Add int4 vector scoring benchmarks

benchmarks/src/main/java/org/elasticsearch/benchmark/vector/scorer/BenchmarkUtils.java

@@ -11,6 +11,7 @@
 
 import org.apache.lucene.backward_codecs.lucene99.OffHeapQuantizedByteVectorValues;
 import org.apache.lucene.codecs.hnsw.FlatVectorScorerUtil;
+import org.apache.lucene.codecs.lucene104.Lucene104ScalarQuantizedVectorScorer;
 import org.apache.lucene.codecs.lucene95.OffHeapByteVectorValues;
 import org.apache.lucene.codecs.lucene95.OffHeapFloatVectorValues;
 import org.apache.lucene.codecs.lucene99.Lucene99ScalarQuantizedVectorScorer;
@@ -134,6 +135,21 @@ static RandomVectorScorer luceneScorer(QuantizedByteVectorValues values, VectorS
         return new Lucene99ScalarQuantizedVectorScorer(null).getRandomVectorScorer(sim, values, queryVec);
     }
 
+    static RandomVectorScorerSupplier lucene104ScoreSupplier(
+        org.apache.lucene.codecs.lucene104.QuantizedByteVectorValues values,
+        VectorSimilarityFunction sim
+    ) throws IOException {
+        return new Lucene104ScalarQuantizedVectorScorer(null).getRandomVectorScorerSupplier(sim, values);
+    }
+
+    static RandomVectorScorer lucene104Scorer(
+        org.apache.lucene.codecs.lucene104.QuantizedByteVectorValues values,
+        VectorSimilarityFunction sim,
+        float[] queryVec
+    ) throws IOException {
+        return new Lucene104ScalarQuantizedVectorScorer(null).getRandomVectorScorer(sim, values, queryVec);
+    }
+
     static RuntimeException rethrow(Throwable t) {
         if (t instanceof Error err) {
             throw err;

benchmarks/src/main/java/org/elasticsearch/benchmark/vector/scorer/Int4BenchmarkUtils.java

@@ -0,0 +1,199 @@
+/*
+ * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one
+ * or more contributor license agreements. Licensed under the "Elastic License
+ * 2.0", the "GNU Affero General Public License v3.0 only", and the "Server Side
+ * Public License v 1"; you may not use this file except in compliance with, at
+ * your election, the "Elastic License 2.0", the "GNU Affero General Public
+ * License v3.0 only", or the "Server Side Public License, v 1".
+ */
+
+package org.elasticsearch.benchmark.vector.scorer;
+
+import org.apache.lucene.codecs.lucene104.Lucene104ScalarQuantizedVectorsFormat;
+import org.apache.lucene.codecs.lucene104.QuantizedByteVectorValues;
+import org.apache.lucene.index.VectorSimilarityFunction;
+import org.apache.lucene.search.VectorScorer;
+import org.apache.lucene.util.VectorUtil;
+import org.apache.lucene.util.hnsw.RandomVectorScorer;
+import org.apache.lucene.util.hnsw.UpdateableRandomVectorScorer;
+import org.apache.lucene.util.quantization.OptimizedScalarQuantizer;
+
+import java.io.IOException;
+import java.util.Arrays;
+import java.util.concurrent.ThreadLocalRandom;
+
+import static org.elasticsearch.benchmark.vector.scorer.ScalarOperations.applyI4Corrections;
+import static org.elasticsearch.benchmark.vector.scorer.ScalarOperations.dotProductI4SinglePacked;
+import static org.elasticsearch.simdvec.internal.vectorization.VectorScorerTestUtils.unpackNibbles;
+
+public class Int4BenchmarkUtils {
+
+    /**
+     * In-memory implementation of {@link QuantizedByteVectorValues} for int4 (PACKED_NIBBLE) benchmarks.
+     * Stores pre-quantized packed nibble vectors with synthetic corrective terms.
+     */
+    static class InMemoryInt4QuantizedByteVectorValues extends QuantizedByteVectorValues {
+
+        private final int dims;
+        private final byte[][] packedVectors;
+        private final OptimizedScalarQuantizer.QuantizationResult[] correctiveTerms;
+        private final float[] centroid;
+        private final float centroidDP;
+        private final OptimizedScalarQuantizer quantizer;
+
+        InMemoryInt4QuantizedByteVectorValues(
+            int dims,
+            byte[][] packedVectors,
+            OptimizedScalarQuantizer.QuantizationResult[] correctiveTerms,
+            float[] centroid,
+            float centroidDP
+        ) {
+            this.dims = dims;
+            this.packedVectors = packedVectors;
+            this.correctiveTerms = correctiveTerms;
+            this.centroid = centroid;
+            this.centroidDP = centroidDP;
+            this.quantizer = new OptimizedScalarQuantizer(VectorSimilarityFunction.DOT_PRODUCT);
+        }
+
+        @Override
+        public int dimension() {
+            return dims;
+        }
+
+        @Override
+        public int size() {
+            return packedVectors.length;
+        }
+
+        @Override
+        public byte[] vectorValue(int ord) throws IOException {
+            return packedVectors[ord];
+        }
+
+        @Override
+        public OptimizedScalarQuantizer.QuantizationResult getCorrectiveTerms(int vectorOrd) throws IOException {
+            return correctiveTerms[vectorOrd];
+        }
+
+        @Override
+        public OptimizedScalarQuantizer getQuantizer() {
+            return quantizer;
+        }
+
+        @Override
+        public Lucene104ScalarQuantizedVectorsFormat.ScalarEncoding getScalarEncoding() {
+            return Lucene104ScalarQuantizedVectorsFormat.ScalarEncoding.PACKED_NIBBLE;
+        }
+
+        @Override
+        public float[] getCentroid() throws IOException {
+            return centroid;
+        }
+
+        @Override
+        public float getCentroidDP() throws IOException {
+            return centroidDP;
+        }
+
+        @Override
+        public VectorScorer scorer(float[] query) throws IOException {
+            return null;
+        }
+
+        @Override
+        public InMemoryInt4QuantizedByteVectorValues copy() throws IOException {
+            return new InMemoryInt4QuantizedByteVectorValues(dims, packedVectors, correctiveTerms, centroid, centroidDP);
+        }
+    }
+
+    private static class ScalarScorer implements UpdateableRandomVectorScorer {
+        private final QuantizedByteVectorValues values;
+        private final int dims;
+        private final VectorSimilarityFunction similarityFunction;
+
+        private byte[] queryUnpacked;
+        private OptimizedScalarQuantizer.QuantizationResult queryCorrections;
+
+        ScalarScorer(QuantizedByteVectorValues values, VectorSimilarityFunction similarityFunction) {
+            this.values = values;
+            this.dims = values.dimension();
+            this.similarityFunction = similarityFunction;
+        }
+
+        @Override
+        public float score(int node) throws IOException {
+            byte[] packed = values.vectorValue(node);
+            int rawDot = dotProductI4SinglePacked(queryUnpacked, packed);
+            var nodeCorrections = values.getCorrectiveTerms(node);
+            return applyI4Corrections(rawDot, dims, nodeCorrections, queryCorrections, values.getCentroidDP(), similarityFunction);
+        }
+
+        @Override
+        public int maxOrd() {
+            return values.size();
+        }
+
+        @Override
+        public void setScoringOrdinal(int node) throws IOException {
+            byte[] packed = values.vectorValue(node);
+            queryUnpacked = unpackNibbles(packed, dims);
+            queryCorrections = values.getCorrectiveTerms(node);
+        }
+    }
+
+    static QuantizedByteVectorValues createI4QuantizedVectorValues(int dims, byte[][] packedVectors) {
+        var random = ThreadLocalRandom.current();
+        var correctiveTerms = new OptimizedScalarQuantizer.QuantizationResult[packedVectors.length];
+        for (int i = 0; i < packedVectors.length; i++) {
+            correctiveTerms[i] = new OptimizedScalarQuantizer.QuantizationResult(
+                random.nextFloat(-1f, 1f),
+                random.nextFloat(-1f, 1f),
+                random.nextFloat(-1f, 1f),
+                random.nextInt(0, dims * 15)
+            );
+        }
+        float[] centroid = new float[dims];
+        for (int i = 0; i < dims; i++) {
+            centroid[i] = random.nextFloat();
+        }
+        float centroidDP = random.nextFloat();
+        return new InMemoryInt4QuantizedByteVectorValues(dims, packedVectors, correctiveTerms, centroid, centroidDP);
+    }
+
+    static UpdateableRandomVectorScorer createI4ScalarScorer(
+        QuantizedByteVectorValues values,
+        VectorSimilarityFunction similarityFunction
+    ) {
+        return new ScalarScorer(values, similarityFunction);
+    }
+
+    static RandomVectorScorer createI4ScalarQueryScorer(
+        QuantizedByteVectorValues values,
+        VectorSimilarityFunction similarityFunction,
+        float[] queryVector
+    ) throws IOException {
+        int dims = values.dimension();
+        OptimizedScalarQuantizer quantizer = values.getQuantizer();
+        float[] centroid = values.getCentroid();
+        Lucene104ScalarQuantizedVectorsFormat.ScalarEncoding encoding = values.getScalarEncoding();
+
+        byte[] queryQuantized = new byte[encoding.getDiscreteDimensions(dims)];
+        float[] queryCopy = Arrays.copyOf(queryVector, queryVector.length);
+        if (similarityFunction == VectorSimilarityFunction.COSINE) {
+            VectorUtil.l2normalize(queryCopy);
+        }
+        var queryCorrections = quantizer.scalarQuantize(queryCopy, queryQuantized, encoding.getQueryBits(), centroid);
+        float centroidDP = values.getCentroidDP();
+
+        return new RandomVectorScorer.AbstractRandomVectorScorer(values) {
+            @Override
+            public float score(int node) throws IOException {
+                byte[] packed = values.vectorValue(node);
+                int rawDot = dotProductI4SinglePacked(queryQuantized, packed);
+                var nodeCorrections = values.getCorrectiveTerms(node);
+                return applyI4Corrections(rawDot, dims, nodeCorrections, queryCorrections, centroidDP, similarityFunction);
+            }
+        };
+    }
+}

benchmarks/src/main/java/org/elasticsearch/benchmark/vector/scorer/ScalarOperations.java

@@ -9,6 +9,13 @@
 
 package org.elasticsearch.benchmark.vector.scorer;
 
+import org.apache.lucene.index.VectorSimilarityFunction;
+import org.apache.lucene.util.VectorUtil;
+import org.apache.lucene.util.quantization.OptimizedScalarQuantizer;
+
+import static org.apache.lucene.index.VectorSimilarityFunction.EUCLIDEAN;
+import static org.apache.lucene.index.VectorSimilarityFunction.MAXIMUM_INNER_PRODUCT;
+
 /**
  * Basic scalar implementations of similarity operations.
  * <p>
@@ -17,6 +24,8 @@
  */
 class ScalarOperations {
 
+    private static final float FOUR_BIT_SCALE = 1f / ((1 << 4) - 1);
+
     static float cosine(byte[] a, byte[] b) {
         int sum = 0;
         int norm1 = 0;
@@ -65,4 +74,54 @@ static int squareDistance(byte[] a, byte[] b) {
         }
         return res;
     }
+
+    static int dotProductI4SinglePacked(byte[] unpacked, byte[] packed) {
+        int total = 0;
+        for (int i = 0; i < packed.length; i++) {
+            byte packedByte = packed[i];
+            byte unpacked1 = unpacked[i];
+            byte unpacked2 = unpacked[i + packed.length];
+            total += (packedByte & 0x0F) * unpacked2;
+            total += ((packedByte & 0xFF) >> 4) * unpacked1;
+        }
+        return total;
+    }
+
+    public static float applyI4Corrections(
+        int rawDot,
+        int dims,
+        OptimizedScalarQuantizer.QuantizationResult nodeCorrections,
+        OptimizedScalarQuantizer.QuantizationResult queryCorrections,
+        float centroidDP,
+        VectorSimilarityFunction similarityFunction
+    ) {
+        float ax = nodeCorrections.lowerInterval();
+        float lx = (nodeCorrections.upperInterval() - ax) * FOUR_BIT_SCALE;
+        float ay = queryCorrections.lowerInterval();
+        float ly = (queryCorrections.upperInterval() - ay) * FOUR_BIT_SCALE;
+        float x1 = nodeCorrections.quantizedComponentSum();
+        float y1 = queryCorrections.quantizedComponentSum();
+
+        float score = ax * ay * dims + ay * lx * x1 + ax * ly * y1 + lx * ly * rawDot;
+
+        // For euclidean, we need to invert the score and apply the additional correction, which is
+        // assumed to be the squared l2norm of the centroid centered vectors.
+        if (similarityFunction == EUCLIDEAN) {
+            score = queryCorrections.additionalCorrection() + nodeCorrections.additionalCorrection() - 2 * score;
+            // Ensure that 'score' (the squared euclidean distance) is non-negative. The computed value
+            // may be negative as a result of quantization loss.
+            return VectorUtil.normalizeDistanceToUnitInterval(Math.max(score, 0f));
+        } else {
+            // For cosine and max inner product, we need to apply the additional correction, which is
+            // assumed to be the non-centered dot-product between the vector and the centroid
+            score += queryCorrections.additionalCorrection() + nodeCorrections.additionalCorrection() - centroidDP;
+            if (similarityFunction == MAXIMUM_INNER_PRODUCT) {
+                return VectorUtil.scaleMaxInnerProductScore(score);
+            }
+            // Ensure that 'score' (a normalized dot product) is in [-1,1]. The computed value may be out
+            // of bounds as a result of quantization loss.
+            score = Math.clamp(score, -1, 1);
+            return (1f + score) / 2f;
+        }
+    }
 }