maybe we can have a good speed up by unrolling the full patch
multiplication instead of just the multiplication by a row? This could be
faster? What is the impact on the number of register?