4 |
Notes on tags used in MITgcmUV |
Notes on tags used in MITgcmUV |
5 |
============================== |
============================== |
6 |
|
|
7 |
|
o adjoint: |
8 |
|
global_sum_ad was incomplete (noted many times by many people ;) |
9 |
|
nice description by Jean U. as follows: |
10 |
|
* allreduce(var) in global_sum.F can be interpreted as |
11 |
|
reduce(sum) followed by broadcast of sum. |
12 |
|
* The adjoint of that is |
13 |
|
sum(adjoint) followed by broadcast(summed adjoint value) |
14 |
|
or in other words, allreduce(sum, adjoint vars) |
15 |
|
We were missing the sum(adjoint)! |
16 |
|
Benign, except(!) for "global" scalars appearing in tile loops (bi,bj) |
17 |
|
(noted by J.M. Campin, e.g., in cost_ssh_mean, variable offset). |
18 |
|
The fix should produce 16-digit accuracy between |
19 |
|
single-CPU multi-tile (e.g. nSx=2, nPx=1) and |
20 |
|
multi-CPU single-tile (e.g. nSx=1, nPx-2) |
21 |
|
We'll probably change this yet again shortly |
22 |
o pkg/icefront |
o pkg/icefront |
23 |
- ALLOW_SUBGLACIAL_RUNOFF option adds capability for time-evolving |
- ALLOW_SUBGLACIAL_RUNOFF option adds capability for time-evolving |
24 |
specification of addmass array based on a 2D field, which is added |
specification of addmass array based on a 2D field, which is added |