gpu: optimize encodeQuadTo

name             old time/op  new time/op  delta
EncodeQuadTo-32  35.4ns ± 1%  11.9ns ± 3%  -66.34%  (p=0.008 n=5+5)

Signed-off-by: Egon Elbre <egonelbre@gmail.com>

gpu/clip.go

@@ -1,6 +1,9 @@
 package gpu
 
 import (
+	"encoding/binary"
+	"math"
+
 	"gioui.org/internal/f32"
 	"gioui.org/internal/stroke"
 )
@@ -12,16 +15,42 @@ type quadSplitter struct {
 }
 
 func encodeQuadTo(data []byte, meta uint32, from, ctrl, to f32.Point) {
-	// NW.
-	encodeVertex(data, meta, -1, 1, from, ctrl, to)
-	// NE.
-	encodeVertex(data[vertStride:], meta, 1, 1, from, ctrl, to)
-	// SW.
-	encodeVertex(data[vertStride*2:], meta, -1, -1, from, ctrl, to)
-	// SE.
-	encodeVertex(data[vertStride*3:], meta, 1, -1, from, ctrl, to)
+	// inlined code:
+	//   encodeVertex(data, meta, -1, 1, from, ctrl, to)
+	//   encodeVertex(data[vertStride:], meta, 1, 1, from, ctrl, to)
+	//   encodeVertex(data[vertStride*2:], meta, -1, -1, from, ctrl, to)
+	//   encodeVertex(data[vertStride*3:], meta, 1, -1, from, ctrl, to)
+	// this code needs to stay in sync with `vertex.encode`.
+
+	bo := binary.LittleEndian
+	data = data[:vertStride*4]
+
+	// encode the main template
+	bo.PutUint32(data[4:8], meta)
+	bo.PutUint32(data[8:12], math.Float32bits(from.X))
+	bo.PutUint32(data[12:16], math.Float32bits(from.Y))
+	bo.PutUint32(data[16:20], math.Float32bits(ctrl.X))
+	bo.PutUint32(data[20:24], math.Float32bits(ctrl.Y))
+	bo.PutUint32(data[24:28], math.Float32bits(to.X))
+	bo.PutUint32(data[28:32], math.Float32bits(to.Y))
+
+	copy(data[vertStride*1:vertStride*2], data[vertStride*0:vertStride*1])
+	copy(data[vertStride*2:vertStride*3], data[vertStride*0:vertStride*1])
+	copy(data[vertStride*3:vertStride*4], data[vertStride*0:vertStride*1])
+
+	bo.PutUint32(data[vertStride*0:vertStride*0+4], math.Float32bits(nwCorner))
+	bo.PutUint32(data[vertStride*1:vertStride*1+4], math.Float32bits(neCorner))
+	bo.PutUint32(data[vertStride*2:vertStride*2+4], math.Float32bits(swCorner))
+	bo.PutUint32(data[vertStride*3:vertStride*3+4], math.Float32bits(seCorner))
 }
 
+const (
+	nwCorner = 1*0.25 + 0*0.5
+	neCorner = 1*0.25 + 1*0.5
+	swCorner = 0*0.25 + 0*0.5
+	seCorner = 0*0.25 + 1*0.5
+)
+
 func encodeVertex(data []byte, meta uint32, cornerx, cornery int16, from, ctrl, to f32.Point) {
 	var corner float32
 	if cornerx == 1 {

gpu/clip_test.go

@@ -0,0 +1,21 @@
+// SPDX-License-Identifier: Unlicense OR MIT
+
+package gpu
+
+import (
+	"testing"
+
+	"gioui.org/internal/f32"
+)
+
+func BenchmarkEncodeQuadTo(b *testing.B) {
+	var data [vertStride * 4]byte
+	for i := 0; i < b.N; i++ {
+		v := float32(i)
+		encodeQuadTo(data[:], 123,
+			f32.Point{X: v, Y: v},
+			f32.Point{X: v, Y: v},
+			f32.Point{X: v, Y: v},
+		)
+	}
+}

gpu/path.go

@@ -113,16 +113,18 @@ type vertex struct {
 	ToX, ToY     float32
 }
 
+// encode needs to stay in-sync with the code in clip.go encodeQuadTo.
 func (v vertex) encode(d []byte, maxy uint32) {
+	d = d[0:32]
 	bo := binary.LittleEndian
-	bo.PutUint32(d[0:], math.Float32bits(v.Corner))
-	bo.PutUint32(d[4:], maxy)
-	bo.PutUint32(d[8:], math.Float32bits(v.FromX))
-	bo.PutUint32(d[12:], math.Float32bits(v.FromY))
-	bo.PutUint32(d[16:], math.Float32bits(v.CtrlX))
-	bo.PutUint32(d[20:], math.Float32bits(v.CtrlY))
-	bo.PutUint32(d[24:], math.Float32bits(v.ToX))
-	bo.PutUint32(d[28:], math.Float32bits(v.ToY))
+	bo.PutUint32(d[0:4], math.Float32bits(v.Corner))
+	bo.PutUint32(d[4:8], maxy)
+	bo.PutUint32(d[8:12], math.Float32bits(v.FromX))
+	bo.PutUint32(d[12:16], math.Float32bits(v.FromY))
+	bo.PutUint32(d[16:20], math.Float32bits(v.CtrlX))
+	bo.PutUint32(d[20:24], math.Float32bits(v.CtrlY))
+	bo.PutUint32(d[24:28], math.Float32bits(v.ToX))
+	bo.PutUint32(d[28:32], math.Float32bits(v.ToY))
 }
 
 const (