3DTV promises to become the mainstream of next-generation TV systems. Highresolution 3DTV provides users with a vivid watching experience. Moreover, free-viewpoint view synthesis (FVVS) extends the common two-view stereo 3D vision into virtual reality by generating unlimited views from any desired viewpoint. In the next-generation 3DTV systems, the set-top box (STB) SoC requires both a high-definition (HD) multiview video-coding (MVC) decoder to reconstruct the real camera-captured scenes and a free-viewpoint view synthesizer to generate the virtual scenes [1-2].There are three main challenges to design an efficient high-resolution 3DTV STB SoC: (1) High processing capability of both the real-view decoder and virtualview synthesizer is required to support various 2D/3D applications in HDTV. For example, 66.2TOPS of computation is consumed to synthesize one virtual view in quad full-HD (QFHD) at 30fps. (2) In order to support FVVS, including 3D translation and 3D rotation (6D), matrix-based warping is needed for every pixel. In this case, two adjacent pixels in the reference view may be warped to the arbitrary positions in the virtual view according to their depth and epipolar geometry. Horizontal raster-scan-based scheduling [4][5] cannot deal with these irregular pixel relationships and can only support 1D horizontal shifts in the view synthesis.(3) The block-based memory accessing of reference pixels is not suitable for FVVS because of the processing nature of the irregular pixel access. 31.5GB/s system memory bandwidth is thus required for each virtual view in QFHD. State-of-the-art 3DTV chips cannot solve the issues above [4-6].Our 3DTV STB SoC is summarized as follows. First, a hardware-oriented 6D FVVS flow is introduced along with the corresponding architecture to solve the first two design challenges. A maximum 1911MPixel/s throughput is achieved, and is 9-to-40.5× higher than the previous works [4][5][6]. Second, the cachebased texture reorder architecture with the dynamic warping reference frame selection (DWRFS) scheme reduces the external memory bandwidth by 95.7% in view synthesis. Finally, the precision-optimized Homographic Transform (HT) and the single-iteration inpainting save 68% area in the warping engine and 93.3% of computing cycles, respectively. Figure 7.1.1 shows the 6D FVVS flow and the target applications. The MVC decoder reconstructs the real-view videos, the corresponding camera matrix, and the depth values from the bitstream. The view synthesizer then generates the virtual views from the real views. As shown in the left of Fig. 7.1.1, the pixels on one specific epipolar line of the virtual-view frame can only be found along the corresponding epipolar line in the reference frame. Therefore, the processing schedule along the epipolar lines avoids the conflict of memory accessing compared with the traditional horizontal raster-scan schedule. To support the various epipolar geometries, 7 kinds of block patterns are used to access reference pixels with slopes between ±45°. The accessed pix...