prefix sum

This commit is contained in:
Richard Thier 2021-03-11 22:05:55 +01:00
parent f8d4f597c6
commit e076ab662b

View File

@ -9,7 +9,6 @@
* LICENCE: CC3 - look it up, you need to mention me but that is all
*/
#include <vector>
#include <cstdio>
#include <cstdint>
@ -27,14 +26,45 @@ namespace MagyarSort {
static constexpr int BITS_PER_DIGIT = 8; // "bit / helyiérték"
static constexpr int DIGIT_RANGE = 256; // "helyiérték állapottér"
template<int digitChoice>
template<int DIGIT_CHOICE>
static inline uint32_t getDigit(uint32_t num) noexcept {
static constexpr int SHIFT = digitChoice * BITS_PER_DIGIT;
static constexpr int SHIFT = DIGIT_CHOICE * BITS_PER_DIGIT;
uint32_t shifted = num >> SHIFT;
return shifted & (DIGIT_RANGE - 1);
}
static inline void calcOccurences(uint32_t arr[], size_t size, size_t *radicsOut) {
for(size_t i = 0; i < size; ++i) {
auto d0 = getDigit<0>(arr[i]);
auto d1 = getDigit<1>(arr[i]);
auto d2 = getDigit<2>(arr[i]);
auto d3 = getDigit<3>(arr[i]);
/* // DEBUG:
printf("d0:%u, arr[i]: %u\n", d0, arr[i]);
printf("d1:%u, arr[i]: %u\n", d1, arr[i]);
printf("d2:%u, arr[i]: %u\n", d2, arr[i]);
printf("d3:%u, arr[i]: %u\n", d3, arr[i]);
*/
++radicsOut[d0];
++radicsOut[d1 + DIGIT_RANGE * 1];
++radicsOut[d2 + DIGIT_RANGE * 2];
++radicsOut[d3 + DIGIT_RANGE * 3];
}
}
template<int DIGIT_CHOICE>
static inline void prefixSum(size_t *radics) {
static constexpr int DSTART = DIGIT_CHOICE * DIGIT_RANGE;
size_t prev = 0;
for(int i = DSTART; i < (DSTART + DIGIT_RANGE); ++i) {
radics[i] += prev;
prev = radics[i];
}
}
/** Sort the given array (in-place sorting) with the given size */
inline void sort(uint32_t arr[], size_t size) noexcept {
// Holds "digit" occurences, prefix sums, whatevers
@ -43,24 +73,17 @@ namespace MagyarSort {
for(int i = 0; i < (DIGITS * DIGIT_RANGE); ++i) { radics[i] = 0; }
// Calculate occurences of digits
for(size_t i = 0; i < size; ++i) {
auto d0 = getDigit<0>(arr[i]);
auto d1 = getDigit<1>(arr[i]);
auto d2 = getDigit<2>(arr[i]);
auto d3 = getDigit<3>(arr[i]);
calcOccurences(arr, size, radics);
printf("d0:%u, arr[i]: %u\n", d0, arr[i]);
printf("d1:%u, arr[i]: %u\n", d1, arr[i]);
printf("d2:%u, arr[i]: %u\n", d2, arr[i]);
printf("d3:%u, arr[i]: %u\n", d3, arr[i]);
// Calculate prefix sums
// TODO: Maybe should use better ILP here?
// but maybe this is more cache friendly?
prefixSum<0>(radics);
prefixSum<1>(radics);
prefixSum<2>(radics);
prefixSum<3>(radics);
++radics[d0];
++radics[d1 + DIGIT_RANGE * 1];
++radics[d2 + DIGIT_RANGE * 2];
++radics[d3 + DIGIT_RANGE * 3];
}
// TODO: remove debug stuffz
/* // DEBUG:
for(size_t j = 0; j < DIGITS; ++j) {
printf("d%d: ", j);
for(size_t i = 0; i < DIGIT_RANGE; ++i) {
@ -68,6 +91,7 @@ namespace MagyarSort {
}
printf("\n\n");
}
*/
}
};