Add doesn't broadcast

[ojwoodford]

MWE:

using TensorOperations

function test_tensor()
    A = rand(2, 2, 3)
    B = rand(2, 4)
    C = rand(2, 3)
    @tensor D[i,j,k] := A[i,l,k] * B[l,j] + C[i,k]
end

test_tensor()

ERROR: TensorOperations.IndexError{String}("add: (:i, :k) to (:i, :j, :k))")

I was expecting this to work, and for the values in C to be broadcast across the rows (dim 2) of D. Is there a reason why it doesn't?

[lkdvos]

Hi!

Unfortunately, this is not expected to work with TensorOperations. The functionality of this package is currently limited to strict einsum notation convention: for each separate term, every index appears exactly twice. Either once on the rhs and once on the lhs for uncontracted indices, or twice on the rhs for contracted indices.
Broadcasting or multi-indices are not supported.

This being said, you can trick the system by manually creating a view that broadcasts at zero overhead. The idea being that if you create a stride 0 index this is equivalent to broadcasting.

using TensorOperations
using Strided

function test_tensor()
    A = rand(2, 2, 3)
    B = rand(2, 4)
    C = rand(2, 3)
    sz = (size(C)..., 4)
    str = (strides(C)..., 0)
    offset = 0
    C_expanded = StridedView(C, sz, str, offset)
    @tensor D[i,j,k] := A[i,l,k] * B[l,j] + C_expanded[i,k,j]
end

test_tensor()