// Copyright 2015 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.

// +build ignore

package main

import (
	"bytes"
	"flag"
	"fmt"
	"go/format"
	"io/ioutil"
	"log"
	"os"
	"strings"
)

var debug = flag.Bool("debug", false, "")

func main() {
	flag.Parse()

	w := new(bytes.Buffer)
	w.WriteString("// generated by \"go run gen.go\". DO NOT EDIT.\n\n" +
		"package draw\n\nimport (\n" +
		"\"image\"\n" +
		"\"image/color\"\n" +
		"\"math\"\n" +
		"\n" +
		"\"golang.org/x/image/math/f64\"\n" +
		")\n")

	gen(w, "nnInterpolator", codeNNScaleLeaf, codeNNTransformLeaf)
	gen(w, "ablInterpolator", codeABLScaleLeaf, codeABLTransformLeaf)
	genKernel(w)

	if *debug {
		os.Stdout.Write(w.Bytes())
		return
	}
	out, err := format.Source(w.Bytes())
	if err != nil {
		log.Fatal(err)
	}
	if err := ioutil.WriteFile("impl.go", out, 0660); err != nil {
		log.Fatal(err)
	}
}

var (
	// dsTypes are the (dst image type, src image type) pairs to generate
	// scale_DType_SType implementations for. The last element in the slice
	// should be the fallback pair ("Image", "image.Image").
	//
	// TODO: add *image.CMYK src type after Go 1.5 is released.
	dsTypes = []struct{ dType, sType string }{
		{"*image.RGBA", "*image.Gray"},
		{"*image.RGBA", "*image.NRGBA"},
		{"*image.RGBA", "*image.RGBA"},
		{"*image.RGBA", "*image.Uniform"},
		{"*image.RGBA", "*image.YCbCr"},
		{"*image.RGBA", "image.Image"},
		{"Image", "image.Image"},
	}
	dTypes, sTypes []string
	sTypesForDType = map[string][]string{}
)

func init() {
	dTypesSeen := map[string]bool{}
	sTypesSeen := map[string]bool{}
	for _, t := range dsTypes {
		if !sTypesSeen[t.sType] {
			sTypesSeen[t.sType] = true
			sTypes = append(sTypes, t.sType)
		}
		if !dTypesSeen[t.dType] {
			dTypesSeen[t.dType] = true
			dTypes = append(dTypes, t.dType)
		}
		sTypesForDType[t.dType] = append(sTypesForDType[t.dType], t.sType)
	}
	sTypesForDType["anyDType"] = sTypes
}

type data struct {
	dType    string
	sType    string
	receiver string
}

func gen(w *bytes.Buffer, receiver string, codes ...string) {
	expn(w, codeRoot, &data{receiver: receiver})
	for _, code := range codes {
		for _, t := range dsTypes {
			expn(w, code, &data{
				dType:    t.dType,
				sType:    t.sType,
				receiver: receiver,
			})
		}
	}
}

func genKernel(w *bytes.Buffer) {
	expn(w, codeKernelRoot, &data{})
	for _, sType := range sTypes {
		expn(w, codeKernelScaleLeafX, &data{
			sType: sType,
		})
	}
	for _, dType := range dTypes {
		expn(w, codeKernelScaleLeafY, &data{
			dType: dType,
		})
	}
	for _, t := range dsTypes {
		expn(w, codeKernelTransformLeaf, &data{
			dType: t.dType,
			sType: t.sType,
		})
	}
}

func expn(w *bytes.Buffer, code string, d *data) {
	for _, line := range strings.Split(code, "\n") {
		line = expnLine(line, d)
		if line == ";" {
			continue
		}
		fmt.Fprintln(w, line)
	}
}

func expnLine(line string, d *data) string {
	for {
		i := strings.IndexByte(line, '$')
		if i < 0 {
			break
		}
		prefix, s := line[:i], line[i+1:]

		i = len(s)
		for j, c := range s {
			if !('A' <= c && c <= 'Z' || 'a' <= c && c <= 'z') {
				i = j
				break
			}
		}
		dollar, suffix := s[:i], s[i:]

		e := expnDollar(prefix, dollar, suffix, d)
		if e == "" {
			log.Fatalf("couldn't expand %q", line)
		}
		line = e
	}
	return line
}

// expnDollar expands a "$foo" fragment in a line of generated code. It returns
// the empty string if there was a problem. It returns ";" if the generated
// code is a no-op.
func expnDollar(prefix, dollar, suffix string, d *data) string {
	switch dollar {
	case "dType":
		return prefix + d.dType + suffix
	case "dTypeRN":
		return prefix + relName(d.dType) + suffix
	case "sType":
		return prefix + d.sType + suffix
	case "sTypeRN":
		return prefix + relName(d.sType) + suffix
	case "receiver":
		return prefix + d.receiver + suffix

	case "switch":
		return expnSwitch("", true, suffix)
	case "switchD":
		return expnSwitch("", false, suffix)
	case "switchS":
		return expnSwitch("anyDType", false, suffix)

	case "preOuter":
		switch d.dType {
		default:
			return ";"
		case "Image":
			return "" +
				"dstColorRGBA64 := &color.RGBA64{}\n" +
				"dstColor := color.Color(dstColorRGBA64)"
		}

	case "preInner":
		switch d.dType {
		default:
			return ";"
		case "*image.Gray", "*image.RGBA":
			return "d := dst.PixOffset(dr.Min.X+adr.Min.X, dr.Min.Y+int(dy))"
		}

	case "preKernelInner":
		switch d.dType {
		default:
			return ";"
		case "*image.Gray", "*image.RGBA":
			return "d := dst.PixOffset(dr.Min.X+int(dx), dr.Min.Y+adr.Min.Y)"
		}

	case "blend":
		args, _ := splitArgs(suffix)
		if len(args) != 4 {
			return ""
		}
		switch d.sType {
		default:
			return fmt.Sprintf(""+
				"%sr = %s*%sr + %s*%sr\n"+
				"%sg = %s*%sg + %s*%sg\n"+
				"%sb = %s*%sb + %s*%sb\n"+
				"%sa = %s*%sa + %s*%sa",
				args[3], args[0], args[1], args[2], args[3],
				args[3], args[0], args[1], args[2], args[3],
				args[3], args[0], args[1], args[2], args[3],
				args[3], args[0], args[1], args[2], args[3],
			)
		case "*image.Gray":
			return fmt.Sprintf(""+
				"%sr = %s*%sr + %s*%sr",
				args[3], args[0], args[1], args[2], args[3],
			)
		}

	case "outputu":
		args, _ := splitArgs(suffix)
		if len(args) != 3 {
			return ""
		}
		switch d.dType {
		default:
			log.Fatalf("bad dType %q", d.dType)
		case "Image":
			switch d.sType {
			default:
				return fmt.Sprintf(""+
					"dstColorRGBA64.R = uint16(%sr)\n"+
					"dstColorRGBA64.G = uint16(%sg)\n"+
					"dstColorRGBA64.B = uint16(%sb)\n"+
					"dstColorRGBA64.A = uint16(%sa)\n"+
					"dst.Set(%s, %s, dstColor)",
					args[2], args[2], args[2], args[2],
					args[0], args[1],
				)
			case "*image.Gray":
				return fmt.Sprintf(""+
					"out := uint16(%sr)\n"+
					"dstColorRGBA64.R = out\n"+
					"dstColorRGBA64.G = out\n"+
					"dstColorRGBA64.B = out\n"+
					"dstColorRGBA64.A = 0xffff\n"+
					"dst.Set(%s, %s, dstColor)",
					args[2],
					args[0], args[1],
				)
			}
		case "*image.RGBA":
			switch d.sType {
			default:
				return fmt.Sprintf(""+
					"dst.Pix[d+0] = uint8(uint32(%sr) >> 8)\n"+
					"dst.Pix[d+1] = uint8(uint32(%sg) >> 8)\n"+
					"dst.Pix[d+2] = uint8(uint32(%sb) >> 8)\n"+
					"dst.Pix[d+3] = uint8(uint32(%sa) >> 8)",
					args[2], args[2], args[2], args[2],
				)
			case "*image.Gray":
				return fmt.Sprintf(""+
					"out := uint8(uint32(%sr) >> 8)\n"+
					"dst.Pix[d+0] = out\n"+
					"dst.Pix[d+1] = out\n"+
					"dst.Pix[d+2] = out\n"+
					"dst.Pix[d+3] = 0xff",
					args[2],
				)
			}
		}

	case "outputf":
		args, _ := splitArgs(suffix)
		if len(args) != 5 {
			return ""
		}
		ret := ""
		switch d.dType {
		default:
			log.Fatalf("bad dType %q", d.dType)
		case "Image":
			switch d.sType {
			default:
				ret = fmt.Sprintf(""+
					"dstColorRGBA64.R = %s(%sr * %s)\n"+
					"dstColorRGBA64.G = %s(%sg * %s)\n"+
					"dstColorRGBA64.B = %s(%sb * %s)\n"+
					"dstColorRGBA64.A = %s(%sa * %s)\n"+
					"dst.Set(%s, %s, dstColor)",
					args[2], args[3], args[4],
					args[2], args[3], args[4],
					args[2], args[3], args[4],
					args[2], args[3], args[4],
					args[0], args[1],
				)
			case "*image.Gray":
				ret = fmt.Sprintf(""+
					"out := %s(%sr * %s)\n"+
					"dstColorRGBA64.R = out\n"+
					"dstColorRGBA64.G = out\n"+
					"dstColorRGBA64.B = out\n"+
					"dstColorRGBA64.A = 0xffff\n"+
					"dst.Set(%s, %s, dstColor)",
					args[2], args[3], args[4],
					args[0], args[1],
				)
			}
		case "*image.RGBA":
			switch d.sType {
			default:
				ret = fmt.Sprintf(""+
					"dst.Pix[d+0] = uint8(%s(%sr * %s) >> 8)\n"+
					"dst.Pix[d+1] = uint8(%s(%sg * %s) >> 8)\n"+
					"dst.Pix[d+2] = uint8(%s(%sb * %s) >> 8)\n"+
					"dst.Pix[d+3] = uint8(%s(%sa * %s) >> 8)",
					args[2], args[3], args[4],
					args[2], args[3], args[4],
					args[2], args[3], args[4],
					args[2], args[3], args[4],
				)
			case "*image.Gray":
				ret = fmt.Sprintf(""+
					"out := uint8(%s(%sr * %s) >> 8)\n"+
					"dst.Pix[d+0] = out\n"+
					"dst.Pix[d+1] = out\n"+
					"dst.Pix[d+2] = out\n"+
					"dst.Pix[d+3] = 0xff",
					args[2], args[3], args[4],
				)
			}
		}
		return strings.Replace(ret, " * 1)", ")", -1)

	case "srcf", "srcu":
		lhs, eqOp := splitEq(prefix)
		if lhs == "" {
			return ""
		}
		args, extra := splitArgs(suffix)
		if len(args) != 2 {
			return ""
		}

		tmp := ""
		if dollar == "srcf" {
			tmp = "u"
		}

		buf := new(bytes.Buffer)
		switch d.sType {
		default:
			log.Fatalf("bad sType %q", d.sType)
		case "image.Image", "*image.NRGBA", "*image.Uniform", "*image.YCbCr": // TODO: separate code for concrete types.
			fmt.Fprintf(buf, "%sr%s, %sg%s, %sb%s, %sa%s := "+
				"src.At(%s, %s).RGBA()\n",
				lhs, tmp, lhs, tmp, lhs, tmp, lhs, tmp,
				args[0], args[1],
			)
		case "*image.Gray":
			// TODO: there's no need to multiply by 0x101 if the next thing
			// we're going to do is shift right by 8.
			fmt.Fprintf(buf, "%si := src.PixOffset(%s, %s)\n"+
				"%sr%s := uint32(src.Pix[%si]) * 0x101\n",
				lhs, args[0], args[1],
				lhs, tmp, lhs,
			)
		case "*image.RGBA":
			// TODO: there's no need to multiply by 0x101 if the next thing
			// we're going to do is shift right by 8.
			fmt.Fprintf(buf, "%si := src.PixOffset(%s, %s)\n"+
				"%sr%s := uint32(src.Pix[%si+0]) * 0x101\n"+
				"%sg%s := uint32(src.Pix[%si+1]) * 0x101\n"+
				"%sb%s := uint32(src.Pix[%si+2]) * 0x101\n"+
				"%sa%s := uint32(src.Pix[%si+3]) * 0x101\n",
				lhs, args[0], args[1],
				lhs, tmp, lhs,
				lhs, tmp, lhs,
				lhs, tmp, lhs,
				lhs, tmp, lhs,
			)
		}

		if dollar == "srcf" {
			switch d.sType {
			default:
				fmt.Fprintf(buf, ""+
					"%sr %s float64(%sru)%s\n"+
					"%sg %s float64(%sgu)%s\n"+
					"%sb %s float64(%sbu)%s\n"+
					"%sa %s float64(%sau)%s\n",
					lhs, eqOp, lhs, extra,
					lhs, eqOp, lhs, extra,
					lhs, eqOp, lhs, extra,
					lhs, eqOp, lhs, extra,
				)
			case "*image.Gray":
				fmt.Fprintf(buf, ""+
					"%sr %s float64(%sru)%s\n",
					lhs, eqOp, lhs, extra,
				)
			}
		}

		return strings.TrimSpace(buf.String())

	case "tweakD":
		if d.dType == "*image.RGBA" {
			return "d += dst.Stride"
		}
		return ";"

	case "tweakDx":
		if d.dType == "*image.RGBA" {
			return strings.Replace(suffix, "dx++", "dx, d = dx+1, d+4", 1)
		}
		return suffix

	case "tweakDy":
		if d.dType == "*image.RGBA" {
			return strings.Replace(suffix, "for dy, s", "for _, s", 1)
		}
		return suffix

	case "tweakP":
		if d.sType == "*image.Gray" {
			if strings.HasPrefix(strings.TrimSpace(suffix), "pa * ") {
				return "1,"
			}
			return "pr,"
		}
		return suffix

	case "tweakPr":
		if d.sType == "*image.Gray" {
			return "pr *= s.invTotalWeightFFFF"
		}
		return ";"

	case "tweakVarP":
		if d.sType == "*image.Gray" {
			return strings.Replace(suffix, "var pr, pg, pb, pa", "var pr", 1)
		}
		return suffix
	}
	return ""
}

func expnSwitch(dType string, expandBoth bool, template string) string {
	switchVar := "dst"
	if dType != "" {
		switchVar = "src"
	}
	lines := []string{fmt.Sprintf("switch %s := %s.(type) {", switchVar, switchVar)}

	fallback, values := "Image", dTypes
	if dType != "" {
		fallback, values = "image.Image", sTypesForDType[dType]
	}
	for _, v := range values {
		if v == fallback {
			lines = append(lines, "default:")
		} else {
			lines = append(lines, fmt.Sprintf("case %s:", v))
		}

		if dType != "" {
			lines = append(lines, expnLine(template, &data{dType: dType, sType: v}))
		} else if !expandBoth {
			lines = append(lines, expnLine(template, &data{dType: v}))
		} else {
			lines = append(lines, expnSwitch(v, false, template))
		}
	}

	lines = append(lines, "}")
	return strings.Join(lines, "\n")
}

func split(s, sep string) (string, string) {
	if i := strings.Index(s, sep); i >= 0 {
		return strings.TrimSpace(s[:i]), strings.TrimSpace(s[i+len(sep):])
	}
	return "", ""
}

func splitEq(s string) (lhs, eqOp string) {
	s = strings.TrimSpace(s)
	if lhs, _ = split(s, ":="); lhs != "" {
		return lhs, ":="
	}
	if lhs, _ = split(s, "+="); lhs != "" {
		return lhs, "+="
	}
	return "", ""
}

func splitArgs(s string) (args []string, extra string) {
	s = strings.TrimSpace(s)
	if s == "" || s[0] != '[' {
		return nil, ""
	}
	s = s[1:]

	i := strings.IndexByte(s, ']')
	if i < 0 {
		return nil, ""
	}
	args, extra = strings.Split(s[:i], ","), s[i+1:]
	for i := range args {
		args[i] = strings.TrimSpace(args[i])
	}
	return args, extra
}

func relName(s string) string {
	if i := strings.LastIndex(s, "."); i >= 0 {
		return s[i+1:]
	}
	return s
}

const (
	codeRoot = `
		func (z $receiver) Scale(dst Image, dr image.Rectangle, src image.Image, sr image.Rectangle, opts *Options) {
			// adr is the affected destination pixels, relative to dr.Min.
			adr := dst.Bounds().Intersect(dr).Sub(dr.Min)
			if adr.Empty() || sr.Empty() {
				return
			}
			// sr is the source pixels. If it extends beyond the src bounds,
			// we cannot use the type-specific fast paths, as they access
			// the Pix fields directly without bounds checking.
			if !sr.In(src.Bounds()) {
				z.scale_Image_Image(dst, dr, adr, src, sr)
			} else {
				$switch z.scale_$dTypeRN_$sTypeRN(dst, dr, adr, src, sr)
			}
		}

		func (z $receiver) Transform(dst Image, s2d *f64.Aff3, src image.Image, sr image.Rectangle, opts *Options) {
			dr := transformRect(s2d, &sr)
			// adr is the affected destination pixels, relative to dr.Min.
			adr := dst.Bounds().Intersect(dr).Sub(dr.Min)
			if adr.Empty() || sr.Empty() {
				return
			}
			d2s := invert(s2d)
			// sr is the source pixels. If it extends beyond the src bounds,
			// we cannot use the type-specific fast paths, as they access
			// the Pix fields directly without bounds checking.
			if !sr.In(src.Bounds()) {
				z.transform_Image_Image(dst, dr, adr, &d2s, src, sr)
			} else {
				$switch z.transform_$dTypeRN_$sTypeRN(dst, dr, adr, &d2s, src, sr)
			}
		}
	`

	codeNNScaleLeaf = `
		func (nnInterpolator) scale_$dTypeRN_$sTypeRN(dst $dType, dr, adr image.Rectangle, src $sType, sr image.Rectangle) {
			dw2 := uint64(dr.Dx()) * 2
			dh2 := uint64(dr.Dy()) * 2
			sw := uint64(sr.Dx())
			sh := uint64(sr.Dy())
			$preOuter
			for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
				sy := (2*uint64(dy) + 1) * sh / dh2
				$preInner
				$tweakDx for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
					sx := (2*uint64(dx) + 1) * sw / dw2
					p := $srcu[sr.Min.X + int(sx), sr.Min.Y + int(sy)]
					$outputu[dr.Min.X + int(dx), dr.Min.Y + int(dy), p]
				}
			}
		}
	`

	codeNNTransformLeaf = `
		func (nnInterpolator) transform_$dTypeRN_$sTypeRN(dst $dType, dr, adr image.Rectangle, d2s *f64.Aff3, src $sType, sr image.Rectangle) {
			$preOuter
			for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
				dyf := float64(dr.Min.Y + int(dy)) + 0.5
				$preInner
				$tweakDx for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
					dxf := float64(dr.Min.X + int(dx)) + 0.5
					// TODO: change the src origin so that we can say int(f) instead of int(math.Floor(f)).
					sx0 := int(math.Floor(d2s[0]*dxf + d2s[1]*dyf + d2s[2]))
					sy0 := int(math.Floor(d2s[3]*dxf + d2s[4]*dyf + d2s[5]))
					if !(image.Point{sx0, sy0}).In(sr) {
						continue
					}
					p := $srcu[sx0, sy0]
					$outputu[dr.Min.X + int(dx), dr.Min.Y + int(dy), p]
				}
			}
		}
	`

	codeABLScaleLeaf = `
		func (ablInterpolator) scale_$dTypeRN_$sTypeRN(dst $dType, dr, adr image.Rectangle, src $sType, sr image.Rectangle) {
			sw := int32(sr.Dx())
			sh := int32(sr.Dy())
			yscale := float64(sh) / float64(dr.Dy())
			xscale := float64(sw) / float64(dr.Dx())
			swMinus1, shMinus1 := sw - 1, sh - 1
			$preOuter

			for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
				sy := (float64(dy)+0.5)*yscale - 0.5
				// If sy < 0, we will clamp sy0 to 0 anyway, so it doesn't matter if
				// we say int32(sy) instead of int32(math.Floor(sy)). Similarly for
				// sx, below.
				sy0 := int32(sy)
				yFrac0 := sy - float64(sy0)
				yFrac1 := 1 - yFrac0
				sy1 := sy0 + 1
				if sy < 0 {
					sy0, sy1 = 0, 0
					yFrac0, yFrac1 = 0, 1
				} else if sy1 > shMinus1 {
					sy0, sy1 = shMinus1, shMinus1
					yFrac0, yFrac1 = 1, 0
				}
				$preInner

				$tweakDx for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
					sx := (float64(dx)+0.5)*xscale - 0.5
					sx0 := int32(sx)
					xFrac0 := sx - float64(sx0)
					xFrac1 := 1 - xFrac0
					sx1 := sx0 + 1
					if sx < 0 {
						sx0, sx1 = 0, 0
						xFrac0, xFrac1 = 0, 1
					} else if sx1 > swMinus1 {
						sx0, sx1 = swMinus1, swMinus1
						xFrac0, xFrac1 = 1, 0
					}

					s00 := $srcf[sr.Min.X + int(sx0), sr.Min.Y + int(sy0)]
					s10 := $srcf[sr.Min.X + int(sx1), sr.Min.Y + int(sy0)]
					$blend[xFrac1, s00, xFrac0, s10]
					s01 := $srcf[sr.Min.X + int(sx0), sr.Min.Y + int(sy1)]
					s11 := $srcf[sr.Min.X + int(sx1), sr.Min.Y + int(sy1)]
					$blend[xFrac1, s01, xFrac0, s11]
					$blend[yFrac1, s10, yFrac0, s11]
					$outputu[dr.Min.X + int(dx), dr.Min.Y + int(dy), s11]
				}
			}
		}
	`

	codeABLTransformLeaf = `
		func (ablInterpolator) transform_$dTypeRN_$sTypeRN(dst $dType, dr, adr image.Rectangle, d2s *f64.Aff3, src $sType, sr image.Rectangle) {
			$preOuter
			for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
				dyf := float64(dr.Min.Y + int(dy)) + 0.5
				$preInner
				$tweakDx for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
					dxf := float64(dr.Min.X + int(dx)) + 0.5
					// TODO: change the src origin so that we can say int(f) instead of int(math.Floor(f)).
					sx := d2s[0]*dxf + d2s[1]*dyf + d2s[2]
					sy := d2s[3]*dxf + d2s[4]*dyf + d2s[5]
					if !(image.Point{int(math.Floor(sx)), int(math.Floor(sy))}).In(sr) {
						continue
					}

					sx -= 0.5
					sxf := math.Floor(sx)
					xFrac0 := sx - sxf
					xFrac1 := 1 - xFrac0
					sx0 := int(sxf)
					sx1 := sx0 + 1
					if sx0 < sr.Min.X {
						sx0, sx1 = sr.Min.X, sr.Min.X
						xFrac0, xFrac1 = 0, 1
					} else if sx1 >= sr.Max.X {
						sx0, sx1 = sr.Max.X-1, sr.Max.X-1
						xFrac0, xFrac1 = 1, 0
					}

					sy -= 0.5
					syf := math.Floor(sy)
					yFrac0 := sy - syf
					yFrac1 := 1 - yFrac0
					sy0 := int(syf)
					sy1 := sy0 + 1
					if sy0 < sr.Min.Y {
						sy0, sy1 = sr.Min.Y, sr.Min.Y
						yFrac0, yFrac1 = 0, 1
					} else if sy1 >= sr.Max.Y {
						sy0, sy1 = sr.Max.Y-1, sr.Max.Y-1
						yFrac0, yFrac1 = 1, 0
					}

					s00 := $srcf[sx0, sy0]
					s10 := $srcf[sx1, sy0]
					$blend[xFrac1, s00, xFrac0, s10]
					s01 := $srcf[sx0, sy1]
					s11 := $srcf[sx1, sy1]
					$blend[xFrac1, s01, xFrac0, s11]
					$blend[yFrac1, s10, yFrac0, s11]
					$outputu[dr.Min.X + int(dx), dr.Min.Y + int(dy), s11]
				}
			}
		}
	`

	codeKernelRoot = `
		func (z *kernelScaler) Scale(dst Image, dr image.Rectangle, src image.Image, sr image.Rectangle, opts *Options) {
			if z.dw != int32(dr.Dx()) || z.dh != int32(dr.Dy()) || z.sw != int32(sr.Dx()) || z.sh != int32(sr.Dy()) {
				z.kernel.Scale(dst, dr, src, sr, opts)
				return
			}
			// adr is the affected destination pixels, relative to dr.Min.
			adr := dst.Bounds().Intersect(dr).Sub(dr.Min)
			if adr.Empty() || sr.Empty() {
				return
			}
			// Create a temporary buffer:
			// scaleX distributes the source image's columns over the temporary image.
			// scaleY distributes the temporary image's rows over the destination image.
			// TODO: is it worth having a sync.Pool for this temporary buffer?
			tmp := make([][4]float64, z.dw*z.sh)

			// sr is the source pixels. If it extends beyond the src bounds,
			// we cannot use the type-specific fast paths, as they access
			// the Pix fields directly without bounds checking.
			if !sr.In(src.Bounds()) {
				z.scaleX_Image(tmp, src, sr)
			} else {
				$switchS z.scaleX_$sTypeRN(tmp, src, sr)
			}

			$switchD z.scaleY_$dTypeRN(dst, dr, adr, tmp)
		}

		func (q *Kernel) Transform(dst Image, s2d *f64.Aff3, src image.Image, sr image.Rectangle, opts *Options) {
			dr := transformRect(s2d, &sr)
			// adr is the affected destination pixels, relative to dr.Min.
			adr := dst.Bounds().Intersect(dr).Sub(dr.Min)
			if adr.Empty() || sr.Empty() {
				return
			}
			d2s := invert(s2d)

			xscale := abs(d2s[0])
			if s := abs(d2s[1]); xscale < s {
				xscale = s
			}
			yscale := abs(d2s[3])
			if s := abs(d2s[4]); yscale < s {
				yscale = s
			}

			// sr is the source pixels. If it extends beyond the src bounds,
			// we cannot use the type-specific fast paths, as they access
			// the Pix fields directly without bounds checking.
			if !sr.In(src.Bounds()) {
				q.transform_Image_Image(dst, dr, adr, &d2s, src, sr, xscale, yscale)
			} else {
				$switch q.transform_$dTypeRN_$sTypeRN(dst, dr, adr, &d2s, src, sr, xscale, yscale)
			}
		}
	`

	codeKernelScaleLeafX = `
		func (z *kernelScaler) scaleX_$sTypeRN(tmp [][4]float64, src $sType, sr image.Rectangle) {
			t := 0
			for y := int32(0); y < z.sh; y++ {
				for _, s := range z.horizontal.sources {
					$tweakVarP var pr, pg, pb, pa float64
					for _, c := range z.horizontal.contribs[s.i:s.j] {
						p += $srcf[sr.Min.X + int(c.coord), sr.Min.Y + int(y)] * c.weight
					}
					$tweakPr
					tmp[t] = [4]float64{
						$tweakP pr * s.invTotalWeightFFFF,
						$tweakP pg * s.invTotalWeightFFFF,
						$tweakP pb * s.invTotalWeightFFFF,
						$tweakP pa * s.invTotalWeightFFFF,
					}
					t++
				}
			}
		}
	`

	codeKernelScaleLeafY = `
		func (z *kernelScaler) scaleY_$dTypeRN(dst $dType, dr, adr image.Rectangle, tmp [][4]float64) {
			$preOuter
			for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
				$preKernelInner
				$tweakDy for dy, s := range z.vertical.sources[adr.Min.Y:adr.Max.Y] {
					var pr, pg, pb, pa float64
					for _, c := range z.vertical.contribs[s.i:s.j] {
						p := &tmp[c.coord*z.dw+dx]
						pr += p[0] * c.weight
						pg += p[1] * c.weight
						pb += p[2] * c.weight
						pa += p[3] * c.weight
					}
					$outputf[dr.Min.X + int(dx), dr.Min.Y + int(adr.Min.Y + dy), ftou, p, s.invTotalWeight]
					$tweakD
				}
			}
		}
	`

	codeKernelTransformLeaf = `
		func (q *Kernel) transform_$dTypeRN_$sTypeRN(dst $dType, dr, adr image.Rectangle, d2s *f64.Aff3, src $sType, sr image.Rectangle, xscale, yscale float64) {
			// When shrinking, broaden the effective kernel support so that we still
			// visit every source pixel.
			xHalfWidth, xKernelArgScale := q.Support, 1.0
			if xscale > 1 {
				xHalfWidth *= xscale
				xKernelArgScale = 1 / xscale
			}
			yHalfWidth, yKernelArgScale := q.Support, 1.0
			if yscale > 1 {
				yHalfWidth *= yscale
				yKernelArgScale = 1 / yscale
			}

			xWeights := make([]float64, 1 + 2*int(math.Ceil(xHalfWidth)))
			yWeights := make([]float64, 1 + 2*int(math.Ceil(yHalfWidth)))

			$preOuter
			for dy := int32(adr.Min.Y); dy < int32(adr.Max.Y); dy++ {
				dyf := float64(dr.Min.Y + int(dy)) + 0.5
				$preInner
				$tweakDx for dx := int32(adr.Min.X); dx < int32(adr.Max.X); dx++ {
					dxf := float64(dr.Min.X + int(dx)) + 0.5
					// TODO: change the src origin so that we can say int(f) instead of int(math.Floor(f)).
					sx := d2s[0]*dxf + d2s[1]*dyf + d2s[2]
					sy := d2s[3]*dxf + d2s[4]*dyf + d2s[5]
					if !(image.Point{int(math.Floor(sx)), int(math.Floor(sy))}).In(sr) {
						continue
					}

					sx -= 0.5
					ix := int(math.Floor(sx - xHalfWidth))
					if ix < sr.Min.X {
						ix = sr.Min.X
					}
					jx := int(math.Ceil(sx + xHalfWidth))
					if jx > sr.Max.X {
						jx = sr.Max.X
					}

					totalXWeight := 0.0
					for kx := ix; kx < jx; kx++ {
						xWeight := 0.0
						if t := abs((sx - float64(kx)) * xKernelArgScale); t < q.Support {
							xWeight = q.At(t)
						}
						xWeights[kx - ix] = xWeight
						totalXWeight += xWeight
					}
					for x := range xWeights[:jx-ix] {
						xWeights[x] /= totalXWeight
					}

					sy -= 0.5
					iy := int(math.Floor(sy - yHalfWidth))
					if iy < sr.Min.Y {
						iy = sr.Min.Y
					}
					jy := int(math.Ceil(sy + yHalfWidth))
					if jy > sr.Max.Y {
						jy = sr.Max.Y
					}

					totalYWeight := 0.0
					for ky := iy; ky < jy; ky++ {
						yWeight := 0.0
						if t := abs((sy - float64(ky)) * yKernelArgScale); t < q.Support {
							yWeight = q.At(t)
						}
						yWeights[ky - iy] = yWeight
						totalYWeight += yWeight
					}
					for y := range yWeights[:jy-iy] {
						yWeights[y] /= totalYWeight
					}

					$tweakVarP var pr, pg, pb, pa float64
					for ky := iy; ky < jy; ky++ {
						yWeight := yWeights[ky - iy]
						for kx := ix; kx < jx; kx++ {
							p += $srcf[kx, ky] * xWeights[kx - ix] * yWeight
						}
					}
					$outputf[dr.Min.X + int(dx), dr.Min.Y + int(dy), fffftou, p, 1]
				}
			}
		}
	`
)