Closed FredericWantiez closed 2 months ago
milankl
commented
2 months ago Weird, this doesn't look like the error originates from within SpeedyWeather.jl though. Could you check whether
BarotropicModel, ShallowWaterModel, PrimitiveDryModel) makes a difference?SpectralTransform, could you try creating one individually SpectralTransform(spectral_grid)?
FredericWantiez
commented
2 months ago BarotropicModel and ShallowWaterModel work, same issue with PrimitiveDryModelSpectralTransform when running SpectralTransform(SpectralGrid(trunc=1)) or similarIt's hard to debug, but seems like julia gets lost in the type inference, this doesn't even return:
julia +1.9 --project -e "using SpeedyWeather; PrimitiveDryModel()"When creating a PrimitiveDryModel the following steps are executed, can you copy & paste them in?
spectral_grid = SpectralGrid()
geometry = Geometry(spectral_grid)
# DYNAMICS
dynamics = true
planet = Earth(spectral_grid)
atmosphere = EarthAtmosphere(spectral_grid)
coriolis = Coriolis(spectral_grid)
geopotential = Geopotential(spectral_grid)
adiabatic_conversion = AdiabaticConversion(spectral_grid)
particle_advection = NoParticleAdvection()
initial_conditions = InitialConditions(PrimitiveDry)
# BOUNDARY CONDITIONS
orography = EarthOrography(spectral_grid)
land_sea_mask = LandSeaMask(spectral_grid)
ocean = SeasonalOceanClimatology(spectral_grid)
land = SeasonalLandTemperature(spectral_grid)
solar_zenith = WhichZenith(spectral_grid, planet)
albedo = AlbedoClimatology(spectral_grid)
# PHYSICS/PARAMETERIZATIONS
physics = true
boundary_layer_drag = BulkRichardsonDrag(spectral_grid)
temperature_relaxation = NoTemperatureRelaxation(spectral_grid)
vertical_diffusion = BulkRichardsonDiffusion(spectral_grid)
surface_thermodynamics = SurfaceThermodynamicsConstant(spectral_grid)
surface_wind = SurfaceWind(spectral_grid)
surface_heat_flux = SurfaceSensibleHeat(spectral_grid)
convection = DryBettsMiller(spectral_grid)
shortwave_radiation = TransparentShortwave(spectral_grid)
longwave_radiation = JeevanjeeRadiation(spectral_grid)
# NUMERICS
device_setup = DeviceSetup(CPUDevice())
time_stepping = Leapfrog(spectral_grid)
spectral_transform = SpectralTransform(spectral_grid)
implicit = ImplicitPrimitiveEquation(spectral_grid)
horizontal_diffusion = HyperDiffusion(spectral_grid)
vertical_advection = CenteredVerticalAdvection(spectral_grid)
# OUTPUT
output = OutputWriter(spectral_grid, PrimitiveDry)
callbacks = Dict{Symbol, AbstractCallback}()
feedback = Feedback()(some are not exported, if you hit a UndefVarError just add SpeedWeather. in front of it. Depending on your version (this is from main) if you still hit an UndefVarError then just skip it, some fields were introduced after the latest release)
FredericWantiez
commented
2 months ago Tried something similar earlier, it all runs fine if I run each of the steps independently. Happens on both the tip of main and the latest release.
milankl
commented
2 months ago Then it must be the wrapping into PrimitiveDryModel, because technically all of these are executed and then PrimtivieDryModel(spectral_grid, geometry, dynamics, ...). A PrimitiveDryModel has a lot of parameters (in the Julia sense, i.e. PrimitiveDryModel{A,B,C,...}) which are inferred at that stage. Maybe that's what the intersect_unionall is referring to. On the other hand that logic is the same as for BarotropicModel, so I'm wondering whether there's just something in 1.9 that's causing the problem.
Thanks for flagging this, but also happy to hear it's not an issue on 1.10, is it a problem for you to use that version?
FredericWantiez
commented
2 months ago That's also my guess, something gets lost when infering the Primitive type. I'll use 1.10 for my experiments, mostly wanted to document the weird behavior somewhere. Thanks for looking into it
milankl
commented
2 months ago @giordano the creation of large parametric structs seems to hang on 1.9 and apple M1/2 👆 but not on 1.10, you think this could be related to your aarch64-related fixes for 1.10? Both M1/2 are armv8 right?
giordano
commented
2 months ago you think this could be related to your aarch64-related fixes for 1.10?
If you're referring to https://julialang.org/blog/2023/12/julia-1.10-highlights/#linux_aarch64_stability_improvements, that was specific to aarch64-linux, as the JITLink linker was already used on aarch64-darwin, the PR mentioned there only switched that on also on Linux. So I don't think that was it.
Both M1/2 are armv8 right?
Yes, but aarch64 is a slightly more accurate name for the architecture
milankl
commented
2 months ago @FredericWantiez I'm closing this as I see no reason not to use 1.10 (?) and there doesn't seem to be anything we can do here. Feel free to reopen anyone if that shouldn't be the case
milankl
commented
2 months ago I think I found the culprit on v1.9, if I remove the type parameter constraints, i.e.
Base.@kwdef mutable struct PrimitiveWetModel{
DS,
GE,
PL,
AT,
CO,
GO,
OR,
AC,
PA,
IC,
LS,
OC,
LA,
ZE,
AL,
SO,
VG,
CC,
BL,
TR,
VD,
SUT,
SUW,
SH,
EV,
LSC,
CV,
SW,
LW,
TS,
ST,
IM,
HD,
VA,
HF,
OW,
FB,
} <: PrimitiveWetinstead of
Base.@kwdef mutable struct PrimitiveWetModel{
NF<:AbstractFloat,
DS<:DeviceSetup,
PL<:AbstractPlanet,
AT<:AbstractAtmosphere,
CO<:AbstractCoriolis,
GO<:AbstractGeopotential,
OR<:AbstractOrography,
AC<:AbstractAdiabaticConversion,
PA<:AbstractParticleAdvection,
IC<:AbstractInitialConditions,
LS<:AbstractLandSeaMask,
OC<:AbstractOcean,
LA<:AbstractLand,
ZE<:AbstractZenith,
AL<:AbstractAlbedo,
SO<:AbstractSoil,
VG<:AbstractVegetation,
CC<:AbstractClausiusClapeyron,
BL<:AbstractBoundaryLayer,
TR<:AbstractTemperatureRelaxation,
VD<:AbstractVerticalDiffusion,
SUT<:AbstractSurfaceThermodynamics,
SUW<:AbstractSurfaceWind,
SH<:AbstractSurfaceHeatFlux,
EV<:AbstractSurfaceEvaporation,
LSC<:AbstractCondensation,
CV<:AbstractConvection,
SW<:AbstractShortwave,
LW<:AbstractLongwave,
TS<:AbstractTimeStepper,
ST<:SpectralTransform{NF},
IM<:AbstractImplicit,
HD<:AbstractHorizontalDiffusion,
VA<:AbstractVerticalAdvection,
HF<:AbstractHoleFilling,
GE<:AbstractGeometry,
OW<:AbstractOutputWriter,
FB<:AbstractFeedback,
} <: PrimitiveWetthen it works. While this would allow you to create a model with, say, orography = SpectralTransform(...) which I wanted to avoid with the <: constraints, model initialization would probably already fail.
Hello,
I've been trying to run some toy models but julia gets stuck on the model creation step. For something like the intro model:
the process never gets past
model = PrimitiveWetModel(...). When I kill the process manually, pages of the following error are displayed:Some details on my setup, running on a M1 Max:
Everything runs smoothly on the release version:
julia-1.10