diff --git a/ypsu.cpp b/ypsu.cpp index 0f778d7..81d579f 100644 --- a/ypsu.cpp +++ b/ypsu.cpp @@ -150,9 +150,6 @@ static inline uint32_t morgrab(uint32_t elem) noexcept { } template static inline void mormord_sort_impl(uint32_t *a, int n) noexcept { - /* Edge-case */ - if(n == 0) return; - /* Preparation */ uint32_t radics[256] = {0}; uint32_t radics2[256] = {0}; @@ -187,41 +184,34 @@ static inline void mormord_sort_impl(uint32_t *a, int n) noexcept { uint32_t reali = 0; #pragma GCC unroll 16 for(int i = 0; i < 256; ++i) { - radics[i] += prev; if(radics[i] != 0) { + radics[i] += prev; real_radics[reali] = prev; real_radics[reali + 1] = radics[i]; + prev = radics[i]; reali += 2; + } else { + radics[i] += prev; + prev = radics[i]; } - prev = radics[i]; } - - // Inplace swap - own ideas + some ideas based-on "famous" ct-swap (for branchless / more ILP): - // void ct-swap(bool secret, uint64_t a[], uint64_t b[], size_t len) { - // uint64_t mask = ~((uint64_t)secret - 1); // 1->111....111; 0->000....000 - // for (size_t i = 0; i < len; i++) { - // uint64_t delta = (a[i] ^ b[i]) & mask; - // a[i] = a[i] ^ delta; // b[i], ha secret - amúgy a[i] - // b[i] = b[i] ^ delta; // fordítva - // } - // } + // Inplace swap uint32_t pivoti = 0; while(pivoti < n) { uint32_t radixval = morgrab(a[pivoti]); - uint32_t targeti = --radics[radixval]; // dec index (!) - - // Bitmask: true -> 11.....1; false -> 00.....0 - uint32_t mask = ~((targeti > pivoti) - 1); - - // Branchless swap (using bitmask) - uint32_t delta = (a[pivoti] ^ a[targeti]) & mask; - a[pivoti] = a[pivoti] ^ delta; - a[targeti] = a[targeti] ^ delta; - - // "else" branch - pivoti += !mask; - radics[radixval] += !mask; // undec index (!) + uint32_t targeti = radics[radixval] - 1; + if(targeti > pivoti) { + // swap + uint32_t tmp = a[pivoti]; + a[pivoti] = a[targeti]; + a[targeti] = tmp; + // dec index + --radics[radixval]; + } else { + // progress pivot + ++pivoti; + } } // Ends recursion