Playground/1brc
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

low collision + fast mixer, more optimization, less if because if is slow (#474)

Browse Source
This commit is contained in:
Van Phu DO 2024-01-20 06:03:51 +09:00 committed by GitHub
parent 586def3620
commit e67920f4af
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 44 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

102
src/main/java/dev/morling/onebrc/CalculateAverage_abeobk.java
View File

@ -60,14 +60,15 @@ public class CalculateAverage_abeobk {
static class Node { static class Node {
long addr; long addr;
long word0;
long tail; long tail;
int keylen;
int min, max; int min, max;
int count; int count;
long sum; long sum;
String key() { String key() {
byte[] sbuf = new byte[MAX_STR_LEN]; byte[] sbuf = new byte[MAX_STR_LEN];
int keylen = (int) (tail >>> 56);
UNSAFE.copyMemory(null, addr, sbuf, Unsafe.ARRAY_BYTE_BASE_OFFSET, keylen); UNSAFE.copyMemory(null, addr, sbuf, Unsafe.ARRAY_BYTE_BASE_OFFSET, keylen);
return new String(sbuf, 0, keylen, StandardCharsets.UTF_8); return new String(sbuf, 0, keylen, StandardCharsets.UTF_8);
} }
@ -76,18 +77,29 @@ public class CalculateAverage_abeobk {
return String.format("%.1f/%.1f/%.1f", min * 0.1, sum * 0.1 / count, max * 0.1); return String.format("%.1f/%.1f/%.1f", min * 0.1, sum * 0.1 / count, max * 0.1);
} }
Node(long a, long t, int val) { Node(long a, long t, int val, int kl) {
addr = a; addr = a;
tail = t; tail = t;
keylen = kl;
sum = min = max = val; sum = min = max = val;
count = 1; count = 1;
} }
Node(long a, long t, int val, int kl, long w0) {
this(a, t, val, kl);
word0 = w0;
}
void add(int val) { void add(int val) {
min = Math.min(min, val);
max = Math.max(max, val);
sum += val; sum += val;
count++; count++;
if (val >= max) {
max = val;
return;
}
if (val < min) {
min = val;
}
} }
void merge(Node other) { void merge(Node other) {
@ -102,7 +114,7 @@ public class CalculateAverage_abeobk {
return false; return false;
// this is faster than comparision if key is short // this is faster than comparision if key is short
long xsum = 0; long xsum = 0;
int n = ((int) (tail >>> 56)) & 0xF8; int n = keylen & 0xF8;
for (int i = 0; i < n; i += 8) { for (int i = 0; i < n; i += 8) {
xsum |= (UNSAFE.getLong(addr + i) ^ UNSAFE.getLong(other_addr + i)); xsum |= (UNSAFE.getLong(addr + i) ^ UNSAFE.getLong(other_addr + i));
} }
@ -130,21 +142,13 @@ public class CalculateAverage_abeobk {
return (xor_semi - 0x0101010101010101L) & (~xor_semi & 0x8080808080808080L); return (xor_semi - 0x0101010101010101L) & (~xor_semi & 0x8080808080808080L);
} }
// very low collision mixer // speed/collision balance
// idea from https://github.com/Cyan4973/xxHash/tree/dev
// zero collision on test data
static final int xxh32(long hash) { static final int xxh32(long hash) {
final int p1 = 0x85EBCA77; // prime final int p1 = 0x85EBCA77; // prime
final int p2 = 0x165667B1; // prime
int low = (int) hash; int low = (int) hash;
int high = (int) (hash >>> 31); int high = (int) (hash >>> 33);
int h = low + high; int h = (low * p1) ^ high;
h ^= h >> 15; return h ^ (h >>> 17);
h *= p1;
h ^= h >> 13;
h *= p2;
h ^= h >> 11;
return h;
} }
// great idea from merykitty (Quan Anh Mai) // great idea from merykitty (Quan Anh Mai)
@ -172,26 +176,23 @@ public class CalculateAverage_abeobk {
int val = 0; int val = 0;
int bucket = 0; int bucket = 0;
long word = UNSAFE.getLong(addr); long word0 = UNSAFE.getLong(addr);
long semipos_code = getSemiPosCode(word); long semipos_code = getSemiPosCode(word0);
// about 50% chance key < 8 chars // about 50% chance key < 8 chars
if (semipos_code != 0) { if (semipos_code != 0) {
int semi_pos = Long.numberOfTrailingZeros(semipos_code) >>> 3; int semi_pos = Long.numberOfTrailingZeros(semipos_code) >>> 3;
addr += semi_pos; addr += semi_pos;
tail = (word & HASH_MASKS[semi_pos]); tail = (word0 & HASH_MASKS[semi_pos]);
bucket = xxh32(tail) & BUCKET_MASK; bucket = xxh32(tail) & BUCKET_MASK;
long keylen = (addr - row_addr);
tail |= (keylen << 56);
long num_word = UNSAFE.getLong(++addr); long num_word = UNSAFE.getLong(++addr);
int dot_pos = Long.numberOfTrailingZeros(~num_word & 0x10101000); int dot_pos = Long.numberOfTrailingZeros(~num_word & 0x10101000);
val = parseNum(num_word, dot_pos); val = parseNum(num_word, dot_pos);
addr += (dot_pos >>> 3) + 3; addr += (dot_pos >>> 3) + 3;
while (true) { while (true) {
var node = map[bucket]; var node = map[bucket];
if (node == null) { if (node == null) {
map[bucket] = new Node(row_addr, tail, val); map[bucket] = new Node(row_addr, tail, val, semi_pos);
break; break;
} }
if (node.tail == tail) { if (node.tail == tail) {
@ -205,26 +206,30 @@ public class CalculateAverage_abeobk {
continue; continue;
} }
hash ^= word; hash ^= word0;
addr += 8; addr += 8;
word = UNSAFE.getLong(addr); long word = UNSAFE.getLong(addr);
semipos_code = getSemiPosCode(word); semipos_code = getSemiPosCode(word);
// frist byte semicolon ~13% // 43% chance
if (semipos_code == 0x80) { if (semipos_code != 0) {
int semi_pos = Long.numberOfTrailingZeros(semipos_code) >>> 3;
addr += semi_pos;
tail = (word & HASH_MASKS[semi_pos]);
hash ^= tail;
bucket = xxh32(hash) & BUCKET_MASK; bucket = xxh32(hash) & BUCKET_MASK;
tail = 8L << 56; int keylen = (int) (addr - row_addr);
long num_word = word >>> 8; long num_word = UNSAFE.getLong(++addr);
int dot_pos = Long.numberOfTrailingZeros(~num_word & 0x10101000); int dot_pos = Long.numberOfTrailingZeros(~num_word & 0x10101000);
val = parseNum(num_word, dot_pos); val = parseNum(num_word, dot_pos);
addr += (dot_pos >>> 3) + 4; addr += (dot_pos >>> 3) + 3;
while (true) { while (true) {
var node = map[bucket]; var node = map[bucket];
if (node == null) { if (node == null) {
map[bucket] = new Node(row_addr, tail, val); map[bucket] = new Node(row_addr, tail, val, keylen, word0);
break; break;
} }
if (UNSAFE.getLong(node.addr) == UNSAFE.getLong(row_addr)) { if (node.word0 == word0 && node.tail == tail) {
node.add(val); node.add(val);
break; break;
} }
@ -247,38 +252,18 @@ public class CalculateAverage_abeobk {
tail = (word & HASH_MASKS[semi_pos]); tail = (word & HASH_MASKS[semi_pos]);
hash ^= tail; hash ^= tail;
bucket = xxh32(hash) & BUCKET_MASK; bucket = xxh32(hash) & BUCKET_MASK;
long keylen = (addr - row_addr); int keylen = (int) (addr - row_addr);
tail |= (keylen << 56);
long num_word = UNSAFE.getLong(++addr);
++addr;
long num_word = UNSAFE.getLong(addr);
int dot_pos = Long.numberOfTrailingZeros(~num_word & 0x10101000); int dot_pos = Long.numberOfTrailingZeros(~num_word & 0x10101000);
val = parseNum(num_word, dot_pos); val = parseNum(num_word, dot_pos);
addr += (dot_pos >>> 3) + 3; addr += (dot_pos >>> 3) + 3;
if (keylen < 16) {
while (true) {
var node = map[bucket];
if (node == null) {
map[bucket] = new Node(row_addr, tail, val);
break;
}
if (node.tail == tail && (UNSAFE.getLong(node.addr) == UNSAFE.getLong(row_addr))) {
node.add(val);
break;
}
bucket++;
if (SHOW_ANALYSIS)
cls[thread_id]++;
}
continue;
}
// longer key
while (true) { while (true) {
var node = map[bucket]; var node = map[bucket];
if (node == null) { if (node == null) {
map[bucket] = new Node(row_addr, tail, val); map[bucket] = new Node(row_addr, tail, val, keylen);
break; break;
} }
if (node.contentEquals(row_addr, tail)) { if (node.contentEquals(row_addr, tail)) {
@ -335,7 +320,7 @@ public class CalculateAverage_abeobk {
if (node == null) if (node == null)
continue; continue;
if (SHOW_ANALYSIS) { if (SHOW_ANALYSIS) {
int kl = (int) (node.tail >>> 56) & (lenhist.length - 1); int kl = node.keylen & (lenhist.length - 1);
lenhist[kl] += node.count; lenhist[kl] += node.count;
} }
var stat = ms.putIfAbsent(node.key(), node); var stat = ms.putIfAbsent(node.key(), node);
@ -353,4 +338,5 @@ public class CalculateAverage_abeobk {
System.out.println(ms); System.out.println(ms);
} }
} }
} }