Since matrix-matrix multiplication can be (and has been in many libraries) highly optimized, it's an easy and decently-performant way to implement the convolutions without having to write your own optimized code specifically for the convolutions you're doing.