A is (N, D) and B is (M, D). L2-normalise every row of each, then return the (N, M) matrix of cosine similarities between every pair.

No loops. This is the retrieval step behind every embedding search and every image-text matching model.

Input

A =
tensor([[1., 0.],
        [0., 2.]])
B =
tensor([[ 1.,  0.],
        [ 1.,  1.],
        [ 0., -3.]])

Output

tensor([[ 1.0000,  0.7071,  0.0000],
        [ 0.0000,  0.7071, -1.0000]])