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

Revert "tpbx: tried removal of relative addressing but it does not help, just makes n be int instead of uint32_t so probably will be reverted. Sad because this actually looked beneficial"

Browse Source 
This reverts commit 5ecb48815b57c51527f2c55c3555fb40ffe48f6b.
This commit is contained in:
Richard Thier 2025-10-01 01:53:38 +02:00
parent 5ecb48815b
commit c032109110
1 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
threepass_xbit.h
View File

@ -96,11 +96,11 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
// Bottom digit a->buf // Bottom digit a->buf
// right-to-left to ensure already sorted digits order we keep for iterations // right-to-left to ensure already sorted digits order we keep for iterations
#pragma GCC unroll 48 #pragma GCC unroll 48
for(int i = n - 1; i >= 0; --i) { for(uint32_t i = n; i > 0; --i) {
// Prefetch caches // Prefetch caches
//__builtin_prefetch(&a[i-8]); //__builtin_prefetch(&a[i-8]);
// Get num and its new offset / location // Get num and its new offset / location
auto num = a[i]; auto num = a[i - 1];
auto bkeyni = (num >> shr3) & mask3; auto bkeyni = (num >> shr3) & mask3;
auto offset = --bucket3[bkeyni]; auto offset = --bucket3[bkeyni];
@ -114,11 +114,11 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
// Mid digit buf->a // Mid digit buf->a
// right-to-left to ensure already sorted digits order we keep for iterations // right-to-left to ensure already sorted digits order we keep for iterations
#pragma GCC unroll 48 #pragma GCC unroll 48
for(int i = n - 1; i >= 0; --i) { for(uint32_t i = n; i > 0; --i) {
// Prefetch caches // Prefetch caches
//__builtin_prefetch(&buf[i-8]); //__builtin_prefetch(&buf[i-8]);
// Get num and its new offset / location // Get num and its new offset / location
auto num = buf[i]; auto num = buf[i - 1];
auto bkeyni = (num >> shr2) & mask2; auto bkeyni = (num >> shr2) & mask2;
auto offset = --bucket2[bkeyni]; auto offset = --bucket2[bkeyni];
@ -132,11 +132,11 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
// Top digit a->buf // Top digit a->buf
// right-to-left to ensure already sorted digits order we keep for iterations // right-to-left to ensure already sorted digits order we keep for iterations
#pragma GCC unroll 48 #pragma GCC unroll 48
for(int i = n - 1; i >= 0; --i) { for(uint32_t i = n; i > 0; --i) {
// Prefetch caches // Prefetch caches
// __builtin_prefetch(&a[i-16]); // __builtin_prefetch(&a[i-16]);
// Get num and its new offset / location // Get num and its new offset / location
auto num = a[i]; auto num = a[i - 1];
auto bkeyni = (num >> shr1) & mask1; auto bkeyni = (num >> shr1) & mask1;
auto offset = --bucket1[bkeyni]; auto offset = --bucket1[bkeyni];