rif <[EMAIL PROTECTED]> writes:
> I'm trying to optimize the following function (yes, profiling
> indicates that this function is the bottleneck, it's taking the lion's
> share of the time and the consing).
>
> (defun find-closest-index (point centers)
> (declare (type simple-array centers)
> (type (simple-array double-float) point))
> (let ((best-dist most-positive-double-float)
> (dist 0.0d0)
> (index -1))
> (declare (type fixnum index)
> (type double-float best-dist)
> (type double-float dist))
> (fixtimes (i (array-dimension centers 0))
> (let ((c (aref centers i)))
> (declare (type (simple-array double-float) c))
> (setf dist (the double-float (l2-distance-squared point c)))
> (when (< dist best-dist)
> (progn
> (setf index i)
> (setf best-dist dist)))))
> index))
>
>
> One of the compiler notes (settings were speed 3, debug and safety 0)
> for this function is:
>
> ; (SETF BEST-DIST DIST)
> ; ==>
> ; (SETQ BEST-DIST DIST)
> ; Note: Doing float to pointer coercion (cost 13) from DIST to BEST-DIST.
>
> And I cannot for the life of me figure out why I'm getting this. Any
> ideas?
Under CMUCL (not SBCL) this variant seems to work:
(defun find-closest-index (point centers)
(declare (optimize (speed 3) (debug 0) (safety 0)))
(declare (type (simple-array * 1) centers)
(type (simple-array double-float 1) point))
(let ((best-dist most-positive-double-float)
(dist 0.0d0)
(index -1))
(declare (type fixnum index)
(type double-float best-dist)
(type double-float dist))
(setq best-dist (- (- best-dist))) ; <-----------------------------!!!
(flet ((l2-distance-squared (a1 a2)
(loop for x double-float across a1
and y double-float across a2
summing (expt (- x y) 2) of-type double-float)))
(dotimes (i (array-dimension centers 0))
(declare (fixnum i))
(let ((c (aref centers i)))
(declare (type (simple-array double-float 1) c))
(setf dist (the double-float (l2-distance-squared point c)))
(when (< dist best-dist)
(progn
(setf index i)
(setf best-dist dist))))))
index))
In general try to add some simple FP operations with the problematic
variable and hope the compiler will not be able to optimize them away.
--
Regards,
Alexey Dejneka
"Alas, the spheres of truth are less transparent than those of
illusion." -- L.E.J. Brouwer
