4 |
Notes on tags used in MITgcmUV |
Notes on tags used in MITgcmUV |
5 |
============================== |
============================== |
6 |
|
|
7 |
|
o pkg/seaice: more clean up |
8 |
|
- rename TAF common block tapes to be more consistent and remove |
9 |
|
unnecessary duplicates |
10 |
|
- add store directives to remove more recomputations, |
11 |
|
but use smaller tapes (and adjust keys) when |
12 |
|
SEAICE_LSR_ADJOINT_ITER is undefined. The idea behind this flag is |
13 |
|
to reduce the memory footprint when the adjoint of the LSR is not |
14 |
|
required. As of now, defining the flag SEAICE_LSR_ADJOINT_ITER |
15 |
|
increase 4 common block tapes for 2D field by a factor of |
16 |
|
SOLV_MAX_FIXED (=500 by default in SEAICE_SIZE.h). For experiment |
17 |
|
offline_exf_seaice, this means on a linux box with gfortran nearly |
18 |
|
a factor of 10 (but keep in mind there is mostly seaice code |
19 |
|
in this experiment) |
20 |
|
> size mitgcmuv_ad |
21 |
|
text data bss dec hex |
22 |
|
5100659 1216 220458952 225560827 d71c8fb |
23 |
|
5339425 1200 1868783272 1874123897 6fb4dc79 (SEAICE_LSR_ADJOINT_ITER) |
24 |
|
but only a moderate increase of 36% for global_ocean.cs32x15 |
25 |
|
text data bss dec hex |
26 |
|
6695027 1648 1014578664 1021275339 3cdf6ccb |
27 |
|
6695203 1648 1382505320 1389202171 52cd8afb (SEAICE_LSR_ADJOINT_ITER) |
28 |
|
- note: there are still recomputations in global_ocean.cs32x15 and |
29 |
|
the gradient agreement changes (gets a little worse) |
30 |
|
- in general, the AD gradients still are wrong |
31 |
o OpenAD |
o OpenAD |
32 |
- clean up code_oad/ directories somewhat by moving files |
- clean up code_oad/ directories somewhat by moving files |
33 |
dontCompile, dontTransform, keepOriginal to OAD_support/ |
dontCompile, dontTransform, keepOriginal to OAD_support/ |