1

Highly Parallel Rate-Distortion Optimized Intra-Mode Decision on Multicore Graphics ProcessorsNgai-Man Cheung,Oscar C. Au, Senior Member, IEEE, Man-Cheung Kung,Peter H.W. Wong, Senior Member, IEEE, and Chun Hung Liu

CSVT NOVEMBER 2009

2

Outline

•Introduction•Intra-Prediction•Parallel RD Optimized Intra-Mode

Decision•Experiments•Conclusion

3

Introduction

•Multicore Graphics Processors▫Graphic Processing Unit (GPUs)▫Coprocessing units for CPUs to accelerate

numerical and signal processing applications , thanks to high-performance multicore and pipeline architectures

•Investigate the use of GPUs to perform RD

optimized intra-mode selection in AVS and H.264

4

Difficulties

•Intra-Mode Decision▫Dependency between current block and adjacent block▫Determine the encoding bit-rate for each of

the candidate modes, some conditional branching

may be needed

5

Contributions

•Analyze the dependency constraints inintra-mode decision

•Propose a strategy to determine the mode decisions of video blocks in parallel▫Encode the blocks in novel orders

•Extend a bit-rate approximation methodto estimate the rate in RD cost computation

6

Intra-prediction in H.264

M A B C D

I a b c d

J e f g h

K i j k l

L m n o p

E F G H

(a) 4 × 4 current blocks and their neighboring reconstructed pixels.

(b) Prediction directions and their corresponding modes.

0 54

6

1

8

73

2: DC mode4x4

7

Intra-prediction in AVS 1.0Vertical mode Horizontal mode

DC mode Down-right mode

Down-left mode:bidirectional prediction

8x8

8

Dependency Analysis•Dependency constraints on block encoding

order▫Prediction Direction

Determine the RD costs of the current block is hard before all the candidate reference blocks have been encoded and reconstructed

▫Pixel Filtering(AVS) Filtering may be applied to the reconstructed

pixels of the adjacent blocks before they are used in prediction, and this filtering may involve pixels from several blocks, leading to additional block dependency

9

Dependency Analysis

• Dependency between the four 8 × 8 blocks (K1-K4) in the current macroblock and their spatially adjacent neighbor blocks (T 1-T4,L1,L2), in AVS intra-prediction

10

Dependency Analysis

• Dependency between the four 4 × 4 blocks (K1-K4)in the current 8 × 8 block and their spatially adjacent neighbor blocks(T 1-T4,L1,L2), in H.264 intra-prediction

11

Dependency Analysis

•The dependency relationships form directed acyclic graphs. ▫Parallelize the RD cost computation of the

four constituent blocks of the same 16x16MB

▫Compute in parallel RD costs of the blocks from different 16x16MB

12

Greedy-Based Block Encoding Order

•Encode those blocks of which all the reference reconstructed pixels are available.

13

Greedy-Based Block Encoding Order

• AVS Example

14

Greedy-Based Block Encoding Order• AVS Example • AVS Example modify

version

• Postpone the encoding of several blocks along the left frame boundary, • All the four constituent blocks of any MB could be encoded consecutively• Does not incur any execution time penalty

15

Greedy-Based Block Encoding Order

• AVS Example

16

Optimality • Lemma 1: The proposed greedy-based encoding order can process all bottleneck path(s) P∗ with exactly n∗ iterations

• Proof of Lemma 1:▫ Suppose the greedy-based order

requires more that n* iterations to process

▫ At least one processing gap of length w which P* is not being processed between

Ki and Ki+1▫ There would exist an immediate parent

block Bm of Ki+1▫ Continuing with backtracking

eventually one would reach some block Kj in P*

▫ P1 has no processing gap >P0▫ P* replace Po to P1 would be longer

than p*

P*

P1P0

17

Optimality

•Theorem 1: The proposed greedy-based order can process all the video blocks in a frame in n∗ iterations

•P∗ = {K1,K2, ...,Kn∗ }, Kn∗ , would be processed in the n∗th iteration by Lemma 1. Since all the paths would also end in Kn∗ , all the blocks could be processed with n∗ iterations

18

Performance estimation• One of the longest paths

in H.264 4 × 4 intra-prediction

• The length can be found to be

• n*=((V/4)/2)x2+H/4-2 =V/4+H/4-2

• n*=(V/4)x2+(H/4)/2-2 =(V/4)x2 +H/8-2

19

Bit-Rate Estimation

•Lagrangian cost function

•Entropy coding may involve many branching instructions, hard to implement on pipeline architecture

20

Fast Bit Rate Estimation for Mode Decision• Tc : number of nonzero coefficients• Tz : number of zeros before the last nonzero coefficients• |Lk| : the absolute value of kth nonzero coefficient• Fk : the frequency of kth nonzero coefficient

[33] M. G. Sarwer and L.-M. Po, “Fast bit rate estimation for mode decision of H.264/AVC,” IEEE Trans. Circuits Syst. Video Technol., vol. 17, no. 10, pp. 1402–1407, Oct. 2007.

21

Experiments

•PC equipped with one GeForce 8800 GTS PCIe graphics card with 96 stream processors

•Intel Pentium 4 3.2 GHz processor with 1GB DDR2 memory

•H.264 JM 14.0•AVS RM 6.2 reference software

22

Encoding Bit-Rate Estimation

23

Parallel RD Optimized Intra-Mode Decision

More than 80 times reduction

QP has no significant effect

•H.264

24

Parallel RD Optimized Intra-Mode Decision

Parallelism within a MB

•H.264

25

Parallel RD Optimized Intra-Mode Decision

Similar speedups when RDO is disabled

•H.264

26

Parallel RD Optimized Intra-Mode Decision

•H.264

27

Parallel RD Optimized Intra-Mode Decision

•H.264

28

Parallel RD Optimized Intra-Mode Decision

•AVS

29

Parallel RD Optimized Intra-Mode Decision

•AVS

30

Parallel RD Optimized Intra-Mode Decision

39

96(processors)x2(threads)/5(modes) = 38.4

31

Conclusion• Based on the dependency analysis of intra-mode

decision , encode the video blocks following the greedy orders, leading to highly parallel RD cost computations.

• More than 80 times speedup for GPU based intra-prediction, GPU can be utilized to offload intra-prediction from CPU.

• To facilitate implementation on GPU, use a bitrate approximation method to estimate the rate in RD cost

computation. • The approximation errors only a small impact to

the coding performance: no more than 0.12 dB loss in PSNR and 0.98% bit-rate increase.