(defmethod mult ((m1 matrix-r) (m2 matrix-r))
  (if (not (= (number-of-cols m1) (number-of-rows m2)))
      (error "Can't perform matrix multiplication")
      (let* ((elements-m1 (elements m1)) (elements-m2 (elements m2))
	     (n (number-of-cols m1))
	     (n-rows (number-of-rows m1))
	     (n-cols (number-of-cols m2))
	     (m (make-instance 'matrix-r :number-of-rows n-rows :number-of-cols n-cols))
	     (elements-m (elements m))
	     (width (lane-width m1))
	     (r-lane (make-array width :element-type 'double-float))
	     (c-lane (make-array width :element-type 'double-float))
	     (sum 0.0d0))
	(declare
	 (type fixnum n-rows n-cols n width)
	 (type (simple-vector) elements-m1 elements-m2 elements-m)
	 (type (simple-array double-float (4)) r-lane c-lane)
	 (type double-float sum)
	 (optimize (debug 0) (safety 0) (speed 3)))
	(loop for i fixnum from 0 below n-rows do
	     (loop for j fixnum from 0 below n-cols do
		  (setf sum 0.0d0)
		  (loop for k fixnum from 0 below (number-of-lanes m1) do
		       (setf r-lane (aref elements-m1 k))
		       (setf c-lane (aref elements-m2 k))
		       (loop for l fixnum below 4 do
			  ;;(incf sum (* (aref elements-m1 (+ k (the fixnum (* i n))))
			  ;;(aref elements-m2 (+ j (the fixnum (* k n)))))))
			    (incf sum (* (aref r-lane l) (aref c-lane l)))))
		  (setf (aref (aref elements-m 0) 0) sum)))
	m)))