Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* Fixed some coding mistakes from the last commit.
* Added new variable FrictWork_diss to account for the energy dissipated by shear
production (excluding the energy diffusion term which can be positive).
* Added a new flag to MOM_lateral_mixing_coeffs called "USE_VISBECK", which
calculates the GM coefficient using the Visbeck et al. (1997) formulation. Previously
this was being done whenever VarMix was enabled, but I wanted to separate VarMix from
this diffusivity call.
* Re-factor of Travis-CI usage
- Uses stages to parallelize testing process
- Adds a doxygen job that tests for doxygen errors
- Reduces dependency on other repositories
- No longer uses scripst from MOM6-examples/tools/tests/Travis-MOM6/
- Still uses configurations from MOM6-examples/ocean_only/
- Adds test building/using a debug executable
- Adds a .testing/ directory
- Uses the latest xenial image on Travis (with openmpi)
Todo:
- Add code coverage (needs more experiments)
- Break dependence on MOM6-examples (long term project)
- Add other portable tests (e.g. parameter scaling, rotation, etc)
* Set ca13csed to zero at k>1
- Like the original cased tracer, the ca13csed tracer used in
generic_BLING.F90 needs to be initialized to zero at all
subsurface layers (when do_13c = .true.)
* Fixed problems with FrictWorkMax and FrictWorkDiss in MOM_hor_visc.F. More specifically,
I moved the calculation of dudx and dvdy outside of the Leith loop so now it is
being calculated with sh_xx.
* +Remove unneeded MOM6 coupler_types.F90 files
With the xanadu release of FMS, the coupler_types module within FMS has
finally been properly scoped without adding unnecssary dependencies. The
coupler_types.F90 files in the MOM6 config_src/solo_driver directories are no
longer needed. In addition, the routines inside coupler_util.F90 files have not
been used anywhere in MOM6 for some time, their functionality being provided via
interfaces inside the coupler_types_mod. These 5 files have now been removed.
All answers are bitwise identical.
* +Added missing US arguments to nuopc_cap code
Added unit_scale_type arguments to several subroutines in the nuopc_driver
directory, added a unit_scale_type element to the nuopc_driver versoin of
ocean_state_type, and use this argument to appropriately rescale forces%ustar
for dimensional consistency testing. These changes are required to go with the
dev-master-candidate-2018-04-22 updates to MOM6. With these changes, the code
in origin/dev-master-candidate-2018-04-22 compiles up through the nuopc_driver
version of MOM_ocean_model.o.
* Checksum support for Depth_list.nc
This patch appends a checksum for the dependencies of the depth and area
lists stored in the Depth_list.nc file, which are used to compute
diagnostics based on APE.
The data in Depth_list.nc depends on the grid fields, and may not be
reproducible when such grids are constructed internally using compiled
code within the executable. This issue was observed in the
'double_gyre' experiment when a PGI-compiled executable was tested using
a Depth_list.nc file generated by a GNU-compiled executable.
By appending a checksum for the grid fields used to compute
Depth_list.nc, we can ensure that the data is consistent with the
experiment grid data. Grid data which is read from external files, such
as mosaic or topography fields, are unaffected by this issue.
This patch improves the reproducibilty of standard diagnostics, such as
total energy, but has no impact on the reproducibility of the internal
model dynamics, which does not depend on Depth_list.nc.
Checksums are computed for the G%bathyT and masked G%areaT grid fields
using the FMS mpp_checksum subroutine, which require collective
operations, and are stored as hex strings in global attributes of the
netCDF file. Strings are used to remain consistent with FMS restart
checksums, and to avoid an observed re-casting of 8-byte integers to
4-bytes by the netCDF library. Attribute names are based on the grid
variable names.
Two flags have been introduced to control this behavior:
REQUIRE_DEPTH_LIST_CHECKSUMS (default: True)
This flag will abort the run if the Depth_list.nc file is present
and checksums are absent from the file. Although this could impose
greater restrictions on existing runs, few runs are configured to
save the depth list file (READ_DEPTH_LIST) and the default behavior
is to reconstruct these lists on every run.
UPDATE_DEPTH_LIST_CHECKSUMS (default: False)
When REQUIRE_DEPTH_LIST_CHECKSUMS is set to false, this flag will
automatically update the checksums of the Depth_list.nc file. While
this can affect the reproducibility of APE diagnostics, it will
ensure the reproducibility of such diagnostics in subsequent runs.
* Further documentation of Depth_list.nc checksums
Additional documentation of the parameters used to store Depth_list.nc
attribute names was added.
* Depth list documentation typo fix
* make uppercase DEBUG lowercase
* fix uninitialized oneOrTwo
* Add CS%US argument lost during merge conflict
* Masked depth; style conformance
The depth checksum is now replaced with masked depth, mask2dT * bathyT,
and the calculation of the depth list has also been updated to use the
masked depth.
Various style conformance changes, such as contraction of do and if
terminations (enddo, endif) and reduction of whitespace in various
multiline function call, has also been applied.
Finally, the attribute name docstrings were updated for clarity.
* Token bugfix
* use ESMF operators
* gitlab: specify MOM6-examples and script commits
- There is no source code change in this commit.
- The switch to the xanadu version of FMS and coupler moves source code
between repositories, and therefore requires new build paths which are
wired into the testing scripts. This commit checks out a specific
version of the gitlab testing scripts along with a xanadu commit of
MOM6-examples (which is on branch xanadu-fms).
- Once the dev/gfdl branch of MOM6-examples has been rolled forward to
xanadu we will return the commit used for gitlab testing to the
HEAD of dev/gfdl.
* Revert masking of depth
Using the masked depth (mask2dT * bathyT) was observed to change energy
values within floating point precision, so the changes have been
reverted. This may be revised at a later time, when we are prepared to
update the energy stats to the new values in the regression tests.
The depth checksum attribute has also been renamed to reflect this
change. This will allow us to re-define the variable as masked at some
later date, and can distinguish between the masked and unmasked
checksums during testing.
* Bugfix: Checksum the unmasked depth
* Added Jansen et al. (2015) version of MEKE dissipation. Added an alternative
way of calculating PE-to_MEKE energy conversion.
* Fixed if statement in MOM_MEKE.F90. Added option to use alternative
form of GM PE-to_MEKE energy conversion in MOM_thickness_diffuse.
* Added option to calculate MKE-to-MEKE energy conversion by GME.
* Cleaned up MOM_hor_visc.F90 a bit, especially parts having to do with
calculating FrictWork.
* Added biharmonic option for MEKE viscosity.
* Added a GME sink term to the MEKE budget. Cleaned up MOM_hor_visc.F90.
* Commented out thickness diffusivity calls for GME, since it is not
presently being used.
* Add seaice_melt and seaice_melt_heat
This commit adds two new fluxes to the nuopc cap. The MOM (coupler)
definition of these terms is seaice_melt (meltw) and seaice_melt_heat (melth).
Currently, the alias name for these terms in the fd.yaml file is
mean_fresh_water_to_ocean_rate and net_heat_flx_to_ocn. We need to change these
to more meningful names and this will require changes in CICE.
* MEKE: Prevent div-by-zero by bathyT in beta calc
The calculation of beta in the MEKE module had an explicit division by
zero when computing the lateral topography derivatives, which were
raising floating point exceptions in the debug builds. This occurs in
the MEKE_equilibrium and MEKE_lengthScales functions.
This issue was not observed in the production tests due to
MEKE_TOPOGRAPHIC_BETA always being set to zero. When FPEs are disabled,
the 0 * (1./0.) operation produces a NaN which is passed to a max()
function, which ignores the NaN and always returns the other value.
We resolve this by explicitly checking for zero values in bathyT and
setting the topographic beta to zero when this term is zero. While this
could potentially change the value of the Rhines scale, these values
only occur over land, which are in general masked, and should not affect
the calculation.
The unoptimized expressions were retained, but recommended changes which
reduce the number of divisions were included in comments.
No value changes were observed in our test suite, and the patch should
be bitwise reproducible.
Minor changes:
- We do not calculate the topographic beta term if the scaling factor,
MEKE_TOPOGRAPHIC_BETA is zero
- the default value of beta was unset in MEKE_lengthScales was unset
when CS%use_old_lscale is True, so we set this to zero.
- Minor whitespace and index syntax changes
* Add FMS/coupler_types.F90 to build Travis-CI build path
* Add missing US entries and fix typos introduced during merge
* Use xanadu-fms version of MRS in gitlab pipeline
* get calendar from CESM
* Revert "Added biharmonic option for MEKE viscosity."
This reverts commit 6b32174.
Conflicts:
src/parameterizations/lateral/MOM_MEKE.F90
src/parameterizations/lateral/MOM_hor_visc.F90
* Willy was silly.
* Working configuration of MEKE/GME. Has biharmonic MEKE viscosity and a
GME sink term.
* Added a "next-to_boundary" mask that should help with fixing abnormally large
values of the velocity gradient tensor magnitude next to land.
* MOM_set_diffusivity: OMEGA scaling
Implementation of OMEGA parameter rescaling (s-1 -> T-1) in the
MOM_set_diffusivity module.
* MOM_set_diffusivity: Rescaling of Omega2, G_Rho0
* MOM_set_diffusivity: N, dissipation scaling
This commit rescales the internal buoyancy frequencies (N2) and
associated parameters in the MOM_set_diffusivity module. Coefficients
used in computation of dissipation have also been scaled relative to W
m-3, e.g. W m-3 s2 is rescaled to W m-3 T2.
Scaled parameters:
* DISSIPATION_N1 (partial): W m-3 T
* DISSIPATION_N2 (partial): W m-3 T2
* DISSIPATION_KD_MIN: Z2 T-1
* dissip_N2 dimension update
* (*)Revised propagate_int_tide code for symmetry
Revised propagate_int_tide code for symmetry, and removed the old and unused
debugging code setting an internal tide point source. These could change
answers at the level of roundoff and MOM_parameter_doc files in cases that have
INTERNAL_TIDES=True, but all answers are bitwise identical in the MOM6-examples
test cases.
* +(*)Revised the diagnostic subML_N2
Revised the calculation of the diagnostic subML_N2 inside of
diagnoseMLDbyDensityDifference, and added as an optional argument the depth
extent over which this stratification was calculated. The previous code used
inconsistent pressures in calculating the densities and was simply wrong. This
changes a diagnostic (subML_N2) and adds a new optional interface. All solutions
are bitwise identical but there are new entries in MOM_parameter_doc files.
* +Added MIN_SALINITY as a new runtime parameter.
Added the new runtime parameter MIN_SALINITY, with a default value that
duplicates the current hard-coded value. Also stored this value in a new
element in the thermo_vars type, and corrected a few spelling errors. By
default all answers are bitwise identical but there are new entries in the
MOM_parameter_doc.all files.
* +Corrected documentation of HALF_STRAT_DEPTH
Corrected documentation of HALF_STRAT_DEPTH written to MOM_parameter_doc files
by Phillips_initialize_thickness. All answers are bitwise identical, but
MOM_parameter_doc files change.
* +Corrected spelling errors in documentation
Corrected numerous spelling errors in comments, including the descriptions
that are written to the MOM_parameter_doc files. All answers are bitwise
identical, but there are numerous minor changes to the MOM_parameter_doc files.
* MOM_set_diffusivity: dissip scaling
Dissipation parameters are now fully rescaled with respect to time:
* DISSIPATION_MIN
* DISSIPATION_N[012]
as are internal dissipation variables.
* MOM_set_diffusivity scaling: N2, Kd_work diags
Dimensional scaling support for the N2 and Kd_work diagnostics in the
MOM_set_diffusivity module, as wel as a hN2pO2 internal variable.
* MOM_set_diffusivity: TKE_to_Kd time rescaling
* MOM_set_diffusivity: scaling of ML radiation terms
Time rescaling of internal variables for add_MLrad_diffusivity,
including the following parameter:
* ML_RAD_KD_MAX
* MOM_set_diffusivity: maxTKE time scaling
The calculation of maxTKE is rescaled inside set_diffusivity, along with
various other internal variables. add_drag_diffusivity not yet scaled.
* MOM_set_diffusivity: more TKE time scaling
* MOM_set_diffusivity time scaling: more params
The following parameters have been time scaled:
* Kd_min
* Kd_max
The following physical constants have been scaled:
* kpp_fill
* dt_fill
Internal variables:
* delta_kd
Some units in comments have also been fixed
* +Added newlines in long parameter descriptions
Added newlines in long parameter descriptions in 6 modules. All answers are
bitwise identical, but there are changes in comments in the MOM_parameter_doc
files.
* +Revised writeMessageAndDesc to split long lines
Revised writeMessageAndDesc to split comment lines that are longer than 112
characters long (usually this is a 32 character indent, with an 80 character
description block), including aligned indents of continuing lines in indented
lists. This will change some MOM_parameter doc files with very long lines, and
it will eliminate the need for explicit newlines in get_param, log_param, and
log_version calls, although they still work as before in shorter lines. All
answers are bitwise identical, but a line or two in some MOM_parameter_doc files
and some SIS_parameter_doc files change.
* * Bugfix: Double diffusive correction to Kd
The double-diffusive salinity correction to diffusivity was incorrectly
using an exponent in its term:
Kd_lay(i,j,k-1) = Kd_lay(i,j,k-1) + 0.5**KS_extra(i,K)
The exponent has been replaced with a multiplication.
* Add FMS/coupler_types.F90 to build Travis-CI build path
* Use xanadu-fms version of MRS in gitlab pipeline
* MOM_set_diffusivity LOTW time scaling
Mostly rescaling of internal variables for LOTW diffusivity.
* Set default of USE_NET_FW_ADJUSTMENT_SIGN_BUG to false in NUOPC cap
* MOM_set_diffusivity time scaling: DD, bg diff, zint
The double diffusive calculations were time-scaled, as well as the
diagnostics.
Background diffusivities and viscosities were also time-scaled, although
a few could not be tested due to lack of experiments depending on the
output.
Two unused double-diffusive parameters were re-enabled and integrated
into the source, replacing fixed dimensional constants:
* MAX_SALT_DIFF_SALT_FINGERS
* KV_MOLECULAR
The legacy "zint" diagnostic scaling was implemented, but it is untested
since few experiments still use these tests.
* MOM_set_diffusivity time scaling: subroutines
Some de-scaling of variables passed to other subroutines has been
shifted out of MOM_set_diffusivity.F90 and into the subroutine source
files.
* MOM_set_diffusivity time scaling: Kd_lay
* MOM_set_diffusivity time scaling: Kd_int
* MOM_set_diffusivity time scaling: Kd_BBl + diag
* Time scaling: entrainment_diffusive kd_(lay|int)
Array operations for rescaling kd_lay and kd_int in the
Entrainment_diffusive call were replaced with internal rescaling of the
variables.
* MOM_tidal mixing time scaling: input arguments
Rescaling of input arguments which pass through to MOM_set_diffusivity.
Calls to CVMix cannot be re-scaled, in particular the N2_int vector, so
we pre-scale this into a temporary variable before passing to CVMix.
Note that the CVMix scaling is untested, since we currently do not have
any CVMix tidal tests.
* MOM_set_diffusivity scaling: use kd scaling factor
* Switching from moab to slurm queue scheduler
* skip first tstep and double second tstep
* apply ocean lag to cesm runs only
* User diffusivity time rescaling
* MOM_set_diffusivity comment/whitespace updates
* Avoid division by zero when Ah_time_scale=0
- New viscosity option had potential division by zero
- No answer changes
* Add control of MEKE viscosity scaling to recover old answers
- Commit db98e9e rearranged the order of limiting and resolution scaling
in order to have the MEKE contribution to viscosity be scaled by the
resolution functon. This changed answers in many configurations even
those not using MEKE.
- This adds a flag that defaults to the non-answer changing mode of not
resolution scaling the MEKE contribution but does recover the desired
result when set to Tree.
* fix ocean lag logic for restart runs
* Corrected indentation to fit code style guide
- As pointed out by @Hallberg-NOAA, some indentations were off.
(flint is coming)
* (*) MOM_mixed_layer_restrat [uv]Dml_diag index fix
This patch resolves a few issues related to the uDml_diag and vDml
diagnostics in the MOM_mixed_layer_restrat module. In both cases, the
index loops were not including the starting values (is-1 for uDml and
js-1 for vDml).
In `mixedlayer_restrat_BML`, uDml_diag(is-1,:) was never assigned a
value, since only the uDml(is:ie,:) values were copied to uDml_diag.
This produced volatile answers along these records.
There were no known issues with vDml, since is:ie spans the vector, but
we have nonetheless moved the vector copy inside of the loop.
We have introduced the following changes:
* [uv]Dml_diag is now evaluated in the loop, rather than relying on a
vector copy, since there are no dependencies within the loop.
* [uv]timescale_diag are no longer initialised within the outer loop,
since their values are explicitly set.
* The I and J loops are now fully nested (with whitespace changes)
In `mixedlayer_restrat_general`, we now include the edge cases
uDml(is-1,:) and vDml(:,js-1) in surface velocity estimates.
In both functions, some half-index notation (ij->IJ) was corrected
within the loops.
The following diagnostics may change value:
* udml_restrat
* vdml_restrat
* uml_restrat
* vml_restrat
* Rescaled 20 TKE-related variables
Applied unit rescaling for dimensional consistency analysis to about 20
TKE-related variables on MOM_set_diffusivity and MOM_tidal_mixing. All answers
are bitwise identical and these revised expressions pass the dimensional
consistency tests for vertical height and time units.
* Kd_interface diag unit fix
Time scaling of Kd_interface diagnostic in the MOM diabatic driver was
fixed to include its time scaling.
A couple diffusive unit scalings were also replaced with diffusion
macro.
* MOM_vert_friction tau[xy]_bot unit fix
Bottom friction forces tau[xy]_bot are computed as Rho0 * u * a_u, and
in this case a_u has been scaled to units of Z s-1, rather than m s-1,
so the tau[xy]_bot diagnostic must be rescaled by Z_to_m.
* +Removed hard newlines in get_param calls
Removed hard newlines from numerous get_param and log_param descriptions
throughout the MOM6 codebase, relying instead on the new automatic newline
capability in writeMessageAndDesc. In the case of formatted lists of options,
the hard newlines were retained. This cleans up the code and gives more standard
messages and smaller parameter_doc files, but it changes the comments in every
MOM_parameter_doc and SIS_parameter_doc file. All answers are bitwise identical.
* + Excise of legacy z-interpolated diagnostics
This patch removes the older method for computing diagnostics onto
z-levels (depth space), associated with modules ending in `_zold`.
These diagnostics are being removed for two reasons:
1. The tests are now integrated into the diagnostic mediator, and are now redundant.
2. The diagnostics did not consistently handle masked values correctly,
which was causing issues with checksum calculations.
This patch also removes the diag_to_Z_CS control structure and its
reference in some functions, causing an API change.
The parameter MIN_Z_DIAG_INTERVAL has also been made obsolete, since it
was used to control the output of some depth-space diagnostics.
Two subroutines in MOM_diag_to_Z.F90 (to be removed) were used in
MOM_tracer_flow_control.F90:
* find_overlap
* find_limited_slope
so they have been moved to this file.
We also note that there are nearly identical subroutines in
midas_vertmap.F90, which could be merged at some point.
This patch will break existing tests, since many use MIN_Z_DIAG_INTERVAL
and several `_zold` diagnostics, and will need to be modified before this
can be merged.
* Z_OUTPUT_GRID_FILE parameter obsolescence
* Updating the MOM_CVMix_KPP code
- Remove explicit dependence on optional WAVES control structure to use Langmuir parameterization
- The Langmuir term can be used w/ forms of Langmuir number that shouldn't require the WAVE module to be present.
* Changes to MOM_wave_interface
- Changes to facilitate running MOM6 with Langmuir turbulence but without WAVES control structure.
- Changes to DHH85 spectra calculation to improve comparison w/ LF17 and GOTM simulations.
* Adding ePBL tuning relationships needed to use versions from RH18/RL19 to MOM6
* Undo changing default La_depth_ratio from 0.04 to 0.2
* Moving new ePBL feature Kd_enhance entirely inside of if-block
* Clean-up of previous commit
* Setting ePBL enhance_K variables to parameters
- The values are hard-coded into variable declaration statements and therefore are meant to be declared as parameters.
* Fixing bug to WaveMethod inserted in WaveInterface by previous commit.
* Renaming Elizabeth tracer and sponge to RGC
* Adding changes to RGC
* +(*) u,v tendendy diagnostic fix
Recent testing showed a difference in the time derivative diagnostics
dudt, dvdt, dhdt for symmetric and non-symmetric grids. The issue
appears to be be due to assumed 1-indexing within the
`register_time_deriv` functions of MOM_diagnostics, due to use of
pointers without explicit start index. This was causing an error in
`calculate_derivs`, which assume standard grid indexing.
There was no error in non-symmetric mode, which assumes 1-indexing (when
global indexing is disabled, as default), but was creating off-by-one
errors in symmetric mode for the u and v time derivaties, which
typically start at 0 for the x and y axes, respectively. The dhdt
diagnostic was unaffected, since h-points always use 1-indexing.
This also affected the dKE/dt diagnostic, since they used dudt and dvdt.
We resolve this by introducing a lower bound input argument to
`register_time_deriv` and use this to define the start index for the
various p3d pointers.
This is a minor API change which only affects the dudt, dvdt, and dhdt
diagnostics, and may change the dudt and dvdt symmetric grid diagnostics.
* (*) speed diagnostic fix on nonsymmetric grids
The speed diagnostic currently requires an interpolation of u and v onto
the h-points. On non-symmetric grids, this was causing an error on the
west and south boundaries, since the compute loops in
`extract_surface_state` did not update the boundaries, which are used in
the speed calculation.
We resolve this by extending the loops from G%IscB, which only includes
the boundary on symmetric grids, to isc-1, which always include the
boundaries. This assumes a nonzero halo, and that the halos of u and v
have been updated.
* Added GEOMETRIC as an option for MEKE. Also added biharmonic MEKE viscosity and the option to use
a non-MEKE thickness diffusivity to diffuse MEKE. Split MEKE viscosity options into parts concerning
Ku and Au.
* Ensures that MEKE%Kh_diff is only used when USE_KH_IN_MEKE = True
* Fixes doxygen syntax
* Fixes additional doxygen syntax
* Removes trailing space
* Fixes unregistered diagnostic id
* Edited RGC_tracer.F90
* dKEdt diagnostic nonsymmetric bugfix
This patch fixes a minor error in the d(KE)/dt diagnostic, specifically
any values on the western and southern boundaries (per PE domain).
d(KE)/dt depends on the dudt and dvdt diagnostics. However, the western
and southern boundaries of these two diagnostics were not updated inside
`calculate_derivs`, which caused an error on the western and southern
boundaries of d(KE)/dt since it requires an interpolation of dudt and
dvdt onto h-points.
This was not an issue in symmetric grids, since the west/south values
would have been included in the update, but were not updated in
non-symmetric mode.
We resolve this by extending the loops to update an additional grid
point to the west/south. This slightly increases the computation, and
assumes an updated halo, but this change has restored consistency of
d(KE)/dt in symmetric and nonsymmetric mode.
* Changes logic of an if statement in calc_slope_functions to avoid change in answers
* Deletes code related to GME_Varmix
* Updating RGC codes
* Changes in MEKE
* Adds missing drag term to second part of split timestep.
* Removes limiter preventing MEKE from going negative.
* Reverts back to original method of initial, equilibrium solution.
* Changes the way of counting the work
* (*) Vertical diagnostic remap symmetric bugfix
This patch fixes an error in the vertial diagnostic remapping on
symmetric grids.
The following functions associated with vertical remapping assume
grid-based indexing, but receive arrays with 1-based indexing:
* diag_remap_do_remap
* vertically_reintegrate_diag_field
* vertically_interpolate_diag_field
Since the calculations typically include a mix of grid-based and 1-based
indexing, we retain the grid index loops but introduce 1-based indices,
i1 and j1, when addressing these arrays.
Note that while the symmetric grid u- and v-points are now defined
relative to the global grid (G%isg, G%jsg), we still assume that the
h-points are locally 1-based. This is currently true in MOM6 but may
require modification if we move to global grid indexing.
We have included a very verbose explanation for the gridding issue, in
case it ever gets addressed in the future.
Symmetric grid results now match non-symmetric grid results, and
non-symmetric grid results are unchanged. Model state and dynamics are
unaffected by this change and should be bitwise reproducible.
* Vertical remap index syntax adjustment
Indices on u and v points were capitalised, as is convention.
* Reverts biharmonic constants to the origial definition
* Remove trailing space
* Updating sponge and tracer
* Correcting code comments
* Fixing formatting
* (*) Bugfix: reproducing_sum, bin equals prec
In reproducing_sum, if a bin inside int_sum is exactly equal to prec
(currently 2**46 to some integer power) then this value will not be
transferred up to the next bin, due to the > check rather than a >=
check. See code block below:
562 do i=ni,2,-1 ; if (abs(int_sum(i)) > prec) then
563 num_carry = int(int_sum(i) * I_prec)
564 int_sum(i) = int_sum(i) - num_carry*prec
565 int_sum(i-1) = int_sum(i-1) + num_carry
566 endif ; enddo
This can cause inconsistency in reproducibility across CPUs,
particularly on 1 rank, since the construction of int_sum will not
produce filled bins of size prec.
We fix this by replacing the > check with a >=, so that both overfilled
and exactly full values in a bin are also transferred up to the next
bin. We specifically fix this in regularize_ints, but also in
carry_overflow for the sake of consistency.
This was observed in the mean calculations of a few sparse diagnostics
at isolated timesteps for certain CPU layouts, and caused variation in
the least significant bit (LSB). When the patch was applied, results
were consistent with the 1-rank result.
It seems like a phenomenal coincidence that a bin value managed to
exactly equal prec, which may indicate a further systematic error
somewhere, but that can be treated as a separate issue.
While this change has the potential to modify answers, since we are
modifying the results of `reproducing_sum`, no changes were observed in
the regression tests.
* (*) Z-remap interface mask{Cu,Cv,B}i bugfix
Z-interpolated diagnostics on interfaces were failing to produce
consistent answers when CPU layout was changed. This was tracked down
to an error in the eastern boundaries of their mask fields.
These mask fields are constructed from the T-point mask, maskTi, whose
halo values were not being constructed. By extending the loops from
[ij]sc..[ij]ec to [ij]sc-1 .. [ij]ec+1, we restored the masks and could
reproduce the values.
This will affect the following diagnostics when z-remapped:
* neutral_slope_[xy]
* KHTH_[uv]
and potentially any other diagnostics on the `axesCui` axis.
* (*) Remapped d[uv]dt_dia, [uv]hml diagnostic fixes
The remapped diagnostics d[uv]dt_dia, and uhml were reporting different
results for different CPU layouts. This was due to a dependence on the
east and north halo values of the h and remap_cs%h thicknesses, which
were not being updated for these variables.
The remap_cs%h thickness directly depends on h, and will properly fill
in its halo values, but is only updated when a `diag_update_remap_grids`
function is called. If h has missing halo values prior to this call,
then remap_cs%h will also be missing.
In most of the code, h has been properly updated, and remap_cs%h only
needs an update when h changes, which is restricted to a few functions.
This appears to not be an issue for most diagnostics, with the ones
above being an exception.
We resolve this by applying a halo update to h prior to any
`diag_update_remap_grids` calls which are needed before the diagnostic
`post_data` call. We only apply this halo update if the diagnostic has
been registered.
Ideally we would only apply the halo update when the remapped halo has
been registered, but it is currently not possible to query this
information, and ought to be done in a future release.
It is likely that there are more diagnostics with this issue, so
consider this a first step.
* (*) Horizontal remapping re-enabled, bugfixes
This patch re-enabled the `post_xy_average` calls in `post_data_3d_low`,
which are used to generate and write the horizontally averaged
diagnostics.
We have also fixed a number of issues in the existing methods, and we
have changed how masked values are handled.
The first major change is that average of a field is now computed on
the native grid points, rather than interpolation onto h-points. This
is due to a dependence on halo values of `h` which would be potentially
undefined in symmetric grids. In most cases the calculation will be
mathematically equivalent, although there will be bitwise differences.
Non-periodic boundaries may also yield different edge values which will
modify the average.
The second major change is the introduction of a reproducible sum for
the calculation of the total volume and field at each vertical level,
which will introduce bitwise changes but is now reproducible across
layouts.
The final change is the removal of explicit `missing_value` checks,
which have been replaced with a new column mask. Masked values are now
explicitly set to zero. This simplifies the calculation of any averages
or checksums, and is also consistent with the overall removal of
explicit missing_value usage in MOM.
These changes have removed depenendies on unassigned values in symmetric
grids, and produce results which are now reproducible across layouts and
for symmetric and nonsymmetric grids. Diagnostic values are expected to
change, but the model state should be bitwise reproducible.
* Commenting salt_flux adjustment in net_FW
* Adds MEKE_VISCOSITY_COEFF_AU and MEKE_VISCOSITY_COEFF_KU into the MEKE documentation
* Obsolete parameter MEKE_VISCOSITY_COEFF
* Add MEKE_GEOMETRIC_EPSILON
Delete hard-coded epsilon and added a new user parameter to specify
this constant.
* move OMP calls below GME and above the K-loop
* Fix hint syntax in MOM_obsolete_param
* Commenting out salt_flux adjustment in net_FW in NUOPC cap
* changes needed for mom6 validation
* (+) Enable checksums for all diagnostics
This patch consists of three major changes.
1. We have modified the `DIAG_AS_CHKSUM` feature as a "checksum" mode,
where all diagnostics registered via these functions:
* `register_diag_field`
* `register_static_field`
* `register_scalar_field`
are caught before they are formally registered with FMS. We make two
changes:
a. We do not formally register the diagnostic with FMS, and instead
assign them dummy FMS ids, used to enable checksums.
b. We enable all diagnostics, regardless of the contents of
`diag_table`
Whenever data is formally written to FMS via `post_data`, we instead
calculate the checksum data for the fields via the `MOM_checksum`
functions and log the output to the `chksum_diag` file.
2. To more fully support these checksums, we have made several changes
to the checksum module.
a. We have introduced two new checksum functions, `chksum0` for
scalars and `zchksum` for 1-d arrays (typically vertical columns)
b. Checksum functions now support an optional IO unit, which defaults
to stderr (now formally defined via `iso_fortran_env`).
c. `subStats` functions now return their mean, min, and max values,
rather than logging the results directly.
3. The `chksum_general` interface and its functions were removed from
`MOM_checksums`, and we use the existing gridded checksums (hcksum,
etc.)
Due to removal of the `chksum_general` functions and addition of IO unit
inputs to the other functions, this patch is an API change to MOM. But
existing checksum functions should be unaffected.
* Do not attach cell methods in diag chksum mode
The diag mediator register functions were using the `attach_cell_method`
and `diag_field_add_attribute` functions to attach attributes to the FMS
diagnostics. This required referencing of internal FMS field tables,
which are not created when DIAG_AS_CHKSUM is enabled.
While this did not raise any errors in the Baltic tests (for unclear
reasons), this was causing issues in nearly every other test,
predominantly when not in debug mode.
We now only conditionally call these functions when `diag_as_chksum` is
disabled.
* Set default AH_VEL_SCALE back to zero
* Repro sum increment_ints_faster overflow check
While reproducing_sum was able to detect an overflow for values
exceeding its maximum allow value, `prec**N_max * (2**63 - 1)`, this
check occurs after the first bin decomposition, which requires an
integer conversion, `int(r * (1./prec_max), 8)`.
When floating point exceptions (FPE) were disabled, this would often
produce a large negative integer, and the overflow would have been
caught later and handled as prescribed by the user. But when FPEs are
enabled, there is a fatal error and the model aborts. (NOTE: The FPE
handler in GCC tags this as "invalid", rather than "overflow")
This patch adds a check inside `increment_ints_faster`, so that the
`int()` operation is avoided in the case of overflow.
* Remove thickness_diffuse_CS from horizontal_viscosity
* Issue FATAL error if SPLIT=False and USE_GME=True
* Removes Barotropic and thickness_diffuse from MOM_dynamics_unsplit*
* Changes Barotropic to BT in horizontal_viscosity
* Passes optional arg. (BT=CS%barotropic_CSp) in calls to horizontal_viscosity
* Flux allocation check in diagnostics
Currently the model will attempt to compute any enabled surface flux
diagnosics, even if the model has not been configured to support this
flux. This would raise memory access errors if the flux had not been
allocated.
We now conditionally check if a particular surface flux has been
allocated before adding it to any cumulative flux diagnostic.
In the future, we should probably implement some framework around all
this and associate the diagnostic with the creation of the flux.
* Disable diabatic KE flux for adiabatic runs
This patch disables the `KE_dia` diagnostic if the flow has been
configured for adiabatic flow.
* Conditional thermodynamic diags; Flux diag bugfix
Several thermodynamic diabatic diagnostics in the legacy_diabatic update
could be registered and computed when the thermodynamics was disabled,
so these are now only conditionally checked and registered when
`ENABLE_THERMODYNAMICS` is true.
We also fixed a bug in the flux diagnostic registration by splitting the
logical association and memory access into two separate stages, since it
was raising errors in builds with stricter memory access.
* Removes trailing space
* Conditionally enable MEKE Kh diagnostics
It is possible to enable the KhMEKE_u and KhMEKE_v diagnostics when
their values have not been updated, which will either write nonsense
values or raise errors/exceptions if checksums are enabled.
This patch introduces a general flag to check if lateral
(non-biharmonic) MEKE fluxes have been turned on, and only registers the
KhMEKE_[uv] variables if they are present.
* fixes to get taux and tauy roundoff between mct and nuopc
* Fix for static compiles
- Arrays CS%Laplac3_const_u, CS%Laplac3_const_v, CS%KH_u_QG, CS%KH_v_QG
are declared use the ALLOCABLE_ macro and so must be allocated with the
ALLOC_ macro. Otherwise in static mode we are allocating a non-allocatable
array.
* Correct declaration of h in calc_QG_Leith_viscosity()
- The shape of dummy argument h was wrong. This would only change answers in
non-symmetric mode so we lucked out!
* Split long comments in RGC_tracer.F90
RGC_tracer.F90 previously had some very long comments at the end of some
lines. These have now been split onto multiple lines to respect the MOM6
standards for line-length. All answers are bitwise identical.
* Split excessively long lines in 2 files
Split excessively long lines and corrected the syntax for unit documentation
in MOM_lateral_mixing_coeffs.F90 and MOM_thickness_diffuse.F90. All answers are
bitwise identical.
* (*)Multiply fmax by US%s_to_T in MOM_hor_visc.F90
Added a dimensional scaling factor for fmax in MOM_hor_visc.F90 that was
dropped at some point in the merging of the dev/ncar code into dev/gfdl. All
answers are bitwise identical and now pass the dimensional scaling test.
* Travis tests for lines exceeding 120 characters
Added the 120 character line limit into the travis testing script.
* removed trailing whitespace
* Deletes comments
* changes to the interfaces to bring them more in line with dev/ncar latest
* more updates to have caps consistent with dev/ncar
* changes to get MOM_ocean_model.F90 closer to nuopc_driver
* removed trailing whitespace
* changes to have minimal differences between nuopc and mct
* more changes to have minimal differences between nuopc and mct
* Rename modules in MCT
* Rename modules in NUOPC
* Rename modules for MCT
* Rename modules for NUOPC
* Capitalize MOM inside MCT modules
* Deletes contribution from kv%slow when computing a_cpl
* Make sure that MEKE%GME_snk is only used when the array is allocated
* Deletes all code related to Jansen15_drag
* Change default way to computing FrictWork in MEKE
In the calculation of MEKE%mom_src, made the old method (using FrictWork) the default
and made the new calculation of FrictWork only available when using GME.- Loading branch information
1 parent
29b1fa7
commit 8fc61ac