extrema gives different results on Fields vs parent arrays

[akshaysridhar]

When running the box_density_current.yml example, and querying the extrema of a scalar field ( in this instance, density)

julia> simulation.integrator.u.c.ρ |> extrema
(0.0f0, 1.1547949f0)
julia> parent(simulation.integrator.u.c.ρ) |> extrema
(0.65239275f0, 1.1547949f0)

The latter is expected, so I suspect there is a bug in our minimum method for Fields

For reference typeof returns

julia> typeof(simulation.integrator.u.c.ρ)
ClimaCore.Fields.Field{ClimaCore.DataLayouts.VIJFH{Float32, 4, SubArray{Float32, 5, CUDA.CuArray{Float32, 5, CUDA.Mem.DeviceBuffer}, Tuple{Base.Slice{Base.OneTo{Int64}}, Base.Slice{Base.OneTo{Int64}}, Base.Slice{Base.OneTo{Int64}}, UnitRange{Int64}, Base.Slice{Base.OneTo{Int64}}}, false}}, ClimaCore.Spaces.ExtrudedFiniteDifferenceSpace{ClimaCore.Grids.ExtrudedFiniteDifferenceGrid{ClimaCore.Grids.SpectralElementGrid2D{ClimaCore.Topologies.Topology2D{ClimaComms.SingletonCommsContext{ClimaComms.CUDADevice}, ClimaCore.Meshes.RectilinearMesh{ClimaCore.Meshes.IntervalMesh{ClimaCore.Domains.IntervalDomain{ClimaCore.Geometry.XPoint{Float32}, Nothing}, LinRange{ClimaCore.Geometry.XPoint{Float32}, Int64}}, ClimaCore.Meshes.IntervalMesh{ClimaCore.Domains.IntervalDomain{ClimaCore.Geometry.YPoint{Float32}, Nothing}, LinRange{ClimaCore.Geometry.YPoint{Float32}, Int64}}}, Vector{CartesianIndex{2}}, Matrix{Int64}, CUDA.CuArray{Tuple{Int64, Int64, Int64, Int64, Bool}, 1, CUDA.Mem.DeviceBuffer}, Vector{Tuple{Int64, Int64, Int64, Int64, Bool}}, CUDA.CuArray{Tuple{Int64, Int64}, 1, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Tuple{Bool, Int64, Int64}, 1, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}, CUDA.CuArray{Int64, 1, CUDA.Mem.DeviceBuffer}, @NamedTuple{}, CUDA.CuArray{Tuple{Int64, Int64}, 1, CUDA.Mem.DeviceBuffer}}, ClimaCore.Quadratures.GLL{4}, ClimaCore.Geometry.CartesianGlobalGeometry, ClimaCore.DataLayouts.IJFH{ClimaCore.Geometry.LocalGeometry{(1, 2), ClimaCore.Geometry.XYPoint{Float32}, Float32, StaticArraysCore.SMatrix{2, 2, Float32, 4}}, 4, CUDA.CuArray{Float32, 4, CUDA.Mem.DeviceBuffer}}, ClimaCore.DataLayouts.IJFH{Float32, 4, CUDA.CuArray{Float32, 4, CUDA.Mem.DeviceBuffer}}, ClimaCore.DataLayouts.IFH{ClimaCore.Geometry.SurfaceGeometry{Float32, ClimaCore.Geometry.UVVector{Float32}}, 4, CUDA.CuArray{Float32, 3, CUDA.Mem.DeviceBuffer}}, @NamedTuple{}}, ClimaCore.Grids.FiniteDifferenceGrid{ClimaCore.Topologies.IntervalTopology{ClimaComms.SingletonCommsContext{ClimaComms.CUDADevice}, ClimaCore.Meshes.IntervalMesh{ClimaCore.Domains.IntervalDomain{ClimaCore.Geometry.ZPoint{Float32}, Tuple{Symbol, Symbol}}, LinRange{ClimaCore.Geometry.ZPoint{Float32}, Int64}}, @NamedTuple{bottom::Int64, top::Int64}}, ClimaCore.Geometry.CartesianGlobalGeometry, ClimaCore.DataLayouts.VF{ClimaCore.Geometry.LocalGeometry{(3,), ClimaCore.Geometry.ZPoint{Float32}, Float32, StaticArraysCore.SMatrix{1, 1, Float32, 1}}, CUDA.CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}}}, ClimaCore.Grids.Flat, ClimaCore.Geometry.CartesianGlobalGeometry, ClimaCore.DataLayouts.VIJFH{ClimaCore.Geometry.LocalGeometry{(1, 2, 3), ClimaCore.Geometry.XYZPoint{Float32}, Float32, StaticArraysCore.SMatrix{3, 3, Float32, 9}}, 4, CUDA.CuArray{Float32, 5, CUDA.Mem.DeviceBuffer}}}, ClimaCore.Grids.CellCenter}}

Doing the same to the parent array gives

julia> typeof(parent(simulation.integrator.u.c.ρ))
SubArray{Float32, 5, CUDA.CuArray{Float32, 5, CUDA.Mem.DeviceBuffer}, Tuple{Base.Slice{Base.OneTo{Int64}}, Base.Slice{Base.OneTo{Int64}}, Base.Slice{Base.OneTo{Int64}}, UnitRange{Int64}, Base.Slice{Base.OneTo{Int64}}}, false}

[akshaysridhar]

Following checks with @charleskawczynski this appears to be a custom GPU method implementation issue in our DataLayouts.

[charleskawczynski]

Hi @akshaysridhar, could you please add a reproducer, like:

# julia --project=examples
ENV["CLIMACOMMS_DEVICE"] = "CUDA";
empty!(ARGS)
push!(ARGS, "--config_file", "path/to/config")
using Revise; include(joinpath("examples", "hybrid", "driver.jl"))

?

[charleskawczynski]

bump! @akshaysridhar, was this with 1 gpu? or multiple? (do you happen to remember)

[akshaysridhar]

bump! @akshaysridhar, was this with 1 gpu? or multiple? (do you happen to remember)

This was a single GPU run (clima A100) @charleskawczynski

[Sbozzolo]

This should be fixed here: CliMA/ClimaCore.jl#2106