prenex/magyarsort
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

added frewr comment - because it becames fastest in oct 1 - 2025

Browse Source
This commit is contained in:
Richard Thier 2025-10-01 19:09:18 +02:00
parent fb0b8ce255
commit b5aeaa1bdb
1 changed files with 9 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
ypsu.cpp
View File

@ -639,6 +639,15 @@ void fourrots(uint32_t *arr, int n) {
} }
// frewr - four rewrites. // frewr - four rewrites.
// Rem.: I realized this changes the keys (rerwrites) by rotating them by 8 in each pass
// So in pass 0 we just & 0xff and put in place, but stored v gets rotated right by 8!
// So in pass 1 we can just & 0xff again - and store again rotated
// Because there are 4 rotates and 4 radices - in the end the "real" values "came back" as keys.
// Currently (in October 1, 2025) this is fastest, but I KNOW this was slower than Magyar in 2021 dec (slightly)
// Interestingly if I build that old build now, its already faster than my Magyarsort (and it misses some later opt)
// So I imagine this is because compiler technology became better and favors this algorithm more now!
// This algorithm is made by Ypsu - my friend. He made it originally as a response to Magyarsort...
// I added and measured unrolls to this and some minor optimizations - but not understanding the alg back then...
void frewr(uint32_t *arr, int n) { void frewr(uint32_t *arr, int n) {
uint32_t *tmpbuf = (uint32_t *)malloc(n * 4); uint32_t *tmpbuf = (uint32_t *)malloc(n * 4);
mlock(tmpbuf, n * 4); mlock(tmpbuf, n * 4);