thier3 / tpxb: int->uint32, but this loses a little perf because likely compiler uses the UB of signed overflow to optimize out stuff so will be reverted as it is not a practical thing anyways
This commit is contained in:
parent
c032109110
commit
808b87f266
44
thiersort3.h
44
thiersort3.h
@ -52,13 +52,13 @@ static inline uint32_t witch_bucket3(uint32_t key) {
|
|||||||
* @param n Number of elements in arr and temparr
|
* @param n Number of elements in arr and temparr
|
||||||
* @param rstate Create with sch_rand_state rstate = schwab_rand_state(junk_uint32_t);
|
* @param rstate Create with sch_rand_state rstate = schwab_rand_state(junk_uint32_t);
|
||||||
*/
|
*/
|
||||||
static inline void thiersort3(uint32_t *arr, uint32_t *temparr, int n) {
|
static inline void thiersort3(uint32_t *arr, uint32_t *temparr, uint32_t n) {
|
||||||
int bucket[256]; /* Inclusive */
|
uint32_t bucket[256]; /* Inclusive */
|
||||||
int bucket_end[256]; /* Not inclusive */
|
uint32_t bucket_end[256]; /* Not inclusive */
|
||||||
|
|
||||||
#ifndef NO_EXTRA_BIT
|
#ifndef NO_EXTRA_BIT
|
||||||
int bucket_leftend[256]; /* for extra 1bit split processing */
|
uint32_t bucket_leftend[256]; /* for extra 1bit split processing */
|
||||||
int bucket_left[256]; /* for extra 1bit split processing */
|
uint32_t bucket_left[256]; /* for extra 1bit split processing */
|
||||||
#endif /* NO_EXTRA_BIT */
|
#endif /* NO_EXTRA_BIT */
|
||||||
|
|
||||||
/* Check if need to sort at all - needed for invariants later */
|
/* Check if need to sort at all - needed for invariants later */
|
||||||
@ -68,25 +68,25 @@ static inline void thiersort3(uint32_t *arr, uint32_t *temparr, int n) {
|
|||||||
|
|
||||||
/* Count */
|
/* Count */
|
||||||
#pragma GCC unroll 64
|
#pragma GCC unroll 64
|
||||||
for(int i = 0; i < 256; ++i) {
|
for(uint32_t i = 0; i < 256; ++i) {
|
||||||
bucket[i] = 0;
|
bucket[i] = 0;
|
||||||
}
|
}
|
||||||
#pragma GCC unroll 128
|
#pragma GCC unroll 128
|
||||||
for(int i = 0; i < n; ++i) {
|
for(uint32_t i = 0; i < n; ++i) {
|
||||||
++bucket[witch_bucket3(arr[i])];
|
++bucket[witch_bucket3(arr[i])];
|
||||||
}
|
}
|
||||||
|
|
||||||
/* Prefix sum (like in Magyarsort) */
|
/* Prefix sum (like in Magyarsort) */
|
||||||
uint32_t prev = 0;
|
uint32_t prev = 0;
|
||||||
#pragma GCC unroll 4
|
#pragma GCC unroll 4
|
||||||
for (int i = 0; i < 256; i++) {
|
for (uint32_t i = 0; i < 256; i++) {
|
||||||
bucket[i] += prev;
|
bucket[i] += prev;
|
||||||
prev = bucket[i];
|
prev = bucket[i];
|
||||||
}
|
}
|
||||||
|
|
||||||
/* Save end-offsets */
|
/* Save end-offsets */
|
||||||
#pragma GCC unroll 64
|
#pragma GCC unroll 64
|
||||||
for(int i = 0; i < 256; ++i) {
|
for(uint32_t i = 0; i < 256; ++i) {
|
||||||
bucket_end[i] = bucket[i];
|
bucket_end[i] = bucket[i];
|
||||||
}
|
}
|
||||||
|
|
||||||
@ -95,7 +95,7 @@ static inline void thiersort3(uint32_t *arr, uint32_t *temparr, int n) {
|
|||||||
bucket_left[0] = 0;
|
bucket_left[0] = 0;
|
||||||
bucket_leftend[0] = 0;
|
bucket_leftend[0] = 0;
|
||||||
#pragma GCC unroll 64
|
#pragma GCC unroll 64
|
||||||
for(int i = 0; i < 4095; ++i) {
|
for(uint32_t i = 0; i < 255; ++i) {
|
||||||
bucket_left[1 + i] = bucket[i];
|
bucket_left[1 + i] = bucket[i];
|
||||||
bucket_leftend[1 + i] = bucket[i];
|
bucket_leftend[1 + i] = bucket[i];
|
||||||
}
|
}
|
||||||
@ -106,18 +106,18 @@ static inline void thiersort3(uint32_t *arr, uint32_t *temparr, int n) {
|
|||||||
/* Move to the buckets */
|
/* Move to the buckets */
|
||||||
/* Rem.: This also changes bucket[i] so they will point to bucket beginnings */
|
/* Rem.: This also changes bucket[i] so they will point to bucket beginnings */
|
||||||
#pragma GCC unroll 128
|
#pragma GCC unroll 128
|
||||||
for(int i = 0; i < n; ++i) {
|
for(uint32_t i = 0; i < n; ++i) {
|
||||||
uint32_t num = arr[i];
|
uint32_t num = arr[i];
|
||||||
uint32_t witch = witch_bucket3(num);
|
uint32_t witch = witch_bucket3(num);
|
||||||
int offset = (--bucket[witch]);
|
uint32_t offset = (--bucket[witch]);
|
||||||
temparr[offset] = num;
|
temparr[offset] = num;
|
||||||
}
|
}
|
||||||
|
|
||||||
/* temparr -> arr each bucket and sort them in-place */
|
/* temparr -> arr each bucket and sort them in-place */
|
||||||
#pragma GCC unroll 2
|
#pragma GCC unroll 2
|
||||||
for(int b = 0; b < 256; ++b) {
|
for(uint32_t b = 0; b < 256; ++b) {
|
||||||
int begin = bucket[b];
|
uint32_t begin = bucket[b];
|
||||||
int end = bucket_end[b];
|
uint32_t end = bucket_end[b];
|
||||||
|
|
||||||
/* Ensure exists */
|
/* Ensure exists */
|
||||||
if(begin >= end) {
|
if(begin >= end) {
|
||||||
@ -137,10 +137,10 @@ static inline void thiersort3(uint32_t *arr, uint32_t *temparr, int n) {
|
|||||||
/* Rem.: This also changes bucket[i] so they will point to bucket.right beginnings */
|
/* Rem.: This also changes bucket[i] so they will point to bucket.right beginnings */
|
||||||
/* Rem.: This also changes bucket_leftend[i] so they will point to bucket.left endings (needed in-process only) */
|
/* Rem.: This also changes bucket_leftend[i] so they will point to bucket.left endings (needed in-process only) */
|
||||||
#pragma GCC unroll 64
|
#pragma GCC unroll 64
|
||||||
for(int i = 0; i < n; ++i) {
|
for(uint32_t i = 0; i < n; ++i) {
|
||||||
uint32_t num = arr[i];
|
uint32_t num = arr[i];
|
||||||
uint32_t witch = witch_bucket3(num);
|
uint32_t witch = witch_bucket3(num);
|
||||||
int offset = (num & (1 << 27)) ?
|
uint32_t offset = (num & (1 << 27)) ?
|
||||||
(--bucket[witch]) :
|
(--bucket[witch]) :
|
||||||
(bucket_leftend[witch]++);
|
(bucket_leftend[witch]++);
|
||||||
|
|
||||||
@ -149,12 +149,12 @@ static inline void thiersort3(uint32_t *arr, uint32_t *temparr, int n) {
|
|||||||
|
|
||||||
/* temparr -> arr each bucket and sort them in-place */
|
/* temparr -> arr each bucket and sort them in-place */
|
||||||
#pragma GCC unroll 2
|
#pragma GCC unroll 2
|
||||||
for(int b = 0; b < 256; ++b) {
|
for(uint32_t b = 0; b < 256; ++b) {
|
||||||
assert(bucket_leftend[b] == bucket[b]);
|
assert(bucket_leftend[b] == bucket[b]);
|
||||||
int lbegin = bucket_left[b];
|
uint32_t lbegin = bucket_left[b];
|
||||||
int lend = bucket[b]; /* non-inclusive */
|
uint32_t lend = bucket[b]; /* non-inclusive */
|
||||||
int rbegin = bucket[b];
|
uint32_t rbegin = bucket[b];
|
||||||
int rend = bucket_end[b]; /* non-inclusive */
|
uint32_t rend = bucket_end[b]; /* non-inclusive */
|
||||||
|
|
||||||
/* Ensure exists and process left part */
|
/* Ensure exists and process left part */
|
||||||
if(lbegin < lend) {
|
if(lbegin < lend) {
|
||||||
|
|||||||
@ -31,17 +31,14 @@ static inline constexpr uint32_t min3u32_xb(uint32_t a, uint32_t b, uint32_t c)
|
|||||||
* @param buf Result array with the same size - result will be here
|
* @param buf Result array with the same size - result will be here
|
||||||
* @param n The number of elements
|
* @param n The number of elements
|
||||||
*/
|
*/
|
||||||
static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
|
static inline void threepass_xb(uint32_t *a, uint32_t *buf, uint32_t n) noexcept {
|
||||||
assert(buf != NULL);
|
assert(buf != NULL);
|
||||||
constexpr int shr1 = TPBX3 + TPBX2;
|
constexpr uint32_t shr1 = TPBX3 + TPBX2;
|
||||||
constexpr int shr2 = TPBX3;
|
constexpr uint32_t shr2 = TPBX3;
|
||||||
constexpr int shr3 = 0;
|
constexpr uint32_t shr3 = 0;
|
||||||
constexpr int mask1 = (1 << TPBX1) - 1;
|
constexpr uint32_t mask1 = (1 << TPBX1) - 1;
|
||||||
constexpr int mask2 = (1 << TPBX2) - 1;
|
constexpr uint32_t mask2 = (1 << TPBX2) - 1;
|
||||||
constexpr int mask3 = (1 << TPBX3) - 1;
|
constexpr uint32_t mask3 = (1 << TPBX3) - 1;
|
||||||
|
|
||||||
/* helper buffers. */
|
|
||||||
int sz = n * sizeof(a[0]);
|
|
||||||
|
|
||||||
static thread_local uint32_t bucket1[1 << TPBX1];
|
static thread_local uint32_t bucket1[1 << TPBX1];
|
||||||
memset(bucket1, 0, (1 << TPBX1) * sizeof(uint32_t));
|
memset(bucket1, 0, (1 << TPBX1) * sizeof(uint32_t));
|
||||||
@ -67,7 +64,7 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
|
|||||||
(1 << TPBX2),
|
(1 << TPBX2),
|
||||||
(1 << TPBX3)
|
(1 << TPBX3)
|
||||||
);
|
);
|
||||||
int i = 0;
|
uint32_t i = 0;
|
||||||
#pragma GCC unroll 8
|
#pragma GCC unroll 8
|
||||||
for (; i < common; ++i) {
|
for (; i < common; ++i) {
|
||||||
bucket1[i] += prev1;
|
bucket1[i] += prev1;
|
||||||
@ -78,17 +75,17 @@ static inline void threepass_xb(uint32_t *a, uint32_t *buf, int n) noexcept {
|
|||||||
prev3 = bucket3[i];
|
prev3 = bucket3[i];
|
||||||
}
|
}
|
||||||
/* Do remaining 1 */
|
/* Do remaining 1 */
|
||||||
for (int j = i; j < (1 << TPBX1); ++j) {
|
for (uint32_t j = i; j < (1 << TPBX1); ++j) {
|
||||||
bucket1[j] += prev1;
|
bucket1[j] += prev1;
|
||||||
prev1 = bucket1[j];
|
prev1 = bucket1[j];
|
||||||
}
|
}
|
||||||
/* Do remaining 2 */
|
/* Do remaining 2 */
|
||||||
for (int j = i; j< (1 << TPBX2); ++j) {
|
for (uint32_t j = i; j< (1 << TPBX2); ++j) {
|
||||||
bucket2[j] += prev2;
|
bucket2[j] += prev2;
|
||||||
prev2 = bucket2[j];
|
prev2 = bucket2[j];
|
||||||
}
|
}
|
||||||
/* Do remaining 3 */
|
/* Do remaining 3 */
|
||||||
for (int j = i; j < (1 << TPBX3); ++j) {
|
for (uint32_t j = i; j < (1 << TPBX3); ++j) {
|
||||||
bucket3[j] += prev3;
|
bucket3[j] += prev3;
|
||||||
prev3 = bucket3[j];
|
prev3 = bucket3[j];
|
||||||
}
|
}
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user