Calling ncon with a view throws MethodError

[votroto]

Is it possible to support views into arrays? It seems the api might be too specific.

e.g. a Matrix works as expected:

A = [1 2 3; 4 5 6]
B = [0, 1, 0]
C = [1, 1]

@show ncon([A, B, C], [[2,1], [1], [2]])

An equivalent view into a matrix throws a MethodError: no method matching sview error

A = view([1 2 3 1; 4 5 6 1; 7 8 9 1], [1,2], [1,2,3])
B = [0, 1, 0]
C = [1, 1]

@show ncon([A, B, C], [[2,1], [1], [2]])

Simply doing collect works as expected.

A = collect(view([1 2 3 1; 4 5 6 1; 7 8 9 1], [1,2], [1,2,3]))
B = [0, 1, 0]
C = [1, 1]

@show ncon([A, B, C], [[2,1], [1], [2]])

Thank You

[Jutho]

If you were to replace A with A = view([1 2 3 1; 4 5 6 1; 7 8 9 1], 1:2, 1:3) you would find that it works. The general rule is that views are supported, if they are sliced with ranges. If you slice with arbitrary vectors, you could do things like view(some_array, [1,3,4,13,14,6], ...) with completely unpredictable memory layout for the elements of the resulting view. Technically, such a view would not have a strided memory layout. Only views with a strided memory layout are supported, which are those that you can make by only using ranges (and single integers, and colons) for indexing/slicing.

[votroto]

Rigth, then I guess what I'm asking for is a fallback implementation for objects without a strided memory layout. Tensor operations are still well defined even on general views, right?

Forgetting the performance implications for a moment, would that be possible?

[Jutho]

That would indeed be possible, a simple implementation based on loops could indeed be provided. The need for this was never very high. There are also some tricky (but not insurmountable) issues with dispatch.