Performance: use parallel accumulators to speed up PanamaFloatVectorOps dotProduct and l2Distance (96% recall at 204 qps)

.github/workflows/ci.yaml

@@ -31,7 +31,7 @@ jobs:
         timeout-minutes: 1
         with:
           distribution: 'adopt'
-          java-version: 19
+          java-version: 21
           cache: 'sbt'
       - uses: arduino/setup-task@v1
         timeout-minutes: 1
@@ -78,7 +78,7 @@ jobs:
         timeout-minutes: 1
         with:
           distribution: 'adopt'
-          java-version: 19
+          java-version: 21
           cache: 'sbt'
       - uses: actions/setup-python@v4
         timeout-minutes: 1
@@ -122,7 +122,7 @@ jobs:
         timeout-minutes: 1
         with:
           distribution: 'adopt'
-          java-version: 19
+          java-version: 21
           cache: 'sbt'
       - uses: arduino/setup-task@v1
         timeout-minutes: 1

.github/workflows/release.yml

@@ -27,7 +27,7 @@ jobs:
       - uses: actions/setup-java@v3
         with:
           distribution: 'adopt'
-          java-version: 19
+          java-version: 21
           cache: 'sbt'
       - uses: arduino/setup-task@v1
         with:

.tool-versions

@@ -1,2 +1,2 @@
-java oracle-19.0.2
+java oracle-21.0.1
 python 3.10.8

docs/pages/performance/fashion-mnist/results.md

@@ -1,10 +1,10 @@
 |Model|Parameters|Recall|Queries per Second|
 |---|---|---|---|
-|eknn-l2lsh|L=100 k=4 w=1024 candidates=500 probes=0|0.378|349.851|
-|eknn-l2lsh|L=100 k=4 w=1024 candidates=1000 probes=0|0.446|296.219|
-|eknn-l2lsh|L=100 k=4 w=1024 candidates=500 probes=3|0.635|286.468|
-|eknn-l2lsh|L=100 k=4 w=1024 candidates=1000 probes=3|0.716|244.536|
-|eknn-l2lsh|L=100 k=4 w=2048 candidates=500 probes=0|0.767|315.023|
-|eknn-l2lsh|L=100 k=4 w=2048 candidates=1000 probes=0|0.847|264.479|
-|eknn-l2lsh|L=100 k=4 w=2048 candidates=500 probes=3|0.922|220.714|
-|eknn-l2lsh|L=100 k=4 w=2048 candidates=1000 probes=3|0.960|193.597|
+|eknn-l2lsh|L=100 k=4 w=1024 candidates=500 probes=0|0.378|377.213|
+|eknn-l2lsh|L=100 k=4 w=1024 candidates=1000 probes=0|0.447|320.000|
+|eknn-l2lsh|L=100 k=4 w=1024 candidates=500 probes=3|0.635|302.384|
+|eknn-l2lsh|L=100 k=4 w=1024 candidates=1000 probes=3|0.717|259.201|
+|eknn-l2lsh|L=100 k=4 w=2048 candidates=500 probes=0|0.767|335.978|
+|eknn-l2lsh|L=100 k=4 w=2048 candidates=1000 probes=0|0.846|287.923|
+|eknn-l2lsh|L=100 k=4 w=2048 candidates=500 probes=3|0.921|229.917|
+|eknn-l2lsh|L=100 k=4 w=2048 candidates=1000 probes=3|0.960|204.375|

elastiknn-jmh-benchmarks/src/main/scala/com/klibisz/elastiknn/jmhbenchmarks/FloatVectorOpsBenchmark.scala

@@ -0,0 +1,116 @@
+package com.klibisz.elastiknn.jmhbenchmarks
+
+import com.klibisz.elastiknn.api.Vec
+import com.klibisz.elastiknn.vectors._
+import org.apache.lucene.util.VectorUtil
+import org.openjdk.jmh.annotations._
+
+import scala.util.Random
+
+@State(Scope.Benchmark)
+class FloatVectorOpsBenchmarkState {
+  implicit private val rng: Random = new Random(0)
+  val v1 = Vec.DenseFloat.random(999).values
+  val v2 = Vec.DenseFloat.random(999).values
+  val panama = new PanamaFloatVectorOps
+  val default = new DefaultFloatVectorOps
+}
+
+class FloatVectorOpsBenchmark {
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def cosineSimilarityPanama(state: FloatVectorOpsBenchmarkState): Double =
+    state.panama.cosineSimilarity(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def cosineSimilarityDefault(state: FloatVectorOpsBenchmarkState): Double =
+    state.default.cosineSimilarity(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def dotProductDefault(state: FloatVectorOpsBenchmarkState): Double =
+    state.default.dotProduct(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def dotProductPanama(state: FloatVectorOpsBenchmarkState): Double =
+    state.panama.dotProduct(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def dotProductPanamaSimple(state: FloatVectorOpsBenchmarkState): Double =
+    state.panama.dotProductSimple(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def dotProductLucene(state: FloatVectorOpsBenchmarkState): Float =
+    VectorUtil.dotProduct(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def l1DistancePanama(state: FloatVectorOpsBenchmarkState): Double =
+    state.panama.l1Distance(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def l1DistanceDefault(state: FloatVectorOpsBenchmarkState): Double =
+    state.default.l1Distance(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def l2DistancePanama(state: FloatVectorOpsBenchmarkState): Double =
+    state.panama.l2Distance(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def l2DistancePanamaSimple(state: FloatVectorOpsBenchmarkState): Double =
+    state.panama.l2DistanceSimple(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def l2DistanceDefault(state: FloatVectorOpsBenchmarkState): Double =
+    state.default.l2Distance(state.v1, state.v2)
+
+  @Benchmark
+  @BenchmarkMode(Array(Mode.Throughput))
+  @Fork(value = 1)
+  @Warmup(time = 1, iterations = 5)
+  @Measurement(time = 1, iterations = 5)
+  def l2DistanceLucene(state: FloatVectorOpsBenchmarkState): Double =
+    Math.sqrt(VectorUtil.squareDistance(state.v1, state.v2) * 1d)
+}