Standard function name to skip intermediate step piping getindex call into vec

General Usage
1

I am wondering about Julia conventions - is there a function or kwarg for skipping an intermediate type when taking an array through getindex and then vec?

More concretely, suppose we are doing a getindex call on a SparseMatrixCSC and convert the result into a vector:

using SparseArrays 
sa = sparse([1 0 1; 0 1 0]) 
x = vec(sa[1:2, 1:2])

What is returned by getindex is a SparseMatrixCSC:

typeof(sa[1:2, 1:2]) <: SparseMatrixCSC

I want to skip the intermediate allocation to a SparseMatrixCSC, but I want to do it in a way that accords with Julia conventions. Note that getindex(args…; kwargs…) does not allow using kwargs with the bracket syntactic sugar.

The reason why I am asking is because I have defined my own array type such that it would be computationally cheaper if I could skip the intermediate getindex type. I would like something like:

typeof(my_arr) <: MyArrayType
typeof(getindex_as_vec(my_arr, 1:2, 1:2)) <: Vector

But I don’t want to write the function name getindex_as_vec if there is some convention I should be adhering to.

2

I think I am an idiot. I think I merely need to use views. e.g. vec(view(myarra, 1:2, 1:2))

1 Like
3

You probably want to be careful with views of sparse arrays, they don’t have great performance. I’d compare both options, depending on the setting either one may be faster

4

Here is a solution using CartesianIndices and broadcasting:

using SparseArrays
A = sparse([1 0 1; 0 1 0 ; 0 1 1])
Is = CartesianIndices(A)
@show getindex.(Ref(A), Is[1:2, 1:2])

However looking up a lot of indices in a sparse matrix is not the fastest thing to do. Maybe if you describe your situation a bit more, we could see whether there is a faster approach?

5

With Sparse arrays, it can sometimes be advantageous to create a copy. You could try the vec∘view, or you could try explicitly collecting it into a new SparseVector. Depending upon your subsequent usage patterns, this may or may not be a win.

1 Like