[MITgcm-devel] Re: [MITgcm-cvs] MITgcm/tools/build_options CVS Commit

[Martin Losch]

I would be interested in which part of the seaice_evp routine causes  
the slowdown with -fp-model precise, but I do not have access to  
Columbia. Will I be able to reproduce this with "any" ifort compiler  
(i.e. even on simple linux boxes)?

Martin

On Sep 14, 2009, at 8:23 PM, Constantinos Evangelinos wrote:

> On Friday 11 September 2009 11:03:26 am David Ferreira wrote:
>> Hi Dimitris,
>> We are looking into this. At this point, with -O3, "-fp-model  
>> precise"
>> or "-mp"
>> is needed otherwise the models crashes (Chris tells me "-ftz" is
>> included into "-mp").
>> Also, it turns out that the verification experiments with cube-sphere
>> fail the restart
>> test with "-fp-model precise", but not with "-mp".
>> I'm running testreports with -O1/2/3, to see clarify a bit where it  
>> gets
>> bad.
>> "-O3 -ftz" or even "-O2" could end up being good compromise.
>> david
>
> On Itanium systems -O3 implies -ftz. It may very well be that is the  
> case for
> -O2 as well. -fp-model-precise appeared to be equivalent to -mp but  
> in fact
> seems to set -no-ftz explicitly so that -O3 -fp-model-precise (-ftz
> overridden) ends up being different from -O3 -mp (-ftz not  
> overridden). -mp/-
> fp-model-precise slows things down anyway (but not by a factor of 50  
> that
> denormal handling causes) so ideally one should find the source code  
> file that
> requires it and only use it there. We are forced to use -mp on ECCO  
> as well.
>
>> Dimitris Menemenlis wrote:
>>> Would it be make sense to also remove "-fp-model precise",
>>> i.e., make the second line the default, and possibly add "-ftz", to
>>> linux_ia64_ifort+mpi_altix_nas ?
>>>
>>> As Chris discovered, the "-fp-model precise" slows down
>>> seaice_evp.F by a factor of ~50.
>
> Constantinos
> -- 
> Dr. Constantinos Evangelinos
> Department of Earth, Atmospheric and Planetary Sciences
> Massachusetts Institute of Technology
>
>
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel