Reduction PR #881 breaks Oceananigans

[giordano]

#881 broke reductions inside Oceananigans.

PRONTOLab/GB-25#45 (comment)

ERROR: LoadError: InexactError: Int64(Inf)
Stacktrace:
  [1] Int64
    @ ./float.jl:994 [inlined]
  [2] convert
    @ ./number.jl:7 [inlined]
  [3] _round_convert
    @ ./rounding.jl:480 [inlined]
  [4] round
    @ ./rounding.jl:479 [inlined]
  [5] ceil(::Type{Int64}, x::Float64)
    @ Base ./rounding.jl:476
  [6] split_explicit_substepping

Code in Oceananigans:
https://github.com/CliMA/Oceananigans.jl/blob/875a06a0964115c7f6ea1ca14c12ede5272d5ab1/src/Fields/field.jl#L671
https://github.com/CliMA/Oceananigans.jl/blob/875a06a0964115c7f6ea1ca14c12ede5272d5ab1/src/Fields/field.jl#L620
https://github.com/CliMA/Oceananigans.jl/blob/875a06a0964115c7f6ea1ca14c12ede5272d5ab1/src/Fields/field.jl#L715

[simone-silvestri]

This is a MWE that shows the difference in behavior between a ReactantState() architecture and a CPU() architecture

using Reactant, Oceananigans

grid = RectilinearGrid(CPU(), size=(2, 4, 8), extent=(1, 1, 1))
c = CenterField(grid)
set!(c, 1)
r = Field{Nothing, Nothing, Nothing}(grid)
c = Oceananigans.Fields.condition_operand(c, nothing, Inf)

Base.mapfirst!(identity, interior(r), interior(c))
@show interior(r)

grid = RectilinearGrid(Oceananigans.Architectures.ReactantState(), size=(2, 4, 8), extent=(1, 1, 1))
rc = CenterField(grid)
set!(rc, 1)
rr = Field{Nothing, Nothing, Nothing}(grid)
rc = Oceananigans.Fields.condition_operand(c, nothing, Inf)

Base.mapfirst!(identity, interior(rr), interior(rc))
@show interior(rr)

In particular interior(r) has, correctly, a value of 1, while interior(rr) is not being initialized correctly.

[wsmoses]

to confirm, this is now fixed on main by the PR reversion right?

[giordano]

Yes, main and v0.2.40 are good.