# HG changeset patch # User Steve Losh # Date 1555818328 14400 # Node ID 9c4c16c0c6a1b23239de17d892c6316093a7d5f9 # Parent 098f018f4b84ac700ed9c7b47e07269d13d93d03 Update diff -r 098f018f4b84 -r 9c4c16c0c6a1 README.markdown --- a/README.markdown Sat Apr 20 21:36:12 2019 -0400 +++ b/README.markdown Sat Apr 20 23:45:28 2019 -0400 @@ -495,3 +495,17 @@ machines (go figure). All the test suites (e.g. TestU01 and PractRand) immediately fail small PRNGs, so it's hard to know if I've got something decent or not. Oh well, it's just for games, it's not life and death. + +Random notes: + +* The "Randograms" in the blog post are confusing because they're sampling the + entire period of the generator, but still get odd-numbered occurence counts. + This confused the hell out of me until I realized that by "pair up the + numbers" they mean a sliding two element window: `(1 2 3 4)` produces `((1 2) + (2 3) (3 4))`. +* The 256x256 size works perfectly for the 8-bit generators. I wasn't able to + extend it for 16-bit generators because exhausting their period would take + *much* longer. +* It's still not clear to me how to decide what's a "good" multiplier for + smaller-sized PRNGs. They fail the tests right away because they're so small, + so I can't really go by that.