UnionArrays: storage-agnostic array type with Union elements

UnionArrays.jl provides an array type with Union element types that is generic over the data storage type.

julia> using UnionArrays

julia> xs = UnionVector(undef, Vector, Union{Float32,Tuple{},UInt8}, 3);

julia> fill!(xs, ());

julia> xs[1]
()

julia> xs[2] = 1.0f0;

julia> xs[3] = UInt8(2);

julia> collect(xs)
3-element Vector{Union{Tuple{}, Float32, UInt8}}:
     ()
    1.0f0
 0x02

For example, it can be used for bringing Union element types to GPU:

julia> using CUDA

julia> xs = UnionVector(undef, CuVector, Union{Float32,Nothing}, 3);

julia> fill!(xs, nothing);

Packages like Transducers.jl and Folds.jl support computations with UnionArrays on GPU:

julia> using Folds, FoldsCUDA

julia> Folds.all(==(nothing), xs)
true

julia> CUDA.@allowscalar begin
           xs[2] = 1.0f0
           xs[3] = 2.0f0
       end;

julia> Folds.sum(x -> x === nothing ? 0.0f0 : x, xs; init = 0.0f0)
3.0f0