| 4 |
Notes on tags used in MITgcmUV |
Notes on tags used in MITgcmUV |
| 5 |
============================== |
============================== |
| 6 |
|
|
| 7 |
|
o eesupp/src (exch-1): |
| 8 |
|
- split exch_rx.F into Forward S/R (exch1_rx.F) and Adjoint S/R (exch1_rx_ad.F) |
| 9 |
|
and remove argument "simulationMode"; same thing for all exch_*rx_cube.F |
| 10 |
|
(renamed to exch1_*rx_cube.F). |
| 11 |
|
- change accordingly exch_*_rx.template and AD version in pkg/autodiff |
| 12 |
|
to call new exch1 S/R ; update Makefile and .cvsignore. |
| 13 |
|
- replace call to exch2_xy_rx S/R with exch2_*3d_rx(myNz=1) (will remove |
| 14 |
|
exch2_xy_rx later) |
| 15 |
|
Note: after updating your code, need to run "make clean" in eesupp/src |
| 16 |
|
o test experiment bottom_ctrl_5x5: |
| 17 |
|
new output_adm.txt with fewer grdchk & less monitor output. |
| 18 |
|
o model/src: |
| 19 |
|
remove the call to ini_cori in initialise_varia.F (just keep the usual one |
| 20 |
|
in initialise_fixed.F). More consistent with coriolis & horizontal grid |
| 21 |
|
beeing treated in the same way (ini_grid is not called a 2nd time). |
| 22 |
|
o pkg/exf: |
| 23 |
|
- rewrite exf_bulk_largeyeager04.F for better vectorization, changes |
| 24 |
|
global_ocean.cs32x15/input.icedyn (11 digits of agreement), updated |
| 25 |
|
- fix adjoint (used worked for ALLOW_ATM_WIND defined, works for this now) |
| 26 |
|
- cosmetic changes to exf_bulkformulae.F to make it look more like |
| 27 |
|
exf_bulk_largeyeager04.F |
| 28 |
|
o eesupp/src (exch-1): |
| 29 |
|
- Fix AD EXCH code for 2-D domain special cases (Nx=1 or Ny=1). |
| 30 |
|
- Separate buffer filling and MPI sending: allow EXCH-1 to work for local |
| 31 |
|
array (non-shared) when using MPI+MTH. Also reduces number of BARRIER |
| 32 |
|
(even without using MPI). |
| 33 |
|
- Message mode: move RecvAck setting (indicator of buffer being ready) |
| 34 |
|
from send_put to recv_get S/R (was useless before, but not sure if |
| 35 |
|
it's much more useful now); |
| 36 |
|
- Switch the order of sync: MPI-proc 1rst and then threads; |
| 37 |
|
- take out spin-waiting code (#undef EXCH_USE_SPINNING), use BARRIER instead. |
| 38 |
|
- fix AD code for communication mode ='get' (but not used). |
| 39 |
|
o model/src: |
| 40 |
|
initialise dPhiHydX,dPhiHydY inside calc_grad_phi_hyd.F (so that TAF does |
| 41 |
|
not drop it in S/R mddynamics, producing NaN in hs94.1x64x5 AD-test exp). |
| 42 |
|
o utils/matlab/rdmnc.m: adapt to be able to use the generic matlab netcdf API |
| 43 |
|
o pkg/ctrl: require a specific set of control variables and exf-flags for |
| 44 |
|
adjoint of seaice (stop in ctrl_check.F if conditions are not satisfied) |
| 45 |
|
o model/src: |
| 46 |
|
a) 1rst version of a new driver S/R to turn-off output freq and flags after the |
| 47 |
|
1rst forward sweep, to avoid over-writing output files during recomputations |
| 48 |
|
recomputations (& grdchk); called from cost_final.F (& ctrl_init_ctrlvar.F ?) |
| 49 |
|
b) 2nd version: |
| 50 |
|
- add specific pkg S/R for seaice & ptracers (called from S/R TURNOFF_MODEL_IO); |
| 51 |
|
- switch-off more output flags (exf_monFreq, chkPtFreq, dumpInitAndLast, |
| 52 |
|
writePickupAtEnd) + seaice & ptracers output; |
| 53 |
|
- do not switch off MNC flags (otherwise, won't get admonitor netcdf file) |
| 54 |
|
o model/src: |
| 55 |
|
- new S/R rotate_uv2en_rl (..._rs) that rotates vector fields. |
| 56 |
|
o pkg/exch2: |
| 57 |
|
- remove "simulationMode" from arg. list of S/R EXCH2_RX1,2_CUBE ; |
| 58 |
|
- add argument "signOption" to EXCH2_RX1_CUBE (will be needed for SM exch). |
| 59 |
|
o pkg/autodiff & exch2: |
| 60 |
|
process exch_uv_agrid_3d_rx.f & exch2_uv_agrid_3d_rx.f through TAF |
| 61 |
|
(for now, only if ALLOW_EXCH2 is defined). |
| 62 |
|
o pkg/autodiff: |
| 63 |
|
- include PACKAGES_CONFIG.h in exch*ad.flow & exch*ad.F |
| 64 |
|
(not supposed to be in CPP_OPTIONS.h) to get ALLOW_EXCH2 defined correctly. |
| 65 |
|
+ clean-up comments (_RS is not always real*4) |
| 66 |
|
- fix bug in ADEXCH_UV_3D_RX (without EXCH2, was calling the forward exch S/R) |
| 67 |
|
- add a stop in ADEXCH_UV_AGRID_3D_RX if trying to use it with EXCH2: |
| 68 |
|
was calling a forward exch2, which is wrong ; used in SEAICE & EXF pkgs. |
| 69 |
|
o testreport (AD-tests): |
| 70 |
|
no longer link files from dir "../input" (1rst link files from |
| 71 |
|
"../input_ad.[xx]" if secondary test "xx" and then from "../input_ad") |
| 72 |
|
|
| 73 |
|
checkpoint62f (2010/04/29) |
| 74 |
|
o model/src/the_main_loop.F: |
| 75 |
|
adjust store directives for seaice evp and multicategory according to |
| 76 |
|
pkg/ecco/the_main_loop.F |
| 77 |
|
o adjoint: |
| 78 |
|
global_sum_ad was incomplete (noted many times by many people ;) |
| 79 |
|
nice description by Jean U. as follows: |
| 80 |
|
* allreduce(var) in global_sum.F can be interpreted as |
| 81 |
|
reduce(sum) followed by broadcast of sum. |
| 82 |
|
* The adjoint of that is |
| 83 |
|
sum(adjoint) followed by broadcast(summed adjoint value) |
| 84 |
|
or in other words, allreduce(sum, adjoint vars) |
| 85 |
|
We were missing the sum(adjoint)! |
| 86 |
|
Benign, except(!) for "global" scalars appearing in tile loops (bi,bj) |
| 87 |
|
(noted by J.M. Campin, e.g., in cost_ssh_mean, variable offset). |
| 88 |
|
The fix should produce 16-digit accuracy between |
| 89 |
|
single-CPU multi-tile (e.g. nSx=2, nPx=1) and |
| 90 |
|
multi-CPU single-tile (e.g. nSx=1, nPx-2) |
| 91 |
|
We'll probably change this yet again shortly |
| 92 |
|
o eesupp/src: |
| 93 |
|
- fix global_sum_r4 (wrong from previous modif in Jun 2009; ever used ?) |
| 94 |
|
o pkg/icefront |
| 95 |
|
- ALLOW_SUBGLACIAL_RUNOFF option adds capability for time-evolving |
| 96 |
|
specification of addmass array based on a 2D field, which is added |
| 97 |
|
at bottommost wet level at each horizontal location. |
| 98 |
|
o model/src: |
| 99 |
|
horizontal grid initialisation (other than curvilinear): add code for Exch2 IO |
| 100 |
|
layout; (not always compatible with delX,delY setting; commented out for now). |
| 101 |
|
o pkg/exch2: |
| 102 |
|
change tile to processor setting (case nSy*nPy > 1): allows default model mapping |
| 103 |
|
(myX and myYGlobalLow) to be used for simple (1 facet, no blank tile) config. |
| 104 |
|
|
| 105 |
|
checkpoint62e (2010/04/15) |
| 106 |
|
o pkg/obcs: |
| 107 |
|
- fix wrong index in OBCS_SPONGE_V for Eastern-OB (reported by David Hebert). |
| 108 |
o pkg/dic: |
o pkg/dic: |
| 109 |
-fix bug in fe_chem.F for multi-tile (nSx*nSy > 1) set-up; |
-fix bug in fe_chem.F for multi-tile (nSx*nSy > 1) set-up; |
| 110 |
(note: not presently tested in any forward verification exp.) |
(note: not presently tested in any forward verification exp.) |
Parent Directory
| 
Revision Log
| 
Revision Graph
| 
Patch