|
| 1 | +/* |
| 2 | + * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one |
| 3 | + * or more contributor license agreements. Licensed under the "Elastic License |
| 4 | + * 2.0", the "GNU Affero General Public License v3.0 only", and the "Server Side |
| 5 | + * Public License v 1"; you may not use this file except in compliance with, at |
| 6 | + * your election, the "Elastic License 2.0", the "GNU Affero General Public |
| 7 | + * License v3.0 only", or the "Server Side Public License, v 1". |
| 8 | + */ |
| 9 | +package org.elasticsearch.benchmark.vector; |
| 10 | + |
| 11 | +import org.apache.lucene.index.VectorSimilarityFunction; |
| 12 | +import org.apache.lucene.store.Directory; |
| 13 | +import org.apache.lucene.store.IOContext; |
| 14 | +import org.apache.lucene.store.IndexInput; |
| 15 | +import org.apache.lucene.store.IndexOutput; |
| 16 | +import org.apache.lucene.store.MMapDirectory; |
| 17 | +import org.apache.lucene.util.quantization.OptimizedScalarQuantizer; |
| 18 | +import org.elasticsearch.common.logging.LogConfigurator; |
| 19 | +import org.elasticsearch.core.IOUtils; |
| 20 | +import org.elasticsearch.simdvec.ES91Int4VectorsScorer; |
| 21 | +import org.elasticsearch.simdvec.ES92Int7VectorsScorer; |
| 22 | +import org.elasticsearch.simdvec.internal.vectorization.ESVectorizationProvider; |
| 23 | +import org.openjdk.jmh.annotations.Benchmark; |
| 24 | +import org.openjdk.jmh.annotations.BenchmarkMode; |
| 25 | +import org.openjdk.jmh.annotations.Fork; |
| 26 | +import org.openjdk.jmh.annotations.Measurement; |
| 27 | +import org.openjdk.jmh.annotations.Mode; |
| 28 | +import org.openjdk.jmh.annotations.OutputTimeUnit; |
| 29 | +import org.openjdk.jmh.annotations.Param; |
| 30 | +import org.openjdk.jmh.annotations.Scope; |
| 31 | +import org.openjdk.jmh.annotations.Setup; |
| 32 | +import org.openjdk.jmh.annotations.State; |
| 33 | +import org.openjdk.jmh.annotations.TearDown; |
| 34 | +import org.openjdk.jmh.annotations.Warmup; |
| 35 | +import org.openjdk.jmh.infra.Blackhole; |
| 36 | + |
| 37 | +import java.io.IOException; |
| 38 | +import java.nio.file.Files; |
| 39 | +import java.util.concurrent.ThreadLocalRandom; |
| 40 | +import java.util.concurrent.TimeUnit; |
| 41 | + |
| 42 | +@BenchmarkMode(Mode.Throughput) |
| 43 | +@OutputTimeUnit(TimeUnit.MILLISECONDS) |
| 44 | +@State(Scope.Benchmark) |
| 45 | +// first iteration is complete garbage, so make sure we really warmup |
| 46 | +@Warmup(iterations = 4, time = 1) |
| 47 | +// real iterations. not useful to spend tons of time here, better to fork more |
| 48 | +@Measurement(iterations = 5, time = 1) |
| 49 | +// engage some noise reduction |
| 50 | +@Fork(value = 1) |
| 51 | +public class Int7ScorerBenchmark { |
| 52 | + |
| 53 | + static { |
| 54 | + LogConfigurator.configureESLogging(); // native access requires logging to be initialized |
| 55 | + } |
| 56 | + |
| 57 | + @Param({ "384", "782", "1024" }) |
| 58 | + int dims; |
| 59 | + |
| 60 | + int numVectors = 20 * ES92Int7VectorsScorer.BULK_SIZE; |
| 61 | + int numQueries = 5; |
| 62 | + |
| 63 | + byte[] scratch; |
| 64 | + byte[][] binaryVectors; |
| 65 | + byte[][] binaryQueries; |
| 66 | + float[] scores = new float[ES92Int7VectorsScorer.BULK_SIZE]; |
| 67 | + |
| 68 | + ES92Int7VectorsScorer scorer; |
| 69 | + Directory dir; |
| 70 | + IndexInput in; |
| 71 | + |
| 72 | + OptimizedScalarQuantizer.QuantizationResult queryCorrections; |
| 73 | + float centroidDp; |
| 74 | + |
| 75 | + @Setup |
| 76 | + public void setup() throws IOException { |
| 77 | + binaryVectors = new byte[numVectors][dims]; |
| 78 | + dir = new MMapDirectory(Files.createTempDirectory("vectorData")); |
| 79 | + try (IndexOutput out = dir.createOutput("vectors", IOContext.DEFAULT)) { |
| 80 | + for (byte[] binaryVector : binaryVectors) { |
| 81 | + for (int i = 0; i < dims; i++) { |
| 82 | + // 4-bit quantization |
| 83 | + binaryVector[i] = (byte) ThreadLocalRandom.current().nextInt(128); |
| 84 | + } |
| 85 | + out.writeBytes(binaryVector, 0, binaryVector.length); |
| 86 | + ThreadLocalRandom.current().nextBytes(binaryVector); |
| 87 | + out.writeBytes(binaryVector, 0, 16); // corrections |
| 88 | + } |
| 89 | + } |
| 90 | + |
| 91 | + queryCorrections = new OptimizedScalarQuantizer.QuantizationResult( |
| 92 | + ThreadLocalRandom.current().nextFloat(), |
| 93 | + ThreadLocalRandom.current().nextFloat(), |
| 94 | + ThreadLocalRandom.current().nextFloat(), |
| 95 | + Short.toUnsignedInt((short) ThreadLocalRandom.current().nextInt()) |
| 96 | + ); |
| 97 | + centroidDp = ThreadLocalRandom.current().nextFloat(); |
| 98 | + |
| 99 | + in = dir.openInput("vectors", IOContext.DEFAULT); |
| 100 | + binaryQueries = new byte[numVectors][dims]; |
| 101 | + for (byte[] binaryVector : binaryVectors) { |
| 102 | + for (int i = 0; i < dims; i++) { |
| 103 | + // 7-bit quantization |
| 104 | + binaryVector[i] = (byte) ThreadLocalRandom.current().nextInt(128); |
| 105 | + } |
| 106 | + } |
| 107 | + |
| 108 | + scratch = new byte[dims]; |
| 109 | + scorer = ESVectorizationProvider.getInstance().newES92Int7VectorsScorer(in, dims); |
| 110 | + } |
| 111 | + |
| 112 | + @TearDown |
| 113 | + public void teardown() throws IOException { |
| 114 | + IOUtils.close(dir, in); |
| 115 | + } |
| 116 | + |
| 117 | + @Benchmark |
| 118 | + @Fork(jvmArgsPrepend = { "--add-modules=jdk.incubator.vector" }) |
| 119 | + public void scoreFromMemorySegment(Blackhole bh) throws IOException { |
| 120 | + for (int j = 0; j < numQueries; j++) { |
| 121 | + in.seek(0); |
| 122 | + for (int i = 0; i < numVectors; i++) { |
| 123 | + bh.consume( |
| 124 | + scorer.score( |
| 125 | + binaryQueries[j], |
| 126 | + queryCorrections.lowerInterval(), |
| 127 | + queryCorrections.upperInterval(), |
| 128 | + queryCorrections.quantizedComponentSum(), |
| 129 | + queryCorrections.additionalCorrection(), |
| 130 | + VectorSimilarityFunction.EUCLIDEAN, |
| 131 | + centroidDp |
| 132 | + ) |
| 133 | + ); |
| 134 | + } |
| 135 | + } |
| 136 | + } |
| 137 | + |
| 138 | + @Benchmark |
| 139 | + @Fork(jvmArgsPrepend = { "--add-modules=jdk.incubator.vector" }) |
| 140 | + public void scoreFromMemorySegmentBulk(Blackhole bh) throws IOException { |
| 141 | + for (int j = 0; j < numQueries; j++) { |
| 142 | + in.seek(0); |
| 143 | + for (int i = 0; i < numVectors; i += ES91Int4VectorsScorer.BULK_SIZE) { |
| 144 | + scorer.scoreBulk( |
| 145 | + binaryQueries[j], |
| 146 | + queryCorrections.lowerInterval(), |
| 147 | + queryCorrections.upperInterval(), |
| 148 | + queryCorrections.quantizedComponentSum(), |
| 149 | + queryCorrections.additionalCorrection(), |
| 150 | + VectorSimilarityFunction.EUCLIDEAN, |
| 151 | + centroidDp, |
| 152 | + scores |
| 153 | + ); |
| 154 | + for (float score : scores) { |
| 155 | + bh.consume(score); |
| 156 | + } |
| 157 | + } |
| 158 | + } |
| 159 | + } |
| 160 | +} |
0 commit comments