[MITgcm-cvs] MITgcm/doc CVS Commit
Martin Losch
mlosch at forge.csail.mit.edu
Mon Jun 15 11:48:41 EDT 2015
Update of /u/gcmpack/MITgcm/doc
In directory forge:/tmp/cvs-serv1351/doc
Modified Files:
tag-index
Log Message:
document the following changes
o pkg/seaice: modify seaice_lsr.F in order to improve the gradient
computations; for SEAICE_VECTORIZE_LSR
- move the loops over which the tridiagonal solvers (seaice_lsr_tridiagu/v)
sweep to the outside,
- remove store directives and add "CADJ loop sequential" directives
in analogy to model/src/solve_tridiagonal.F
- replace many "#ifdef SEAICE_VECTORIZE_LSR_ZEBRA" by variable loop
boundaries to yield more readable code. This has also the charming side
effect that your can use SEAICE_VECTORIZE_LSR_ZEBRA without
SEAICE_VECTORIZE_LSR (i.e. adjoint without recomputations in
seaice_lsr_tridiagu/v now requires either of these flags, vectorization
still requires SEAICE_VECTORIZE_LSR with SEAICE_VECTORIZE_LSR_ZEBRA as
an additional option)
The tridiagonal solvers are now completely analoguous to solve_tridiagonal.F
but the adjoint gradients (which are affected by this change) still explode.
- Rename SEAICE_VECTORIZE_LSR_ZEBRA to SEAICE_LSR_ZEBRA and define it
in lab_sea forward experiment (changes results -> updated)
- updating adjoint experiment will follow
More information about the MITgcm-cvs
mailing list