Adding upstream version 1.20.14.upstream/1.20.14upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

5 files changed, 2424 insertions, 0 deletions

diff --git a/src/image/draw/bench_test.go b/src/image/draw/bench_test.go
new file mode 100644
index 0000000..55d25b8
--- /dev/null
+++ b/src/image/draw/bench_test.go
@@ -0,0 +1,275 @@
+// Copyright 2011 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package draw
+
+import (
+	"image"
+	"image/color"
+	"reflect"
+	"testing"
+)
+
+const (
+	dstw, dsth = 640, 480
+	srcw, srch = 400, 300
+)
+
+var palette = color.Palette{
+	color.Black,
+	color.White,
+}
+
+// bench benchmarks drawing src and mask images onto a dst image with the
+// given op and the color models to create those images from.
+// The created images' pixels are initialized to non-zero values.
+func bench(b *testing.B, dcm, scm, mcm color.Model, op Op) {
+	b.StopTimer()
+
+	var dst Image
+	switch dcm {
+	case color.RGBAModel:
+		dst1 := image.NewRGBA(image.Rect(0, 0, dstw, dsth))
+		for y := 0; y < dsth; y++ {
+			for x := 0; x < dstw; x++ {
+				dst1.SetRGBA(x, y, color.RGBA{
+					uint8(5 * x % 0x100),
+					uint8(7 * y % 0x100),
+					uint8((7*x + 5*y) % 0x100),
+					0xff,
+				})
+			}
+		}
+		dst = dst1
+	case color.RGBA64Model:
+		dst1 := image.NewRGBA64(image.Rect(0, 0, dstw, dsth))
+		for y := 0; y < dsth; y++ {
+			for x := 0; x < dstw; x++ {
+				dst1.SetRGBA64(x, y, color.RGBA64{
+					uint16(53 * x % 0x10000),
+					uint16(59 * y % 0x10000),
+					uint16((59*x + 53*y) % 0x10000),
+					0xffff,
+				})
+			}
+		}
+		dst = dst1
+	default:
+		// The == operator isn't defined on a color.Palette (a slice), so we
+		// use reflection.
+		if reflect.DeepEqual(dcm, palette) {
+			dst1 := image.NewPaletted(image.Rect(0, 0, dstw, dsth), palette)
+			for y := 0; y < dsth; y++ {
+				for x := 0; x < dstw; x++ {
+					dst1.SetColorIndex(x, y, uint8(x^y)&1)
+				}
+			}
+			dst = dst1
+		} else {
+			b.Fatal("unknown destination color model", dcm)
+		}
+	}
+
+	var src image.Image
+	switch scm {
+	case nil:
+		src = &image.Uniform{C: color.RGBA{0x11, 0x22, 0x33, 0x44}}
+	case color.CMYKModel:
+		src1 := image.NewCMYK(image.Rect(0, 0, srcw, srch))
+		for y := 0; y < srch; y++ {
+			for x := 0; x < srcw; x++ {
+				src1.SetCMYK(x, y, color.CMYK{
+					uint8(13 * x % 0x100),
+					uint8(11 * y % 0x100),
+					uint8((11*x + 13*y) % 0x100),
+					uint8((31*x + 37*y) % 0x100),
+				})
+			}
+		}
+		src = src1
+	case color.GrayModel:
+		src1 := image.NewGray(image.Rect(0, 0, srcw, srch))
+		for y := 0; y < srch; y++ {
+			for x := 0; x < srcw; x++ {
+				src1.SetGray(x, y, color.Gray{
+					uint8((11*x + 13*y) % 0x100),
+				})
+			}
+		}
+		src = src1
+	case color.RGBAModel:
+		src1 := image.NewRGBA(image.Rect(0, 0, srcw, srch))
+		for y := 0; y < srch; y++ {
+			for x := 0; x < srcw; x++ {
+				src1.SetRGBA(x, y, color.RGBA{
+					uint8(13 * x % 0x80),
+					uint8(11 * y % 0x80),
+					uint8((11*x + 13*y) % 0x80),
+					0x7f,
+				})
+			}
+		}
+		src = src1
+	case color.RGBA64Model:
+		src1 := image.NewRGBA64(image.Rect(0, 0, srcw, srch))
+		for y := 0; y < srch; y++ {
+			for x := 0; x < srcw; x++ {
+				src1.SetRGBA64(x, y, color.RGBA64{
+					uint16(103 * x % 0x8000),
+					uint16(101 * y % 0x8000),
+					uint16((101*x + 103*y) % 0x8000),
+					0x7fff,
+				})
+			}
+		}
+		src = src1
+	case color.NRGBAModel:
+		src1 := image.NewNRGBA(image.Rect(0, 0, srcw, srch))
+		for y := 0; y < srch; y++ {
+			for x := 0; x < srcw; x++ {
+				src1.SetNRGBA(x, y, color.NRGBA{
+					uint8(13 * x % 0x100),
+					uint8(11 * y % 0x100),
+					uint8((11*x + 13*y) % 0x100),
+					0x7f,
+				})
+			}
+		}
+		src = src1
+	case color.YCbCrModel:
+		yy := make([]uint8, srcw*srch)
+		cb := make([]uint8, srcw*srch)
+		cr := make([]uint8, srcw*srch)
+		for i := range yy {
+			yy[i] = uint8(3 * i % 0x100)
+			cb[i] = uint8(5 * i % 0x100)
+			cr[i] = uint8(7 * i % 0x100)
+		}
+		src = &image.YCbCr{
+			Y:              yy,
+			Cb:             cb,
+			Cr:             cr,
+			YStride:        srcw,
+			CStride:        srcw,
+			SubsampleRatio: image.YCbCrSubsampleRatio444,
+			Rect:           image.Rect(0, 0, srcw, srch),
+		}
+	default:
+		b.Fatal("unknown source color model", scm)
+	}
+
+	var mask image.Image
+	switch mcm {
+	case nil:
+		// No-op.
+	case color.AlphaModel:
+		mask1 := image.NewAlpha(image.Rect(0, 0, srcw, srch))
+		for y := 0; y < srch; y++ {
+			for x := 0; x < srcw; x++ {
+				a := uint8((23*x + 29*y) % 0x100)
+				// Glyph masks are typically mostly zero,
+				// so we only set a quarter of mask1's pixels.
+				if a >= 0xc0 {
+					mask1.SetAlpha(x, y, color.Alpha{a})
+				}
+			}
+		}
+		mask = mask1
+	default:
+		b.Fatal("unknown mask color model", mcm)
+	}
+
+	b.StartTimer()
+	for i := 0; i < b.N; i++ {
+		// Scatter the destination rectangle to draw into.
+		x := 3 * i % (dstw - srcw)
+		y := 7 * i % (dsth - srch)
+
+		DrawMask(dst, dst.Bounds().Add(image.Pt(x, y)), src, image.ZP, mask, image.ZP, op)
+	}
+}
+
+// The BenchmarkFoo functions exercise a drawFoo fast-path function in draw.go.
+
+func BenchmarkFillOver(b *testing.B) {
+	bench(b, color.RGBAModel, nil, nil, Over)
+}
+
+func BenchmarkFillSrc(b *testing.B) {
+	bench(b, color.RGBAModel, nil, nil, Src)
+}
+
+func BenchmarkCopyOver(b *testing.B) {
+	bench(b, color.RGBAModel, color.RGBAModel, nil, Over)
+}
+
+func BenchmarkCopySrc(b *testing.B) {
+	bench(b, color.RGBAModel, color.RGBAModel, nil, Src)
+}
+
+func BenchmarkNRGBAOver(b *testing.B) {
+	bench(b, color.RGBAModel, color.NRGBAModel, nil, Over)
+}
+
+func BenchmarkNRGBASrc(b *testing.B) {
+	bench(b, color.RGBAModel, color.NRGBAModel, nil, Src)
+}
+
+func BenchmarkYCbCr(b *testing.B) {
+	bench(b, color.RGBAModel, color.YCbCrModel, nil, Over)
+}
+
+func BenchmarkGray(b *testing.B) {
+	bench(b, color.RGBAModel, color.GrayModel, nil, Over)
+}
+
+func BenchmarkCMYK(b *testing.B) {
+	bench(b, color.RGBAModel, color.CMYKModel, nil, Over)
+}
+
+func BenchmarkGlyphOver(b *testing.B) {
+	bench(b, color.RGBAModel, nil, color.AlphaModel, Over)
+}
+
+func BenchmarkRGBAMaskOver(b *testing.B) {
+	bench(b, color.RGBAModel, color.RGBAModel, color.AlphaModel, Over)
+}
+
+func BenchmarkGrayMaskOver(b *testing.B) {
+	bench(b, color.RGBAModel, color.GrayModel, color.AlphaModel, Over)
+}
+
+func BenchmarkRGBA64ImageMaskOver(b *testing.B) {
+	bench(b, color.RGBAModel, color.RGBA64Model, color.AlphaModel, Over)
+}
+
+func BenchmarkRGBA(b *testing.B) {
+	bench(b, color.RGBAModel, color.RGBA64Model, nil, Src)
+}
+
+func BenchmarkPalettedFill(b *testing.B) {
+	bench(b, palette, nil, nil, Src)
+}
+
+func BenchmarkPalettedRGBA(b *testing.B) {
+	bench(b, palette, color.RGBAModel, nil, Src)
+}
+
+// The BenchmarkGenericFoo functions exercise the generic, slow-path code.
+
+func BenchmarkGenericOver(b *testing.B) {
+	bench(b, color.RGBA64Model, color.RGBA64Model, nil, Over)
+}
+
+func BenchmarkGenericMaskOver(b *testing.B) {
+	bench(b, color.RGBA64Model, color.RGBA64Model, color.AlphaModel, Over)
+}
+
+func BenchmarkGenericSrc(b *testing.B) {
+	bench(b, color.RGBA64Model, color.RGBA64Model, nil, Src)
+}
+
+func BenchmarkGenericMaskSrc(b *testing.B) {
+	bench(b, color.RGBA64Model, color.RGBA64Model, color.AlphaModel, Src)
+}
diff --git a/src/image/draw/clip_test.go b/src/image/draw/clip_test.go
new file mode 100644
index 0000000..0abf53e
--- /dev/null
+++ b/src/image/draw/clip_test.go
@@ -0,0 +1,205 @@
+// Copyright 2011 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package draw
+
+import (
+	"image"
+	"testing"
+)
+
+type clipTest struct {
+	desc          string
+	r, dr, sr, mr image.Rectangle
+	sp, mp        image.Point
+	nilMask       bool
+	r0            image.Rectangle
+	sp0, mp0      image.Point
+}
+
+var clipTests = []clipTest{
+	// The following tests all have a nil mask.
+	{
+		"basic",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 100, 100),
+		image.ZR,
+		image.ZP,
+		image.ZP,
+		true,
+		image.Rect(0, 0, 100, 100),
+		image.ZP,
+		image.ZP,
+	},
+	{
+		"clip dr",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(40, 40, 60, 60),
+		image.Rect(0, 0, 100, 100),
+		image.ZR,
+		image.ZP,
+		image.ZP,
+		true,
+		image.Rect(40, 40, 60, 60),
+		image.Pt(40, 40),
+		image.ZP,
+	},
+	{
+		"clip sr",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 100, 100),
+		image.Rect(20, 20, 80, 80),
+		image.ZR,
+		image.ZP,
+		image.ZP,
+		true,
+		image.Rect(20, 20, 80, 80),
+		image.Pt(20, 20),
+		image.ZP,
+	},
+	{
+		"clip dr and sr",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 50, 100),
+		image.Rect(20, 20, 80, 80),
+		image.ZR,
+		image.ZP,
+		image.ZP,
+		true,
+		image.Rect(20, 20, 50, 80),
+		image.Pt(20, 20),
+		image.ZP,
+	},
+	{
+		"clip dr and sr, sp outside sr (top-left)",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 50, 100),
+		image.Rect(20, 20, 80, 80),
+		image.ZR,
+		image.Pt(15, 8),
+		image.ZP,
+		true,
+		image.Rect(5, 12, 50, 72),
+		image.Pt(20, 20),
+		image.ZP,
+	},
+	{
+		"clip dr and sr, sp outside sr (middle-left)",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 50, 100),
+		image.Rect(20, 20, 80, 80),
+		image.ZR,
+		image.Pt(15, 66),
+		image.ZP,
+		true,
+		image.Rect(5, 0, 50, 14),
+		image.Pt(20, 66),
+		image.ZP,
+	},
+	{
+		"clip dr and sr, sp outside sr (bottom-left)",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 50, 100),
+		image.Rect(20, 20, 80, 80),
+		image.ZR,
+		image.Pt(15, 91),
+		image.ZP,
+		true,
+		image.ZR,
+		image.Pt(15, 91),
+		image.ZP,
+	},
+	{
+		"clip dr and sr, sp inside sr",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 50, 100),
+		image.Rect(20, 20, 80, 80),
+		image.ZR,
+		image.Pt(44, 33),
+		image.ZP,
+		true,
+		image.Rect(0, 0, 36, 47),
+		image.Pt(44, 33),
+		image.ZP,
+	},
+
+	// The following tests all have a non-nil mask.
+	{
+		"basic mask",
+		image.Rect(0, 0, 80, 80),
+		image.Rect(20, 0, 100, 80),
+		image.Rect(0, 0, 50, 49),
+		image.Rect(0, 0, 46, 47),
+		image.ZP,
+		image.ZP,
+		false,
+		image.Rect(20, 0, 46, 47),
+		image.Pt(20, 0),
+		image.Pt(20, 0),
+	},
+	{
+		"clip sr and mr",
+		image.Rect(0, 0, 100, 100),
+		image.Rect(0, 0, 100, 100),
+		image.Rect(23, 23, 55, 86),
+		image.Rect(44, 44, 87, 58),
+		image.Pt(10, 10),
+		image.Pt(11, 11),
+		false,
+		image.Rect(33, 33, 45, 47),
+		image.Pt(43, 43),
+		image.Pt(44, 44),
+	},
+}
+
+func TestClip(t *testing.T) {
+	dst0 := image.NewRGBA(image.Rect(0, 0, 100, 100))
+	src0 := image.NewRGBA(image.Rect(0, 0, 100, 100))
+	mask0 := image.NewRGBA(image.Rect(0, 0, 100, 100))
+	for _, c := range clipTests {
+		dst := dst0.SubImage(c.dr).(*image.RGBA)
+		src := src0.SubImage(c.sr).(*image.RGBA)
+		r, sp, mp := c.r, c.sp, c.mp
+		if c.nilMask {
+			clip(dst, &r, src, &sp, nil, nil)
+		} else {
+			clip(dst, &r, src, &sp, mask0.SubImage(c.mr), &mp)
+		}
+
+		// Check that the actual results equal the expected results.
+		if !c.r0.Eq(r) {
+			t.Errorf("%s: clip rectangle want %v got %v", c.desc, c.r0, r)
+			continue
+		}
+		if !c.sp0.Eq(sp) {
+			t.Errorf("%s: sp want %v got %v", c.desc, c.sp0, sp)
+			continue
+		}
+		if !c.nilMask {
+			if !c.mp0.Eq(mp) {
+				t.Errorf("%s: mp want %v got %v", c.desc, c.mp0, mp)
+				continue
+			}
+		}
+
+		// Check that the clipped rectangle is contained by the dst / src / mask
+		// rectangles, in their respective coordinate spaces.
+		if !r.In(c.dr) {
+			t.Errorf("%s: c.dr %v does not contain r %v", c.desc, c.dr, r)
+		}
+		// sr is r translated into src's coordinate space.
+		sr := r.Add(c.sp.Sub(c.dr.Min))
+		if !sr.In(c.sr) {
+			t.Errorf("%s: c.sr %v does not contain sr %v", c.desc, c.sr, sr)
+		}
+		if !c.nilMask {
+			// mr is r translated into mask's coordinate space.
+			mr := r.Add(c.mp.Sub(c.dr.Min))
+			if !mr.In(c.mr) {
+				t.Errorf("%s: c.mr %v does not contain mr %v", c.desc, c.mr, mr)
+			}
+		}
+	}
+}
diff --git a/src/image/draw/draw.go b/src/image/draw/draw.go
new file mode 100644
index 0000000..920ebb9
--- /dev/null
+++ b/src/image/draw/draw.go
@@ -0,0 +1,1086 @@
+// Copyright 2009 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package draw provides image composition functions.
+//
+// See "The Go image/draw package" for an introduction to this package:
+// https://golang.org/doc/articles/image_draw.html
+package draw
+
+import (
+	"image"
+	"image/color"
+	"image/internal/imageutil"
+)
+
+// m is the maximum color value returned by image.Color.RGBA.
+const m = 1<<16 - 1
+
+// Image is an image.Image with a Set method to change a single pixel.
+type Image interface {
+	image.Image
+	Set(x, y int, c color.Color)
+}
+
+// RGBA64Image extends both the Image and image.RGBA64Image interfaces with a
+// SetRGBA64 method to change a single pixel. SetRGBA64 is equivalent to
+// calling Set, but it can avoid allocations from converting concrete color
+// types to the color.Color interface type.
+type RGBA64Image interface {
+	image.RGBA64Image
+	Set(x, y int, c color.Color)
+	SetRGBA64(x, y int, c color.RGBA64)
+}
+
+// Quantizer produces a palette for an image.
+type Quantizer interface {
+	// Quantize appends up to cap(p) - len(p) colors to p and returns the
+	// updated palette suitable for converting m to a paletted image.
+	Quantize(p color.Palette, m image.Image) color.Palette
+}
+
+// Op is a Porter-Duff compositing operator.
+type Op int
+
+const (
+	// Over specifies ``(src in mask) over dst''.
+	Over Op = iota
+	// Src specifies ``src in mask''.
+	Src
+)
+
+// Draw implements the Drawer interface by calling the Draw function with this
+// Op.
+func (op Op) Draw(dst Image, r image.Rectangle, src image.Image, sp image.Point) {
+	DrawMask(dst, r, src, sp, nil, image.Point{}, op)
+}
+
+// Drawer contains the Draw method.
+type Drawer interface {
+	// Draw aligns r.Min in dst with sp in src and then replaces the
+	// rectangle r in dst with the result of drawing src on dst.
+	Draw(dst Image, r image.Rectangle, src image.Image, sp image.Point)
+}
+
+// FloydSteinberg is a Drawer that is the Src Op with Floyd-Steinberg error
+// diffusion.
+var FloydSteinberg Drawer = floydSteinberg{}
+
+type floydSteinberg struct{}
+
+func (floydSteinberg) Draw(dst Image, r image.Rectangle, src image.Image, sp image.Point) {
+	clip(dst, &r, src, &sp, nil, nil)
+	if r.Empty() {
+		return
+	}
+	drawPaletted(dst, r, src, sp, true)
+}
+
+// clip clips r against each image's bounds (after translating into the
+// destination image's coordinate space) and shifts the points sp and mp by
+// the same amount as the change in r.Min.
+func clip(dst Image, r *image.Rectangle, src image.Image, sp *image.Point, mask image.Image, mp *image.Point) {
+	orig := r.Min
+	*r = r.Intersect(dst.Bounds())
+	*r = r.Intersect(src.Bounds().Add(orig.Sub(*sp)))
+	if mask != nil {
+		*r = r.Intersect(mask.Bounds().Add(orig.Sub(*mp)))
+	}
+	dx := r.Min.X - orig.X
+	dy := r.Min.Y - orig.Y
+	if dx == 0 && dy == 0 {
+		return
+	}
+	sp.X += dx
+	sp.Y += dy
+	if mp != nil {
+		mp.X += dx
+		mp.Y += dy
+	}
+}
+
+func processBackward(dst image.Image, r image.Rectangle, src image.Image, sp image.Point) bool {
+	return dst == src &&
+		r.Overlaps(r.Add(sp.Sub(r.Min))) &&
+		(sp.Y < r.Min.Y || (sp.Y == r.Min.Y && sp.X < r.Min.X))
+}
+
+// Draw calls DrawMask with a nil mask.
+func Draw(dst Image, r image.Rectangle, src image.Image, sp image.Point, op Op) {
+	DrawMask(dst, r, src, sp, nil, image.Point{}, op)
+}
+
+// DrawMask aligns r.Min in dst with sp in src and mp in mask and then replaces the rectangle r
+// in dst with the result of a Porter-Duff composition. A nil mask is treated as opaque.
+func DrawMask(dst Image, r image.Rectangle, src image.Image, sp image.Point, mask image.Image, mp image.Point, op Op) {
+	clip(dst, &r, src, &sp, mask, &mp)
+	if r.Empty() {
+		return
+	}
+
+	// Fast paths for special cases. If none of them apply, then we fall back
+	// to general but slower implementations.
+	//
+	// For NRGBA and NRGBA64 image types, the code paths aren't just faster.
+	// They also avoid the information loss that would otherwise occur from
+	// converting non-alpha-premultiplied color to and from alpha-premultiplied
+	// color. See TestDrawSrcNonpremultiplied.
+	switch dst0 := dst.(type) {
+	case *image.RGBA:
+		if op == Over {
+			if mask == nil {
+				switch src0 := src.(type) {
+				case *image.Uniform:
+					sr, sg, sb, sa := src0.RGBA()
+					if sa == 0xffff {
+						drawFillSrc(dst0, r, sr, sg, sb, sa)
+					} else {
+						drawFillOver(dst0, r, sr, sg, sb, sa)
+					}
+					return
+				case *image.RGBA:
+					drawCopyOver(dst0, r, src0, sp)
+					return
+				case *image.NRGBA:
+					drawNRGBAOver(dst0, r, src0, sp)
+					return
+				case *image.YCbCr:
+					// An image.YCbCr is always fully opaque, and so if the
+					// mask is nil (i.e. fully opaque) then the op is
+					// effectively always Src. Similarly for image.Gray and
+					// image.CMYK.
+					if imageutil.DrawYCbCr(dst0, r, src0, sp) {
+						return
+					}
+				case *image.Gray:
+					drawGray(dst0, r, src0, sp)
+					return
+				case *image.CMYK:
+					drawCMYK(dst0, r, src0, sp)
+					return
+				}
+			} else if mask0, ok := mask.(*image.Alpha); ok {
+				switch src0 := src.(type) {
+				case *image.Uniform:
+					drawGlyphOver(dst0, r, src0, mask0, mp)
+					return
+				case *image.RGBA:
+					drawRGBAMaskOver(dst0, r, src0, sp, mask0, mp)
+					return
+				case *image.Gray:
+					drawGrayMaskOver(dst0, r, src0, sp, mask0, mp)
+					return
+				// Case order matters. The next case (image.RGBA64Image) is an
+				// interface type that the concrete types above also implement.
+				case image.RGBA64Image:
+					drawRGBA64ImageMaskOver(dst0, r, src0, sp, mask0, mp)
+					return
+				}
+			}
+		} else {
+			if mask == nil {
+				switch src0 := src.(type) {
+				case *image.Uniform:
+					sr, sg, sb, sa := src0.RGBA()
+					drawFillSrc(dst0, r, sr, sg, sb, sa)
+					return
+				case *image.RGBA:
+					d0 := dst0.PixOffset(r.Min.X, r.Min.Y)
+					s0 := src0.PixOffset(sp.X, sp.Y)
+					drawCopySrc(
+						dst0.Pix[d0:], dst0.Stride, r, src0.Pix[s0:], src0.Stride, sp, 4*r.Dx())
+					return
+				case *image.NRGBA:
+					drawNRGBASrc(dst0, r, src0, sp)
+					return
+				case *image.YCbCr:
+					if imageutil.DrawYCbCr(dst0, r, src0, sp) {
+						return
+					}
+				case *image.Gray:
+					drawGray(dst0, r, src0, sp)
+					return
+				case *image.CMYK:
+					drawCMYK(dst0, r, src0, sp)
+					return
+				}
+			}
+		}
+		drawRGBA(dst0, r, src, sp, mask, mp, op)
+		return
+	case *image.Paletted:
+		if op == Src && mask == nil {
+			if src0, ok := src.(*image.Uniform); ok {
+				colorIndex := uint8(dst0.Palette.Index(src0.C))
+				i0 := dst0.PixOffset(r.Min.X, r.Min.Y)
+				i1 := i0 + r.Dx()
+				for i := i0; i < i1; i++ {
+					dst0.Pix[i] = colorIndex
+				}
+				firstRow := dst0.Pix[i0:i1]
+				for y := r.Min.Y + 1; y < r.Max.Y; y++ {
+					i0 += dst0.Stride
+					i1 += dst0.Stride
+					copy(dst0.Pix[i0:i1], firstRow)
+				}
+				return
+			} else if !processBackward(dst, r, src, sp) {
+				drawPaletted(dst0, r, src, sp, false)
+				return
+			}
+		}
+	case *image.NRGBA:
+		if op == Src && mask == nil {
+			if src0, ok := src.(*image.NRGBA); ok {
+				d0 := dst0.PixOffset(r.Min.X, r.Min.Y)
+				s0 := src0.PixOffset(sp.X, sp.Y)
+				drawCopySrc(
+					dst0.Pix[d0:], dst0.Stride, r, src0.Pix[s0:], src0.Stride, sp, 4*r.Dx())
+				return
+			}
+		}
+	case *image.NRGBA64:
+		if op == Src && mask == nil {
+			if src0, ok := src.(*image.NRGBA64); ok {
+				d0 := dst0.PixOffset(r.Min.X, r.Min.Y)
+				s0 := src0.PixOffset(sp.X, sp.Y)
+				drawCopySrc(
+					dst0.Pix[d0:], dst0.Stride, r, src0.Pix[s0:], src0.Stride, sp, 8*r.Dx())
+				return
+			}
+		}
+	}
+
+	x0, x1, dx := r.Min.X, r.Max.X, 1
+	y0, y1, dy := r.Min.Y, r.Max.Y, 1
+	if processBackward(dst, r, src, sp) {
+		x0, x1, dx = x1-1, x0-1, -1
+		y0, y1, dy = y1-1, y0-1, -1
+	}
+
+	// FALLBACK1.17
+	//
+	// Try the draw.RGBA64Image and image.RGBA64Image interfaces, part of the
+	// standard library since Go 1.17. These are like the draw.Image and
+	// image.Image interfaces but they can avoid allocations from converting
+	// concrete color types to the color.Color interface type.
+
+	if dst0, _ := dst.(RGBA64Image); dst0 != nil {
+		if src0, _ := src.(image.RGBA64Image); src0 != nil {
+			if mask == nil {
+				sy := sp.Y + y0 - r.Min.Y
+				my := mp.Y + y0 - r.Min.Y
+				for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+					sx := sp.X + x0 - r.Min.X
+					mx := mp.X + x0 - r.Min.X
+					for x := x0; x != x1; x, sx, mx = x+dx, sx+dx, mx+dx {
+						if op == Src {
+							dst0.SetRGBA64(x, y, src0.RGBA64At(sx, sy))
+						} else {
+							srgba := src0.RGBA64At(sx, sy)
+							a := m - uint32(srgba.A)
+							drgba := dst0.RGBA64At(x, y)
+							dst0.SetRGBA64(x, y, color.RGBA64{
+								R: uint16((uint32(drgba.R)*a)/m) + srgba.R,
+								G: uint16((uint32(drgba.G)*a)/m) + srgba.G,
+								B: uint16((uint32(drgba.B)*a)/m) + srgba.B,
+								A: uint16((uint32(drgba.A)*a)/m) + srgba.A,
+							})
+						}
+					}
+				}
+				return
+
+			} else if mask0, _ := mask.(image.RGBA64Image); mask0 != nil {
+				sy := sp.Y + y0 - r.Min.Y
+				my := mp.Y + y0 - r.Min.Y
+				for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+					sx := sp.X + x0 - r.Min.X
+					mx := mp.X + x0 - r.Min.X
+					for x := x0; x != x1; x, sx, mx = x+dx, sx+dx, mx+dx {
+						ma := uint32(mask0.RGBA64At(mx, my).A)
+						switch {
+						case ma == 0:
+							if op == Over {
+								// No-op.
+							} else {
+								dst0.SetRGBA64(x, y, color.RGBA64{})
+							}
+						case ma == m && op == Src:
+							dst0.SetRGBA64(x, y, src0.RGBA64At(sx, sy))
+						default:
+							srgba := src0.RGBA64At(sx, sy)
+							if op == Over {
+								drgba := dst0.RGBA64At(x, y)
+								a := m - (uint32(srgba.A) * ma / m)
+								dst0.SetRGBA64(x, y, color.RGBA64{
+									R: uint16((uint32(drgba.R)*a + uint32(srgba.R)*ma) / m),
+									G: uint16((uint32(drgba.G)*a + uint32(srgba.G)*ma) / m),
+									B: uint16((uint32(drgba.B)*a + uint32(srgba.B)*ma) / m),
+									A: uint16((uint32(drgba.A)*a + uint32(srgba.A)*ma) / m),
+								})
+							} else {
+								dst0.SetRGBA64(x, y, color.RGBA64{
+									R: uint16(uint32(srgba.R) * ma / m),
+									G: uint16(uint32(srgba.G) * ma / m),
+									B: uint16(uint32(srgba.B) * ma / m),
+									A: uint16(uint32(srgba.A) * ma / m),
+								})
+							}
+						}
+					}
+				}
+				return
+			}
+		}
+	}
+
+	// FALLBACK1.0
+	//
+	// If none of the faster code paths above apply, use the draw.Image and
+	// image.Image interfaces, part of the standard library since Go 1.0.
+
+	var out color.RGBA64
+	sy := sp.Y + y0 - r.Min.Y
+	my := mp.Y + y0 - r.Min.Y
+	for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+		sx := sp.X + x0 - r.Min.X
+		mx := mp.X + x0 - r.Min.X
+		for x := x0; x != x1; x, sx, mx = x+dx, sx+dx, mx+dx {
+			ma := uint32(m)
+			if mask != nil {
+				_, _, _, ma = mask.At(mx, my).RGBA()
+			}
+			switch {
+			case ma == 0:
+				if op == Over {
+					// No-op.
+				} else {
+					dst.Set(x, y, color.Transparent)
+				}
+			case ma == m && op == Src:
+				dst.Set(x, y, src.At(sx, sy))
+			default:
+				sr, sg, sb, sa := src.At(sx, sy).RGBA()
+				if op == Over {
+					dr, dg, db, da := dst.At(x, y).RGBA()
+					a := m - (sa * ma / m)
+					out.R = uint16((dr*a + sr*ma) / m)
+					out.G = uint16((dg*a + sg*ma) / m)
+					out.B = uint16((db*a + sb*ma) / m)
+					out.A = uint16((da*a + sa*ma) / m)
+				} else {
+					out.R = uint16(sr * ma / m)
+					out.G = uint16(sg * ma / m)
+					out.B = uint16(sb * ma / m)
+					out.A = uint16(sa * ma / m)
+				}
+				// The third argument is &out instead of out (and out is
+				// declared outside of the inner loop) to avoid the implicit
+				// conversion to color.Color here allocating memory in the
+				// inner loop if sizeof(color.RGBA64) > sizeof(uintptr).
+				dst.Set(x, y, &out)
+			}
+		}
+	}
+}
+
+func drawFillOver(dst *image.RGBA, r image.Rectangle, sr, sg, sb, sa uint32) {
+	// The 0x101 is here for the same reason as in drawRGBA.
+	a := (m - sa) * 0x101
+	i0 := dst.PixOffset(r.Min.X, r.Min.Y)
+	i1 := i0 + r.Dx()*4
+	for y := r.Min.Y; y != r.Max.Y; y++ {
+		for i := i0; i < i1; i += 4 {
+			dr := &dst.Pix[i+0]
+			dg := &dst.Pix[i+1]
+			db := &dst.Pix[i+2]
+			da := &dst.Pix[i+3]
+
+			*dr = uint8((uint32(*dr)*a/m + sr) >> 8)
+			*dg = uint8((uint32(*dg)*a/m + sg) >> 8)
+			*db = uint8((uint32(*db)*a/m + sb) >> 8)
+			*da = uint8((uint32(*da)*a/m + sa) >> 8)
+		}
+		i0 += dst.Stride
+		i1 += dst.Stride
+	}
+}
+
+func drawFillSrc(dst *image.RGBA, r image.Rectangle, sr, sg, sb, sa uint32) {
+	sr8 := uint8(sr >> 8)
+	sg8 := uint8(sg >> 8)
+	sb8 := uint8(sb >> 8)
+	sa8 := uint8(sa >> 8)
+	// The built-in copy function is faster than a straightforward for loop to fill the destination with
+	// the color, but copy requires a slice source. We therefore use a for loop to fill the first row, and
+	// then use the first row as the slice source for the remaining rows.
+	i0 := dst.PixOffset(r.Min.X, r.Min.Y)
+	i1 := i0 + r.Dx()*4
+	for i := i0; i < i1; i += 4 {
+		dst.Pix[i+0] = sr8
+		dst.Pix[i+1] = sg8
+		dst.Pix[i+2] = sb8
+		dst.Pix[i+3] = sa8
+	}
+	firstRow := dst.Pix[i0:i1]
+	for y := r.Min.Y + 1; y < r.Max.Y; y++ {
+		i0 += dst.Stride
+		i1 += dst.Stride
+		copy(dst.Pix[i0:i1], firstRow)
+	}
+}
+
+func drawCopyOver(dst *image.RGBA, r image.Rectangle, src *image.RGBA, sp image.Point) {
+	dx, dy := r.Dx(), r.Dy()
+	d0 := dst.PixOffset(r.Min.X, r.Min.Y)
+	s0 := src.PixOffset(sp.X, sp.Y)
+	var (
+		ddelta, sdelta int
+		i0, i1, idelta int
+	)
+	if r.Min.Y < sp.Y || r.Min.Y == sp.Y && r.Min.X <= sp.X {
+		ddelta = dst.Stride
+		sdelta = src.Stride
+		i0, i1, idelta = 0, dx*4, +4
+	} else {
+		// If the source start point is higher than the destination start point, or equal height but to the left,
+		// then we compose the rows in right-to-left, bottom-up order instead of left-to-right, top-down.
+		d0 += (dy - 1) * dst.Stride
+		s0 += (dy - 1) * src.Stride
+		ddelta = -dst.Stride
+		sdelta = -src.Stride
+		i0, i1, idelta = (dx-1)*4, -4, -4
+	}
+	for ; dy > 0; dy-- {
+		dpix := dst.Pix[d0:]
+		spix := src.Pix[s0:]
+		for i := i0; i != i1; i += idelta {
+			s := spix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			sr := uint32(s[0]) * 0x101
+			sg := uint32(s[1]) * 0x101
+			sb := uint32(s[2]) * 0x101
+			sa := uint32(s[3]) * 0x101
+
+			// The 0x101 is here for the same reason as in drawRGBA.
+			a := (m - sa) * 0x101
+
+			d := dpix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			d[0] = uint8((uint32(d[0])*a/m + sr) >> 8)
+			d[1] = uint8((uint32(d[1])*a/m + sg) >> 8)
+			d[2] = uint8((uint32(d[2])*a/m + sb) >> 8)
+			d[3] = uint8((uint32(d[3])*a/m + sa) >> 8)
+		}
+		d0 += ddelta
+		s0 += sdelta
+	}
+}
+
+// drawCopySrc copies bytes to dstPix from srcPix. These arguments roughly
+// correspond to the Pix fields of the image package's concrete image.Image
+// implementations, but are offset (dstPix is dst.Pix[dpOffset:] not dst.Pix).
+func drawCopySrc(
+	dstPix []byte, dstStride int, r image.Rectangle,
+	srcPix []byte, srcStride int, sp image.Point,
+	bytesPerRow int) {
+
+	d0, s0, ddelta, sdelta, dy := 0, 0, dstStride, srcStride, r.Dy()
+	if r.Min.Y > sp.Y {
+		// If the source start point is higher than the destination start
+		// point, then we compose the rows in bottom-up order instead of
+		// top-down. Unlike the drawCopyOver function, we don't have to check
+		// the x coordinates because the built-in copy function can handle
+		// overlapping slices.
+		d0 = (dy - 1) * dstStride
+		s0 = (dy - 1) * srcStride
+		ddelta = -dstStride
+		sdelta = -srcStride
+	}
+	for ; dy > 0; dy-- {
+		copy(dstPix[d0:d0+bytesPerRow], srcPix[s0:s0+bytesPerRow])
+		d0 += ddelta
+		s0 += sdelta
+	}
+}
+
+func drawNRGBAOver(dst *image.RGBA, r image.Rectangle, src *image.NRGBA, sp image.Point) {
+	i0 := (r.Min.X - dst.Rect.Min.X) * 4
+	i1 := (r.Max.X - dst.Rect.Min.X) * 4
+	si0 := (sp.X - src.Rect.Min.X) * 4
+	yMax := r.Max.Y - dst.Rect.Min.Y
+
+	y := r.Min.Y - dst.Rect.Min.Y
+	sy := sp.Y - src.Rect.Min.Y
+	for ; y != yMax; y, sy = y+1, sy+1 {
+		dpix := dst.Pix[y*dst.Stride:]
+		spix := src.Pix[sy*src.Stride:]
+
+		for i, si := i0, si0; i < i1; i, si = i+4, si+4 {
+			// Convert from non-premultiplied color to pre-multiplied color.
+			s := spix[si : si+4 : si+4] // Small cap improves performance, see https://golang.org/issue/27857
+			sa := uint32(s[3]) * 0x101
+			sr := uint32(s[0]) * sa / 0xff
+			sg := uint32(s[1]) * sa / 0xff
+			sb := uint32(s[2]) * sa / 0xff
+
+			d := dpix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			dr := uint32(d[0])
+			dg := uint32(d[1])
+			db := uint32(d[2])
+			da := uint32(d[3])
+
+			// The 0x101 is here for the same reason as in drawRGBA.
+			a := (m - sa) * 0x101
+
+			d[0] = uint8((dr*a/m + sr) >> 8)
+			d[1] = uint8((dg*a/m + sg) >> 8)
+			d[2] = uint8((db*a/m + sb) >> 8)
+			d[3] = uint8((da*a/m + sa) >> 8)
+		}
+	}
+}
+
+func drawNRGBASrc(dst *image.RGBA, r image.Rectangle, src *image.NRGBA, sp image.Point) {
+	i0 := (r.Min.X - dst.Rect.Min.X) * 4
+	i1 := (r.Max.X - dst.Rect.Min.X) * 4
+	si0 := (sp.X - src.Rect.Min.X) * 4
+	yMax := r.Max.Y - dst.Rect.Min.Y
+
+	y := r.Min.Y - dst.Rect.Min.Y
+	sy := sp.Y - src.Rect.Min.Y
+	for ; y != yMax; y, sy = y+1, sy+1 {
+		dpix := dst.Pix[y*dst.Stride:]
+		spix := src.Pix[sy*src.Stride:]
+
+		for i, si := i0, si0; i < i1; i, si = i+4, si+4 {
+			// Convert from non-premultiplied color to pre-multiplied color.
+			s := spix[si : si+4 : si+4] // Small cap improves performance, see https://golang.org/issue/27857
+			sa := uint32(s[3]) * 0x101
+			sr := uint32(s[0]) * sa / 0xff
+			sg := uint32(s[1]) * sa / 0xff
+			sb := uint32(s[2]) * sa / 0xff
+
+			d := dpix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			d[0] = uint8(sr >> 8)
+			d[1] = uint8(sg >> 8)
+			d[2] = uint8(sb >> 8)
+			d[3] = uint8(sa >> 8)
+		}
+	}
+}
+
+func drawGray(dst *image.RGBA, r image.Rectangle, src *image.Gray, sp image.Point) {
+	i0 := (r.Min.X - dst.Rect.Min.X) * 4
+	i1 := (r.Max.X - dst.Rect.Min.X) * 4
+	si0 := (sp.X - src.Rect.Min.X) * 1
+	yMax := r.Max.Y - dst.Rect.Min.Y
+
+	y := r.Min.Y - dst.Rect.Min.Y
+	sy := sp.Y - src.Rect.Min.Y
+	for ; y != yMax; y, sy = y+1, sy+1 {
+		dpix := dst.Pix[y*dst.Stride:]
+		spix := src.Pix[sy*src.Stride:]
+
+		for i, si := i0, si0; i < i1; i, si = i+4, si+1 {
+			p := spix[si]
+			d := dpix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			d[0] = p
+			d[1] = p
+			d[2] = p
+			d[3] = 255
+		}
+	}
+}
+
+func drawCMYK(dst *image.RGBA, r image.Rectangle, src *image.CMYK, sp image.Point) {
+	i0 := (r.Min.X - dst.Rect.Min.X) * 4
+	i1 := (r.Max.X - dst.Rect.Min.X) * 4
+	si0 := (sp.X - src.Rect.Min.X) * 4
+	yMax := r.Max.Y - dst.Rect.Min.Y
+
+	y := r.Min.Y - dst.Rect.Min.Y
+	sy := sp.Y - src.Rect.Min.Y
+	for ; y != yMax; y, sy = y+1, sy+1 {
+		dpix := dst.Pix[y*dst.Stride:]
+		spix := src.Pix[sy*src.Stride:]
+
+		for i, si := i0, si0; i < i1; i, si = i+4, si+4 {
+			s := spix[si : si+4 : si+4] // Small cap improves performance, see https://golang.org/issue/27857
+			d := dpix[i : i+4 : i+4]
+			d[0], d[1], d[2] = color.CMYKToRGB(s[0], s[1], s[2], s[3])
+			d[3] = 255
+		}
+	}
+}
+
+func drawGlyphOver(dst *image.RGBA, r image.Rectangle, src *image.Uniform, mask *image.Alpha, mp image.Point) {
+	i0 := dst.PixOffset(r.Min.X, r.Min.Y)
+	i1 := i0 + r.Dx()*4
+	mi0 := mask.PixOffset(mp.X, mp.Y)
+	sr, sg, sb, sa := src.RGBA()
+	for y, my := r.Min.Y, mp.Y; y != r.Max.Y; y, my = y+1, my+1 {
+		for i, mi := i0, mi0; i < i1; i, mi = i+4, mi+1 {
+			ma := uint32(mask.Pix[mi])
+			if ma == 0 {
+				continue
+			}
+			ma |= ma << 8
+
+			// The 0x101 is here for the same reason as in drawRGBA.
+			a := (m - (sa * ma / m)) * 0x101
+
+			d := dst.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			d[0] = uint8((uint32(d[0])*a + sr*ma) / m >> 8)
+			d[1] = uint8((uint32(d[1])*a + sg*ma) / m >> 8)
+			d[2] = uint8((uint32(d[2])*a + sb*ma) / m >> 8)
+			d[3] = uint8((uint32(d[3])*a + sa*ma) / m >> 8)
+		}
+		i0 += dst.Stride
+		i1 += dst.Stride
+		mi0 += mask.Stride
+	}
+}
+
+func drawGrayMaskOver(dst *image.RGBA, r image.Rectangle, src *image.Gray, sp image.Point, mask *image.Alpha, mp image.Point) {
+	x0, x1, dx := r.Min.X, r.Max.X, 1
+	y0, y1, dy := r.Min.Y, r.Max.Y, 1
+	if r.Overlaps(r.Add(sp.Sub(r.Min))) {
+		if sp.Y < r.Min.Y || sp.Y == r.Min.Y && sp.X < r.Min.X {
+			x0, x1, dx = x1-1, x0-1, -1
+			y0, y1, dy = y1-1, y0-1, -1
+		}
+	}
+
+	sy := sp.Y + y0 - r.Min.Y
+	my := mp.Y + y0 - r.Min.Y
+	sx0 := sp.X + x0 - r.Min.X
+	mx0 := mp.X + x0 - r.Min.X
+	sx1 := sx0 + (x1 - x0)
+	i0 := dst.PixOffset(x0, y0)
+	di := dx * 4
+	for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+		for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+			mi := mask.PixOffset(mx, my)
+			ma := uint32(mask.Pix[mi])
+			ma |= ma << 8
+			si := src.PixOffset(sx, sy)
+			sy := uint32(src.Pix[si])
+			sy |= sy << 8
+			sa := uint32(0xffff)
+
+			d := dst.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			dr := uint32(d[0])
+			dg := uint32(d[1])
+			db := uint32(d[2])
+			da := uint32(d[3])
+
+			// dr, dg, db and da are all 8-bit color at the moment, ranging in [0,255].
+			// We work in 16-bit color, and so would normally do:
+			// dr |= dr << 8
+			// and similarly for dg, db and da, but instead we multiply a
+			// (which is a 16-bit color, ranging in [0,65535]) by 0x101.
+			// This yields the same result, but is fewer arithmetic operations.
+			a := (m - (sa * ma / m)) * 0x101
+
+			d[0] = uint8((dr*a + sy*ma) / m >> 8)
+			d[1] = uint8((dg*a + sy*ma) / m >> 8)
+			d[2] = uint8((db*a + sy*ma) / m >> 8)
+			d[3] = uint8((da*a + sa*ma) / m >> 8)
+		}
+		i0 += dy * dst.Stride
+	}
+}
+
+func drawRGBAMaskOver(dst *image.RGBA, r image.Rectangle, src *image.RGBA, sp image.Point, mask *image.Alpha, mp image.Point) {
+	x0, x1, dx := r.Min.X, r.Max.X, 1
+	y0, y1, dy := r.Min.Y, r.Max.Y, 1
+	if dst == src && r.Overlaps(r.Add(sp.Sub(r.Min))) {
+		if sp.Y < r.Min.Y || sp.Y == r.Min.Y && sp.X < r.Min.X {
+			x0, x1, dx = x1-1, x0-1, -1
+			y0, y1, dy = y1-1, y0-1, -1
+		}
+	}
+
+	sy := sp.Y + y0 - r.Min.Y
+	my := mp.Y + y0 - r.Min.Y
+	sx0 := sp.X + x0 - r.Min.X
+	mx0 := mp.X + x0 - r.Min.X
+	sx1 := sx0 + (x1 - x0)
+	i0 := dst.PixOffset(x0, y0)
+	di := dx * 4
+	for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+		for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+			mi := mask.PixOffset(mx, my)
+			ma := uint32(mask.Pix[mi])
+			ma |= ma << 8
+			si := src.PixOffset(sx, sy)
+			sr := uint32(src.Pix[si+0])
+			sg := uint32(src.Pix[si+1])
+			sb := uint32(src.Pix[si+2])
+			sa := uint32(src.Pix[si+3])
+			sr |= sr << 8
+			sg |= sg << 8
+			sb |= sb << 8
+			sa |= sa << 8
+			d := dst.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			dr := uint32(d[0])
+			dg := uint32(d[1])
+			db := uint32(d[2])
+			da := uint32(d[3])
+
+			// dr, dg, db and da are all 8-bit color at the moment, ranging in [0,255].
+			// We work in 16-bit color, and so would normally do:
+			// dr |= dr << 8
+			// and similarly for dg, db and da, but instead we multiply a
+			// (which is a 16-bit color, ranging in [0,65535]) by 0x101.
+			// This yields the same result, but is fewer arithmetic operations.
+			a := (m - (sa * ma / m)) * 0x101
+
+			d[0] = uint8((dr*a + sr*ma) / m >> 8)
+			d[1] = uint8((dg*a + sg*ma) / m >> 8)
+			d[2] = uint8((db*a + sb*ma) / m >> 8)
+			d[3] = uint8((da*a + sa*ma) / m >> 8)
+		}
+		i0 += dy * dst.Stride
+	}
+}
+
+func drawRGBA64ImageMaskOver(dst *image.RGBA, r image.Rectangle, src image.RGBA64Image, sp image.Point, mask *image.Alpha, mp image.Point) {
+	x0, x1, dx := r.Min.X, r.Max.X, 1
+	y0, y1, dy := r.Min.Y, r.Max.Y, 1
+	if image.Image(dst) == src && r.Overlaps(r.Add(sp.Sub(r.Min))) {
+		if sp.Y < r.Min.Y || sp.Y == r.Min.Y && sp.X < r.Min.X {
+			x0, x1, dx = x1-1, x0-1, -1
+			y0, y1, dy = y1-1, y0-1, -1
+		}
+	}
+
+	sy := sp.Y + y0 - r.Min.Y
+	my := mp.Y + y0 - r.Min.Y
+	sx0 := sp.X + x0 - r.Min.X
+	mx0 := mp.X + x0 - r.Min.X
+	sx1 := sx0 + (x1 - x0)
+	i0 := dst.PixOffset(x0, y0)
+	di := dx * 4
+	for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+		for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+			mi := mask.PixOffset(mx, my)
+			ma := uint32(mask.Pix[mi])
+			ma |= ma << 8
+			srgba := src.RGBA64At(sx, sy)
+			d := dst.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			dr := uint32(d[0])
+			dg := uint32(d[1])
+			db := uint32(d[2])
+			da := uint32(d[3])
+
+			// dr, dg, db and da are all 8-bit color at the moment, ranging in [0,255].
+			// We work in 16-bit color, and so would normally do:
+			// dr |= dr << 8
+			// and similarly for dg, db and da, but instead we multiply a
+			// (which is a 16-bit color, ranging in [0,65535]) by 0x101.
+			// This yields the same result, but is fewer arithmetic operations.
+			a := (m - (uint32(srgba.A) * ma / m)) * 0x101
+
+			d[0] = uint8((dr*a + uint32(srgba.R)*ma) / m >> 8)
+			d[1] = uint8((dg*a + uint32(srgba.G)*ma) / m >> 8)
+			d[2] = uint8((db*a + uint32(srgba.B)*ma) / m >> 8)
+			d[3] = uint8((da*a + uint32(srgba.A)*ma) / m >> 8)
+		}
+		i0 += dy * dst.Stride
+	}
+}
+
+func drawRGBA(dst *image.RGBA, r image.Rectangle, src image.Image, sp image.Point, mask image.Image, mp image.Point, op Op) {
+	x0, x1, dx := r.Min.X, r.Max.X, 1
+	y0, y1, dy := r.Min.Y, r.Max.Y, 1
+	if image.Image(dst) == src && r.Overlaps(r.Add(sp.Sub(r.Min))) {
+		if sp.Y < r.Min.Y || sp.Y == r.Min.Y && sp.X < r.Min.X {
+			x0, x1, dx = x1-1, x0-1, -1
+			y0, y1, dy = y1-1, y0-1, -1
+		}
+	}
+
+	sy := sp.Y + y0 - r.Min.Y
+	my := mp.Y + y0 - r.Min.Y
+	sx0 := sp.X + x0 - r.Min.X
+	mx0 := mp.X + x0 - r.Min.X
+	sx1 := sx0 + (x1 - x0)
+	i0 := dst.PixOffset(x0, y0)
+	di := dx * 4
+
+	// Try the image.RGBA64Image interface, part of the standard library since
+	// Go 1.17.
+	//
+	// This optimization is similar to how FALLBACK1.17 optimizes FALLBACK1.0
+	// in DrawMask, except here the concrete type of dst is known to be
+	// *image.RGBA.
+	if src0, _ := src.(image.RGBA64Image); src0 != nil {
+		if mask == nil {
+			if op == Over {
+				for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+					for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+						srgba := src0.RGBA64At(sx, sy)
+						d := dst.Pix[i : i+4 : i+4]
+						dr := uint32(d[0])
+						dg := uint32(d[1])
+						db := uint32(d[2])
+						da := uint32(d[3])
+						a := (m - uint32(srgba.A)) * 0x101
+						d[0] = uint8((dr*a/m + uint32(srgba.R)) >> 8)
+						d[1] = uint8((dg*a/m + uint32(srgba.G)) >> 8)
+						d[2] = uint8((db*a/m + uint32(srgba.B)) >> 8)
+						d[3] = uint8((da*a/m + uint32(srgba.A)) >> 8)
+					}
+					i0 += dy * dst.Stride
+				}
+			} else {
+				for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+					for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+						srgba := src0.RGBA64At(sx, sy)
+						d := dst.Pix[i : i+4 : i+4]
+						d[0] = uint8(srgba.R >> 8)
+						d[1] = uint8(srgba.G >> 8)
+						d[2] = uint8(srgba.B >> 8)
+						d[3] = uint8(srgba.A >> 8)
+					}
+					i0 += dy * dst.Stride
+				}
+			}
+			return
+
+		} else if mask0, _ := mask.(image.RGBA64Image); mask0 != nil {
+			if op == Over {
+				for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+					for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+						ma := uint32(mask0.RGBA64At(mx, my).A)
+						srgba := src0.RGBA64At(sx, sy)
+						d := dst.Pix[i : i+4 : i+4]
+						dr := uint32(d[0])
+						dg := uint32(d[1])
+						db := uint32(d[2])
+						da := uint32(d[3])
+						a := (m - (uint32(srgba.A) * ma / m)) * 0x101
+						d[0] = uint8((dr*a + uint32(srgba.R)*ma) / m >> 8)
+						d[1] = uint8((dg*a + uint32(srgba.G)*ma) / m >> 8)
+						d[2] = uint8((db*a + uint32(srgba.B)*ma) / m >> 8)
+						d[3] = uint8((da*a + uint32(srgba.A)*ma) / m >> 8)
+					}
+					i0 += dy * dst.Stride
+				}
+			} else {
+				for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+					for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+						ma := uint32(mask0.RGBA64At(mx, my).A)
+						srgba := src0.RGBA64At(sx, sy)
+						d := dst.Pix[i : i+4 : i+4]
+						d[0] = uint8(uint32(srgba.R) * ma / m >> 8)
+						d[1] = uint8(uint32(srgba.G) * ma / m >> 8)
+						d[2] = uint8(uint32(srgba.B) * ma / m >> 8)
+						d[3] = uint8(uint32(srgba.A) * ma / m >> 8)
+					}
+					i0 += dy * dst.Stride
+				}
+			}
+			return
+		}
+	}
+
+	// Use the image.Image interface, part of the standard library since Go
+	// 1.0.
+	//
+	// This is similar to FALLBACK1.0 in DrawMask, except here the concrete
+	// type of dst is known to be *image.RGBA.
+	for y := y0; y != y1; y, sy, my = y+dy, sy+dy, my+dy {
+		for i, sx, mx := i0, sx0, mx0; sx != sx1; i, sx, mx = i+di, sx+dx, mx+dx {
+			ma := uint32(m)
+			if mask != nil {
+				_, _, _, ma = mask.At(mx, my).RGBA()
+			}
+			sr, sg, sb, sa := src.At(sx, sy).RGBA()
+			d := dst.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+			if op == Over {
+				dr := uint32(d[0])
+				dg := uint32(d[1])
+				db := uint32(d[2])
+				da := uint32(d[3])
+
+				// dr, dg, db and da are all 8-bit color at the moment, ranging in [0,255].
+				// We work in 16-bit color, and so would normally do:
+				// dr |= dr << 8
+				// and similarly for dg, db and da, but instead we multiply a
+				// (which is a 16-bit color, ranging in [0,65535]) by 0x101.
+				// This yields the same result, but is fewer arithmetic operations.
+				a := (m - (sa * ma / m)) * 0x101
+
+				d[0] = uint8((dr*a + sr*ma) / m >> 8)
+				d[1] = uint8((dg*a + sg*ma) / m >> 8)
+				d[2] = uint8((db*a + sb*ma) / m >> 8)
+				d[3] = uint8((da*a + sa*ma) / m >> 8)
+
+			} else {
+				d[0] = uint8(sr * ma / m >> 8)
+				d[1] = uint8(sg * ma / m >> 8)
+				d[2] = uint8(sb * ma / m >> 8)
+				d[3] = uint8(sa * ma / m >> 8)
+			}
+		}
+		i0 += dy * dst.Stride
+	}
+}
+
+// clamp clamps i to the interval [0, 0xffff].
+func clamp(i int32) int32 {
+	if i < 0 {
+		return 0
+	}
+	if i > 0xffff {
+		return 0xffff
+	}
+	return i
+}
+
+// sqDiff returns the squared-difference of x and y, shifted by 2 so that
+// adding four of those won't overflow a uint32.
+//
+// x and y are both assumed to be in the range [0, 0xffff].
+func sqDiff(x, y int32) uint32 {
+	// This is an optimized code relying on the overflow/wrap around
+	// properties of unsigned integers operations guaranteed by the language
+	// spec. See sqDiff from the image/color package for more details.
+	d := uint32(x - y)
+	return (d * d) >> 2
+}
+
+func drawPaletted(dst Image, r image.Rectangle, src image.Image, sp image.Point, floydSteinberg bool) {
+	// TODO(nigeltao): handle the case where the dst and src overlap.
+	// Does it even make sense to try and do Floyd-Steinberg whilst
+	// walking the image backward (right-to-left bottom-to-top)?
+
+	// If dst is an *image.Paletted, we have a fast path for dst.Set and
+	// dst.At. The dst.Set equivalent is a batch version of the algorithm
+	// used by color.Palette's Index method in image/color/color.go, plus
+	// optional Floyd-Steinberg error diffusion.
+	palette, pix, stride := [][4]int32(nil), []byte(nil), 0
+	if p, ok := dst.(*image.Paletted); ok {
+		palette = make([][4]int32, len(p.Palette))
+		for i, col := range p.Palette {
+			r, g, b, a := col.RGBA()
+			palette[i][0] = int32(r)
+			palette[i][1] = int32(g)
+			palette[i][2] = int32(b)
+			palette[i][3] = int32(a)
+		}
+		pix, stride = p.Pix[p.PixOffset(r.Min.X, r.Min.Y):], p.Stride
+	}
+
+	// quantErrorCurr and quantErrorNext are the Floyd-Steinberg quantization
+	// errors that have been propagated to the pixels in the current and next
+	// rows. The +2 simplifies calculation near the edges.
+	var quantErrorCurr, quantErrorNext [][4]int32
+	if floydSteinberg {
+		quantErrorCurr = make([][4]int32, r.Dx()+2)
+		quantErrorNext = make([][4]int32, r.Dx()+2)
+	}
+	pxRGBA := func(x, y int) (r, g, b, a uint32) { return src.At(x, y).RGBA() }
+	// Fast paths for special cases to avoid excessive use of the color.Color
+	// interface which escapes to the heap but need to be discovered for
+	// each pixel on r. See also https://golang.org/issues/15759.
+	switch src0 := src.(type) {
+	case *image.RGBA:
+		pxRGBA = func(x, y int) (r, g, b, a uint32) { return src0.RGBAAt(x, y).RGBA() }
+	case *image.NRGBA:
+		pxRGBA = func(x, y int) (r, g, b, a uint32) { return src0.NRGBAAt(x, y).RGBA() }
+	case *image.YCbCr:
+		pxRGBA = func(x, y int) (r, g, b, a uint32) { return src0.YCbCrAt(x, y).RGBA() }
+	}
+
+	// Loop over each source pixel.
+	out := color.RGBA64{A: 0xffff}
+	for y := 0; y != r.Dy(); y++ {
+		for x := 0; x != r.Dx(); x++ {
+			// er, eg and eb are the pixel's R,G,B values plus the
+			// optional Floyd-Steinberg error.
+			sr, sg, sb, sa := pxRGBA(sp.X+x, sp.Y+y)
+			er, eg, eb, ea := int32(sr), int32(sg), int32(sb), int32(sa)
+			if floydSteinberg {
+				er = clamp(er + quantErrorCurr[x+1][0]/16)
+				eg = clamp(eg + quantErrorCurr[x+1][1]/16)
+				eb = clamp(eb + quantErrorCurr[x+1][2]/16)
+				ea = clamp(ea + quantErrorCurr[x+1][3]/16)
+			}
+
+			if palette != nil {
+				// Find the closest palette color in Euclidean R,G,B,A space:
+				// the one that minimizes sum-squared-difference.
+				// TODO(nigeltao): consider smarter algorithms.
+				bestIndex, bestSum := 0, uint32(1<<32-1)
+				for index, p := range palette {
+					sum := sqDiff(er, p[0]) + sqDiff(eg, p[1]) + sqDiff(eb, p[2]) + sqDiff(ea, p[3])
+					if sum < bestSum {
+						bestIndex, bestSum = index, sum
+						if sum == 0 {
+							break
+						}
+					}
+				}
+				pix[y*stride+x] = byte(bestIndex)
+
+				if !floydSteinberg {
+					continue
+				}
+				er -= palette[bestIndex][0]
+				eg -= palette[bestIndex][1]
+				eb -= palette[bestIndex][2]
+				ea -= palette[bestIndex][3]
+
+			} else {
+				out.R = uint16(er)
+				out.G = uint16(eg)
+				out.B = uint16(eb)
+				out.A = uint16(ea)
+				// The third argument is &out instead of out (and out is
+				// declared outside of the inner loop) to avoid the implicit
+				// conversion to color.Color here allocating memory in the
+				// inner loop if sizeof(color.RGBA64) > sizeof(uintptr).
+				dst.Set(r.Min.X+x, r.Min.Y+y, &out)
+
+				if !floydSteinberg {
+					continue
+				}
+				sr, sg, sb, sa = dst.At(r.Min.X+x, r.Min.Y+y).RGBA()
+				er -= int32(sr)
+				eg -= int32(sg)
+				eb -= int32(sb)
+				ea -= int32(sa)
+			}
+
+			// Propagate the Floyd-Steinberg quantization error.
+			quantErrorNext[x+0][0] += er * 3
+			quantErrorNext[x+0][1] += eg * 3
+			quantErrorNext[x+0][2] += eb * 3
+			quantErrorNext[x+0][3] += ea * 3
+			quantErrorNext[x+1][0] += er * 5
+			quantErrorNext[x+1][1] += eg * 5
+			quantErrorNext[x+1][2] += eb * 5
+			quantErrorNext[x+1][3] += ea * 5
+			quantErrorNext[x+2][0] += er * 1
+			quantErrorNext[x+2][1] += eg * 1
+			quantErrorNext[x+2][2] += eb * 1
+			quantErrorNext[x+2][3] += ea * 1
+			quantErrorCurr[x+2][0] += er * 7
+			quantErrorCurr[x+2][1] += eg * 7
+			quantErrorCurr[x+2][2] += eb * 7
+			quantErrorCurr[x+2][3] += ea * 7
+		}
+
+		// Recycle the quantization error buffers.
+		if floydSteinberg {
+			quantErrorCurr, quantErrorNext = quantErrorNext, quantErrorCurr
+			for i := range quantErrorNext {
+				quantErrorNext[i] = [4]int32{}
+			}
+		}
+	}
+}
diff --git a/src/image/draw/draw_test.go b/src/image/draw/draw_test.go
new file mode 100644
index 0000000..a34d1c3
--- /dev/null
+++ b/src/image/draw/draw_test.go
@@ -0,0 +1,810 @@
+// Copyright 2010 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package draw
+
+import (
+	"image"
+	"image/color"
+	"image/png"
+	"os"
+	"testing"
+	"testing/quick"
+)
+
+// slowestRGBA is a draw.Image like image.RGBA but it is a different type and
+// therefore does not trigger the draw.go fastest code paths.
+//
+// Unlike slowerRGBA, it does not implement the draw.RGBA64Image interface.
+type slowestRGBA struct {
+	Pix    []uint8
+	Stride int
+	Rect   image.Rectangle
+}
+
+func (p *slowestRGBA) ColorModel() color.Model { return color.RGBAModel }
+
+func (p *slowestRGBA) Bounds() image.Rectangle { return p.Rect }
+
+func (p *slowestRGBA) At(x, y int) color.Color {
+	return p.RGBA64At(x, y)
+}
+
+func (p *slowestRGBA) RGBA64At(x, y int) color.RGBA64 {
+	if !(image.Point{x, y}.In(p.Rect)) {
+		return color.RGBA64{}
+	}
+	i := p.PixOffset(x, y)
+	s := p.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+	r := uint16(s[0])
+	g := uint16(s[1])
+	b := uint16(s[2])
+	a := uint16(s[3])
+	return color.RGBA64{
+		(r << 8) | r,
+		(g << 8) | g,
+		(b << 8) | b,
+		(a << 8) | a,
+	}
+}
+
+func (p *slowestRGBA) Set(x, y int, c color.Color) {
+	if !(image.Point{x, y}.In(p.Rect)) {
+		return
+	}
+	i := p.PixOffset(x, y)
+	c1 := color.RGBAModel.Convert(c).(color.RGBA)
+	s := p.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+	s[0] = c1.R
+	s[1] = c1.G
+	s[2] = c1.B
+	s[3] = c1.A
+}
+
+func (p *slowestRGBA) PixOffset(x, y int) int {
+	return (y-p.Rect.Min.Y)*p.Stride + (x-p.Rect.Min.X)*4
+}
+
+func convertToSlowestRGBA(m image.Image) *slowestRGBA {
+	if rgba, ok := m.(*image.RGBA); ok {
+		return &slowestRGBA{
+			Pix:    append([]byte(nil), rgba.Pix...),
+			Stride: rgba.Stride,
+			Rect:   rgba.Rect,
+		}
+	}
+	rgba := image.NewRGBA(m.Bounds())
+	Draw(rgba, rgba.Bounds(), m, m.Bounds().Min, Src)
+	return &slowestRGBA{
+		Pix:    rgba.Pix,
+		Stride: rgba.Stride,
+		Rect:   rgba.Rect,
+	}
+}
+
+func init() {
+	var p any = (*slowestRGBA)(nil)
+	if _, ok := p.(RGBA64Image); ok {
+		panic("slowestRGBA should not be an RGBA64Image")
+	}
+}
+
+// slowerRGBA is a draw.Image like image.RGBA but it is a different type and
+// therefore does not trigger the draw.go fastest code paths.
+//
+// Unlike slowestRGBA, it still implements the draw.RGBA64Image interface.
+type slowerRGBA struct {
+	Pix    []uint8
+	Stride int
+	Rect   image.Rectangle
+}
+
+func (p *slowerRGBA) ColorModel() color.Model { return color.RGBAModel }
+
+func (p *slowerRGBA) Bounds() image.Rectangle { return p.Rect }
+
+func (p *slowerRGBA) At(x, y int) color.Color {
+	return p.RGBA64At(x, y)
+}
+
+func (p *slowerRGBA) RGBA64At(x, y int) color.RGBA64 {
+	if !(image.Point{x, y}.In(p.Rect)) {
+		return color.RGBA64{}
+	}
+	i := p.PixOffset(x, y)
+	s := p.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+	r := uint16(s[0])
+	g := uint16(s[1])
+	b := uint16(s[2])
+	a := uint16(s[3])
+	return color.RGBA64{
+		(r << 8) | r,
+		(g << 8) | g,
+		(b << 8) | b,
+		(a << 8) | a,
+	}
+}
+
+func (p *slowerRGBA) Set(x, y int, c color.Color) {
+	if !(image.Point{x, y}.In(p.Rect)) {
+		return
+	}
+	i := p.PixOffset(x, y)
+	c1 := color.RGBAModel.Convert(c).(color.RGBA)
+	s := p.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+	s[0] = c1.R
+	s[1] = c1.G
+	s[2] = c1.B
+	s[3] = c1.A
+}
+
+func (p *slowerRGBA) SetRGBA64(x, y int, c color.RGBA64) {
+	if !(image.Point{x, y}.In(p.Rect)) {
+		return
+	}
+	i := p.PixOffset(x, y)
+	s := p.Pix[i : i+4 : i+4] // Small cap improves performance, see https://golang.org/issue/27857
+	s[0] = uint8(c.R >> 8)
+	s[1] = uint8(c.G >> 8)
+	s[2] = uint8(c.B >> 8)
+	s[3] = uint8(c.A >> 8)
+}
+
+func (p *slowerRGBA) PixOffset(x, y int) int {
+	return (y-p.Rect.Min.Y)*p.Stride + (x-p.Rect.Min.X)*4
+}
+
+func convertToSlowerRGBA(m image.Image) *slowerRGBA {
+	if rgba, ok := m.(*image.RGBA); ok {
+		return &slowerRGBA{
+			Pix:    append([]byte(nil), rgba.Pix...),
+			Stride: rgba.Stride,
+			Rect:   rgba.Rect,
+		}
+	}
+	rgba := image.NewRGBA(m.Bounds())
+	Draw(rgba, rgba.Bounds(), m, m.Bounds().Min, Src)
+	return &slowerRGBA{
+		Pix:    rgba.Pix,
+		Stride: rgba.Stride,
+		Rect:   rgba.Rect,
+	}
+}
+
+func init() {
+	var p any = (*slowerRGBA)(nil)
+	if _, ok := p.(RGBA64Image); !ok {
+		panic("slowerRGBA should be an RGBA64Image")
+	}
+}
+
+func eq(c0, c1 color.Color) bool {
+	r0, g0, b0, a0 := c0.RGBA()
+	r1, g1, b1, a1 := c1.RGBA()
+	return r0 == r1 && g0 == g1 && b0 == b1 && a0 == a1
+}
+
+func fillBlue(alpha int) image.Image {
+	return image.NewUniform(color.RGBA{0, 0, uint8(alpha), uint8(alpha)})
+}
+
+func fillAlpha(alpha int) image.Image {
+	return image.NewUniform(color.Alpha{uint8(alpha)})
+}
+
+func vgradGreen(alpha int) image.Image {
+	m := image.NewRGBA(image.Rect(0, 0, 16, 16))
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Set(x, y, color.RGBA{0, uint8(y * alpha / 15), 0, uint8(alpha)})
+		}
+	}
+	return m
+}
+
+func vgradAlpha(alpha int) image.Image {
+	m := image.NewAlpha(image.Rect(0, 0, 16, 16))
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Set(x, y, color.Alpha{uint8(y * alpha / 15)})
+		}
+	}
+	return m
+}
+
+func vgradGreenNRGBA(alpha int) image.Image {
+	m := image.NewNRGBA(image.Rect(0, 0, 16, 16))
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Set(x, y, color.RGBA{0, uint8(y * 0x11), 0, uint8(alpha)})
+		}
+	}
+	return m
+}
+
+func vgradCr() image.Image {
+	m := &image.YCbCr{
+		Y:              make([]byte, 16*16),
+		Cb:             make([]byte, 16*16),
+		Cr:             make([]byte, 16*16),
+		YStride:        16,
+		CStride:        16,
+		SubsampleRatio: image.YCbCrSubsampleRatio444,
+		Rect:           image.Rect(0, 0, 16, 16),
+	}
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Cr[y*m.CStride+x] = uint8(y * 0x11)
+		}
+	}
+	return m
+}
+
+func vgradGray() image.Image {
+	m := image.NewGray(image.Rect(0, 0, 16, 16))
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Set(x, y, color.Gray{uint8(y * 0x11)})
+		}
+	}
+	return m
+}
+
+func vgradMagenta() image.Image {
+	m := image.NewCMYK(image.Rect(0, 0, 16, 16))
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Set(x, y, color.CMYK{0, uint8(y * 0x11), 0, 0x3f})
+		}
+	}
+	return m
+}
+
+func hgradRed(alpha int) Image {
+	m := image.NewRGBA(image.Rect(0, 0, 16, 16))
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Set(x, y, color.RGBA{uint8(x * alpha / 15), 0, 0, uint8(alpha)})
+		}
+	}
+	return m
+}
+
+func gradYellow(alpha int) Image {
+	m := image.NewRGBA(image.Rect(0, 0, 16, 16))
+	for y := 0; y < 16; y++ {
+		for x := 0; x < 16; x++ {
+			m.Set(x, y, color.RGBA{uint8(x * alpha / 15), uint8(y * alpha / 15), 0, uint8(alpha)})
+		}
+	}
+	return m
+}
+
+type drawTest struct {
+	desc     string
+	src      image.Image
+	mask     image.Image
+	op       Op
+	expected color.Color
+}
+
+var drawTests = []drawTest{
+	// Uniform mask (0% opaque).
+	{"nop", vgradGreen(255), fillAlpha(0), Over, color.RGBA{136, 0, 0, 255}},
+	{"clear", vgradGreen(255), fillAlpha(0), Src, color.RGBA{0, 0, 0, 0}},
+	// Uniform mask (100%, 75%, nil) and uniform source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is {0, 0, 90, 90}.
+	{"fill", fillBlue(90), fillAlpha(255), Over, color.RGBA{88, 0, 90, 255}},
+	{"fillSrc", fillBlue(90), fillAlpha(255), Src, color.RGBA{0, 0, 90, 90}},
+	{"fillAlpha", fillBlue(90), fillAlpha(192), Over, color.RGBA{100, 0, 68, 255}},
+	{"fillAlphaSrc", fillBlue(90), fillAlpha(192), Src, color.RGBA{0, 0, 68, 68}},
+	{"fillNil", fillBlue(90), nil, Over, color.RGBA{88, 0, 90, 255}},
+	{"fillNilSrc", fillBlue(90), nil, Src, color.RGBA{0, 0, 90, 90}},
+	// Uniform mask (100%, 75%, nil) and variable source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is {0, 48, 0, 90}.
+	{"copy", vgradGreen(90), fillAlpha(255), Over, color.RGBA{88, 48, 0, 255}},
+	{"copySrc", vgradGreen(90), fillAlpha(255), Src, color.RGBA{0, 48, 0, 90}},
+	{"copyAlpha", vgradGreen(90), fillAlpha(192), Over, color.RGBA{100, 36, 0, 255}},
+	{"copyAlphaSrc", vgradGreen(90), fillAlpha(192), Src, color.RGBA{0, 36, 0, 68}},
+	{"copyNil", vgradGreen(90), nil, Over, color.RGBA{88, 48, 0, 255}},
+	{"copyNilSrc", vgradGreen(90), nil, Src, color.RGBA{0, 48, 0, 90}},
+	// Uniform mask (100%, 75%, nil) and variable NRGBA source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is {0, 136, 0, 90} in NRGBA-space, which is {0, 48, 0, 90} in RGBA-space.
+	// The result pixel is different than in the "copy*" test cases because of rounding errors.
+	{"nrgba", vgradGreenNRGBA(90), fillAlpha(255), Over, color.RGBA{88, 46, 0, 255}},
+	{"nrgbaSrc", vgradGreenNRGBA(90), fillAlpha(255), Src, color.RGBA{0, 46, 0, 90}},
+	{"nrgbaAlpha", vgradGreenNRGBA(90), fillAlpha(192), Over, color.RGBA{100, 34, 0, 255}},
+	{"nrgbaAlphaSrc", vgradGreenNRGBA(90), fillAlpha(192), Src, color.RGBA{0, 34, 0, 68}},
+	{"nrgbaNil", vgradGreenNRGBA(90), nil, Over, color.RGBA{88, 46, 0, 255}},
+	{"nrgbaNilSrc", vgradGreenNRGBA(90), nil, Src, color.RGBA{0, 46, 0, 90}},
+	// Uniform mask (100%, 75%, nil) and variable YCbCr source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is {0, 0, 136} in YCbCr-space, which is {11, 38, 0, 255} in RGB-space.
+	{"ycbcr", vgradCr(), fillAlpha(255), Over, color.RGBA{11, 38, 0, 255}},
+	{"ycbcrSrc", vgradCr(), fillAlpha(255), Src, color.RGBA{11, 38, 0, 255}},
+	{"ycbcrAlpha", vgradCr(), fillAlpha(192), Over, color.RGBA{42, 28, 0, 255}},
+	{"ycbcrAlphaSrc", vgradCr(), fillAlpha(192), Src, color.RGBA{8, 28, 0, 192}},
+	{"ycbcrNil", vgradCr(), nil, Over, color.RGBA{11, 38, 0, 255}},
+	{"ycbcrNilSrc", vgradCr(), nil, Src, color.RGBA{11, 38, 0, 255}},
+	// Uniform mask (100%, 75%, nil) and variable Gray source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is {136} in Gray-space, which is {136, 136, 136, 255} in RGBA-space.
+	{"gray", vgradGray(), fillAlpha(255), Over, color.RGBA{136, 136, 136, 255}},
+	{"graySrc", vgradGray(), fillAlpha(255), Src, color.RGBA{136, 136, 136, 255}},
+	{"grayAlpha", vgradGray(), fillAlpha(192), Over, color.RGBA{136, 102, 102, 255}},
+	{"grayAlphaSrc", vgradGray(), fillAlpha(192), Src, color.RGBA{102, 102, 102, 192}},
+	{"grayNil", vgradGray(), nil, Over, color.RGBA{136, 136, 136, 255}},
+	{"grayNilSrc", vgradGray(), nil, Src, color.RGBA{136, 136, 136, 255}},
+	// Same again, but with a slowerRGBA source.
+	{"graySlower", convertToSlowerRGBA(vgradGray()), fillAlpha(255),
+		Over, color.RGBA{136, 136, 136, 255}},
+	{"graySrcSlower", convertToSlowerRGBA(vgradGray()), fillAlpha(255),
+		Src, color.RGBA{136, 136, 136, 255}},
+	{"grayAlphaSlower", convertToSlowerRGBA(vgradGray()), fillAlpha(192),
+		Over, color.RGBA{136, 102, 102, 255}},
+	{"grayAlphaSrcSlower", convertToSlowerRGBA(vgradGray()), fillAlpha(192),
+		Src, color.RGBA{102, 102, 102, 192}},
+	{"grayNilSlower", convertToSlowerRGBA(vgradGray()), nil,
+		Over, color.RGBA{136, 136, 136, 255}},
+	{"grayNilSrcSlower", convertToSlowerRGBA(vgradGray()), nil,
+		Src, color.RGBA{136, 136, 136, 255}},
+	// Same again, but with a slowestRGBA source.
+	{"graySlowest", convertToSlowestRGBA(vgradGray()), fillAlpha(255),
+		Over, color.RGBA{136, 136, 136, 255}},
+	{"graySrcSlowest", convertToSlowestRGBA(vgradGray()), fillAlpha(255),
+		Src, color.RGBA{136, 136, 136, 255}},
+	{"grayAlphaSlowest", convertToSlowestRGBA(vgradGray()), fillAlpha(192),
+		Over, color.RGBA{136, 102, 102, 255}},
+	{"grayAlphaSrcSlowest", convertToSlowestRGBA(vgradGray()), fillAlpha(192),
+		Src, color.RGBA{102, 102, 102, 192}},
+	{"grayNilSlowest", convertToSlowestRGBA(vgradGray()), nil,
+		Over, color.RGBA{136, 136, 136, 255}},
+	{"grayNilSrcSlowest", convertToSlowestRGBA(vgradGray()), nil,
+		Src, color.RGBA{136, 136, 136, 255}},
+	// Uniform mask (100%, 75%, nil) and variable CMYK source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is {0, 136, 0, 63} in CMYK-space, which is {192, 89, 192} in RGB-space.
+	{"cmyk", vgradMagenta(), fillAlpha(255), Over, color.RGBA{192, 89, 192, 255}},
+	{"cmykSrc", vgradMagenta(), fillAlpha(255), Src, color.RGBA{192, 89, 192, 255}},
+	{"cmykAlpha", vgradMagenta(), fillAlpha(192), Over, color.RGBA{178, 67, 145, 255}},
+	{"cmykAlphaSrc", vgradMagenta(), fillAlpha(192), Src, color.RGBA{145, 67, 145, 192}},
+	{"cmykNil", vgradMagenta(), nil, Over, color.RGBA{192, 89, 192, 255}},
+	{"cmykNilSrc", vgradMagenta(), nil, Src, color.RGBA{192, 89, 192, 255}},
+	// Variable mask and uniform source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is {0, 0, 255, 255}.
+	// The mask pixel's alpha is 102, or 40%.
+	{"generic", fillBlue(255), vgradAlpha(192), Over, color.RGBA{81, 0, 102, 255}},
+	{"genericSrc", fillBlue(255), vgradAlpha(192), Src, color.RGBA{0, 0, 102, 102}},
+	// Same again, but with a slowerRGBA mask.
+	{"genericSlower", fillBlue(255), convertToSlowerRGBA(vgradAlpha(192)),
+		Over, color.RGBA{81, 0, 102, 255}},
+	{"genericSrcSlower", fillBlue(255), convertToSlowerRGBA(vgradAlpha(192)),
+		Src, color.RGBA{0, 0, 102, 102}},
+	// Same again, but with a slowestRGBA mask.
+	{"genericSlowest", fillBlue(255), convertToSlowestRGBA(vgradAlpha(192)),
+		Over, color.RGBA{81, 0, 102, 255}},
+	{"genericSrcSlowest", fillBlue(255), convertToSlowestRGBA(vgradAlpha(192)),
+		Src, color.RGBA{0, 0, 102, 102}},
+	// Variable mask and variable source.
+	// At (x, y) == (8, 8):
+	// The destination pixel is {136, 0, 0, 255}.
+	// The source pixel is:
+	//   - {0, 48, 0, 90}.
+	//   - {136} in Gray-space, which is {136, 136, 136, 255} in RGBA-space.
+	// The mask pixel's alpha is 102, or 40%.
+	{"rgbaVariableMaskOver", vgradGreen(90), vgradAlpha(192), Over, color.RGBA{117, 19, 0, 255}},
+	{"grayVariableMaskOver", vgradGray(), vgradAlpha(192), Over, color.RGBA{136, 54, 54, 255}},
+}
+
+func makeGolden(dst image.Image, r image.Rectangle, src image.Image, sp image.Point, mask image.Image, mp image.Point, op Op) image.Image {
+	// Since golden is a newly allocated image, we don't have to check if the
+	// input source and mask images and the output golden image overlap.
+	b := dst.Bounds()
+	sb := src.Bounds()
+	mb := image.Rect(-1e9, -1e9, 1e9, 1e9)
+	if mask != nil {
+		mb = mask.Bounds()
+	}
+	golden := image.NewRGBA(image.Rect(0, 0, b.Max.X, b.Max.Y))
+	for y := r.Min.Y; y < r.Max.Y; y++ {
+		sy := y + sp.Y - r.Min.Y
+		my := y + mp.Y - r.Min.Y
+		for x := r.Min.X; x < r.Max.X; x++ {
+			if !(image.Pt(x, y).In(b)) {
+				continue
+			}
+			sx := x + sp.X - r.Min.X
+			if !(image.Pt(sx, sy).In(sb)) {
+				continue
+			}
+			mx := x + mp.X - r.Min.X
+			if !(image.Pt(mx, my).In(mb)) {
+				continue
+			}
+
+			const M = 1<<16 - 1
+			var dr, dg, db, da uint32
+			if op == Over {
+				dr, dg, db, da = dst.At(x, y).RGBA()
+			}
+			sr, sg, sb, sa := src.At(sx, sy).RGBA()
+			ma := uint32(M)
+			if mask != nil {
+				_, _, _, ma = mask.At(mx, my).RGBA()
+			}
+			a := M - (sa * ma / M)
+			golden.Set(x, y, color.RGBA64{
+				uint16((dr*a + sr*ma) / M),
+				uint16((dg*a + sg*ma) / M),
+				uint16((db*a + sb*ma) / M),
+				uint16((da*a + sa*ma) / M),
+			})
+		}
+	}
+	return golden.SubImage(b)
+}
+
+func TestDraw(t *testing.T) {
+	rr := []image.Rectangle{
+		image.Rect(0, 0, 0, 0),
+		image.Rect(0, 0, 16, 16),
+		image.Rect(3, 5, 12, 10),
+		image.Rect(0, 0, 9, 9),
+		image.Rect(8, 8, 16, 16),
+		image.Rect(8, 0, 9, 16),
+		image.Rect(0, 8, 16, 9),
+		image.Rect(8, 8, 9, 9),
+		image.Rect(8, 8, 8, 8),
+	}
+	for _, r := range rr {
+	loop:
+		for _, test := range drawTests {
+			for i := 0; i < 3; i++ {
+				dst := hgradRed(255).(*image.RGBA).SubImage(r).(Image)
+				// For i != 0, substitute a different-typed dst that will take
+				// us off the fastest code paths. We should still get the same
+				// result, in terms of final pixel RGBA values.
+				switch i {
+				case 1:
+					dst = convertToSlowerRGBA(dst)
+				case 2:
+					dst = convertToSlowestRGBA(dst)
+				}
+
+				// Draw the (src, mask, op) onto a copy of dst using a slow but obviously correct implementation.
+				golden := makeGolden(dst, image.Rect(0, 0, 16, 16), test.src, image.ZP, test.mask, image.ZP, test.op)
+				b := dst.Bounds()
+				if !b.Eq(golden.Bounds()) {
+					t.Errorf("draw %v %s on %T: bounds %v versus %v",
+						r, test.desc, dst, dst.Bounds(), golden.Bounds())
+					continue
+				}
+				// Draw the same combination onto the actual dst using the optimized DrawMask implementation.
+				DrawMask(dst, image.Rect(0, 0, 16, 16), test.src, image.ZP, test.mask, image.ZP, test.op)
+				if image.Pt(8, 8).In(r) {
+					// Check that the resultant pixel at (8, 8) matches what we expect
+					// (the expected value can be verified by hand).
+					if !eq(dst.At(8, 8), test.expected) {
+						t.Errorf("draw %v %s on %T: at (8, 8) %v versus %v",
+							r, test.desc, dst, dst.At(8, 8), test.expected)
+						continue
+					}
+				}
+				// Check that the resultant dst image matches the golden output.
+				for y := b.Min.Y; y < b.Max.Y; y++ {
+					for x := b.Min.X; x < b.Max.X; x++ {
+						if !eq(dst.At(x, y), golden.At(x, y)) {
+							t.Errorf("draw %v %s on %T: at (%d, %d), %v versus golden %v",
+								r, test.desc, dst, x, y, dst.At(x, y), golden.At(x, y))
+							continue loop
+						}
+					}
+				}
+			}
+		}
+	}
+}
+
+func TestDrawOverlap(t *testing.T) {
+	for _, op := range []Op{Over, Src} {
+		for yoff := -2; yoff <= 2; yoff++ {
+		loop:
+			for xoff := -2; xoff <= 2; xoff++ {
+				m := gradYellow(127).(*image.RGBA)
+				dst := m.SubImage(image.Rect(5, 5, 10, 10)).(*image.RGBA)
+				src := m.SubImage(image.Rect(5+xoff, 5+yoff, 10+xoff, 10+yoff)).(*image.RGBA)
+				b := dst.Bounds()
+				// Draw the (src, mask, op) onto a copy of dst using a slow but obviously correct implementation.
+				golden := makeGolden(dst, b, src, src.Bounds().Min, nil, image.ZP, op)
+				if !b.Eq(golden.Bounds()) {
+					t.Errorf("drawOverlap xoff=%d,yoff=%d: bounds %v versus %v", xoff, yoff, dst.Bounds(), golden.Bounds())
+					continue
+				}
+				// Draw the same combination onto the actual dst using the optimized DrawMask implementation.
+				DrawMask(dst, b, src, src.Bounds().Min, nil, image.ZP, op)
+				// Check that the resultant dst image matches the golden output.
+				for y := b.Min.Y; y < b.Max.Y; y++ {
+					for x := b.Min.X; x < b.Max.X; x++ {
+						if !eq(dst.At(x, y), golden.At(x, y)) {
+							t.Errorf("drawOverlap xoff=%d,yoff=%d: at (%d, %d), %v versus golden %v", xoff, yoff, x, y, dst.At(x, y), golden.At(x, y))
+							continue loop
+						}
+					}
+				}
+			}
+		}
+	}
+}
+
+// TestNonZeroSrcPt checks drawing with a non-zero src point parameter.
+func TestNonZeroSrcPt(t *testing.T) {
+	a := image.NewRGBA(image.Rect(0, 0, 1, 1))
+	b := image.NewRGBA(image.Rect(0, 0, 2, 2))
+	b.Set(0, 0, color.RGBA{0, 0, 0, 5})
+	b.Set(1, 0, color.RGBA{0, 0, 5, 5})
+	b.Set(0, 1, color.RGBA{0, 5, 0, 5})
+	b.Set(1, 1, color.RGBA{5, 0, 0, 5})
+	Draw(a, image.Rect(0, 0, 1, 1), b, image.Pt(1, 1), Over)
+	if !eq(color.RGBA{5, 0, 0, 5}, a.At(0, 0)) {
+		t.Errorf("non-zero src pt: want %v got %v", color.RGBA{5, 0, 0, 5}, a.At(0, 0))
+	}
+}
+
+func TestFill(t *testing.T) {
+	rr := []image.Rectangle{
+		image.Rect(0, 0, 0, 0),
+		image.Rect(0, 0, 40, 30),
+		image.Rect(10, 0, 40, 30),
+		image.Rect(0, 20, 40, 30),
+		image.Rect(10, 20, 40, 30),
+		image.Rect(10, 20, 15, 25),
+		image.Rect(10, 0, 35, 30),
+		image.Rect(0, 15, 40, 16),
+		image.Rect(24, 24, 25, 25),
+		image.Rect(23, 23, 26, 26),
+		image.Rect(22, 22, 27, 27),
+		image.Rect(21, 21, 28, 28),
+		image.Rect(20, 20, 29, 29),
+	}
+	for _, r := range rr {
+		m := image.NewRGBA(image.Rect(0, 0, 40, 30)).SubImage(r).(*image.RGBA)
+		b := m.Bounds()
+		c := color.RGBA{11, 0, 0, 255}
+		src := &image.Uniform{C: c}
+		check := func(desc string) {
+			for y := b.Min.Y; y < b.Max.Y; y++ {
+				for x := b.Min.X; x < b.Max.X; x++ {
+					if !eq(c, m.At(x, y)) {
+						t.Errorf("%s fill: at (%d, %d), sub-image bounds=%v: want %v got %v", desc, x, y, r, c, m.At(x, y))
+						return
+					}
+				}
+			}
+		}
+		// Draw 1 pixel at a time.
+		for y := b.Min.Y; y < b.Max.Y; y++ {
+			for x := b.Min.X; x < b.Max.X; x++ {
+				DrawMask(m, image.Rect(x, y, x+1, y+1), src, image.ZP, nil, image.ZP, Src)
+			}
+		}
+		check("pixel")
+		// Draw 1 row at a time.
+		c = color.RGBA{0, 22, 0, 255}
+		src = &image.Uniform{C: c}
+		for y := b.Min.Y; y < b.Max.Y; y++ {
+			DrawMask(m, image.Rect(b.Min.X, y, b.Max.X, y+1), src, image.ZP, nil, image.ZP, Src)
+		}
+		check("row")
+		// Draw 1 column at a time.
+		c = color.RGBA{0, 0, 33, 255}
+		src = &image.Uniform{C: c}
+		for x := b.Min.X; x < b.Max.X; x++ {
+			DrawMask(m, image.Rect(x, b.Min.Y, x+1, b.Max.Y), src, image.ZP, nil, image.ZP, Src)
+		}
+		check("column")
+		// Draw the whole image at once.
+		c = color.RGBA{44, 55, 66, 77}
+		src = &image.Uniform{C: c}
+		DrawMask(m, b, src, image.ZP, nil, image.ZP, Src)
+		check("whole")
+	}
+}
+
+func TestDrawSrcNonpremultiplied(t *testing.T) {
+	var (
+		opaqueGray       = color.NRGBA{0x99, 0x99, 0x99, 0xff}
+		transparentBlue  = color.NRGBA{0x00, 0x00, 0xff, 0x00}
+		transparentGreen = color.NRGBA{0x00, 0xff, 0x00, 0x00}
+		transparentRed   = color.NRGBA{0xff, 0x00, 0x00, 0x00}
+
+		opaqueGray64        = color.NRGBA64{0x9999, 0x9999, 0x9999, 0xffff}
+		transparentPurple64 = color.NRGBA64{0xfedc, 0x0000, 0x7654, 0x0000}
+	)
+
+	// dst and src are 1x3 images but the dr rectangle (and hence the overlap)
+	// is only 1x2. The Draw call should affect dst's pixels at (1, 10) and (2,
+	// 10) but the pixel at (0, 10) should be untouched.
+	//
+	// The src image is entirely transparent (and the Draw operator is Src) so
+	// the two touched pixels should be set to transparent colors.
+	//
+	// In general, Go's color.Color type (and specifically the Color.RGBA
+	// method) works in premultiplied alpha, where there's no difference
+	// between "transparent blue" and "transparent red". It's all "just
+	// transparent" and canonically "transparent black" (all zeroes).
+	//
+	// However, since the operator is Src (so the pixels are 'copied', not
+	// 'blended') and both dst and src images are *image.NRGBA (N stands for
+	// Non-premultiplied alpha which *does* distinguish "transparent blue" and
+	// "transparent red"), we prefer that this distinction carries through and
+	// dst's touched pixels should be transparent blue and transparent green,
+	// not just transparent black.
+	{
+		dst := image.NewNRGBA(image.Rect(0, 10, 3, 11))
+		dst.SetNRGBA(0, 10, opaqueGray)
+		src := image.NewNRGBA(image.Rect(1, 20, 4, 21))
+		src.SetNRGBA(1, 20, transparentBlue)
+		src.SetNRGBA(2, 20, transparentGreen)
+		src.SetNRGBA(3, 20, transparentRed)
+
+		dr := image.Rect(1, 10, 3, 11)
+		Draw(dst, dr, src, image.Point{1, 20}, Src)
+
+		if got, want := dst.At(0, 10), opaqueGray; got != want {
+			t.Errorf("At(0, 10):\ngot  %#v\nwant %#v", got, want)
+		}
+		if got, want := dst.At(1, 10), transparentBlue; got != want {
+			t.Errorf("At(1, 10):\ngot  %#v\nwant %#v", got, want)
+		}
+		if got, want := dst.At(2, 10), transparentGreen; got != want {
+			t.Errorf("At(2, 10):\ngot  %#v\nwant %#v", got, want)
+		}
+	}
+
+	// Check image.NRGBA64 (not image.NRGBA) similarly.
+	{
+		dst := image.NewNRGBA64(image.Rect(0, 0, 1, 1))
+		dst.SetNRGBA64(0, 0, opaqueGray64)
+		src := image.NewNRGBA64(image.Rect(0, 0, 1, 1))
+		src.SetNRGBA64(0, 0, transparentPurple64)
+		Draw(dst, dst.Bounds(), src, image.Point{0, 0}, Src)
+		if got, want := dst.At(0, 0), transparentPurple64; got != want {
+			t.Errorf("At(0, 0):\ngot  %#v\nwant %#v", got, want)
+		}
+	}
+}
+
+// TestFloydSteinbergCheckerboard tests that the result of Floyd-Steinberg
+// error diffusion of a uniform 50% gray source image with a black-and-white
+// palette is a checkerboard pattern.
+func TestFloydSteinbergCheckerboard(t *testing.T) {
+	b := image.Rect(0, 0, 640, 480)
+	// We can't represent 50% exactly, but 0x7fff / 0xffff is close enough.
+	src := &image.Uniform{color.Gray16{0x7fff}}
+	dst := image.NewPaletted(b, color.Palette{color.Black, color.White})
+	FloydSteinberg.Draw(dst, b, src, image.Point{})
+	nErr := 0
+	for y := b.Min.Y; y < b.Max.Y; y++ {
+		for x := b.Min.X; x < b.Max.X; x++ {
+			got := dst.Pix[dst.PixOffset(x, y)]
+			want := uint8(x+y) % 2
+			if got != want {
+				t.Errorf("at (%d, %d): got %d, want %d", x, y, got, want)
+				if nErr++; nErr == 10 {
+					t.Fatal("there may be more errors")
+				}
+			}
+		}
+	}
+}
+
+// embeddedPaletted is an Image that behaves like an *image.Paletted but whose
+// type is not *image.Paletted.
+type embeddedPaletted struct {
+	*image.Paletted
+}
+
+// TestPaletted tests that the drawPaletted function behaves the same
+// regardless of whether dst is an *image.Paletted.
+func TestPaletted(t *testing.T) {
+	f, err := os.Open("../testdata/video-001.png")
+	if err != nil {
+		t.Fatalf("open: %v", err)
+	}
+	defer f.Close()
+	video001, err := png.Decode(f)
+	if err != nil {
+		t.Fatalf("decode: %v", err)
+	}
+	b := video001.Bounds()
+
+	cgaPalette := color.Palette{
+		color.RGBA{0x00, 0x00, 0x00, 0xff},
+		color.RGBA{0x55, 0xff, 0xff, 0xff},
+		color.RGBA{0xff, 0x55, 0xff, 0xff},
+		color.RGBA{0xff, 0xff, 0xff, 0xff},
+	}
+	drawers := map[string]Drawer{
+		"src":             Src,
+		"floyd-steinberg": FloydSteinberg,
+	}
+	sources := map[string]image.Image{
+		"uniform":  &image.Uniform{color.RGBA{0xff, 0x7f, 0xff, 0xff}},
+		"video001": video001,
+	}
+
+	for dName, d := range drawers {
+	loop:
+		for sName, src := range sources {
+			dst0 := image.NewPaletted(b, cgaPalette)
+			dst1 := image.NewPaletted(b, cgaPalette)
+			d.Draw(dst0, b, src, image.Point{})
+			d.Draw(embeddedPaletted{dst1}, b, src, image.Point{})
+			for y := b.Min.Y; y < b.Max.Y; y++ {
+				for x := b.Min.X; x < b.Max.X; x++ {
+					if !eq(dst0.At(x, y), dst1.At(x, y)) {
+						t.Errorf("%s / %s: at (%d, %d), %v versus %v",
+							dName, sName, x, y, dst0.At(x, y), dst1.At(x, y))
+						continue loop
+					}
+				}
+			}
+		}
+	}
+}
+
+func TestSqDiff(t *testing.T) {
+	// This test is similar to the one from the image/color package, but
+	// sqDiff in this package accepts int32 instead of uint32, so test it
+	// for appropriate input.
+
+	// canonical sqDiff implementation
+	orig := func(x, y int32) uint32 {
+		var d uint32
+		if x > y {
+			d = uint32(x - y)
+		} else {
+			d = uint32(y - x)
+		}
+		return (d * d) >> 2
+	}
+	testCases := []int32{
+		0,
+		1,
+		2,
+		0x0fffd,
+		0x0fffe,
+		0x0ffff,
+		0x10000,
+		0x10001,
+		0x10002,
+		0x7ffffffd,
+		0x7ffffffe,
+		0x7fffffff,
+		-0x7ffffffd,
+		-0x7ffffffe,
+		-0x80000000,
+	}
+	for _, x := range testCases {
+		for _, y := range testCases {
+			if got, want := sqDiff(x, y), orig(x, y); got != want {
+				t.Fatalf("sqDiff(%#x, %#x): got %d, want %d", x, y, got, want)
+			}
+		}
+	}
+	if err := quick.CheckEqual(orig, sqDiff, &quick.Config{MaxCountScale: 10}); err != nil {
+		t.Fatal(err)
+	}
+}
diff --git a/src/image/draw/example_test.go b/src/image/draw/example_test.go
new file mode 100644
index 0000000..2ccc2f4
--- /dev/null
+++ b/src/image/draw/example_test.go
@@ -0,0 +1,48 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package draw_test
+
+import (
+	"fmt"
+	"image"
+	"image/color"
+	"image/draw"
+	"math"
+)
+
+func ExampleDrawer_floydSteinberg() {
+	const width = 130
+	const height = 50
+
+	im := image.NewGray(image.Rectangle{Max: image.Point{X: width, Y: height}})
+	for x := 0; x < width; x++ {
+		for y := 0; y < height; y++ {
+			dist := math.Sqrt(math.Pow(float64(x-width/2), 2)/3+math.Pow(float64(y-height/2), 2)) / (height / 1.5) * 255
+			var gray uint8
+			if dist > 255 {
+				gray = 255
+			} else {
+				gray = uint8(dist)
+			}
+			im.SetGray(x, y, color.Gray{Y: 255 - gray})
+		}
+	}
+	pi := image.NewPaletted(im.Bounds(), []color.Color{
+		color.Gray{Y: 255},
+		color.Gray{Y: 160},
+		color.Gray{Y: 70},
+		color.Gray{Y: 35},
+		color.Gray{Y: 0},
+	})
+
+	draw.FloydSteinberg.Draw(pi, im.Bounds(), im, image.ZP)
+	shade := []string{" ", "░", "▒", "▓", "█"}
+	for i, p := range pi.Pix {
+		fmt.Print(shade[p])
+		if (i+1)%width == 0 {
+			fmt.Print("\n")
+		}
+	}
+}