Smart RecursiveArrayTools.VectorOfArray which allows binary operations with AbstractVector

General Usage
1

This is a follow-up question from How to solve set of nonlinear fixed-point equations using Anderson acceleration in NLsolve.jl?. Suppose I have a VectorOfArray object

z = VectorOfArray([rand(4), rand(4)])

Then the following operations are not allowed

t .- rand(8)  # ERROR: DimensionMismatch("arrays could not be broadcast to a common size; got a dimension with lengths 4 and 8")
t - rand(8)  # ERROR: DimensionMismatch("dimensions must match: a has dims (Base.OneTo(32), Base.OneTo(2)), b has dims (Base.OneTo(64),), mismatch at 1")

In practice, it will be more complicated where the elements inside the VectorOfArray shall be any AbstractArray.

It is sensible to do following using copyto!

function Base.copyto!(vf::AbstractOfArray{T}, v::AbstractVector) where {T<:MyAbstractArrayType}
    total_elements = 0
    for i in 1:length(vf)
        total_elements += length(vf[1])
    end
    total_elements == length(v) || error("Length mismatch.")

    j = 1
    for i in 1:length(vf)
        n = length(vf[i])
        vf[i] .= v[j:j+n-1]
        j += n
    end
end

z1 = VectorOfArray([rand(4), rand(4)])
z2 = VectorOfArray([rand(4), rand(4)])
copyto!(z1, rand(8))
# Use RecursiveArrayTools broadcasting
z2 .- z1
# Or
z2 - z1

Or shall I define a new operator method as

function Base.-(vf::AbstractOfArray{T}, v::AbstractVector) where {T<:MyAbstractArrayType}
    total_elements = 0
    for i in 1:length(vf)
        total_elements += length(vf[1])
    end
    total_elements == length(v) || error("Length mismatch.")

    out = deepcopy(vf)

    j = 1
    for i in 1:length(vf)
        n = length(vf[i])
        out[i] = vf[i] - v[j:j+n-1]
        j += n
    end
    return out
end

However, this approach again relies on whether we can do - with vf[i]andv` which may fail.

I want to know the best practice to solve this kind of issue. Thanks!

2

Best thing to do would be to extend the broadcast overload.