Adaptive implicit vertical advection by simone-silvestri · Pull Request #5604 · CliMA/Oceananigans.jl

simone-silvestri · 2026-05-18T20:15:31Z

this PR implements a new advection configuration, an AdaptiveImplicitVerticalAdvection
~~that wraps around an advection scheme (scheme = AdaptiveImplicitVerticalAdvection(explicit_scheme = WENO(), cfl = 0.7) for example)~~
constructed by adding a time_discretization kwarg to the advection schemes (by default ExplicitTimeDiscretization()) enabled (for example) by passing AdaptiveImplicitTimeDiscretization(; maximum_explicit_cfl)

WENO(; time_discretization = AdaptiveImplicitTimeDiscretization(cfl = 0.5))

to treat implicitly part of vertical advection.

In particular it split the advection in an explicit part that advects with the scheme using a vertical velocity we = w * max(cfl / scheme.cfl) and an implicit part that advects with an upwind first order scheme using a velocity of wi = w - we.

We reuse the implicit tridiagonal solver to advect implicitly. A caveat (for the moment), is that the vertical velocity itself does not participate in this split. An example of a comparison between a reference, a fully explicit scheme and AIVA can be found in the validation/implicit_vertical_advection folder. It advects a gaussian, these are the results:

julia> include("validation/implicit_vertical_advection/comparison_implicit_explicit_advection.jl"); fig
[ Info: Δz_min = 0.00023788883447073417, w₀ = 1.0
[ Info: AIVA Δt = 0.0023788883447073417 (CFL ≈ 10.0), 168 steps
[ Info: Ref  Δτ = 0.00011894441723536708 (CFL ≈ 0.5), 3363 steps

=== AIVA stability + correctness validation ===
grid: Nz=128, Δz_min=2.379e-04, Δz_max=2.355e-02
AIVA Δt = 2.3789e-03 (CFL ≈ 10.00), reference Δτ = 1.1894e-04 (CFL ≈ 0.50)
Final time = 3.9965e-01 (168 AIVA steps = 3360 reference steps)

--- Stability ---
Explicit UpwindBiased(order=5) at AIVA Δt: BLEW UP at step 12 (t = 2.855e-02)
AIVA at large Δt: completed 168 steps, max|c| = 9.965e-01 (initial 9.998e-01)

--- Correctness (vs analytical translation) ---
Analytical: max|c| = 9.988e-01, mass = 1.4122e-01, centroid = -0.1510 (target -0.1503)
Reference:  max|c| = 9.966e-01, mass = 1.4124e-01, centroid = -0.1510, L²-rel = 2.008e-03, L∞ = 2.209e-03
AIVA:       max|c| = 9.965e-01, mass = 1.5073e-01, centroid = -0.1447, L²-rel = 9.845e-02, L∞ = 1.267e-01
AIVA vs reference: L² = 3.104e-02

--- Mass conservation (∫c dz) ---
Initial:        AIVA = 1.418014e-01, reference = 1.418014e-01
Final:          AIVA = 1.507308e-01, reference = 1.412434e-01
Drift (final/initial − 1): AIVA = +6.297e-02, reference = -3.935e-03
Min/max during AIVA run: 1.418014e-01 / 1.507308e-01

and the figure it produces:

…vertical-advection

glwagner · 2026-05-19T03:30:43Z

Let's avoid a wrapper if at all possible (instead you can add a property to the model). Using a wrapper makes the user interface hard; we can't make it a default option easily.

Generally wrappers should be a method of last resort, only if they are ABSOLUTELY the worst of many other evil options, in my opinion.

simone-silvestri · 2026-05-19T06:33:40Z

I would avoid a property in the model, it's an advection scheme property. Another option is to add an ExplicitTimeDiscretization or VerticallyImplicitTimeDiscretization directly to the advection scheme, even if the implicit component would be only first order upwind.

…tion scheme

…CliMA/Oceananigans.jl into ss/adaptive-implicit-vertical-advection

glwagner · 2026-05-19T14:36:47Z

I would avoid a property in the model, it's an advection scheme property. Another option is to add an ExplicitTimeDiscretization or VerticallyImplicitTimeDiscretization directly to the advection scheme, even if the implicit component would be only first order upwind.

It's really a model property, because it has to be supported by changing the whole algorithm (ie this cannot be implemented for models that do not have tridiagonal solvers already)

simone-silvestri · 2026-05-19T14:44:37Z

As a model property leads to the usual design problem of having

HydrostaticFreeSurfaceModel(; advection = nothing, implicit_advection = true)

which creates conflicting keyword arguments. I like the wrapper approach which I feel is also quite easy and maintainable but this is a personal preference. The implicit solver is a timestepper property, not a model property, so if we want to be completely clean in the design then the time discretization kwarg should belong to the timestepper. However this requires a redesign of the closures which follow the design I am proposing here for the advection.

simone-silvestri · 2026-05-19T14:46:36Z

So I would be ok in mimicking the closure's approach where the "time discretization" belongs to the advection scheme.

glwagner · 2026-05-19T14:49:36Z

As a model property leads to the usual design problem of having
HydrostaticFreeSurfaceModel(; advection = nothing, implicit_advection = true)
which creates conflicting keyword arguments. I like the wrapper approach which I feel is also quite easy and maintainable but this is a personal preference. The implicit solver is a timestepper property, not a model property, so if we want to be completely clean in the design then the time discretization kwarg should belong to the timestepper. However this requires a redesign of the closures which follow the design I am proposing here for the advection.

The point is that this is FAR lesser evil than the wrapper .

glwagner · 2026-05-19T14:50:16Z

Wrappers have created a lot of issues in the past (compared to a simple validation check) and also impact compile time

simone-silvestri · 2026-05-19T14:50:55Z

Otherwise, I am also ok taking the more principled route of redesigning the timestepper to include the vertical time discretization but will require a couple of more PRs

glwagner · 2026-05-19T14:51:56Z

I would avoid a property in the model, it's an advection scheme property. Another option is to add an ExplicitTimeDiscretization or VerticallyImplicitTimeDiscretization directly to the advection scheme, even if the implicit component would be only first order upwind.

I'm ok with embedding in the advection scheme.

Note, validation is required for ANY approach, so the problem of "invalid combinations" is not an actual differentiator.

glwagner · 2026-05-19T14:52:45Z

Otherwise, I am also ok taking the more principled route of redesigning the timestepper to include the vertical time discretization but will require a couple of more PRs

I don't think you need to change the time-stepper here, or at least our discussion does not bear on whether or not to change the timestepper.

…CliMA/Oceananigans.jl into ss/adaptive-implicit-vertical-advection

github-actions · 2026-05-20T09:52:21Z

Benchmark Comparison

Benchmark Comparison: PR vs Main

Benchmark	PR (pts/s)	Main (pts/s)	Change
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	58051491.205	57800437.820	+0.4%
EarthOcean_tripolar_180x90x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	26427421.952	20318510.690	+30.1%
EarthOcean_tripolar_720x360x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	60263892.743	60263840.425	+0%
EarthOcean_tripolar_360x180x50_F32_WENOVectorInvariantDefault_WENO7_CATKE_2tr	73317686.595	73336721.808	-0%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_nothing_2tr	101080235.96	101189344.41	-0.1%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+Biharmonic_2tr	40931141.269	40882213.754	+0.1%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+GM+Biharmonic_2tr	12490825.024	12495084.490	-0%
EarthOcean_tripolar_360x180x50_F64_nothing_nothing_CATKE_2tr	85682391.282	87057497.431	-1.6%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant5_WENO5_CATKE_2tr	65251253.789	65230590.323	+0%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant9_WENO9_CATKE_2tr	44540466.394	43184925.771	+3.1%
EarthOcean_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	46219177.112	46219034.375	+0%
EarthOcean_immersed_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	52969904.663	52912412.764	+0.1%
EarthOcean_tripolar_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	51784968.362	51730971.217	+0.1%
EarthOcean_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	56417990.502	56400090.285	+0%
EarthOcean_immersed_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	59851618.925	59840902.466	+0%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_3tr	54164230.858	54131327.146	+0.1%

NSYS Kernel Profiling

EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr

Kernel	Median (ms)	Main (ms)	Change	Instances	Avg (ms)	Min (ms)	Max (ms)
`gpu_compute_hydrostatic_free_surface_Gu_`	2.454	2.454	+0.0%	318	2.461	2.447	2.676
`gpu_compute_hydrostatic_free_surface_Gv_`	2.171	2.170	+0.0%	318	2.177	2.164	2.366
`gpu__rk_substep_turbulent_kinetic_energy_`	1.990	1.990	+0.0%	315	1.994	1.986	2.172
`gpu_compute_CATKE_closure_fields_`	1.590	1.590	+0.0%	318	1.594	1.582	1.724
`gpu_compute_hydrostatic_free_surface_Gc_`	0.984	0.979	+0.6%	315	0.987	0.980	1.079
`gpu_compute_hydrostatic_free_surface_Gc_`	0.979	0.979	+0.1%	315	0.982	0.976	1.073
`gpu_compute_hydrostatic_free_surface_Gc_`	0.978	0.979	-0.0%	315	0.981	0.974	1.069
`gpu__compute_w_from_continuity_`	0.340	0.339	+0.5%	633	0.340	0.335	0.350
`gpu_compute_TKE_diffusivity_`	0.641	0.641	-0.0%	315	0.643	0.635	0.691
`gpu__compute_split_explicit_transport_velocities_`	0.487	0.487	+0.0%	315	0.487	0.483	0.492

simone-silvestri · 2026-05-20T09:59:52Z

@glwagner this PR should be ready to review. (I have also checked there is no performance regression for including the time discretization in the advection schemes).

github-actions · 2026-05-20T14:41:55Z

Benchmark Comparison

Benchmark Comparison: PR vs Main

Benchmark	PR (pts/s)	Main (pts/s)	Change
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	57775227.818	55973325.271	+3.2%
EarthOcean_tripolar_180x90x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	25546645.407	25107471.563	+1.7%
EarthOcean_tripolar_720x360x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	60238189.470	60262283.425	-0%
EarthOcean_tripolar_360x180x50_F32_WENOVectorInvariantDefault_WENO7_CATKE_2tr	73198783.587	72789115.797	+0.6%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_nothing_2tr	101262897.50	100976485.78	+0.3%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+Biharmonic_2tr	40848828.092	40777398.431	+0.2%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+GM+Biharmonic_2tr	12497678.867	12496401.683	+0%
EarthOcean_tripolar_360x180x50_F64_nothing_nothing_CATKE_2tr	87074966.783	86460791.435	+0.7%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant5_WENO5_CATKE_2tr	65186814.804	65122910.231	+0.1%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant9_WENO9_CATKE_2tr	44525282.072	43146090.679	+3.2%
EarthOcean_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	46220036.538	46210335.682	+0%
EarthOcean_immersed_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	52089518.662	53062667.053	-1.8%
EarthOcean_tripolar_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	51800471.035	51781174.539	+0%
EarthOcean_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	56433668.370	56409345.874	+0%
EarthOcean_immersed_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	59847051.992	59589003.848	+0.4%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_3tr	54158988.286	53754758.366	+0.8%

NSYS Kernel Profiling

EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr

Kernel	Median (ms)	Main (ms)	Change	Instances	Avg (ms)	Min (ms)	Max (ms)
`gpu_compute_hydrostatic_free_surface_Gu_`	2.454	2.454	+0.0%	318	2.463	2.449	2.679
`gpu_compute_hydrostatic_free_surface_Gv_`	2.170	2.171	-0.0%	318	2.178	2.163	2.364
`gpu__rk_substep_turbulent_kinetic_energy_`	1.989	1.990	-0.0%	315	1.996	1.985	2.171
`gpu_compute_CATKE_closure_fields_`	1.591	1.590	+0.0%	318	1.595	1.584	1.726
`gpu_compute_hydrostatic_free_surface_Gc_`	0.985	0.979	+0.6%	315	0.988	0.980	1.077
`gpu_compute_hydrostatic_free_surface_Gc_`	0.980	0.979	+0.1%	315	0.983	0.976	1.072
`gpu_compute_hydrostatic_free_surface_Gc_`	0.979	0.979	-0.0%	315	0.982	0.974	1.071
`gpu__compute_w_from_continuity_`	0.338	0.338	-0.2%	633	0.338	0.333	0.346
`gpu_compute_TKE_diffusivity_`	0.641	0.641	+0.0%	315	0.643	0.634	0.690
`gpu__compute_split_explicit_transport_velocities_`	0.487	0.487	+0.0%	315	0.487	0.483	0.491

glwagner · 2026-05-20T20:27:15Z

+
+A fully-explicit time-discretization of a `TurbulenceClosure`.
+"""
+struct ExplicitTimeDiscretization <: AbstractTimeDiscretization end


There is also an ETD in TurbulenceClosures? Should this go somewhere more general?

This is the same one that is in TurbulenceClosures, I moved it up to Advection.

It's weird that "VerticallyImplicitTimeDiscretization" (used for closures) is defined in Advection

I suggest putting this into TimeSteppers module, which is loaded before Advection:

Oceananigans.jl/src/Oceananigans.jl

Lines 233 to 234 in c2c4d1f

include("TimeSteppers/TimeSteppers.jl")

include("Advection/Advection.jl")

sounds good to me!

glwagner · 2026-05-20T20:29:26Z

 @inline bias(u::Number) = ifelse(u > 0, LeftBias, RightBias)

-@inline function advective_momentum_flux_Uu(i, j, k, grid, scheme::UpwindScheme, U, u)
+@inline function advective_momentum_flux_Uu(i, j, k, grid, scheme::UpwindScheme, ::ETD, U, u)


I'm slightly confused by this. The time discretization is also inside scheme, right?

yeah but this is to dispatch, so we unwrap the time_discretization(scheme) and these fluxes are the "Explicit" ones.

another strategy is to define ExplicitUpwindScheme right? This seems a bit simpler. Curious what you think.

we could also do like the closures do and just reroute advective_momentum_flux_Uu(i, j, k, grid, scheme, ::ETD, U, u) to advective_momentum_flux_Uu(i, j, k, grid, scheme, U, u) and keep the same code in this file. This would allow us to avoid aliases

Oh I see, is that the benefit of using the extra argument? I don't completely grasp the totality of this design

the idea here is that the _advective_flux (once the order is decided) reroute to
advective_flux(..., td, ...) so that we select what happens, then if it is adaptive vertically implicit we have this extra step of computing the "explicit" vertical velocity otherwise business as usual. It is true that ETD is business as usual so we can remove the extra argument in case of an ETD which leaves the previous functions untouched. This would basically be the same design as

Oceananigans.jl/src/TurbulenceClosures/implicit_explicit_time_discretization.jl

Lines 37 to 45 in 201d431

@inline diffusive_flux_x(i, j, k, grid, ::ATD, args...) = diffusive_flux_x(i, j, k, grid, args...)

@inline diffusive_flux_y(i, j, k, grid, ::ATD, args...) = diffusive_flux_y(i, j, k, grid, args...)

@inline diffusive_flux_z(i, j, k, grid, ::ATD, args...) = diffusive_flux_z(i, j, k, grid, args...)

@inline viscous_flux_ux(i, j, k, grid, ::ATD, args...) = viscous_flux_ux(i, j, k, grid, args...)

@inline viscous_flux_uy(i, j, k, grid, ::ATD, args...) = viscous_flux_uy(i, j, k, grid, args...)

@inline viscous_flux_uz(i, j, k, grid, ::ATD, args...) = viscous_flux_uz(i, j, k, grid, args...)

@inline viscous_flux_vx(i, j, k, grid, ::ATD, args...) = viscous_flux_vx(i, j, k, grid, args...)

In the current state the difference is that we are not removing the extra argument.

So in the current state, if we want to define a new advecitve flux, we define it for the "ETD" case with explicit dispatch -- is that right?

yes that would be the route. Otherwise if we add the rerouting that removes the extra argument when we do not need to explicitly specify ETD

…CliMA/Oceananigans.jl into ss/adaptive-implicit-vertical-advection

github-actions · 2026-05-21T11:54:03Z

Benchmark Comparison

Benchmark Comparison: PR vs Main

Benchmark	PR (pts/s)	Main (pts/s)	Change
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	57788393.676	57834951.544	-0.1%
EarthOcean_tripolar_180x90x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	25997835.956	22999766.598	+13%
EarthOcean_tripolar_720x360x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	60269076.764	60263543.812	+0%
EarthOcean_tripolar_360x180x50_F32_WENOVectorInvariantDefault_WENO7_CATKE_2tr	66600567.398	69869673.831	-4.7%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_nothing_2tr	100943014.09	101215811.79	-0.3%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+Biharmonic_2tr	40903828.745	40926178.673	-0.1%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+GM+Biharmonic_2tr	12503258.019	12508250.639	-0%
EarthOcean_tripolar_360x180x50_F64_nothing_nothing_CATKE_2tr	87092770.150	87105374.069	-0%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant5_WENO5_CATKE_2tr	65243439.340	65213050.291	+0%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant9_WENO9_CATKE_2tr	44541198.724	43191923.795	+3.1%
EarthOcean_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	46208915.872	46208825.236	+0%
EarthOcean_immersed_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	53258924.023	53254949.131	+0%
EarthOcean_tripolar_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	51792956.341	51806339.870	-0%
EarthOcean_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	56413517.017	56397535.747	+0%
EarthOcean_immersed_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	59819489.406	59704921.670	+0.2%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_3tr	54157874.491	54161916.752	-0%

NSYS Kernel Profiling

EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr

Kernel	Median (ms)	Main (ms)	Change	Instances	Avg (ms)	Min (ms)	Max (ms)
`gpu_compute_hydrostatic_free_surface_Gu_`	2.455	2.456	-0.0%	318	2.466	2.448	2.677
`gpu_compute_hydrostatic_free_surface_Gv_`	2.172	2.172	-0.0%	318	2.180	2.164	2.369
`gpu__rk_substep_turbulent_kinetic_energy_`	1.989	1.989	-0.0%	315	1.995	1.983	2.175
`gpu_compute_CATKE_closure_fields_`	1.588	1.589	-0.0%	318	1.594	1.581	1.721
`gpu_compute_hydrostatic_free_surface_Gc_`	0.983	0.978	+0.5%	315	0.987	0.979	1.076
`gpu_compute_hydrostatic_free_surface_Gc_`	0.979	0.978	+0.1%	315	0.983	0.976	1.072
`gpu_compute_hydrostatic_free_surface_Gc_`	0.978	0.978	+0.0%	315	0.982	0.974	1.071
`gpu__compute_w_from_continuity_`	0.339	0.338	+0.1%	633	0.339	0.334	0.348
`gpu_compute_TKE_diffusivity_`	0.642	0.643	-0.0%	315	0.644	0.637	0.693
`gpu__compute_split_explicit_transport_velocities_`	0.485	0.485	-0.0%	315	0.485	0.480	0.489

github-actions · 2026-05-21T19:11:10Z

Benchmark Comparison

Benchmark Comparison: PR vs Main

Benchmark	PR (pts/s)	Main (pts/s)	Change
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	57838998.969	57865538.475	-0%
EarthOcean_tripolar_180x90x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	26442867.822	26266912.756	+0.7%
EarthOcean_tripolar_720x360x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	60265556.917	60206410.721	+0.1%
EarthOcean_tripolar_360x180x50_F32_WENOVectorInvariantDefault_WENO7_CATKE_2tr	73313802.666	73321082.962	-0%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_nothing_2tr	100585614.25	100606407.66	-0%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+Biharmonic_2tr	40921885.839	40758296.041	+0.4%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE+GM+Biharmonic_2tr	12502891.061	12503582.715	-0%
EarthOcean_tripolar_360x180x50_F64_nothing_nothing_CATKE_2tr	87184039.839	87239560.941	-0.1%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant5_WENO5_CATKE_2tr	65138330.232	65237732.831	-0.2%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariant9_WENO9_CATKE_2tr	44532335.214	43132129.368	+3.2%
EarthOcean_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	46239315.022	45476339.071	+1.7%
EarthOcean_immersed_lat_lon_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	53262159.158	53266403.741	-0%
EarthOcean_tripolar_zstar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	51816303.658	51797475.416	+0%
EarthOcean_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	56405053.319	56426105.532	-0%
EarthOcean_immersed_lat_lon_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr	59794489.307	59826194.413	-0.1%
EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_3tr	54084047.204	54085899.505	-0%

NSYS Kernel Profiling

EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr

Kernel	Median (ms)	Main (ms)	Change	Instances	Avg (ms)	Min (ms)	Max (ms)
`gpu_compute_hydrostatic_free_surface_Gu_`	2.455	2.454	+0.0%	318	2.463	2.447	2.680
`gpu_compute_hydrostatic_free_surface_Gv_`	2.170	2.170	+0.0%	318	2.177	2.164	2.366
`gpu__rk_substep_turbulent_kinetic_energy_`	1.989	1.989	-0.0%	315	1.995	1.985	2.174
`gpu_compute_CATKE_closure_fields_`	1.590	1.590	+0.0%	318	1.595	1.582	1.725
`gpu_compute_hydrostatic_free_surface_Gc_`	0.984	0.978	+0.6%	315	0.987	0.979	1.078
`gpu_compute_hydrostatic_free_surface_Gc_`	0.980	0.978	+0.1%	315	0.982	0.976	1.073
`gpu_compute_hydrostatic_free_surface_Gc_`	0.979	0.978	+0.0%	315	0.981	0.975	1.073
`gpu__compute_w_from_continuity_`	0.339	0.339	-0.2%	633	0.339	0.334	0.346
`gpu_compute_TKE_diffusivity_`	0.641	0.641	-0.0%	315	0.643	0.636	0.691
`gpu__compute_split_explicit_transport_velocities_`	0.487	0.487	+0.0%	315	0.487	0.483	0.491

simone-silvestri added 7 commits March 4, 2026 11:05

adaptive implicit vertical advection

7180851

Merge remote-tracking branch 'origin/main' into ss/adaptive-implicit-…

06d9cf4

…vertical-advection

small changes

4d0b1d1

Merge remote-tracking branch 'origin/main' into ss/adaptive-implicit-…

f5d2a49

…vertical-advection

adaptive implicit vertical advection

cda8e16

add a validation

c43e33d

rename test

85c1bf3

simone-silvestri commented May 18, 2026

View reviewed changes

Comment thread src/Advection/adaptive_implicit_vertical_advection.jl Outdated

Apply suggestion from @simone-silvestri

32119fc

simone-silvestri requested review from glwagner and taimoorsohail May 18, 2026 20:27

navidcy added the numerics 🧮 So things don't blow up and boil the lobsters alive label May 19, 2026

simone-silvestri added 4 commits May 19, 2026 09:06

make AdaptiveVerticallyImplicitDiscretization a property of the advec…

9daee3f

…tion scheme

Merge branch 'ss/adaptive-implicit-vertical-advection' of github.com:…

3dba587

…CliMA/Oceananigans.jl into ss/adaptive-implicit-vertical-advection

import correct functions

c8a5983

add the new file

188544f

Merge branch 'main' into ss/adaptive-implicit-vertical-advection

79dbf90

simone-silvestri added 3 commits May 19, 2026 16:56

vertical_discretization -> time_discretization

f13b013

Merge branch 'ss/adaptive-implicit-vertical-advection' of github.com:…

4815664

…CliMA/Oceananigans.jl into ss/adaptive-implicit-vertical-advection

fix everything

6aea1a0

simone-silvestri added 3 commits May 20, 2026 09:33

Merge branch 'main' into ss/adaptive-implicit-vertical-advection

a8feb46

fix mmissing time_discretization

f3a2edc

fix whitespaces

4d1f8c1

Merge branch 'main' into ss/adaptive-implicit-vertical-advection

72ce7ce

glwagner reviewed May 20, 2026

View reviewed changes

Comment thread src/Advection/time_discretization.jl Outdated

glwagner reviewed May 20, 2026

View reviewed changes

Comment thread src/Advection/time_discretization.jl Outdated

glwagner reviewed May 20, 2026

View reviewed changes

Comment thread src/Models/HydrostaticFreeSurfaceModels/update_hydrostatic_free_surface_model_state.jl Outdated

glwagner reviewed May 20, 2026

View reviewed changes

Comment thread src/TimeSteppers/runge_kutta_3.jl Outdated

simone-silvestri added 6 commits May 21, 2026 12:01

fix advection

26b1e83

Merge branch 'ss/adaptive-implicit-vertical-advection' of github.com:…

e3a2610

…CliMA/Oceananigans.jl into ss/adaptive-implicit-vertical-advection

remove whitespace

909cfe8

more fixes

4c2f525

add some tests

f6dea1f

typo

8abc504

simone-silvestri added 7 commits May 21, 2026 14:03

correct VectorInvariant

9fa06cc

make sure we can run the tests

4ef2283

correct vector invariant

e7999ba

move discretizations to TImeSteppers

c05a2b1

fix the tests

2bf6f97

fix some more bugs

57d93c9

remove useless comment

d1daf5d

glwagner approved these changes May 21, 2026

View reviewed changes

rout horizontal advection through ETD

c8259e7

	include("TimeSteppers/TimeSteppers.jl")
	include("Advection/Advection.jl")

	@inline diffusive_flux_x(i, j, k, grid, ::ATD, args...) = diffusive_flux_x(i, j, k, grid, args...)
	@inline diffusive_flux_y(i, j, k, grid, ::ATD, args...) = diffusive_flux_y(i, j, k, grid, args...)
	@inline diffusive_flux_z(i, j, k, grid, ::ATD, args...) = diffusive_flux_z(i, j, k, grid, args...)

	@inline viscous_flux_ux(i, j, k, grid, ::ATD, args...) = viscous_flux_ux(i, j, k, grid, args...)
	@inline viscous_flux_uy(i, j, k, grid, ::ATD, args...) = viscous_flux_uy(i, j, k, grid, args...)
	@inline viscous_flux_uz(i, j, k, grid, ::ATD, args...) = viscous_flux_uz(i, j, k, grid, args...)

	@inline viscous_flux_vx(i, j, k, grid, ::ATD, args...) = viscous_flux_vx(i, j, k, grid, args...)

Conversation

simone-silvestri commented May 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

glwagner commented May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

simone-silvestri commented May 19, 2026

Uh oh!

glwagner commented May 19, 2026

Uh oh!

simone-silvestri commented May 19, 2026

Uh oh!

simone-silvestri commented May 19, 2026

Uh oh!

glwagner commented May 19, 2026

Uh oh!

glwagner commented May 19, 2026

Uh oh!

simone-silvestri commented May 19, 2026

Uh oh!

glwagner commented May 19, 2026

Uh oh!

glwagner commented May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented May 20, 2026

Benchmark Comparison

Benchmark Comparison: PR vs Main

NSYS Kernel Profiling

EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr

Uh oh!

simone-silvestri commented May 20, 2026

Uh oh!

github-actions Bot commented May 20, 2026

Benchmark Comparison

Benchmark Comparison: PR vs Main

NSYS Kernel Profiling

EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

simone-silvestri May 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented May 21, 2026

Benchmark Comparison

Benchmark Comparison: PR vs Main

NSYS Kernel Profiling

EarthOcean_tripolar_360x180x50_F64_WENOVectorInvariantDefault_WENO7_CATKE_2tr

Uh oh!

github-actions Bot commented May 21, 2026

simone-silvestri commented May 18, 2026 •

edited

Loading

glwagner commented May 19, 2026 •

edited

Loading

glwagner commented May 19, 2026 •

edited

Loading

simone-silvestri May 21, 2026 •

edited

Loading