[MITgcm-devel] Problem in yersterday changes

[Gael Forget]

Hi Jean-Michel,

I had a quick look. There is one thing that looks particularly odd, 
which is the addition of the_main_loop in model_ad.flow. 
I am not sure what Patrick had in mine, but the change looks 
like we dont want taf to differentiate the_main_loop.F anymore (?!).
Could you give it a try after reverting to v1.5 for model_ad.flow.

Cheers,
Gael

On Oct 18, 2010, at 10:22 AM, Jean-Michel Campin wrote:

> Hi Martin,
> 
> I've checked that the problem comes from changes in pkg/autodiff 
> made on Sat. Oct 16 (all tests with TAF version 2.0.0).
> I am not going to do more tests now, because I don't know much about
> thoses changes.
> Can someone check those changes ? Or is it clear that they are OK
> in which case, might be a more tricky Pb.
> 
> Cheers,
> Jean-Michel
> 
> On Mon, Oct 18, 2010 at 12:51:56PM +0200, Martin Losch wrote:
>> Jean-Michel
>> 
>> timeint = deltaTclock = modelstep. Modelstep is the "cal" variable name. it is set once in cal_set and then kept in a common block and it should 3600 in this experiment, right? cal_getmonthsrec calls cal_timeinterval(-modelstep,...) so I assume that there is some memory problem. TAF has been changed a lot recently, so it's very well possible that suddenly the order or arguments etc. might be changed in the adjoint routines, so that memory gets overwritten? Can you figure out what happens with older versions of TAF?
>> 
>> Martin
>> 
>> On Oct 17, 2010, at 5:28 PM, Jean-Michel Campin wrote:
>> 
>>> Some update:
>>> I can reproduce the problem on my laptop with ifort and without MPI,
>>> and the problem is still there with seaice_growth.F v.1.89 from
>>> Oct.15 (looks like Gael's changes are not the problem).
>>> 
>>> Here is what I get when I run the noseaice test with idb:
>>>> Program received signal SIGFPE
>>>> cal_timeinterval (timeint=-9.2559631349317831e+61, timeunit=(...), date=(...), mythid=1, .tmp.TIMEUNIT.len_V$10=4) at /home/jmc/mitgcm/gcm_ifc/verification/lab_sea/bld_ad1p/cal_timeinterval.f:628
>>>> 628         date(1) = int(timeint/float(secondsperday))
>>> I guess timeint/secondsperday is too big to fit into an integer.
>>> 
>>> Jean-Michel
>>> 
>>> On Sun, Oct 17, 2010 at 10:51:41AM -0400, Jean-Michel Campin wrote:
>>>> Hi,
>>>> 
>>>> Looks like there is a problem with all lab_sea AD tests
>>>> as seen on aces with ifort+mpi:
>>>> http://mitgcm.org/testing/results/2010_10/tr_aces-adm_20101017_0/summary.txt
>>>> 
>>>> My impression is that it's not directly related to mpi (to be confirmed).
>>>> 
>>>> This was before Martin's lates changes in seaice-solve4temp.F 
>>>> and caused by one of yesterday modifs. 
>>>> 
>>>> I propose to hold on changes until this is fixed.
>>>> 
>>>> Cheers,
>>>> 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
>> 
>> 
>> _______________________________________________
>> 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