Customizing WideSIMD Architectures for H'264

1
Customizing Wide-SIMD Architecturesfor H.264

• Sangwon Seo1, Mark Woh1, Scott Mahlke1, Trevor
  Mudge1
• Vijay Sundaram2, Chaitali Chakrabarti2
• 1 University of Michigan
• 2 Arizona State University

2
Outline

• Motivation
• H.264 Analysis
• Proposed Architecture
• H.264 Kernel Mappings
• Results
• Conclusion

3
Motivation Smart Phone
Reference Images http//www.apple.com/iphone/gal
lery/
4
Motivation Inside Smart Phone
Reference Images http//idannyb.files.wordpress.
com/2008/07/xiuvbfueck3gsdum-large.jpg
5
H.264 Design
H.264 encoder/decoder reference design
Reference Images I. Richardson, H.264 and
MPEG-4 video compression, WILEY, 2003
6
H.264 Analysis

• H.264 Kernel Algorithms
• Heavy SIMD workload
• Different natural SIMD widths
• High Medium Thread Level Parallelism
• Need to support multiple SIMD widths to maximize
  the SIMD utilization

7
H.264 Analysis

• Example Deblocking Filter
• Two dimensional data are used for multimedia
  algorithms.
• Row or column order memory access works well for
  one set of edges, but not for the other.
• Diagonal memory bank system helps to access
  blocks along a row or a column.

Horizontal Filtering
Vertical Filtering
8
H.264 Analysis

• Subgraphs for Innerloops of two kernel algorithms
• Large amount of data locality
• Large RF power consumption (Read/Write)
• Bypass and Temporary buffer support

9
H.264 - Analysis

• Instruction Pairs
• Heavy usage of shuffle and arithmetic operations
• Add-Shift round operation
• Sub-Abs SAD operation
• Need to fuse the frequently used instruction pairs

10
H.264 - Analysis

• Permutation Patterns for Intraprediction
• Fixed set of shuffle patterns
• Need for programmable shuffle network

11
Modified SIMD architecture
12
Modified SIMD architecture
Multiple SIMD widths Thread-Level Parallelism
13
Modified SIMD architecture
Diagonal Memory Organization Memory Bank System
Shuffle Network
14
Modified SIMD architecture
Short-lived values stored in temporary buffers
15
Modified SIMD architecture
Short-lived values Fused Operation
16
Modified SIMD architecture
Shuffle Networks are placed here and there to
align data
17
Mapping of H.264 Kernels

• Intra Prediction

18
Results

• System Breakdown
• H.264 CIF video at 30fps

19
Results

• Speedup Breakdown
• 2.13x performance increase on average

20
Results

• Energy-Delay product comparison
• 29 energy-delay improvement on average

21
Results

• Comparison with latest H.264 encoders
• 17 T. C. Chen et.al, 2.8 to 62.7 mW
  low-power and power-aware H.264 encoder for
  mobile
• applications, 2007 IEEE Symposium on
  VLSI Circuits, pp. 222223, June 2007.
• 18 M. Bhatnagar, TMS320DM6446/3 Power
  Consumption Summary, Texas Instruments
• Application Reports, http//focus.ti.com/
  lit/an/spraad6a/spraad6a.pdf, Feb. 2008.

22
Conclusion

• Key architectural enhancements
• SIMD partitioning
• Diagonal memory bank system
• Bypass and temporary buffer support
• Fused operation support
• Programmable crossbar
• Future work
• Image processing algorithms on SIMD architecture

23
Backup Slides
24
H.264 Analysis

• Diagonal Memory Organization
• Two dimensional data are used for multimedia
  algorithms.
• Blocks along a row or a column need to be
  accessed easily.

25
Mapping of H.264 Kernels

• Deblocking Filter

26
Mapping of H.264 Kernels

• Motion Compensation

27
Mapping of H.264 Kernels

• Motion Estimation