Skip to content

Commit

Permalink
Process two consecutive lines at a time (#651)
Browse files Browse the repository at this point in the history 
Use a better hash function

Don't return index from temperature parsing
extra JVM args

Co-authored-by: Ian Preston <[email protected]>
  • Loading branch information
@ianopolousfast @ianopolous
ianopolousfast and ianopolous authored Jan 31, 2024
1 parent ab2a9a6 commit 0c5c228
Show file tree
Hide file tree
Showing 2 changed files with 65 additions and 87 deletions.
2 changes: 2 additions & 0 deletions calculate_average_ianopolousfast.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,4 +16,6 @@
#

JAVA_OPTS="--enable-preview --add-modules=jdk.incubator.vector"
#-Djdk.incubator.vector.VECTOR_ACCESS_OOB_CHECK=0 -XX:-UseTransparentHugePages"

java $JAVA_OPTS --class-path target/average-1.0.0-SNAPSHOT.jar dev.morling.onebrc.CalculateAverage_ianopolousfast
150 changes: 63 additions & 87 deletions src/main/java/dev/morling/onebrc/CalculateAverage_ianopolousfast.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,6 @@
import jdk.incubator.vector.VectorOperators;
import jdk.incubator.vector.VectorSpecies;

import java.io.IOException;
import java.lang.foreign.Arena;
import java.lang.foreign.MemorySegment;
import java.nio.ByteOrder;
Expand All @@ -39,10 +38,7 @@
* * read chunks in parallel
* * minimise allocation
* * no unsafe
*
* Timings on 4 core i7-7500U CPU @ 2.70GHz:
* average_baseline: 4m48s
* ianopolous: 13.8s
* * process multiple lines in each thread for better ILP
*/
public class CalculateAverage_ianopolousfast {

Expand Down Expand Up @@ -91,11 +87,22 @@ public static boolean matchingStationBytes(long start, long end, MemorySegment b
return true;
}

private static int hashToIndex(long hash, int len) {
// From Thomas Wuerthinger's entry
int hashAsInt = (int) (hash ^ (hash >>> 28));
int finalHash = (hashAsInt ^ (hashAsInt >>> 15));
return (finalHash & (len - 1));
private static final int GOLDEN_RATIO = 0x9E3779B9;
private static final int HASH_LROTATE = 5;

// hash from giovannicuccu
private static int hash(MemorySegment memorySegment, long start, int len) {
int x;
int y;
if (len >= Integer.BYTES) {
x = memorySegment.get(JAVA_INT_UNALIGNED, start);
y = memorySegment.get(JAVA_INT_UNALIGNED, start + len - Integer.BYTES);
}
else {
x = memorySegment.get(JAVA_BYTE, start);
y = memorySegment.get(JAVA_BYTE, start + len - Byte.BYTES);
}
return (Integer.rotateLeft(x * GOLDEN_RATIO, HASH_LROTATE) ^ y) * GOLDEN_RATIO;
}

public static Stat createStation(long start, long end, MemorySegment buffer) {
Expand All @@ -105,8 +112,9 @@ public static Stat createStation(long start, long end, MemorySegment buffer) {
return new Stat(stationBuffer);
}

public static Stat dedupeStation(long start, long end, long hash, MemorySegment buffer, Stat[] stations) {
int index = hashToIndex(hash, MAX_STATIONS);
public static Stat dedupeStation(long start, long end, MemorySegment buffer, Stat[] stations) {
int hash = hash(buffer, start, (int) (end - start));
int index = hash & (MAX_STATIONS - 1);
Stat match = stations[index];
while (match != null) {
if (matchingStationBytes(start, end, buffer, match))
Expand All @@ -119,37 +127,11 @@ public static Stat dedupeStation(long start, long end, long hash, MemorySegment
return res;
}

static long maskHighBytes(long d, int nbytes) {
return d & (-1L << ((8 - nbytes) * 8));
}

public static Stat parseStation(long lineStart, MemorySegment buffer, Stat[] stations) {
ByteVector line = ByteVector.fromMemorySegment(BYTE_SPECIES, buffer, lineStart, ByteOrder.nativeOrder());
int keySize = line.compare(VectorOperators.EQ, ';').firstTrue();

long first8 = buffer.get(LONG_LAYOUT, lineStart);
long second8 = 0;
if (keySize <= 8) {
first8 = maskHighBytes(first8, keySize & 0x07);
}
else if (keySize < 16) {
second8 = maskHighBytes(buffer.get(LONG_LAYOUT, lineStart + 8), keySize & 0x07);
}
else if (keySize == BYTE_SPECIES.vectorByteSize()) {
while (buffer.get(JAVA_BYTE, lineStart + keySize) != ';') {
keySize++;
}
second8 = maskHighBytes(buffer.get(LONG_LAYOUT, lineStart + 8), keySize & 0x07);
}
long hash = first8 ^ second8; // todo include later bytes
return dedupeStation(lineStart, lineStart + keySize, hash, buffer, stations);
}

public static short getMinus(long d) {
return ((d & 0xff00000000000000L) ^ 0x2d00000000000000L) != 0 ? 0 : (short) -1;
}

public static long processTemperature(long lineSplit, int size, MemorySegment buffer, Stat station) {
public static void processTemperature(long lineSplit, int size, MemorySegment buffer, Stat station) {
long d = buffer.get(LONG_LAYOUT, lineSplit);
// negative is either 0 or -1
short negative = getMinus(d);
Expand All @@ -162,10 +144,9 @@ public static long processTemperature(long lineSplit, int size, MemorySegment bu
100 * (((byte) (d >> 24)) - '0'));
temperature = (short) ((temperature ^ negative) - negative); // negative treatment inspired by merkitty
station.add(temperature);
return lineSplit + size + 1;
}

private static long parseLine(long lineStart, MemorySegment buffer, Stat[] stations) {
private static int lineSize(long lineStart, MemorySegment buffer) {
ByteVector line = ByteVector.fromMemorySegment(BYTE_SPECIES, buffer, lineStart, ByteOrder.nativeOrder());
int lineSize = line.compare(VectorOperators.EQ, '\n').firstTrue();
int index = lineSize;
Expand All @@ -174,33 +155,19 @@ private static long parseLine(long lineStart, MemorySegment buffer, Stat[] stati
ByteOrder.nativeOrder()).compare(VectorOperators.EQ, '\n').firstTrue();
lineSize += index;
}
int keySize = lineSize - 6 + ByteVector.fromMemorySegment(BYTE_SPECIES, buffer, lineStart + lineSize - 6,
ByteOrder.nativeOrder()).compare(VectorOperators.EQ, ';').firstTrue();
return lineSize;
}

long first8 = buffer.get(LONG_LAYOUT, lineStart);
long second8 = 0;
if (keySize <= 8) {
first8 = maskHighBytes(first8, keySize & 0x07);
}
else if (keySize < 16) {
second8 = maskHighBytes(buffer.get(LONG_LAYOUT, lineStart + 8), keySize & 0x07);
}
else if (keySize == BYTE_SPECIES.vectorByteSize()) {
while (buffer.get(JAVA_BYTE, lineStart + keySize) != ';') {
keySize++;
}
second8 = maskHighBytes(buffer.get(LONG_LAYOUT, lineStart + 8), keySize & 0x07);
}
long hash = first8 ^ second8; // todo include later bytes
Stat station = dedupeStation(lineStart, lineStart + keySize, hash, buffer, stations);
return processTemperature(lineStart + keySize + 1, lineSize - keySize - 1, buffer, station);
private static int keySize(int lineSize, long lineStart, MemorySegment buffer) {
return lineSize - 6 + ByteVector.fromMemorySegment(BYTE_SPECIES, buffer, lineStart + lineSize - 6,
ByteOrder.nativeOrder()).compare(VectorOperators.EQ, ';').firstTrue();
}

public static Stat[] parseStats(long startByte, long endByte, MemorySegment buffer) {
public static Stat[] parseStats(long start1, long end2, MemorySegment buffer) {
// read first partial line
if (startByte > 0) {
if (start1 > 0) {
for (int i = 0; i < MAX_LINE_LENGTH; i++) {
byte b = buffer.get(JAVA_BYTE, startByte++);
byte b = buffer.get(JAVA_BYTE, start1++);
if (b == '\n') {
break;
}
Expand All @@ -213,38 +180,47 @@ public static Stat[] parseStats(long startByte, long endByte, MemorySegment buff
// this allows us to not worry about reading beyond the end
// in the inner loop (reducing branches)
// We need at least the vector lane size bytes back
if (endByte == buffer.byteSize()) {
if (end2 == buffer.byteSize()) {
// reverse at least vector lane width
endByte = Math.max(buffer.byteSize() - BYTE_SPECIES.vectorByteSize(), 0);
while (endByte > 0 && buffer.get(JAVA_BYTE, endByte) != '\n')
endByte--;
end2 = Math.max(buffer.byteSize() - 2 * BYTE_SPECIES.vectorByteSize(), 0);
while (end2 > 0 && buffer.get(JAVA_BYTE, end2) != '\n')
end2--;

if (endByte > 0)
endByte++;
if (end2 > 0)
end2++;
// copy into a larger buffer to avoid reading off end
MemorySegment end = Arena.global().allocate(MAX_LINE_LENGTH + BYTE_SPECIES.vectorByteSize());
for (long i = endByte; i < buffer.byteSize(); i++)
end.set(JAVA_BYTE, i - endByte, buffer.get(JAVA_BYTE, i));
MemorySegment end = Arena.global().allocate(MAX_LINE_LENGTH + 2 * BYTE_SPECIES.vectorByteSize());
for (long i = end2; i < buffer.byteSize(); i++)
end.set(JAVA_BYTE, i - end2, buffer.get(JAVA_BYTE, i));
int index = 0;
while (endByte + index < buffer.byteSize()) {
Stat station = parseStation(index, end, stations);
int tempSize = 3;
if (end.get(JAVA_BYTE, index + station.namelen + 5) == '\n')
tempSize = 4;
if (end.get(JAVA_BYTE, index + station.namelen + 6) == '\n')
tempSize = 5;
index = (int) processTemperature(index + station.namelen + 1, tempSize, end, station);
while (end2 + index < buffer.byteSize()) {
int lineSize1 = lineSize(index, end);
int semiSearchStart = index + Math.max(0, lineSize1 - 6);
int keySize1 = semiSearchStart - index + ByteVector.fromMemorySegment(BYTE_SPECIES, end, semiSearchStart,
ByteOrder.nativeOrder()).compare(VectorOperators.EQ, ';').firstTrue();
Stat station1 = dedupeStation(index, index + keySize1, end, stations);
processTemperature(index + keySize1 + 1, lineSize1 - keySize1 - 1, end, station1);
index += lineSize1 + 1;
}
}

innerloop(startByte, endByte, buffer, stations);
return stations;
}

private static void innerloop(long startByte, long endByte, MemorySegment buffer, Stat[] stations) {
while (startByte < endByte) {
startByte = parseLine(startByte, buffer, stations);
while (start1 < end2) {
int lineSize1 = lineSize(start1, buffer);
long start2 = start1 + lineSize1 + 1;
int lineSize2 = start2 < end2 ? lineSize(start2, buffer) : 0;
int keySize1 = keySize(lineSize1, start1, buffer);
int keySize2 = keySize(lineSize2, start2, buffer);
Stat station1 = dedupeStation(start1, start1 + keySize1, buffer, stations);
processTemperature(start1 + keySize1 + 1, lineSize1 - keySize1 - 1, buffer, station1);
if (start2 < end2) {
Stat station2 = dedupeStation(start2, start2 + keySize2, buffer, stations);
processTemperature(start2 + keySize2 + 1, lineSize2 - keySize2 - 1, buffer, station2);
start1 = start2 + lineSize2 + 1;
}
else
start1 += lineSize1 + 1;
}
return stations;
}

public static class Stat {
Expand Down

0 comments on commit 0c5c228

Please sign in to comment.