Merge pull request #376 from rshura/knn_haversine

Use haversine instead of distance in knn if distance is not required.

Gopkg.lock

@@ -243,7 +243,7 @@
 
 [[projects]]
   branch = "master"
-  digest = "1:c5ac96e72d3ff6694602f3273dd71ef04a67c9591465aac92dc1aa8c821b8f91"
+  digest = "1:145703130ac1de36086ab350337777161f9c1d791e81a73659ac1f569e15b5e5"
   name = "github.com/tidwall/geojson"
   packages = [
     ".",
@@ -251,7 +251,7 @@
     "geometry",
   ]
   pruneopts = ""
-  revision = "32782c39ca84f98113436a297f14601e4fee527d"
+  revision = "dbcb73c57c65ff784ce2ccaad3f062c9787d6f81"
 
 [[projects]]
   digest = "1:3ddca2bd5496c6922a2a9e636530e178a43c2a534ea6634211acdc7d10222794"
@@ -467,6 +467,7 @@
     "github.com/tidwall/buntdb",
     "github.com/tidwall/evio",
     "github.com/tidwall/geojson",
+    "github.com/tidwall/geojson/geo",
     "github.com/tidwall/geojson/geometry",
     "github.com/tidwall/gjson",
     "github.com/tidwall/lotsa",

internal/server/search.go

@@ -10,6 +10,7 @@ import (
 
 	"github.com/mmcloughlin/geohash"
 	"github.com/tidwall/geojson"
+	"github.com/tidwall/geojson/geo"
 	"github.com/tidwall/geojson/geometry"
 	"github.com/tidwall/resp"
 	"github.com/tidwall/tile38/internal/bing"
@@ -371,14 +372,9 @@ func (server *Server) cmdNearby(msg *Message) (res resp.Value, err error) {
 	if sw.col != nil {
 		var matched uint32
 		iter := func(id string, o geojson.Object, fields []float64, dist *float64) bool {
-			// Calculate distance if we need to
 			distance := 0.0
 			if s.distance {
-				if dist != nil {
-					distance = *dist
-				} else {
-					distance = o.Distance(s.obj)
-				}
+				distance = geo.DistanceFromHaversine(*dist)
 			}
 			return sw.writeObject(ScanWriterParams{
 				id:              id,
@@ -411,6 +407,7 @@ func (server *Server) nearestNeighbors(
 	iter func(id string, o geojson.Object, fields []float64, dist *float64,
 	) bool) {
 	limit := int(sw.cursor + sw.limit)
+	maxDist := target.Haversine()
 	var items []iterItem
 	sw.col.Nearby(target, func(id string, o geojson.Object, fields []float64) bool {
 		if server.hasExpired(s.key, id) {
@@ -423,8 +420,8 @@ func (server *Server) nearestNeighbors(
 		if !match {
 			return true
 		}
-		dist := o.Distance(target)
-		if target.Meters() > 0 && dist > target.Meters() {
+		dist := target.HaversineTo(o.Center())
+		if maxDist > 0 && dist > maxDist {
 			return false
 		}
 		items = append(items, iterItem{id: id, o: o, fields: fields, dist: dist})

vendor/github.com/tidwall/geojson/circle.go

@@ -84,29 +84,37 @@ func (g *Circle) Center() geometry.Point {
 	return g.center
 }
 
+// Haversine returns the haversine corresponding to circle's radius
+func (g *Circle) Haversine() float64 {
+	return g.haversine
+}
+
+// HaversineTo returns the haversine from a given point to circle's center
+func (g *Circle) HaversineTo(p geometry.Point) float64 {
+	return geo.Haversine(p.Y, p.X, g.center.Y, g.center.X)
+}
+
 // Within returns true if circle is contained inside object
 func (g *Circle) Within(obj Object) bool {
 	return obj.Contains(g)
 }
 
-func (g *Circle) contains(p geometry.Point, allowOnEdge bool) bool {
+// containsPoint returns true if circle contains a given point
+func (g *Circle) containsPoint(p geometry.Point) bool {
 	h := geo.Haversine(p.Y, p.X, g.center.Y, g.center.X)
-	if allowOnEdge {
-		return h <= g.haversine
-	}
-	return h < g.haversine
+	return h <= g.haversine
 }
 
 // Contains returns true if the circle contains other object
 func (g *Circle) Contains(obj Object) bool {
 	switch other := obj.(type) {
 	case *Point:
-		return g.contains(other.Center(), false)
+		return g.containsPoint(other.Center())
 	case *Circle:
 		return other.Distance(g) < (other.meters + g.meters)
 	case *LineString:
 		for i := 0; i < other.base.NumPoints(); i++ {
-			if geoDistancePoints(other.base.PointAt(i), g.center) > g.meters {
+			if !g.containsPoint(other.base.PointAt(i)) {
 				return false
 			}
 		}
@@ -124,14 +132,13 @@ func (g *Circle) Contains(obj Object) bool {
 	}
 }
 
+// intersectsSegment returns true if the circle intersects a given segment
 func (g *Circle) intersectsSegment(seg geometry.Segment) bool {
 	start, end := seg.A, seg.B
 
-	// These are faster checks.  If they succeed there's no need do complicate things.
-	if g.contains(start, true) {
-		return true
-	}
-	if g.contains(end, true) {
+	// These are faster checks.
+	// If they succeed there's no need do complicate things.
+	if g.containsPoint(start) || g.containsPoint(end) {
 		return true
 	}
 
@@ -152,14 +159,14 @@ func (g *Circle) intersectsSegment(seg geometry.Segment) bool {
 	}
 
 	// Distance from the closest point to the center
-	return g.contains(geometry.Point{X: px, Y: py}, true)
+	return g.containsPoint(geometry.Point{X: px, Y: py})
 }
 
 // Intersects returns true the circle intersects other object
 func (g *Circle) Intersects(obj Object) bool {
 	switch other := obj.(type) {
 	case *Point:
-		return g.contains(other.Center(), true)
+		return g.containsPoint(other.Center())
 	case *Circle:
 		return other.Distance(g) <= (other.meters + g.meters)
 	case *LineString:

vendor/github.com/tidwall/geojson/circle_test.go

@@ -163,11 +163,11 @@ func TestCircleIntersects(t *testing.T) {
 // This snippet tests 100M comparisons.
 // On my box this takes 24.5s without haversine trick, and 13.7s with the trick.
 //
-//func TestCircle_Performance(t *testing.T) {
+//func TestCirclePerformance(t *testing.T) {
 //	g := NewCircle(P(-122.4412, 37.7335), 1000, 64)
 //	r := rand.New(rand.NewSource(42))
 //	for i:= 0; i < 100000000; i++ {
-//		g.Contains(PO((r.Float64() - 0.5) * 180, r.Float64() * 90))
+//		g.Contains(PO(r.Float64()*360 - 180, r.Float64()*180 - 90))
 //	}
 //	expect(t, true)
 //}

vendor/github.com/tidwall/geojson/geo/geo.go

@@ -45,11 +45,15 @@ func DistanceToHaversine(meters float64) float64 {
 	return sin * sin
 }
 
-// DistanceTo return the distance in meteres between two point.
+// DistanceFromHaversine...
+func DistanceFromHaversine(haversine float64) float64 {
+	return earthRadius * 2 * math.Asin(math.Sqrt(haversine))
+}
+
+// DistanceTo return the distance in meters between two point.
 func DistanceTo(latA, lonA, latB, lonB float64) (meters float64) {
 	a := Haversine(latA, lonA, latB, lonB)
-	c := 2 * math.Atan2(math.Sqrt(a), math.Sqrt(1-a))
-	return earthRadius * c
+	return DistanceFromHaversine(a)
 }
 
 // DestinationPoint return the destination from a point based on a

vendor/github.com/tidwall/geojson/geo/geo_test.go

@@ -77,15 +77,51 @@ func TestHaversine(t *testing.T) {
 
 func TestNormalizeDistance(t *testing.T) {
 	start := time.Now()
-	for time.Since(start) < time.Second/4 {
+	for time.Since(start) < time.Second {
 		for i := 0; i < 1000; i++ {
-			meters1 := rand.Float64() * 100000000
+			meters1 := rand.Float64() * earthRadius * 3 // wrap three times
 			meters2 := NormalizeDistance(meters1)
-			dist1 := math.Floor(DistanceToHaversine(meters2) * 100000000.0)
-			dist2 := math.Floor(DistanceToHaversine(meters1) * 100000000.0)
+			dist1 := math.Floor(DistanceToHaversine(meters2) * 1e8)
+			dist2 := math.Floor(DistanceToHaversine(meters1) * 1e8)
 			if dist1 != dist2 {
 				t.Fatalf("expected %f, got %f", dist2, dist1)
 			}
 		}
 	}
 }
+
+type point struct {
+	lat, lon float64
+}
+
+func BenchmarkHaversine(b *testing.B) {
+	pointA := point{
+		lat: rand.Float64()*180 - 90,
+		lon: rand.Float64()*360 - 180,
+	}
+	points := make([]point, b.N)
+	for i := 0; i < b.N; i++ {
+		points[i].lat = rand.Float64()*180 - 90
+		points[i].lon = rand.Float64()*360 - 180
+	}
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		Haversine(pointA.lat, pointA.lon, points[i].lat, points[i].lon)
+	}
+}
+
+func BenchmarkDistanceTo(b *testing.B) {
+	pointA := point{
+		lat: rand.Float64()*180 - 90,
+		lon: rand.Float64()*360 - 180,
+	}
+	points := make([]point, b.N)
+	for i := 0; i < b.N; i++ {
+		points[i].lat = rand.Float64()*180 - 90
+		points[i].lon = rand.Float64()*360 - 180
+	}
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		DistanceTo(pointA.lat, pointA.lon, points[i].lat, points[i].lon)
+	}
+}