findnz does not work for arrays of arbitrary dimension

[cossio]

It is only defined for matrices. But to me it seems reasonable to define findnz(A) for any higher-dimensional array. It would return (I1, I2, ..., Id, V), where d is the number of dimensions in A, Ii is the indices of non-zero entries along dimension i, and V is the value of those non-zero entries.

[cossio]

The implementation could be something like

function findnz2(A::AbstractArray{T, N}) where {T, N}
    nnzA = count(t -> t != 0, A)
    idxs = ntuple(_ -> zeros(Int, nnzA), N)
    V = Vector{T}(nnzA)
    cnt = 0
    for idx in eachindex(A)
        if A[idx] ≠ 0
            cnt += 1
            indexes::NTuple{N, Int} = ind2sub(A, idx)
            for d = 1 : N
                idxs[d][cnt] = indexes[d]
                V[cnt] = A[idx]
            end
        end
    end
    return (idxs..., V)
end

[nalimilan]

findnz has been moved to SparseMatrices since it's mainly designed for this case. Use findall(iszero, x) instead for general arrays. On 0.7, it returns cartesian indices.

[mbauman]

Note that findnz on dense matrices is a strange operation — and one that we had been considering deprecating (#24910 and #25641). It's really X <: SparseArray ? findstored(X) : find(!iszero, X).

Unfortunately that discussion had gotten closed and forgotten without fully resolving the deprecation.

[cossio]

I guess it's better to deprecate it then, before 0.7 is out.