Re: [Matlisp-users] m.*! not native?

[Jefferson P. <jp...@cs...>]

Raymond Toy wrote:
>>>>>>"Jefferson" == Jefferson Provost <jp...@cs...> writes:
>>>>>
> 
>     Jefferson> I'm not sure.  The short answer is that I would love to have it better
>     Jefferson> optmized, but if it's not a priority for you guys, I understand.
> 
> If you can, please try the CVS version.  I've put in an optimized
> version that should run as fast as a Fortran version should (on
> CMUCL).
> 
> I'd be interested to know if it's fast enough for you.

Yes it's quite fast.  Multiplying two NxN matrices is still a little 
slower than the outer product of two N-length vectors (which should also 
be O(N^2) operations) -- but this is probably from going out of cache on 
the large matrices, while the smaller vectors will stay in cache.

As I mentioned before, I had changed my code so I didn't need to call 
M.* as often, but that was a sub-optimal solution.  But, I realized that 
my main use of M.* was to "mask" a matrix, i.e. set a bunch of elements 
to 0.0 using another matrix as a guide:  (m.*! mask x).  So I wrote a 
special function just to do that masking.  It's MUCH faster, since it 
does no multiplication at all...


(defmethod mmask! ((a real-matrix) (mask real-matrix))

   (let* ((nxm (number-of-elements a))
	 (aa (store a))
	 (bb (store mask)))
     (declare (type fixnum nxm)
	     (type (real-matrix-store-type (*)) aa bb)
	     (optimize (speed 3) (safety 0)))

     (dotimes (k nxm a)
       (declare (type fixnum k))
       (let ((b-val (aref bb k)) )
	(declare (type real-matrix-element-type b-val)
		 (:explain :calls :types :boxing))
	(when (zerop b-val)
	  (setf (aref aa k) 0.0d0))))))

J.