Variable initialisation

[dmey]

Variables (e.g. column_variables.jl) are initialised with zero. Is there any reason for doing so? If not why not using NaN for floats and typemax or typemin for ints?

[milankl]

Does it matter, it will be overwritten anyway? To initialise arrays as with NaNs we would need to do

julia> fill(convert(NF,NaN),2,3)
2×3 Matrix{Float16}:
 NaN  NaN  NaN
 NaN  NaN  NaN

for example, such that for ever NF<:AbstractFloat there needs to a conversion from NaN to whatever NaN corresponds to in NF.

[dmey]

Sometime is hard to distinguish unset or improperly used variables down at the bottom of the stack and for debugging and assertions having NaNs helps... If there is no difference in terms of performance etc, I think it would be safer to use NaNs where possible.

[milankl]

No it's not a performance issue. In both cases memory is allocated and filled with bits. Doesn't matter what those are.

julia> using BenchmarkTools
julia> c = NaN
julia> @btime fill($c,1000,1000);
  1.160 ms (2 allocations: 7.63 MiB)

julia> c = 0.0
julia> @btime fill($c,1000,1000);
  1.163 ms (2 allocations: 7.63 MiB)

Note that in the dynamics we allocate with zeros, and I never had a debugging-related issue with this. Happy for you to convert the initialisation for ColumnVariables to init with NaN, if people prefer that. I would suggest to define

function nans(::Type{T},dims...) where T
     return fill(convert(T,NaN),dims...)
end

in utility_functions.jl such that zeros(NF,...) works as nans(NF,...)

julia> nans(Float16,2)
2-element Vector{Float16}:
 NaN
 NaN

julia> nans(Float32,2)
2-element Vector{Float32}:
 NaN
 NaN

[milankl]

nans has been added now ☝🏼

[dmey]

@milankl great, thanks. Will add this to my current PRs. Feel free to close.