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

Revert "tiny (hopefully) optimization for tpxb"

Browse Source 
This reverts commit 2aa7de0d40bca5c8713c5edb41a2ff3995a2ea01.
This commit is contained in:
Richard Thier 2025-10-02 08:09:44 +02:00
parent 2aa7de0d40
commit 74e24486f4
1 changed files with 9 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
threepass_xbit.h
View File

@ -94,27 +94,27 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
/* right-to-left to ensure already sorted digits order we keep for iterations */
/* Bottom digit a->buf */
#pragma GCC unroll 64
for(int i = n - 1; i >= 0; --i) {
uint32_t num = a[i];
#pragma GCC unroll 80
for(uint32_t i = n; i > 0; --i) {
uint32_t num = a[i - 1];
/* save rotated into the masked, bucketed loc */
buf[bucket3[num & mask3]--] = (num >> TPBX3) | (num << (32 - TPBX3));
//__builtin_prefetch(&buf[bucket3[num & mask3] - 2]);
}
/* Mid digit buf->a */
#pragma GCC unroll 64
for(int i = n - 1; i >= 0; --i) {
uint32_t num = buf[i];
#pragma GCC unroll 80
for(uint32_t i = n; i > 0; --i) {
uint32_t num = buf[i - 1];
/* save rotated into the masked, bucketed loc */
a[bucket2[num & mask2]--] = (num >> TPBX2) | (num << (32 - TPBX2));
//__builtin_prefetch(&a[bucket2[num & mask2] - 2]);
}
/* Top digit a->buf */
#pragma GCC unroll 64
for(int i = n - 1; i >= 0; --i) {
uint32_t num = a[i];
#pragma GCC unroll 80
for(uint32_t i = n; i > 0; --i) {
uint32_t num = a[i - 1];
/* Reconstruct the original key in this element by where its stuff is stored */
constexpr int rot = TPBX1 + (32 - TPBX1 - TPBX2 - TPBX3);
/* save rotated into the masked, bucketed loc */