Skip to content

Commit

Permalink
8341003: [lworld+fp16] Benchmarks for various Float16 operations
Browse files Browse the repository at this point in the history 
Reviewed-by: bkilambi
  • Loading branch information
Jatin Bhateja committed Sep 27, 2024
1 parent fb5b1b1 commit 0ce9f0f
Showing 1 changed file with 235 additions and 0 deletions.
235 changes: 235 additions & 0 deletions test/micro/org/openjdk/bench/java/lang/Float16OpsBenchmark.java
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,235 @@
/*
* Copyright (c) 2024, Oracle and/or its affiliates. All rights vectorReserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood ShovectorRes, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*/
package org.openjdk.bench.java.lang;

import java.util.stream.IntStream;
import java.util.concurrent.TimeUnit;
import org.openjdk.jmh.annotations.*;

@OutputTimeUnit(TimeUnit.MILLISECONDS)
@State(Scope.Thread)
@Fork(jvmArgsPrepend = {"--enable-preview", "-Xbatch", "-XX:-TieredCompilation"})
public class Float16OpsBenchmark {
@Param({"256", "512", "1024", "2048"})
int vectorDim;

int [] rexp;
Float16 [] vectorRes;
Float16 [] vector1;
Float16 [] vector2;
Float16 [] vector3;

@Setup(Level.Trial)
public void BmSetup() {
rexp = new int[vectorDim];
vectorRes = new Float16[vectorDim];
vector1 = new Float16[vectorDim];
vector2 = new Float16[vectorDim];
vector3 = new Float16[vectorDim];

IntStream.range(0, vectorDim).forEach(i -> {vector1[i] = Float16.valueOf((short)i);});
IntStream.range(0, vectorDim).forEach(i -> {vector2[i] = Float16.valueOf((short)i);});
IntStream.range(0, vectorDim).forEach(i -> {vector3[i] = Float16.valueOf((short)i);});

// Special Values
Float16 [] specialValues = {Float16.NaN, Float16.NEGATIVE_INFINITY, Float16.valueOf(0.0), Float16.valueOf(-0.0), Float16.POSITIVE_INFINITY};
IntStream.range(0, vectorDim).forEach(
i -> {
if ((i % 64) == 0) {
int idx1 = i % specialValues.length;
int idx2 = (i + 1) % specialValues.length;
int idx3 = (i + 2) % specialValues.length;
vector1[i] = specialValues[idx1];
vector2[i] = specialValues[idx2];
vector3[i] = specialValues[idx3];
}
}
);
}

@Benchmark
public void addBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.add(vector1[i], vector2[i]);
}
}

@Benchmark
public void subBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.subtract(vector1[i], vector2[i]);
}
}

@Benchmark
public void mulBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.multiply(vector1[i], vector2[i]);
}
}

@Benchmark
public void divBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.divide(vector1[i], vector2[i]);
}
}

@Benchmark
public void fmaBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.fma(vector1[i], vector2[i], vector3[i]);
}
}

@Benchmark
public boolean isInfiniteBenchmark() {
boolean res = true;
for (int i = 0; i < vectorDim; i++) {
res &= Float16.isInfinite(vector1[i]);
}
return res;
}

@Benchmark
public boolean isFiniteBenchmark() {
boolean res = true;
for (int i = 0; i < vectorDim; i++) {
res &= Float16.isFinite(vector1[i]);
}
return res;
}

@Benchmark
public boolean isNaNBenchmark() {
boolean res = true;
for (int i = 0; i < vectorDim; i++) {
res &= Float16.isNaN(vector1[i]);
}
return res;
}

@Benchmark
public void maxBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.max(vector1[i], vector2[i]);
}
}

@Benchmark
public void minBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.min(vector1[i], vector2[i]);
}
}

@Benchmark
public void sqrtBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.sqrt(vector1[i]);
}
}

@Benchmark
public void negateBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.negate(vector1[i]);
}
}

@Benchmark
public void absBenchmark() {
for (int i = 0; i < vectorDim; i++) {
vectorRes[i] = Float16.abs(vector1[i]);
}
}

@Benchmark
public void getExponentBenchmark() {
for (int i = 0; i < vectorDim; i++) {
rexp[i] = Float16.getExponent(vector1[i]);
}
}

@Benchmark
public Float16 cosineSimilarityDoubleRoundingFP16() {
Float16 macRes = Float16.valueOf(0.0f);
Float16 vector1Square = Float16.valueOf(0.0f);
Float16 vector2Square = Float16.valueOf(0.0f);
for (int i = 0; i < vectorDim; i++) {
// Explicit add and multiply operation ensures double rounding.
macRes = Float16.add(Float16.multiply(vector1[i], vector2[i]), macRes);
vector1Square = Float16.fma(vector1[i], vector1[i], vector1Square);
vector2Square = Float16.fma(vector2[i], vector2[i], vector2Square);
}
return Float16.divide(macRes, Float16.add(vector1Square, vector2Square));
}

@Benchmark
public Float16 cosineSimilaritySingleRoundingFP16() {
Float16 macRes = Float16.valueOf(0.0f);
Float16 vector1Square = Float16.valueOf(0.0f);
Float16 vector2Square = Float16.valueOf(0.0f);
for (int i = 0; i < vectorDim; i++) {
macRes = Float16.fma(vector1[i], vector2[i], macRes);
vector1Square = Float16.fma(vector1[i], vector1[i], vector1Square);
vector2Square = Float16.fma(vector2[i], vector2[i], vector2Square);
}
return Float16.divide(macRes, Float16.add(vector1Square, vector2Square));
}

@Benchmark
public Float16 cosineSimilarityDequantizedFP16() {
float macRes = 0.0f;
float vector1Square = 0.0f;
float vector2Square = 0.0f;
for (int i = 0; i < vectorDim; i++) {
macRes = Math.fma(vector1[i].floatValue(), vector2[i].floatValue(), macRes);
vector1Square = Math.fma(vector1[i].floatValue(), vector1[i].floatValue(), vector1Square);
vector2Square = Math.fma(vector2[i].floatValue(), vector2[i].floatValue(), vector2Square);
}
return Float16.valueOf(macRes / (vector1Square + vector2Square));
}

@Benchmark
public Float16 euclideanDistanceFP16() {
Float16 distRes = Float16.valueOf(0.0f);
Float16 squareRes = Float16.valueOf(0.0f);
for (int i = 0; i < vectorDim; i++) {
squareRes = Float16.subtract(vector1[i], vector2[i]);
distRes = Float16.fma(squareRes, squareRes, distRes);
}
return Float16.sqrt(distRes);
}

@Benchmark
public Float16 euclideanDistanceDequantizedFP16() {
float distRes = 0.0f;
float squareRes = 0.0f;
for (int i = 0; i < vectorDim; i++) {
squareRes = vector1[i].floatValue() - vector2[i].floatValue();
distRes = distRes + squareRes * squareRes;
}
return Float16.sqrt(Float16.valueOf(distRes));
}
}

0 comments on commit 0ce9f0f

Please sign in to comment.