1 : HEVC Rough Mode Decision (Ji Hun Jang et al.: Down Sampling for Fast Rough Mode Decision for a Hardware-based HEVC Intra-frame encoder) (Special Paper) 21 3, 2016 5 (JBE Vol. 21, No. 3, May 2016) http://dx.doi.org/10.5909/jbe.2016.21.3.341 ISSN 2287-9137 (Online) ISSN 1226-7953 (Print) HEVC Rough Mode Decision a), a) Down Sampling for Fast Rough Mode Decision for a Hardware-based HEVC Intra-frame encoder Ji Hun Jang a) and Chae Eun Rhee a) HEVC H.264. H.264 50% 1.4.. rough mode decision(rmd). rate-distortion optimization (RDO) RMD RDO. RMD. RMD, RMD 70%. BDBR 0.04%. RMD 33%, 66%. Abstract HEVC is the next compression standard and is expected to be used widely replacing the conventional H.264/AVC standard. The compression ratio of the HEVC is twice times than H.264/AVC, whereas its computational complexity is increased by up to 40%. Many research efforts have been made to reduce the computational complexity and to speed up encoding. For intra coding, the rough mode decision (RMD) is commonly applied. The rate-distortion optimization (RDO) process to decide the best mode is too complex so that RMD chooses the candidate modes with a simple process and sends the candidates to RDO process. However, for large-size blocks, the RMD also requires considerable computations. In this paper, a down-sampling scheme is proposed for the RMD process. The reference pixel loading, predicted pixel generation are performed using the down-sampled pixel data. When the proposed scheme is applied to the RMD, the computational complexity is reduced by 70% with a marginal bitrate increase of 0.04%. In terms of area of hardware-based RMD, the gate count and the buffer size is reduced 33% and 66%, respectively. Keyword : Down-sampling, Rough mode decision, HEVC, intra prediction a) (Dept. of Information and Communication Engineering, Univ. of Inha) Corresponding Author : (Rhee Chae Eun) E-mail: chae.rhee@inha.ac.kr Tel: +82-31-330-6443 ORCID: http://orcid.org/0000-0002-7851-1703 IWAIT 2016. 2015 ( ) (NRF-2015R1C1A1A02037625). Manuscript received March 21, 2016; Revised May 24, 2016; Accepted May 24, 2016.
(JBE Vol. 21, No. 3, May 2016). High Efficiency Video Coding(HEVC) Standard H.264 Standard 50% 1.4 [1]. HEVC Intra Prediction Inter Prediction. Intra Prediction 8x8 pixels 64x64 pixels Coding Unit(CU) 35. H.264 4x4 pixels 9, 3.9. bitrate Rate Distortion Optimization(RDO) [2]. RDO RDO. HEVC, Rough Mode Decision(RMD). RMD HEVC 1. Intra Prediction 35. RMD SATD PU N [3]. RMD Most Probable Mode(MPM) RDO. Intra prediction. PU PU, MPM [4]. MPM R-D cost Intra prediction. RMD N RDO N Prediction Unit(PU) PU 4x4 8x8 N 8, 3. RDO. RMD. RDO, RMD Sum of absolute Transformed differences(satd) Cost. RDO 35 RMD RMD RMD 1. HEVC Fig. 1. Fast Mode Decision Algorithm in HEVC
1 : HEVC Rough Mode Decision (Ji Hun Jang et al.: Down Sampling for Fast Rough Mode Decision for a Hardware-based HEVC Intra-frame encoder)., 32x32 64x64..,. RMD RMD spatial correlation.. 2, 3 2 RMD. 4.. RMD. CU 2. 2 (a) 16x16 CU., 32 CU 65. 32x32 CU 129, 64x64 CU 257. CU. spatial correlation [5]. RMD RMD 70%. spatial correlation. RMD, 2:1. 2 (b) RMD. 2 (a) (b) 16x16 CU RMD (b), 2:1 33 64. 3 2. 16x16 CU (a) (b) Fig. 2. Intra prediction of 16x16 CU (a) without downsampling (b) with downsampling
(JBE Vol. 21, No. 3, May 2016)..,., 3 16x16 CU RMD,. Original RMD SATD. SATD 8x8, 16x16 SATD 8x8 4. SATD 8x8 SATD. 2 1, 4 1.. Original Intra prediction., 8x8 16x16 CU Bi-linear filtering 8x8 16x16 32x32 64x64 CU.. (1) Original HEVC. SATD 4 1. (1) SATD 4 1. (2) SATD 4 RMD. λ λ. RMD 3. Fig. 3. Selected reference pixels and original pixels in Down-sampling scheme 64x64 32x32 CU. 16x16 8x8CU. 16x16 8x8 CU 4 RMD. Original RMD 8x8, 16x16 32x32 CU. 8x8, 16x16 32x32 CU Neighbor Pixel Buffer CU. Original Pixel Buffer CU. Smoothing Reference Pixel Intra prediction smoothing smoothing. Smoothed Pixel Buffer Non smoothed Pixel Buffer smoothing smoothing
1 : HEVC Rough Mode Decision (Ji Hun Jang et al.: Down Sampling for Fast Rough Mode Decision for a Hardware-based HEVC Intra-frame encoder) 4. Original RMD Fig. 4. Original RMD Hardware Architecture CU. 35, SATD Original Pixel Buffer SATD. SATD Mode Cost Buffer, 35 Mode Cost Switch N. 32x32 CTU 8x8 CU 16x16 CU [6]. 32x32 CTU CU 32x32 CU 1, 16x16 CU 4 8x8 CU 16. 4 32x32 CTU CU 5. 5 32x32, 16x16 8x8 CU RMD. 32x32, 16x16 8x8 CU 5. 7 5 35. 32x32 CU 35 RMD 16x16 CU 4 8x8 CU 16 RMD. 8pixels/cycle 8x8, 16x16 32x32 CU RMD. 4 32x32 CU 16x16 32x32.
(JBE Vol. 21, No. 3, May 2016) 5. RMD Fig. 5. Timetable of RMD Hardware 6. RMD Fig. 6. RMD Hardware Architecture of Down-sampling scheme RMD 6. pixel Neighbor Pixel Buffer Original Pixel Buffer RMD hardware, 32x32 CU. 32x32 RMD Hardware. Neighbor Pixel Buffer, Smoothed Pixel Buffer Non Smoothed Pixel Buffer 129 Bytes 387 Bytes. Original Pixel Buffer 32x 32 Original Pixels 1,024 Bytes. Mode Cost Buffer Cost 70 Bytes. 32x32 RMD Hardware, Cost Buffer. 2,234 Bytes 32x32 RMD Hardware 1,471 Bytes.. 32x32 CTU
1 : HEVC Rough Mode Decision (Ji Hun Jang et al.: Down Sampling for Fast Rough Mode Decision for a Hardware-based HEVC Intra-frame encoder). RMD 32x32 CU Gate counts 33%, 73%. RMD SATD 64x64 CU 32x32 CU 75%. 1 y, u v BD-rate. Intra-only configuration 22,27,32 37 100. 1 BD-rate Y, U, V 0.34%, 0.69% 0.82% BD-rate. 1 BD-rate. 1. BD-rate Table 1. BD-rate of Down-sampling scheme Class A B E C D Sequence BD-Rate (%) Y U V Traffic 0.20 0.36 0.40 PeopleOnStreet 0.10 0.39 0.45 ParkScene 0.07 0.08 0.13 Cactus 0.14 0.13 0.15 BasketballDrive 0.64 1.27 1.06 BQTerrace 0.07 0.17 0.25 Kimono 0.37 0.38 0.36 FourPeople 0.13 0.48 0.50 Vidyo1 0.34 0.69 0.82 Johnny 0.41 0.49 0.08 KristenAndSara 0.35 0.51 0.62 BasketballDrill 0.29-0.01 0.18 PartyScene 0.01-0.01 0.06 BQMall 0.05 0.08-0.02 Flowervase 0.22 0.46 0.31 Keiba 0.08 0.22 0.33 RaceHorses 0.04 0.10 0.11 BlowingBubbles 0.00-0.07 0.00 BasketballPass 0.22 0.40 0.34 BQSquare -0.01 0.11 0.21 Mobisode 0.45 0.39 0.76 Average 0.34 0.69 0.82 Reference Software HM13.0 RMD. 2 Intel i5-4460 @ 3.20GHz CPU 1 Sequence 100. 32x32 64x64 SATD 75% 32x32 64x64 RMD 56.5%, 58.4%. RMD (total) 26.4%. 2. HM13.0 RMD Table 2. Average processing time of RMD in HM13.0 average processing time (micro sec) HM 13.0 Down sampling scheme 8x8 RMD 20.64 19.76 16x16 RMD 58.53 56.00 32x32 RMD 204.59 88.92 64x64 RMD 785.93 326.81 Total 3861.73 2843.13 (References) [1] JW Kim, DH Kim, JG Kim, An Early Termination Algorithm of Prediction Unit(PU) Search for Fast HEVC Encoding, JBE, Vol19, No.5, p628-630, Septemver, 2014. [2] EY Won, SI Chae, An Improved RDO Algorithm for the HEVC intra encoder, KSBE Summer Conference, p123-126, June, 2013. [3] Tariq, Junaid, and Sam Kwong. Hybrid Fast Intra Mode Decision And Early Termination of Prediction Unit(PU) Splitting For HEVC, Systems, Man, And Cybernetics (SMC), 2015 IEEE International Conference on. IEEE, 2015. [4] Lainema, Jani, et al. Intra coding of the HEVC standard. Circuits and Systems for Video Technology, IEEE Transactions on 22.12, 2012 [5] Zhang, Qiuwen, et al. A fast and efficient coding unit size decision algorithm based on temporal and spatial correlation. Optik-International Journal for Light and Electron Optics 126.21, 2015 [6] Lee, Jaeho, et al. A fast CU size decision algorithm for HEVC. Circuits and Systems for Video Technology, IEEE Transactions on 25.3, 2015
348 방송공학회논문지 제21권 제3호, 2016년 5월 (JBE Vol. 21, No. 3, May 2016) 저자소개 장지훈 년 인하대학교 정보통신공학부 졸업 학사 년 현재 인하대학교 정보통신공학부 재학 석사) 주관심분야 영상 압축 설계 - 2014 : ( ) - 2014 ~ : ( - ORCID : http://orcid.org/0000-0002-5004-2642 :, SoC 이채은 - 년 : 서울대학교 컴퓨터공학과 졸업(학사) 년 : 서울대학교 전기컴퓨터공학부 졸업(석사) 년 : 서울대학교 전기컴퓨터공학부 졸업(박사) 년 ~ 2005년 : 삼성전자 SYSTEM LSI 사업부 선임연구원 년 ~ 현재 : 인하대학교 정보통신공학과(조교수) : http://orcid.org/0000-0002-7851-1703 주관심분야 : 영상 압축, SoC 설계 2000 2002 2011 2002 2013 ORCID