graphics: Improve speed by using []int16 instead of []uint8

imageparts.go

@@ -18,7 +18,6 @@ import (
 	"image"
 	"math"
 
-	"github.com/hajimehoshi/ebiten/internal/endian"
 	"github.com/hajimehoshi/ebiten/internal/graphics"
 )
 
@@ -77,17 +76,16 @@ func v(y, height2p int) int16 {
 	return int16(math.MaxInt16 * y / height2p)
 }
 
-func vertices(parts ImageParts, width, height int, geo *GeoM) []uint8 {
+func vertices(parts ImageParts, width, height int, geo *GeoM) []int16 {
 	// TODO: This function should be in graphics package?
-	totalSize := graphics.QuadVertexSizeInBytes()
+	totalSize := graphics.QuadVertexSizeInBytes() / 2
 	oneSize := totalSize / 4
 	l := parts.Len()
-	vertices := make([]uint8, l*totalSize)
+	vs := make([]int16, l*totalSize)
 	width2p := graphics.NextPowerOf2Int(width)
 	height2p := graphics.NextPowerOf2Int(height)
 	n := 0
-	vs := make([]int16, 16)
+	geo16 := floatsToInt16s(geo.Element(0, 0),
-	geoBytes := floatBytes(geo.Element(0, 0),
 		geo.Element(0, 1),
 		geo.Element(1, 0),
 		geo.Element(1, 1),
@@ -104,45 +102,36 @@ func vertices(parts ImageParts, width, height int, geo *GeoM) []uint8 {
 			continue
 		}
 		u0, v0, u1, v1 := u(sx0, width2p), v(sy0, height2p), u(sx1, width2p), v(sy1, height2p)
-		vs[0] = x0
+		offset := n * totalSize
-		vs[1] = y0
+		vs[offset] = x0
-		vs[2] = u0
+		vs[offset+1] = y0
-		vs[3] = v0
+		vs[offset+2] = u0
-		vs[4] = x1
+		vs[offset+3] = v0
-		vs[5] = y0
+		for j, g := range geo16 {
-		vs[6] = u1
+			vs[offset+4+j] = g
-		vs[7] = v0
-		vs[8] = x0
-		vs[9] = y1
-		vs[10] = u0
-		vs[11] = v1
-		vs[12] = x1
-		vs[13] = y1
-		vs[14] = u1
-		vs[15] = v1
-		// Use direct assign here. `append` function might be slow on browsers.
-		for j := 0; j < 4; j++ {
-			offset := totalSize*n + oneSize*j
-			if endian.IsLittle() {
-				// Subslicing like vs[4*j:4*j+4] is slow on browsers.
-				// Don't do this.
-				for k := 0; k < 4; k++ {
-					v := vs[4*j+k]
-					vertices[offset+2*k] = uint8(v)
-					vertices[offset+2*k+1] = uint8(v >> 8)
 		}
-			} else {
+		vs[offset+oneSize] = x1
-				for k := 0; k < 4; k++ {
+		vs[offset+oneSize+1] = y0
-					v := vs[4*j+k]
+		vs[offset+oneSize+2] = u1
-					vertices[offset+2*k] = uint8(v >> 8)
+		vs[offset+oneSize+3] = v0
-					vertices[offset+2*k+1] = uint8(v)
+		for j, g := range geo16 {
+			vs[offset+oneSize+4+j] = g
 		}
+		vs[offset+2*oneSize] = x0
+		vs[offset+2*oneSize+1] = y1
+		vs[offset+2*oneSize+2] = u0
+		vs[offset+2*oneSize+3] = v1
+		for j, g := range geo16 {
+			vs[offset+2*oneSize+4+j] = g
 		}
-			for k, g := range geoBytes {
+		vs[offset+3*oneSize] = x1
-				vertices[offset+8+k] = g
+		vs[offset+3*oneSize+1] = y1
-			}
+		vs[offset+3*oneSize+2] = u1
+		vs[offset+3*oneSize+3] = v1
+		for j, g := range geo16 {
+			vs[offset+3*oneSize+4+j] = g
 		}
 		n++
 	}
-	return vertices[:n*totalSize]
+	return vs[:n*totalSize]
 }

internal/graphics/command.go

@@ -126,7 +126,7 @@ func (q *commandQueue) Flush(context *opengl.Context) error {
 	// glViewport must be called at least at every frame on iOS.
 	context.ResetViewportSize()
 	for _, g := range q.commandGroups() {
-		vertices := []uint8{}
+		vertices := []int16{}
 		for _, c := range g {
 			switch c := c.(type) {
 			case *drawImageCommand:
@@ -148,7 +148,8 @@ func (q *commandQueue) Flush(context *opengl.Context) error {
 				return err
 			}
 			if c, ok := c.(*drawImageCommand); ok {
-				indexOffsetInBytes += 6 * len(c.vertices) / QuadVertexSizeInBytes() * 2
+				n := len(c.vertices) * 2 / QuadVertexSizeInBytes()
+				indexOffsetInBytes += 6 * n * 2
 			}
 		}
 		if 0 < numc {
@@ -185,7 +186,7 @@ func (c *fillCommand) Exec(context *opengl.Context, indexOffsetInBytes int) erro
 type drawImageCommand struct {
 	dst      *Image
 	src      *Image
-	vertices []uint8
+	vertices []int16
 	color    Matrix
 	mode     opengl.CompositeMode
 }
@@ -227,8 +228,8 @@ func (c *drawImageCommand) Exec(context *opengl.Context, indexOffsetInBytes int)
 func (c *drawImageCommand) split(quadsNum int) [2]*drawImageCommand {
 	c1 := *c
 	c2 := *c
-	c1.vertices = c.vertices[:quadsNum*QuadVertexSizeInBytes()]
+	c1.vertices = c.vertices[:quadsNum*QuadVertexSizeInBytes()/2]
-	c2.vertices = c.vertices[quadsNum*QuadVertexSizeInBytes():]
+	c2.vertices = c.vertices[quadsNum*QuadVertexSizeInBytes()/2:]
 	return [2]*drawImageCommand{&c1, &c2}
 }
 
@@ -254,14 +255,14 @@ func (c *drawImageCommand) isMergeable(other *drawImageCommand) bool {
 
 func (c *drawImageCommand) merge(other *drawImageCommand) *drawImageCommand {
 	newC := *c
-	newC.vertices = make([]uint8, 0, len(c.vertices)+len(other.vertices))
+	newC.vertices = make([]int16, 0, len(c.vertices)+len(other.vertices))
 	newC.vertices = append(newC.vertices, c.vertices...)
 	newC.vertices = append(newC.vertices, other.vertices...)
 	return &newC
 }
 
 func (c *drawImageCommand) quadsNum() int {
-	return len(c.vertices) / QuadVertexSizeInBytes()
+	return len(c.vertices) * 2 / QuadVertexSizeInBytes()
 }
 
 type replacePixelsCommand struct {

internal/graphics/image.go

@@ -94,7 +94,7 @@ func (i *Image) Fill(clr color.RGBA) error {
 	return nil
 }
 
-func (i *Image) DrawImage(src *Image, vertices []uint8, clr Matrix, mode opengl.CompositeMode) error {
+func (i *Image) DrawImage(src *Image, vertices []int16, clr Matrix, mode opengl.CompositeMode) error {
 	c := &drawImageCommand{
 		dst:      i,
 		src:      src,

internal/graphics/opengl/context_desktop.go

@@ -486,9 +486,9 @@ func (c *Context) BindElementArrayBuffer(b Buffer) {
 	})
 }
 
-func (c *Context) BufferSubData(bufferType BufferType, data []uint8) {
+func (c *Context) BufferSubData(bufferType BufferType, data []int16) {
 	_ = c.runOnContextThread(func() error {
-		gl.BufferSubData(uint32(bufferType), 0, len(data), gl.Ptr(data))
+		gl.BufferSubData(uint32(bufferType), 0, len(data)*2, gl.Ptr(data))
 		return nil
 	})
 }

internal/graphics/opengl/context_js.go

@@ -387,7 +387,7 @@ func (c *Context) BindElementArrayBuffer(b Buffer) {
 	gl.BindBuffer(gl.ELEMENT_ARRAY_BUFFER, b.Object)
 }
 
-func (c *Context) BufferSubData(bufferType BufferType, data []uint8) {
+func (c *Context) BufferSubData(bufferType BufferType, data []int16) {
 	gl := c.gl
 	gl.BufferSubData(int(bufferType), 0, data)
 }

internal/graphics/opengl/context_mobile.go

@@ -386,9 +386,18 @@ func (c *Context) BindElementArrayBuffer(b Buffer) {
 	gl.BindBuffer(mgl.ELEMENT_ARRAY_BUFFER, mgl.Buffer(b))
 }
 
-func (c *Context) BufferSubData(bufferType BufferType, data []uint8) {
+func int16ToBytes(v []int16) []byte {
+	b := make([]byte, len(v)*2)
+	for i, x := range v {
+		b[2*i] = uint8(uint16(x))
+		b[2*i+1] = uint8(uint16(x) >> 8)
+	}
+	return b
+}
+
+func (c *Context) BufferSubData(bufferType BufferType, data []int16) {
 	gl := c.gl
-	gl.BufferSubData(mgl.Enum(bufferType), 0, data)
+	gl.BufferSubData(mgl.Enum(bufferType), 0, int16ToBytes(data))
 }
 
 func (c *Context) DeleteBuffer(b Buffer) {

internal/restorable/image.go

@@ -25,7 +25,7 @@ import (
 
 type drawImageHistoryItem struct {
 	image    *graphics.Image
-	vertices []uint8
+	vertices []int16
 	colorm   graphics.Matrix
 	mode     opengl.CompositeMode
 }
@@ -144,7 +144,7 @@ func (p *Image) ReplacePixels(pixels []uint8) error {
 	return nil
 }
 
-func (p *Image) DrawImage(img *Image, vertices []uint8, colorm graphics.Matrix, mode opengl.CompositeMode) error {
+func (p *Image) DrawImage(img *Image, vertices []int16, colorm graphics.Matrix, mode opengl.CompositeMode) error {
 	if img.stale || img.volatile {
 		p.makeStale()
 	} else {
@@ -156,7 +156,7 @@ func (p *Image) DrawImage(img *Image, vertices []uint8, colorm graphics.Matrix,
 	return nil
 }
 
-func (p *Image) appendDrawImageHistory(image *graphics.Image, vertices []uint8, colorm graphics.Matrix, mode opengl.CompositeMode) {
+func (p *Image) appendDrawImageHistory(image *graphics.Image, vertices []int16, colorm graphics.Matrix, mode opengl.CompositeMode) {
 	if p.stale {
 		return
 	}

math.go

@@ -22,24 +22,16 @@ import (
 	"github.com/hajimehoshi/ebiten/internal/endian"
 )
 
-func floatBytes(xs ...float64) []uint8 {
+func floatsToInt16s(xs ...float64) []int16 {
-	bits := make([]uint8, 0, len(xs)*4)
+	r := make([]int16, 0, len(xs)*2)
 	for _, x := range xs {
 		x32 := float32(x)
 		n := *(*uint32)(unsafe.Pointer(&x32))
 		if endian.IsLittle() {
-			bits = append(bits,
+			r = append(r, int16(n), int16(n>>16))
-				uint8(n),
-				uint8(n>>8),
-				uint8(n>>16),
-				uint8(n>>24))
 		} else {
-			bits = append(bits,
+			r = append(r, int16(n>>16), int16(n))
-				uint8(n>>24),
-				uint8(n>>16),
-				uint8(n>>8),
-				uint8(n))
 		}
 	}
-	return bits
+	return r
 }

math_js.go

@@ -20,12 +20,12 @@ import (
 	"github.com/gopherjs/gopherjs/js"
 )
 
-func floatBytes(xs ...float64) []uint8 {
+func floatsToInt16s(xs ...float64) []int16 {
 	a := js.Global.Get("ArrayBuffer").New(4 * len(xs))
 	af32 := js.Global.Get("Float32Array").New(a)
-	a8 := js.Global.Get("Uint8Array").New(a)
+	a16 := js.Global.Get("Int16Array").New(a)
 	for i, x := range xs {
 		af32.SetIndex(i, x)
 	}
-	return a8.Interface().([]uint8)
+	return a16.Interface().([]int16)
 }