CPU patterns fallback

internal/generator/cpu.go

@@ -11,79 +11,57 @@ type YCbCr10 struct {
 	Cr uint32
 }
 
-type PixelSampler func(x, y, width, height, frameIndex int) YCbCr10
+// FrameRenderer writes a complete frame or patches part of an existing frame.
+type FrameRenderer func(dst []byte, width, height, frameIndex int) error
 
 type CPUGenerator struct {
-	width   int
-	height  int
-	sampler PixelSampler
+	width  int
+	height int
+	base   []byte
+	patch  FrameRenderer
 }
 
 func NewCPUGenerator(
 	width, height uint,
-	sampler PixelSampler,
+	baseRenderer FrameRenderer,
+	patch FrameRenderer,
 ) (*CPUGenerator, error) {
 	if width == 0 || height == 0 {
-		return nil, fmt.Errorf(
-			"width and height must be greater than 0, got: %dx%d",
-			width,
-			height,
-		)
+		return nil, fmt.Errorf("cpu: width and height must be greater than zero, got %dx%d", width, height)
 	}
 	if width%6 != 0 {
-		return nil, fmt.Errorf(
-			"width must be divisible by 6, got: %d",
-			width,
-		)
+		return nil, fmt.Errorf("cpu: width must be divisible by 6, got %d", width)
 	}
-	if sampler == nil {
-		return nil, fmt.Errorf("pixel sampler is nil")
+	if baseRenderer == nil {
+		return nil, fmt.Errorf("cpu: base renderer is nil")
 	}
 
-	return &CPUGenerator{
-		width:   int(width),
-		height:  int(height),
-		sampler: sampler,
-	}, nil
+	g := &CPUGenerator{
+		width:  int(width),
+		height: int(height),
+		base:   make([]byte, int(width/6*height)*16),
+		patch:  patch,
+	}
+	if err := baseRenderer(g.base, g.width, g.height, 0); err != nil {
+		return nil, fmt.Errorf("cpu: render base frame: %w", err)
+	}
+	return g, nil
 }
 
-func (g *CPUGenerator) GenerateFrame(dest []byte, frameIndex int) error {
-	frameSize := (g.width / 6) * g.height * 16
-	if len(dest) < frameSize {
+func (g *CPUGenerator) GenerateFrame(dst []byte, frameIndex int) error {
+	if len(dst) < len(g.base) {
 		return fmt.Errorf(
 			"cpu: destination is too small: got %d bytes, need %d",
-			len(dest),
-			frameSize,
+			len(dst),
+			len(g.base),
 		)
 	}
 
-	const componentMask uint32 = 0x3ff
-	offset := 0
-	for y := 0; y < g.height; y++ {
-		for x := 0; x < g.width; x += 6 {
-			var pixels [6]YCbCr10
-			for i := range pixels {
-				pixels[i] = g.sampler(x+i, y, g.width, g.height, frameIndex)
-			}
+	copy(dst, g.base)
 
-			word0 := pixels[0].Cb&componentMask |
-				(pixels[0].Y&componentMask)<<10 |
-				(pixels[0].Cr&componentMask)<<20
-			word1 := pixels[1].Y&componentMask |
-				(pixels[2].Cb&componentMask)<<10 |
-				(pixels[2].Y&componentMask)<<20
-			word2 := pixels[2].Cr&componentMask |
-				(pixels[3].Y&componentMask)<<10 |
-				(pixels[4].Cb&componentMask)<<20
-			word3 := pixels[4].Y&componentMask |
-				(pixels[4].Cr&componentMask)<<10 |
-				(pixels[5].Y&componentMask)<<20
-
-			binary.LittleEndian.PutUint32(dest[offset:], word0)
-			binary.LittleEndian.PutUint32(dest[offset+4:], word1)
-			binary.LittleEndian.PutUint32(dest[offset+8:], word2)
-			binary.LittleEndian.PutUint32(dest[offset+12:], word3)
-			offset += 16
+	if g.patch != nil {
+		if err := g.patch(dst[:len(g.base)], g.width, g.height, frameIndex); err != nil {
+			return fmt.Errorf("cpu: patch frame %d: %w", frameIndex, err)
 		}
 	}
 
@@ -91,3 +69,16 @@ func (g *CPUGenerator) GenerateFrame(dest []byte, frameIndex int) error {
 }
 
 func (g *CPUGenerator) Close() error { return nil }
+
+func packV210Block(dst []byte, pixels [6]YCbCr10) {
+	const mask uint32 = 0x3ff
+	word0 := pixels[0].Cb&mask | (pixels[0].Y&mask)<<10 | (pixels[0].Cr&mask)<<20
+	word1 := pixels[1].Y&mask | (pixels[2].Cb&mask)<<10 | (pixels[2].Y&mask)<<20
+	word2 := pixels[2].Cr&mask | (pixels[3].Y&mask)<<10 | (pixels[4].Cb&mask)<<20
+	word3 := pixels[4].Y&mask | (pixels[4].Cr&mask)<<10 | (pixels[5].Y&mask)<<20
+
+	binary.LittleEndian.PutUint32(dst, word0)
+	binary.LittleEndian.PutUint32(dst[4:], word1)
+	binary.LittleEndian.PutUint32(dst[8:], word2)
+	binary.LittleEndian.PutUint32(dst[12:], word3)
+}