Data Compression Conference 2013 Chenggang Yan, Yongdong Zhang, Feng Dai and Liang Li 1

Outline  Introduction  Related Work  Proposed Method  Experimental Results  Conclusion 2

Introduction (1/2)  HEVC coding tree unit (CTU) 3

Introduction (2/2)  Local parallel method (LPM) Maximum parallelism of LMP is equal or less than 8. independent Pus (IPUs)  Directed acyclic graph (DAG) 4

Related Work (1/2)  Local parallel method (LPM) [16] Motion estimate region (MER) 5 [16] Minhua Zhou, “AHG10: Configurable and CU-group level parallel merge/skip,” JCTVC-H0082, Feb. 2012

Related Work (2/2)  Local parallel method (LPM) 123 M = 16 or 8 6 8

Proposed Method  A. Data Dependency Analysis  B. DAG for CTUs  C. Highly Parallel Framework 7

Proposed Method.A (1/3)  Independent PUs (IPUs) The IPU’s left boundary and MER’s left boundary do not overlap. The IPU’s upper boundary and MER’s upper boundary do not overlap.  123 8

Proposed Method.A (2/3) 9

Proposed Method.A (3/3)  Neighboring CTUs left upper upper-left upper-right 10

Proposed Method  A. Data Dependency Analysis  B. DAG for CTUs  C. Highly Parallel Framework 11

Proposed Method.B (1/4)  Generate a DAG to capture the dependency relationships of CTUs. 12

Proposed Method.B (2/4)  DAG consists of a set of vertices V and edges E. data dependency an edge. Processed remove

Proposed Method.B (3/4)  Condition matrix (CM) 14

Proposed Method.B (4/4) 15

Proposed Method  A. Data Dependency Analysis  B. DAG for CTUs  C. Highly Parallel Framework 16

Proposed Method.C (1/5) 17

Proposed Method.C (2/5)  Step1 : Initialize DQ and CM. DQ is a waiting queue. CM is designed to record the number of related CTUs for each CTU.  Step2 : When some values in the CM become zero, get the corresponding coordinates and push them into DQ. 18

Proposed Method.C (3/5)  Step3 : Get coordinates from DQ and process corresponding CTUs in parallel on many-core platform.  Step4 : Update CM. When a CTU with coordinate (i, j) in CM is processed, the values of coordinates (i+1, j), (i+1, j-1), (i,j+1) and (i+1,j+1) in CM will minus one operation.  Step5 : Repeat above steps 2~4 until each frame is over. 19

Proposed Method.C (4/5)  Maximum parallelism of CTU 123  Maximum parallelism of highly parallel framework 123  Average parallelism of highly parallel framework

Proposed Method.C (5/5) 21

Experimental Results (1/5) 22

Experimental Results (2/5) 23

Experimental Results (3/5) 24

Experimental Results (4/5) 25

Experimental Results (5/5) 26

Conclusion (1/1)  Highly parallel framework provide sufficient parallelism for many-core platforms.  Use the DAG-based order to parallelize CTUs. 27