//	For 4:2:0, CStride == YStride/2 && len(Cb) == len(Cr) == len(Y)/4.
//	For 4:4:0, CStride == YStride/1 && len(Cb) == len(Cr) == len(Y)/2.
//	For 4:1:1, CStride == YStride/4 && len(Cb) == len(Cr) == len(Y)/4.
//	For 4:1:0, CStride == YStride/4 && len(Cb) == len(Cr) == len(Y)/8.
type YCbCr struct {
	Y, Cb, Cr      []uint8
	YStride        int
	CStride        int
	SubsampleRatio YCbCrSubsampleRatio

pkg image, method (*NYCbCrA) YCbCrAt(int, int) color.YCbCr
pkg image, method (*NYCbCrA) YOffset(int, int) int
pkg image, type NYCbCrA struct
pkg image, type NYCbCrA struct, A []uint8
pkg image, type NYCbCrA struct, AStride int
pkg image, type NYCbCrA struct, embedded YCbCr
pkg image/color, method (NYCbCrA) RGBA() (uint32, uint32, uint32, uint32)
pkg image/color, type NYCbCrA struct
pkg image/color, type NYCbCrA struct, A uint8

	switch src.SubsampleRatio {

	case image.YCbCrSubsampleRatio444:
		for y, sy := y0, sp.Y; y != y1; y, sy = y+1, sy+1 {
			dpix := dst.Pix[y*dst.Stride:]
			yi := (sy-src.Rect.Min.Y)*src.YStride + (sp.X - src.Rect.Min.X)

			ci := (sy-src.Rect.Min.Y)*src.CStride + (sp.X - src.Rect.Min.X)
			for x := x0; x != x1; x, yi, ci = x+4, yi+1, ci+1 {

				// This is an inline version of image/color/ycbcr.go's func YCbCrToRGB.

		b[unzig[zig]] *= qt[zig]
	}
	idct(b)
	dst, stride := []byte(nil), 0
	if d.nComp == 1 {
		dst, stride = d.img1.Pix[8*(by*d.img1.Stride+bx):], d.img1.Stride
	} else {
		switch compIndex {
		case 0:
			dst, stride = d.img3.Y[8*(by*d.img3.YStride+bx):], d.img3.YStride
		case 1:
			dst, stride = d.img3.Cb[8*(by*d.img3.CStride+bx):], d.img3.CStride
		case 2:
			dst, stride = d.img3.Cr[8*(by*d.img3.CStride+bx):], d.img3.CStride

	"444": `
		ci := (sy-src.Rect.Min.Y)*src.CStride + (sp.X - src.Rect.Min.X)
		for x := x0; x != x1; x, yi, ci = x+4, yi+1, ci+1 {
	`,
	"422": `
		ciBase := (sy-src.Rect.Min.Y)*src.CStride - src.Rect.Min.X/2
		for x, sx := x0, sp.X; x != x1; x, sx, yi = x+4, sx+1, yi+1 {
			ci := ciBase + sx/2
	`,
	"420": `
		ciBase := (sy/2-src.Rect.Min.Y/2)*src.CStride - src.Rect.Min.X/2

			m1 := m1.(*image.YCbCr)
			if err := check(m0.Bounds(), m0.Y, m1.Y, m0.YStride, m1.YStride); err != nil {
				t.Errorf("%s (Y): %v", tc, err)
				continue
			}
			if err := check(m0.Bounds(), m0.Cb, m1.Cb, m0.CStride, m1.CStride); err != nil {
				t.Errorf("%s (Cb): %v", tc, err)
				continue
			}
			if err := check(m0.Bounds(), m0.Cr, m1.Cr, m0.CStride, m1.CStride); err != nil {
				t.Errorf("%s (Cr): %v", tc, err)
				continue
			}

kernel\_size, 3, out\_channel]. Space to depth is to transform this first
convolution's input to [batch\_size, height // stride, width // stride, 3 \*
stride \* stride] and the kernel to [kernel\_size // stride, kernel\_size //
stride, 3 \* stride \* stride, out\_channel] to improve TPU MXU utilization.

![drawings](images/space_to_depth_transform.png)

		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 {

	translations := [4]struct {
		src    []byte
		stride int
	}{
		{d.img3.Y, d.img3.YStride},
		{d.img3.Cb, d.img3.CStride},
		{d.img3.Cr, d.img3.CStride},
		{d.blackPix, d.blackStride},
	}
	for t, translation := range translations {
		subsample := d.comp[t].h != d.comp[0].h || d.comp[t].v != d.comp[0].v
		for iBase, y := 0, bounds.Min.Y; y < bounds.Max.Y; iBase, y = iBase+img.Stride, y+1 {
			sy := y - bounds.Min.Y

			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)
	}