From 29d1ec1053694b1c81ec9ae8bb676fa9e084d891 Mon Sep 17 00:00:00 2001
From: John Suykerbuyk <john@syketech.com>
Date: Wed, 8 Apr 2026 17:28:59 -0600
Subject: [PATCH] perf: use vek32 SIMD for EuclideanDistance

EuclideanDistance used a plain Go loop while CosineDistance already used
vek32 with AVX2 SIMD acceleration. Replace the loop with vek32.Distance
which provides the same SIMD optimization.

Benchmark at 1536 dimensions (OpenAI default) on AMD Ryzen 7 7745HX:

  Before: EuclideanDistance  335 ns/op
  After:  EuclideanDistance   47 ns/op  (7.1x faster)
  Cosine: CosineDistance      80 ns/op  (unchanged, already SIMD)

Also adds BenchmarkEuclideanDistance to distance_test.go.
---
 distance.go      | 9 +--------
 distance_test.go | 9 +++++++++
 2 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/distance.go b/distance.go
index 4ab7cad..86024dc 100644
--- a/distance.go
+++ b/distance.go
@@ -1,7 +1,6 @@
 package hnsw
 
 import (
-	"math"
 	"reflect"
 
 	"github.com/viterin/vek/vek32"
@@ -17,13 +16,7 @@ func CosineDistance(a, b []float32) float32 {
 
 // EuclideanDistance computes the Euclidean distance between two vectors.
 func EuclideanDistance(a, b []float32) float32 {
-	// TODO: can we speedup with vek?
-	var sum float32 = 0
-	for i := range a {
-		diff := a[i] - b[i]
-		sum += diff * diff
-	}
-	return float32(math.Sqrt(float64(sum)))
+	return vek32.Distance(a, b)
 }
 
 var distanceFuncs = map[string]DistanceFunc{
diff --git a/distance_test.go b/distance_test.go
index e411ee5..11c2c94 100644
--- a/distance_test.go
+++ b/distance_test.go
@@ -30,6 +30,15 @@ func TestCosineSimilarity(t *testing.T) {
 	require.InDelta(t, 0, CosineDistance(a, b), 0.000001)
 }
 
+func BenchmarkEuclideanDistance(b *testing.B) {
+	v1 := randFloats(1536)
+	v2 := randFloats(1536)
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		EuclideanDistance(v1, v2)
+	}
+}
+
 func BenchmarkCosineSimilarity(b *testing.B) {
 	v1 := randFloats(1536)
 	v2 := randFloats(1536)