4 |
Notes on tags used in MITgcmUV |
Notes on tags used in MITgcmUV |
5 |
============================== |
============================== |
6 |
|
|
7 |
|
o pkg/ptracers & diagnostics: |
8 |
|
- use ptracers (short) name (instead of long one) to make ptracers diagnostics |
9 |
|
title (to avoid some unexpected truncation); sort out ptracers diags units. |
10 |
|
|
11 |
|
checkpoint65 (2014/07/18) |
12 |
|
o pkg/fizhi: |
13 |
|
- add parenthesis around fizhi tendencies in fizhi_tendency_apply.F: this |
14 |
|
affects results at machine truncation level. Update output of experiments |
15 |
|
fizhi-cs-32x32x40 (12 digits for cg2d) & fizhi-cs-aqualev20 (13 digits). |
16 |
o pkg/seaice: more clean up |
o pkg/seaice: more clean up |
17 |
- rename TAF common block tapes to be more consistent and remove |
- rename TAF common block tapes to be more consistent and remove |
18 |
unnecessary duplicates |
unnecessary duplicates |
19 |
- add store directives to remove more recomputations, |
- add store directives to remove more recomputations, but use smaller |
20 |
but use smaller tapes (and adjust keys) when |
tapes (and adjust keys) when SEAICE_LSR_ADJOINT_ITER is undefined. |
21 |
SEAICE_LSR_ADJOINT_ITER is undefined. The idea behind this flag is |
The idea behind this flag is to reduce the memory footprint when the |
22 |
to reduce the memory footprint when the adjoint of the LSR is not |
adjoint of the LSR is not required. As of now, defining the flag |
23 |
required. As of now, defining the flag SEAICE_LSR_ADJOINT_ITER |
SEAICE_LSR_ADJOINT_ITER increases 4 common block tapes for 2D field by |
24 |
increase 4 common block tapes for 2D field by a factor of |
a factor of SOLV_MAX_FIXED (=500 by default in SEAICE_SIZE.h). |
25 |
SOLV_MAX_FIXED (=500 by default in SEAICE_SIZE.h). For experiment |
For experiment offline_exf_seaice, this means on a linux box with gfortran |
26 |
offline_exf_seaice, this means on a linux box with gfortran nearly |
nearly a factor of 10 (but keep in mind there is mostly seaice code in |
27 |
a factor of 10 (but keep in mind there is mostly seaice code |
this experiment): |
|
in this experiment) |
|
28 |
> size mitgcmuv_ad |
> size mitgcmuv_ad |
29 |
text data bss dec hex |
text data bss dec hex |
30 |
5100659 1216 220458952 225560827 d71c8fb |
5100659 1216 220458952 225560827 d71c8fb |
31 |
5339425 1200 1868783272 1874123897 6fb4dc79 (SEAICE_LSR_ADJOINT_ITER) |
5339425 1200 1868783272 1874123897 6fb4dc79 (SEAICE_LSR_ADJOINT_ITER) |
32 |
but only a moderate increase of 36% for global_ocean.cs32x15 |
but only a moderate increase of 36% for global_ocean.cs32x15: |
33 |
text data bss dec hex |
text data bss dec hex |
34 |
6695027 1648 1014578664 1021275339 3cdf6ccb |
6695027 1648 1014578664 1021275339 3cdf6ccb |
35 |
6695203 1648 1382505320 1389202171 52cd8afb (SEAICE_LSR_ADJOINT_ITER) |
6695203 1648 1382505320 1389202171 52cd8afb (SEAICE_LSR_ADJOINT_ITER) |
36 |
- note: there are still recomputations in global_ocean.cs32x15 and |
- note: there are still recomputations in global_ocean.cs32x15 and |
37 |
the gradient agreement changes (gets a little worse) |
offline_exf_seaice and the gradient agreement changes (gets a |
38 |
|
little worse) for global_ocean.cs32x15 (-> update output_adm.seaice.txt) |
39 |
- in general, the AD gradients still are wrong |
- in general, the AD gradients still are wrong |
40 |
o OpenAD |
o OpenAD |
41 |
- clean up code_oad/ directories somewhat by moving files |
- clean up code_oad/ directories somewhat by moving files |