tag:blogger.com,1999:blog-569950264826755092.post1106105762322850063..comments2017-06-21T01:09:53.640-04:00Comments on multigrad: Math evolution and dirty tricksYannick Hold-Geoffroynoreply@blogger.comBlogger3125tag:blogger.com,1999:blog-569950264826755092.post-22058218698079660322014-04-09T17:33:35.737-04:002014-04-09T17:33:35.737-04:00Seems like finding the optimal constant could be b...Seems like finding the optimal constant could be brute forced in < 1 second. Finding a more efficient algo with a GA is a more interesting idea imho.manthraxhttp://www.blogger.com/profile/11262642310898294688noreply@blogger.comtag:blogger.com,1999:blog-569950264826755092.post-73284739184623839692014-04-09T16:25:02.557-04:002014-04-09T16:25:02.557-04:00I tried to concentrate my efforts on the generatio...I tried to concentrate my efforts on the generation of new equations so I haven't dealt with the constant optimization so much. I made too much simplifications in my second stage to find the best constant. If you are interested by the constant, many people concentrated on finding the best constant, among them are Chris Lomont, Charles McEniry (as cited in the wikipedia page) and also Jan Kadlec ( http://rrrola.wz.cz/inv_sqrt.html ).Yannick Hold-Geoffroyhttp://www.blogger.com/profile/17227249978993145288noreply@blogger.comtag:blogger.com,1999:blog-569950264826755092.post-80533211422453858772014-04-09T16:17:10.645-04:002014-04-09T16:17:10.645-04:00It's interesting that there are actually a fai...It's interesting that there are actually a fairly large range of numbers that would fall into the range of acceptable. The hex value chosen in the original code, 0x5f3759df, is 0x2F098 off of 0x5f346947 (or 192,664) appears to be a better value with regards to error. It becomes a lot more obvious with the graphs that the original constant isn't even necessarily ideal but at the same time that doesn't actually matter.Davehttp://www.blogger.com/profile/00824925262798597455noreply@blogger.com