real_arr_type not type stable

[roflmaostc]

Where did you need those functions?

julia> real_arr_type(Array{ComplexF64}, dims=2)
Matrix{Float64} (alias for Array{Float64, 2})

julia> real_arr_type(Array{ComplexF64}, dims=3)
Array{Float64, 3}

julia> @code_warntype real_arr_type(Array{ComplexF64}, dims=3)
MethodInstance for (::NDTools.var"#real_arr_type##kw")(::NamedTuple{(:dims,), Tuple{Int64}}, ::typeof(real_arr_type), ::Type{Array{ComplexF64}})
  from (::NDTools.var"#real_arr_type##kw")(::Any, ::typeof(real_arr_type), ::Type{TA}) where TA<:AbstractArray in NDTools at /home/fxw/.julia/dev/NDTools.jl/src/type_tools.jl:27
Static Parameters
  TA = Array{ComplexF64}
Arguments
  _::Core.Const(NDTools.var"#real_arr_type##kw"())
  @_2::NamedTuple{(:dims,), Tuple{Int64}}
  @_3::Core.Const(NDTools.real_arr_type)
  @_4::Type{Array{ComplexF64}}
Locals
  dims::Int64
  @_6::Int64
Body::DataType
1 ─ %1  = Base.haskey(@_2, :dims)::Core.Const(true)
│         Core.typeassert(%1, Core.Bool)
│         (@_6 = Base.getindex(@_2, :dims))
└──       goto #3
2 ─       Core.Const(:(@_6 = 1))
3 ┄ %6  = @_6::Int64
│         (dims = %6)
│   %8  = (:dims,)::Core.Const((:dims,))
│   %9  = Core.apply_type(Core.NamedTuple, %8)::Core.Const(NamedTuple{(:dims,)})
│   %10 = Base.structdiff(@_2, %9)::Core.Const(NamedTuple())
│   %11 = Base.pairs(%10)::Core.Const(Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}}())
│   %12 = Base.isempty(%11)::Core.Const(true)
│         Core.typeassert(%12, Core.Bool)
└──       goto #5
4 ─       Core.Const(:(Base.kwerr(@_2, @_3, @_4)))
5 ┄ %16 = NDTools.:(var"#real_arr_type#2")(dims, @_3, @_4)::DataType
└──       return %16

[RainerHeintzmann]

e.g. here:
https://github.com/bionanoimaging/SeparableFunctions.jl/blob/main/src/radial.jl#L115

[roflmaostc]

In this specific case you could also pass arr and call inside rr2_sep:

similar(arr, real(eltype(arr)), ...)

[roflmaostc]

similar also works for CUDA arrays, that's the point

[RainerHeintzmann]

Yup. But the point is, that there are functions (like rr2_sep) which request an Array type as input and its good to have a way to create such a type.

[roflmaostc]

But as I said, it needs to perform the runtime dispatch and hence can cause performance issues:

Something like this would be better:

ulia> function similar_arr_type_4(::Type{TA}, ::Type{ET}, dims=Val(2)) where {TA, ET}
           typeof(similar(TA{ET}, ntuple(x->0, dims)))
       end
similar_arr_type_4 (generic function with 2 methods)

julia> similar_arr_type_4(Array, Float64)
Matrix{Float64} (alias for Array{Float64, 2})

julia> similar_arr_type_4(Array, Float64)^C

julia> @code_warntype similar_arr_type_4(Array, Float64)
MethodInstance for similar_arr_type_4(::Type{Array}, ::Type{Float64})
  from similar_arr_type_4(::Type{TA}, ::Type{ET}) where {TA, ET} in Main at REPL[47]:1
Static Parameters
  TA = Array
  ET = Float64
Arguments
  #self#::Core.Const(similar_arr_type_4)
  @_2::Type{Array}
  @_3::Core.Const(Float64)
Body::Type{Matrix{Float64}}
1 ─ %1 = Main.Val(2)::Core.Const(Val{2}())
│   %2 = (#self#)(@_2, @_3, %1)::Core.Const(Matrix{Float64})
└──      return %2