my bad, it's RECIPROCAL square root, not reverse :)
There is also a reciropcal function that is a lot faster than a standard DIV. It also takes 3 cycles just like reciprocal square root. It simply does 1/x. So you would have to do a multiply instead of a division. I will post how it works in more detail on the bounding box thread. There is a packed version and a scalar version. I will use the scalar version.
On the subject of the reciprocal and reciprocal square root, see Phantom's pictures at the bottom of tihis thread: http://ompf.org/forum/viewtopic.php?f=11&t=1021
TBP's reference in his third posting in that thread also shows how to regain the lost accuracy in 3 cycles for reciprocal or 5 for reciprocal square root. Now if I just had the code...
Quote from: dr_eck on October 07, 2008, 02:32:37 PM
On the subject of the reciprocal and reciprocal square root, see Phantom's pictures at the bottom of tihis thread: http://ompf.org/forum/viewtopic.php?f=11&t=1021
TBP's reference in his third posting in that thread also shows how to regain the lost accuracy in 3 cycles for reciprocal or 5 for reciprocal square root. Now if I just had the code...
thanks Doc :)
EDIT: I had actually mentioned doing newton raphson at one point, just not on the reciprocal square root. Doc, you already have some newton raphson code? If not I can write some.
EDIT2: this is my FAVORITE math reference webpage. This is where tons of cool formulas to compute pi ( and newton raphson) are.
http://mathworld.wolfram.com/
EDIT3: tbp posted some sample code from intel for the Newton Raphson. here is the link.
http://sunsite.rediris.es/pub/mirror/intel/design/perftool/cbts/appnotes/ap803/