Type instability in CFtransformdata!

[tiemvanderdeure]

Just following up on JuliaGeo/NCDatasets.jl#273 and moving it over here since it's actually a CommonDataModel problem.

CFtransformdata is not type stable if the eltype of data and out are not the same.

using CommonDataModel
(fv, scale_factor, add_offset, time_origin, time_factor, maskingvalue) = ((1f0,), nothing, nothing, nothing, nothing, missing)
A = rand(1000,1000)
CommonDataModel.CFtransformdata(A, fv, scale_factor, add_offset, time_origin, time_factor, maskingvalue, Union{Float64, Missing});
@time CommonDataModel.CFtransformdata(A, fv, scale_factor, add_offset, time_origin, time_factor, maskingvalue, Union{Float64, Missing});

shows 0.082064 seconds (2.00 M allocations: 39.093 MiB, 18.78% gc time).

For some reason this doesn't always show up on actual file reads, but if there is a time dimension it does:

using Rasters, NCDatasets, Dates
ras = Raster(rand(X(1:100), Y(1:100), Ti(DateTime(2000):Year(1):DateTime(2000)))) |> replace_missing

fname = tempname() * ".nc"
write(fname, ras)
Raster(fname); @time Raster(fname);
write(fname, ras[Ti = 1], force = true)
Raster(fname); @time Raster(fname);

Returns

  0.004901 seconds (21.19 k allocations: 635.992 KiB)
  0.004241 seconds (1.03 k allocations: 298.234 KiB)

Note that the time dimension here has length 1 so this really should be the same.

This is all fixed after dropping the @inline in

CommonDataModel.jl/src/cfvariable.jl

Line 373 in 86a1fce

@inline function CFtransformdata!(out,data::AbstractArray{T,N},fv,scale_factor,add_offset,time_origin,time_factor,maskingvalue) where {T,N}

I'll make a PR with that change, but maybe there are other/better fixes? I don't understand at all why the time dimension makes a difference and neither @code_warntype nor Cthulhu flag this.

[Alexander-Barth]

Thanks again for the PR. I am closing the issue. Feel free to reopen if the issue persists :-)