1. Dynamic Buffering in EBCOT
Paul Zeman
Robert Kenney
ECE 734
Final Project

2. JPEG 2000 Overview
Discrete Wavelet Transform – Frequency domain transformation of input which produces low amplitude coefficients
Quantization – Optional step for lossy (9,7) compression
EBCOT – Embedded Block Coding with Optimized Truncation
Entropy codes DWT coefficients and produces a code stream

3. EBCOT Tier-1 Operation Invented by David Taubman
Takes 50% - 70% of the run time to complete
Context coding performed on each bit plane for a code block
Each bit in a bit plane is coded in one of three passes based on significance:
Significance Propagation Pass
Magnitude Refinement Pass
Clean-up Pass
Significance – bits following the first non-zero bit in coefficient
Scan order
for a single
bit-plane

4. EBCOT Tier-1 Dependencies
Significance Pass:
Codes bits in a bit plane that have a high probability of becoming significant
Based on value and neighborhood
Magnitude Refinement Pass
Codes bits that have already become significant in previous (higher) bit plane
Not dependent on other two passes of current bit plane
Clean-up Pass
Codes bits that have not been coded in SP or MRP
Codes bits in a bit plane that have a high probability of becoming significant based on value and neighborhood
Run-Mode codes entire non-significant stripe with one symbol

5. Previous Work Hardware implementations Serial pass scan order
Parallelizing passes
Each has a context window
Sig. Pass and Mag. Ref. pass share same window
Clean-up Pass’s window is two stripes behind
All three passes forced to move in lockstep

6. Proposed Design
Introduce dynamic buffering allowing for each pass to progress through the bit plane at it’s own pace in parallel with use of a larger data pool
Dependency constraints must be preserved
Take advantage of pixel skipping and column skipping

7. Project Goals Implement Goals Serial scan EBCOT context coder
Parallel scan with passes progressing in lockstep
Parallel scan with dynamic buffering allowing for free movement of passes (our design)
Goals
15% faster execution than lockstep parallel scan
47% faster execution than serial scan
Parallel scan with lockstep is 38% faster than serial scan

8. Any Questions?