It's efficient to do convolution with larger kernels if they are seperable, the sum of seperalable kernels. Suppose each kernel was the outer product of two matrices: R, S. We would like an Op that computes something like conv_seperable(images, R, S, mode='valid').