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

Likely better ILP and no manual digit counts in code

Browse Source
This commit is contained in:
Richard Thier 2021-03-11 23:04:50 +01:00
parent 22e80d4cd5
commit 151b8f398b
1 changed files with 29 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

49
magyarsort.h
View File

@ -11,6 +11,7 @@
#include <cstdio> #include <cstdio>
#include <cstdint> #include <cstdint>
#include <cstring> // memset
namespace MagyarSort { namespace MagyarSort {
@ -34,21 +35,19 @@ namespace MagyarSort {
return shifted & (DIGIT_RANGE - 1); return shifted & (DIGIT_RANGE - 1);
} }
/** Functor: no class should be generated I think (compiler should be smart) */ /** Recursive Functor: no class should be generated I think (compiler should be smart) */
template<int DIGIT> template<int DIGIT>
struct OccurenceMagic : public OccurenceMagic<DIGIT - 1> { struct OccurenceMagic : public OccurenceMagic<DIGIT - 1> {
inline OccurenceMagic(uint32_t arr[], size_t i, size_t *radicsOut) noexcept inline OccurenceMagic(uint32_t arr[], size_t i, size_t *radicsOut) noexcept
: OccurenceMagic<DIGIT -1 >(arr, i, radicsOut) { : OccurenceMagic<DIGIT -1 >(arr, i, radicsOut) {
// parents run first so template recursion runs DIGIT=0 first... // Parents run first so template recursion runs DIGIT=0 first...
++radicsOut[getDigit<DIGIT>(arr[i]) + DIGIT_RANGE * DIGIT]; ++radicsOut[getDigit<DIGIT>(arr[i]) + DIGIT_RANGE * DIGIT];
} }
}; };
/** Ends template recursion */ /** Ends template recursion */
template<> template<>
struct OccurenceMagic<-1> { struct OccurenceMagic<-1> {
inline OccurenceMagic(uint32_t arr[], size_t i, size_t *radicsOut) noexcept { inline OccurenceMagic(uint32_t arr[], size_t i, size_t *radicsOut) noexcept {}
/* empty */
}
}; };
static inline void countOccurences(uint32_t arr[], size_t size, size_t *radicsOut) noexcept { static inline void countOccurences(uint32_t arr[], size_t size, size_t *radicsOut) noexcept {
@ -58,13 +57,29 @@ namespace MagyarSort {
} }
} }
template<int DIGIT_CHOICE> /** Recursive Functor: no class should be generated I think (compiler should be smart) */
static inline void prefixSum(size_t *radics) noexcept { template<int DIGIT>
static constexpr int DSTART = DIGIT_CHOICE * DIGIT_RANGE; struct PrefixMagic : public PrefixMagic<DIGIT - 1> {
size_t prev = 0; inline PrefixMagic(size_t *radics, size_t *prev, int i) noexcept
for(int i = DSTART; i < (DSTART + DIGIT_RANGE); ++i) { : PrefixMagic<DIGIT - 1>(radics, prev, i) {
radics[i] += prev; static constexpr int DSTART = (DIGIT * DIGIT_RANGE);
prev = radics[i]; radics[DSTART + i] += prev[DIGIT];
prev[DIGIT] = radics[DSTART + i];
}
};
/** Ends template recursion */
template<>
struct PrefixMagic<-1> {
inline PrefixMagic(size_t *radics, size_t *prev, int i) noexcept {}
};
static inline void calcPrefixSums(size_t *radics) noexcept {
static thread_local size_t prev[DIGITS];
memset(prev, 0, sizeof(prev));
for(int i = 0; i < DIGIT_RANGE; ++i) {
// This is a template-unrolled loop too
PrefixMagic<DIGITS - 1>(radics, prev, i);
} }
} }
@ -73,19 +88,13 @@ namespace MagyarSort {
// Holds "digit" occurences, prefix sums, whatevers // Holds "digit" occurences, prefix sums, whatevers
// First "DIGIT_RANGE" elem is for MSB "DIGITS", last is for LSB // First "DIGIT_RANGE" elem is for MSB "DIGITS", last is for LSB
static thread_local size_t radics[DIGITS * DIGIT_RANGE]; static thread_local size_t radics[DIGITS * DIGIT_RANGE];
for(int i = 0; i < (DIGITS * DIGIT_RANGE); ++i) { radics[i] = 0; } memset(radics, 0, sizeof(radics));
// Calculate occurences of digits // Calculate occurences of digits
countOccurences(arr, size, radics); countOccurences(arr, size, radics);
// Calculate prefix sums // Calculate prefix sums
// TODO: Maybe should use better ILP here? calcPrefixSums(radics);
// but maybe this is more cache friendly?
// TODO: manual digits!
prefixSum<0>(radics);
prefixSum<1>(radics);
prefixSum<2>(radics);
prefixSum<3>(radics);
/* // DEBUG: /* // DEBUG:
*/ */