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
0x02For 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