[MITgcm-cvs] MITgcm/doc CVS Commit
Martin Losch
mlosch at forge.csail.mit.edu
Wed Jul 16 08:03:54 EDT 2014
Update of /u/gcmpack/MITgcm/doc
In directory forge:/tmp/cvs-serv22885/doc
Modified Files:
tag-index
Log Message:
Add this:
o pkg/seaice: more clean up
- rename TAF common block tapes to be more consistent and remove
unnecessary duplicates
- add store directives to remove more recomputations,
but use smaller tapes (and adjust keys) when
SEAICE_LSR_ADJOINT_ITER is undefined. The idea behind this flag is
to reduce the memory footprint when the adjoint of the LSR is not
required. As of now, defining the flag SEAICE_LSR_ADJOINT_ITER
increase 4 common block tapes for 2D field by a factor of
SOLV_MAX_FIXED (=500 by default in SEAICE_SIZE.h). For experiment
offline_exf_seaice, this means on a linux box with gfortran nearly
a factor of 10 (but keep in mind there is mostly seaice code
in this experiment)
> size mitgcmuv_ad
text data bss dec hex
5100659 1216 220458952 225560827 d71c8fb
5339425 1200 1868783272 1874123897 6fb4dc79 (SEAICE_LSR_ADJOINT_ITER)
but only a moderate increase of 36% for global_ocean.cs32x15
text data bss dec hex
6695027 1648 1014578664 1021275339 3cdf6ccb
6695203 1648 1382505320 1389202171 52cd8afb (SEAICE_LSR_ADJOINT_ITER)
- note: there are still recomputations in global_ocean.cs32x15 and
the gradient agreement changes (gets a little worse)
- in general, the AD gradients still are wrong
More information about the MITgcm-cvs
mailing list