1/SEAICE_deltaTdyn wiht precomputed recip_deltaT) changes some results at truncation level, also for adjoint and tlm. For the adm and tlm seaice_dynmix experiment, however, the FD gradients are, totally unexpectly, broken (only 2 digits remain and the comparision to the adjoint and tlm gradients are 2 orders of magnitude less accurate). Since the adjoint and tangent-linear gradients only change a little (11 digits of agreement remain, so that I get a ???pass"), and the cost function is also not affected as much, I am a little at loss, what may cause this.
>>>> 
>>>> What should I do?
>>>> 1. revert to old code: go back to 1/SEAICE_deltaTdyn instead of recip_deltaT
>>>> 2. update all but the seaice_dynmix experiment
>>>> 3. update all experiments and worry about the accuracy later?
>>>> 
>>>> Martin
>>>> 
>>>>> Begin forwarded message:
>>>>> 
>>>>> From: Martin Losch <mlosch at forge.csail.mit.edu>
>>>>> Subject: [MITgcm-cvs] MITgcm/pkg/seaice CVS Commit
>>>>> Date: 20 Jan 2016 15:02:53 CET
>>>>> To: <mitgcm-cvs at mitgcm.org>
>>>>> Reply-To: <MITgcm-cvs at mitgcm.org>
>>>>> 
>>>>> Update of /u/gcmpack/MITgcm/pkg/seaice
>>>>> In directory forge:/tmp/cvs-serv1185/pkg/seaice
>>>>> 
>>>>> Modified Files:
>>>>> 	seaice_lsr.F 
>>>>> Log Message:
>>>>> replace 1./SEAICE_deltaTdyn with recip_deltaT for better vectorisation
>>>>> affects some experiments at truncation level
>>>>> 
>>>>> 
>>>>> _______________________________________________
>>>>> 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
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel