mmmul   xtlang

;; optimized 4x4 matrix only!!!
;; NEED TO ALSO DO integers and doubles!!
(bind-func mmmul
  (lambda (m1:float* m2:float* result:float*)
    (let ((row1:/4,float/* (cast (pref-ptr m2 0)))
          (row2:/4,float/* (cast (pref-ptr m2 4)))
          (row3:/4,float/* (cast (pref-ptr m2 8)))
          (row4:/4,float/* (cast (pref-ptr m2 12)))
          (out:/4,float/* (cast result))
          (tmp:/4,float/* null) (i 0))
      (dotimes (i 4)
        (set! tmp (cast (pref-ptr m1 (* 4 i))))
        (let ((a (vshuffle tmp null 0 0 0 0))
              (b (vshuffle tmp null 1 1 1 1))
              (c (vshuffle tmp null 2 2 2 2))
              (d (vshuffle tmp null 3 3 3 3)))
          (pset! out i
                 (+ (+ (* a row1)
                       (* b row2))
                    (+ (* c row3)
                       (* d row4))))))
      result)))