Use of ThreadLevel Parallelism in Architectural Complexity Reduction of MPEG2 Encoder

1
Use of Thread-Level Parallelism in Architectural
Complexity Reduction of MPEG2 Encoder

• T. R. Jacobs
• V. A, Chouliaras
• D. J. Mulvaney
• J. L. Nunez
• Loughborough University

2
Contents

• Compression and MPEG2
• Thread-Level Parallelism the hows and whys
• Development environment
• Testing
• Results
• Conclusions
• Questions

3
Compression and MPEG2

• Uncompressed video data 175 Mbit/s
• DVD and DVB max 9.8 Mbit/s
• Motion Picture Export Group 2
• Asymmetric encoding/decoding process
• DCT block based compression
• 8 8 pixel blocks
• 16 16 pixel Macroblocks (MB)

4
MPEG2 Dataflow

• I frame
• fdct 8x8 blocks (lossless)
• Quantize (lossy)
• B / P frame
• Recover previous frame (iQuant, idct)
• Search for Macroblocks (MB), produce Motion
  Vectors (MV)
• Create estimate from recovered MV
• Fdct and quant errors between estimated and real
  frame
• Variable Length Code the quant errors and MV

5
Thread-Level Parallelism 1

• Spread workload over a number of processors
• Decrease number of loop iterations
• Calculate original loop number

6
Thread-Level Parallelism 2

• Control of variables
• Ensure mutual exclusivity
• Uses of local and shared variables
• Shared private space arrays
• static arrayMAX_THREAD
• arraycontext x
• Serial code inside parallel loop
• One thread performs serial code using each
  contexts data
• if (context0)
• for (z0zltMAX_THREADz)
• a arrayz
• Synchronisation
• Barriers

7
Development Environment

• Linux workstation
• MPEG2 Test Model 5
• Open source
• SimpleScalar simulator
• MIPS processor simulator
• Ideal multi-processor environment
• Produces instruction count per processor

8
Testing

• Relative complexity
• Context zero instruction count 100
• Single threaded instruction count
• Three test video sequences
• 25 frames
• 352 288
• Nine different processor counts
• Seven search window ranges

9
Cups, Deadline and Paris
10
Results
11
Optimal Context Number
Threaded loop run once only iteration 1

• Iteration width /
• (16 MAX_THREAD)
• Iteration block_count / MAX_THREAD

• Optimal context number Width
• 
  16
• Optimal context number chroma
• 
  block count
• 4.2.0 6
• 4.2.2 8
• 4.4.4 12

12
Conclusions

• Exploit natural parallelism
• Achieve complexity savings of 96
• More saving possible through vectorising

13
Questions?